Add Batch bb854e6c-ec8c-4079-844d-f5f6ccbc61f9 data

.gitattributes

@@ -5423,3 +5423,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2024/Training-free[[:space:]]Deep[[:space:]]Concept[[:space:]]Injection[[:space:]]Enables[[:space:]]Language[[:space:]]Models[[:space:]]for[[:space:]]Video[[:space:]]Question[[:space:]]Answering/30bc2320-3072-42e8-b045-7ea96ee374bf_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2024/TransferTOD_[[:space:]]A[[:space:]]Generalizable[[:space:]]Chinese[[:space:]]Multi-Domain[[:space:]]Task-Oriented[[:space:]]Dialogue[[:space:]]System[[:space:]]with[[:space:]]Transfer[[:space:]]Capabilities/ac2486b2-5117-4e14-b15a-c57808d7db92_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2024/Transformers[[:space:]]are[[:space:]]Multi-State[[:space:]]RNNs/21a138ef-7a87-4091-bbf0-b72e8fb5c9c9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/TraveLER_[[:space:]]A[[:space:]]Modular[[:space:]]Multi-LMM[[:space:]]Agent[[:space:]]Framework[[:space:]]for[[:space:]]Video[[:space:]]Question-Answering/4933ce75-2f8b-4d94-99a0-a97b4270db9a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Tree[[:space:]]of[[:space:]]Problems_[[:space:]]Improving[[:space:]]structured[[:space:]]problem[[:space:]]solving[[:space:]]with[[:space:]]compositionality/418dcc40-3e0e-4ed6-aa5d-caf94c578b26_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Triad_[[:space:]]A[[:space:]]Framework[[:space:]]Leveraging[[:space:]]a[[:space:]]Multi-Role[[:space:]]LLM-based[[:space:]]Agent[[:space:]]to[[:space:]]Solve[[:space:]]Knowledge[[:space:]]Base[[:space:]]Question[[:space:]]Answering/2bfc46f7-f4ed-4d88-9bf5-1ead6a8c1d52_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/TroL_[[:space:]]Traversal[[:space:]]of[[:space:]]Layers[[:space:]]for[[:space:]]Large[[:space:]]Language[[:space:]]and[[:space:]]Vision[[:space:]]Models/d3959391-dc37-4afa-81df-706ce45ce8e9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Turn[[:space:]]Waste[[:space:]]into[[:space:]]Worth_[[:space:]]Rectifying[[:space:]]Top-k[[:space:]]Router[[:space:]]of[[:space:]]MoE/78b71285-cba8-4c67-9072-33097abe467c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Twists,[[:space:]]Humps,[[:space:]]and[[:space:]]Pebbles_[[:space:]]Multilingual[[:space:]]Speech[[:space:]]Recognition[[:space:]]Models[[:space:]]Exhibit[[:space:]]Gender[[:space:]]Performance[[:space:]]Gaps/75d52286-5b60-4f2d-833b-11b4a49fb28d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/UNICORN_[[:space:]]A[[:space:]]Unified[[:space:]]Causal[[:space:]]Video-Oriented[[:space:]]Language-Modeling[[:space:]]Framework[[:space:]]for[[:space:]]Temporal[[:space:]]Video-Language[[:space:]]Tasks/1d2b554e-cf0c-46d1-9ba0-ee7589838c7b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/UNO[[:space:]]Arena[[:space:]]for[[:space:]]Evaluating[[:space:]]Sequential[[:space:]]Decision-Making[[:space:]]Capability[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models/24977746-f195-40b6-ae76-06cf317c5f92_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/UOUO_[[:space:]]Uncontextualized[[:space:]]Uncommon[[:space:]]Objects[[:space:]]for[[:space:]]Measuring[[:space:]]Knowledge[[:space:]]Horizons[[:space:]]of[[:space:]]Vision[[:space:]]Language[[:space:]]Models/3fd78e6a-da2e-45dd-83ff-a49b6a122da9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Uncertainty[[:space:]]in[[:space:]]Language[[:space:]]Models_[[:space:]]Assessment[[:space:]]through[[:space:]]Rank-Calibration/d8e5286d-2bde-41fc-9990-e94616eb0d3f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Understanding[[:space:]]Higher-Order[[:space:]]Correlations[[:space:]]Among[[:space:]]Semantic[[:space:]]Components[[:space:]]in[[:space:]]Embeddings/ed00e30a-36cc-4a2e-875b-0dcb4f0bf780_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Understanding[[:space:]]Slang[[:space:]]with[[:space:]]LLMs_[[:space:]]Modelling[[:space:]]Cross-Cultural[[:space:]]Nuances[[:space:]]through[[:space:]]Paraphrasing/2cac368e-4fa1-4ae1-90e1-97e1526bf1a4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Understanding[[:space:]]and[[:space:]]Mitigating[[:space:]]Language[[:space:]]Confusion[[:space:]]in[[:space:]]LLMs/a17d276a-4ede-4d17-8ae9-0bb2aabfd0ba_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Understanding[[:space:]]“Democratization”[[:space:]]in[[:space:]]NLP[[:space:]]and[[:space:]]ML[[:space:]]Research/c89f1e1f-0e52-436d-8857-021230fe75c4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/UniFashion_[[:space:]]A[[:space:]]Unified[[:space:]]Vision-Language[[:space:]]Model[[:space:]]for[[:space:]]Multimodal[[:space:]]Fashion[[:space:]]Retrieval[[:space:]]and[[:space:]]Generation/3e796285-bb56-451d-96bb-ccd6a0715b22_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/UniGen_[[:space:]]Universal[[:space:]]Domain[[:space:]]Generalization[[:space:]]for[[:space:]]Sentiment[[:space:]]Classification[[:space:]]via[[:space:]]Zero-shot[[:space:]]Dataset[[:space:]]Generation/bf48bad7-34ec-4a98-b33e-bf2b09fa1389_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Unifying[[:space:]]Multimodal[[:space:]]Retrieval[[:space:]]via[[:space:]]Document[[:space:]]Screenshot[[:space:]]Embedding/26ce9ccc-7065-489f-9318-f1b2e02225ea_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Universal[[:space:]]Vulnerabilities[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models_[[:space:]]Backdoor[[:space:]]Attacks[[:space:]]for[[:space:]]In-context[[:space:]]Learning/0b7a94b2-5288-4cc1-9fe6-c38d61f1ca9a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Unknown[[:space:]]Claims_[[:space:]]Generation[[:space:]]of[[:space:]]Fact-Checking[[:space:]]Training[[:space:]]Examples[[:space:]]from[[:space:]]Unstructured[[:space:]]and[[:space:]]Structured[[:space:]]Data/456b11db-cffd-49db-a2e9-75d096d1ba80_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Unlabeled[[:space:]]Debiasing[[:space:]]in[[:space:]]Downstream[[:space:]]Tasks[[:space:]]via[[:space:]]Class-wise[[:space:]]Low[[:space:]]Variance[[:space:]]Regularization/21a44796-0af0-4b39-9dff-b7f79824f089_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Unleashing[[:space:]]the[[:space:]]Power[[:space:]]of[[:space:]]Emojis[[:space:]]in[[:space:]]Texts[[:space:]]via[[:space:]]Self-supervised[[:space:]]Graph[[:space:]]Pre-Training/3a3d1d25-e00c-4be7-aa6e-c18a814604b1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Unlocking[[:space:]]Anticipatory[[:space:]]Text[[:space:]]Generation_[[:space:]]A[[:space:]]Constrained[[:space:]]Approach[[:space:]]for[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]Decoding/37ebe6b9-99be-4117-976d-ac9494e79145_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Unlocking[[:space:]]Markets_[[:space:]]A[[:space:]]Multilingual[[:space:]]Benchmark[[:space:]]to[[:space:]]Cross-Market[[:space:]]Question[[:space:]]Answering/19bff1ee-9aa9-454d-8066-616391a08b9f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Unlocking[[:space:]]Memorization[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]with[[:space:]]Dynamic[[:space:]]Soft[[:space:]]Prompting/4d23614f-8034-43fe-89f2-b8b14d1cd07c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Unlocking[[:space:]]the[[:space:]]Future_[[:space:]]Exploring[[:space:]]Look-Ahead[[:space:]]Planning[[:space:]]Mechanistic[[:space:]]Interpretability[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models/c70e9fe9-8cc3-4232-bbdf-0acd9fe7b58a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Unraveling[[:space:]]Babel_[[:space:]]Exploring[[:space:]]Multilingual[[:space:]]Activation[[:space:]]Patterns[[:space:]]of[[:space:]]LLMs[[:space:]]and[[:space:]]Their[[:space:]]Applications/20faf9dd-a711-4451-a786-64a8197f3675_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Unsupervised[[:space:]]Discrete[[:space:]]Representations[[:space:]]of[[:space:]]American[[:space:]]Sign[[:space:]]Language/68a1b196-1fb9-4176-9817-a09f4fad1db5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Unsupervised[[:space:]]End-to-End[[:space:]]Task-Oriented[[:space:]]Dialogue[[:space:]]with[[:space:]]LLMs_[[:space:]]The[[:space:]]Power[[:space:]]of[[:space:]]the[[:space:]]Noisy[[:space:]]Channel/dee5a84a-0347-4d5d-a0d2-007e0b51fa63_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Unsupervised[[:space:]]Extraction[[:space:]]of[[:space:]]Dialogue[[:space:]]Policies[[:space:]]from[[:space:]]Conversations/28a8e11c-62f5-4e66-90d8-cbc56c121ba6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Unsupervised[[:space:]]Human[[:space:]]Preference[[:space:]]Learning/d65fa1bb-fb94-40be-b99b-bb60f941644f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Unsupervised[[:space:]]Named[[:space:]]Entity[[:space:]]Disambiguation[[:space:]]for[[:space:]]Low[[:space:]]Resource[[:space:]]Domains/736a9954-3a70-4fe0-acf8-f73180e61c06_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Unveiling[[:space:]]Factual[[:space:]]Recall[[:space:]]Behaviors[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]through[[:space:]]Knowledge[[:space:]]Neurons/da3898cb-a9b5-48b9-93aa-6ff653780c83_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Unveiling[[:space:]]In-Context[[:space:]]Learning_[[:space:]]A[[:space:]]Coordinate[[:space:]]System[[:space:]]to[[:space:]]Understand[[:space:]]Its[[:space:]]Working[[:space:]]Mechanism/65efef7c-e037-42f6-83c8-48d4d3818d34_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Unveiling[[:space:]]Multi-level[[:space:]]and[[:space:]]Multi-modal[[:space:]]Semantic[[:space:]]Representations[[:space:]]in[[:space:]]the[[:space:]]Human[[:space:]]Brain[[:space:]]using[[:space:]]Large[[:space:]]Language[[:space:]]Models/1b772ab9-2fd4-4015-ab91-d517630c9a55_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Unveiling[[:space:]]and[[:space:]]Consulting[[:space:]]Core[[:space:]]Experts[[:space:]]in[[:space:]]Retrieval-Augmented[[:space:]]MoE-based[[:space:]]LLMs/c89b18e4-3233-4a72-abfd-8c8810fda162_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Unveiling[[:space:]]the[[:space:]]Lexical[[:space:]]Sensitivity[[:space:]]of[[:space:]]LLMs_[[:space:]]Combinatorial[[:space:]]Optimization[[:space:]]for[[:space:]]Prompt[[:space:]]Enhancement/d288ed33-e1be-4361-9a3f-54f5d41cb86a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Unveiling[[:space:]]the[[:space:]]Role[[:space:]]of[[:space:]]Pretraining[[:space:]]in[[:space:]]Direct[[:space:]]Speech[[:space:]]Translation/daa1343f-7493-4883-b30d-ad4a0cbe4b54_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Unveiling[[:space:]]the[[:space:]]mystery[[:space:]]of[[:space:]]visual[[:space:]]attributes[[:space:]]of[[:space:]]concrete[[:space:]]and[[:space:]]abstract[[:space:]]concepts_[[:space:]]Variability,[[:space:]]nearest[[:space:]]neighbors,[[:space:]]and[[:space:]]challenging[[:space:]]categories/48bdd261-56d1-4ff8-84c9-fe0f21fa3a2c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Updating[[:space:]]CLIP[[:space:]]to[[:space:]]Prefer[[:space:]]Descriptions[[:space:]]Over[[:space:]]Captions/f0c647ce-69d8-4137-aa93-9d81996d6c7c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/User[[:space:]]Inference[[:space:]]Attacks[[:space:]]on[[:space:]]Large[[:space:]]Language[[:space:]]Models/7fb0a5c7-4f58-4624-b110-c12b5db45e45_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Using[[:space:]]Language[[:space:]]Models[[:space:]]to[[:space:]]Disambiguate[[:space:]]Lexical[[:space:]]Choices[[:space:]]in[[:space:]]Translation/3a3c6065-7af4-4f2b-aafe-06a7a5452851_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/VGBench_[[:space:]]Evaluating[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]on[[:space:]]Vector[[:space:]]Graphics[[:space:]]Understanding[[:space:]]and[[:space:]]Generation/98dc2bce-3483-4d81-88a7-d36e896930d3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/VHASR_[[:space:]]A[[:space:]]Multimodal[[:space:]]Speech[[:space:]]Recognition[[:space:]]System[[:space:]]With[[:space:]]Vision[[:space:]]Hotwords/a8d1073c-00eb-4cb8-a0d7-a720f041c382_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/VIEWS_[[:space:]]Entity-Aware[[:space:]]News[[:space:]]Video[[:space:]]Captioning/3f3ff6e8-02af-4f8f-a333-a5895b619cac_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/VIMI_[[:space:]]Grounding[[:space:]]Video[[:space:]]Generation[[:space:]]through[[:space:]]Multi-modal[[:space:]]Instruction/f4aa2209-e928-47c4-86a7-3e032e5d4a13_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/VIVA_[[:space:]]A[[:space:]]Benchmark[[:space:]]for[[:space:]]Vision-Grounded[[:space:]]Decision-Making[[:space:]]with[[:space:]]Human[[:space:]]Values/9ff2016e-5dd9-4d3d-8928-931410d0ee52_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/VLEU_[[:space:]]a[[:space:]]Method[[:space:]]for[[:space:]]Automatic[[:space:]]Evaluation[[:space:]]for[[:space:]]Generalizability[[:space:]]of[[:space:]]Text-to-Image[[:space:]]Models/0e5d7611-70dd-44f5-ba72-757ecfd1362f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/VLFeedback_[[:space:]]A[[:space:]]Large-Scale[[:space:]]AI[[:space:]]Feedback[[:space:]]Dataset[[:space:]]for[[:space:]]Large[[:space:]]Vision-Language[[:space:]]Models[[:space:]]Alignment/3dbb750c-6a3d-4753-aaa8-1820eb5ac143_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/VPTQ_[[:space:]]Extreme[[:space:]]Low-bit[[:space:]]Vector[[:space:]]Post-Training[[:space:]]Quantization[[:space:]]for[[:space:]]Large[[:space:]]Language[[:space:]]Models/dc1b1285-bd15-42a4-8eeb-7fdd1fc54c89_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Varying[[:space:]]Sentence[[:space:]]Representations[[:space:]]via[[:space:]]Condition-Specified[[:space:]]Routers/94cdf4c7-5041-4b15-9e7d-472bf0be4dc6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Verba[[:space:]]volant,[[:space:]]scripta[[:space:]]volant_[[:space:]]Don’t[[:space:]]worry![[:space:]]There[[:space:]]are[[:space:]]computational[[:space:]]solutions[[:space:]]for[[:space:]]protoword[[:space:]]reconstruction/130c1e55-2f23-46b0-88fd-3063af812fba_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Verifiable,[[:space:]]Debuggable,[[:space:]]and[[:space:]]Repairable[[:space:]]Commonsense[[:space:]]Logical[[:space:]]Reasoning[[:space:]]via[[:space:]]LLM-based[[:space:]]Theory[[:space:]]Resolution/305dec4e-09b9-4585-a76a-73ac2690b3e8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Verification[[:space:]]and[[:space:]]Refinement[[:space:]]of[[:space:]]Natural[[:space:]]Language[[:space:]]Explanations[[:space:]]through[[:space:]]LLM-Symbolic[[:space:]]Theorem[[:space:]]Proving/c0c4dff5-2a30-47d2-ab85-f2f3c064b3ae_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/VerifyMatch_[[:space:]]A[[:space:]]Semi-Supervised[[:space:]]Learning[[:space:]]Paradigm[[:space:]]for[[:space:]]Natural[[:space:]]Language[[:space:]]Inference[[:space:]]with[[:space:]]Confidence-Aware[[:space:]]MixUp/8b7badd7-2ce9-4772-b9f0-4930358cccd1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Video-LLaVA_[[:space:]]Learning[[:space:]]United[[:space:]]Visual[[:space:]]Representation[[:space:]]by[[:space:]]Alignment[[:space:]]Before[[:space:]]Projection/c90213b6-1945-4d4b-b267-39e43d584c58_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Video-Text[[:space:]]Prompting[[:space:]]for[[:space:]]Weakly[[:space:]]Supervised[[:space:]]Spatio-Temporal[[:space:]]Video[[:space:]]Grounding/039a8f2d-ad13-4096-8415-3b51e4f6fd77_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/VideoCLIP-XL_[[:space:]]Advancing[[:space:]]Long[[:space:]]Description[[:space:]]Understanding[[:space:]]for[[:space:]]Video[[:space:]]CLIP[[:space:]]Models/7b1d1562-6261-47c1-bbc7-f9746261c17c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/VideoScore_[[:space:]]Building[[:space:]]Automatic[[:space:]]Metrics[[:space:]]to[[:space:]]Simulate[[:space:]]Fine-grained[[:space:]]Human[[:space:]]Feedback[[:space:]]for[[:space:]]Video[[:space:]]Generation/7abc75a1-5df8-434d-afdd-b3eb2c40558b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Virtual[[:space:]]Personas[[:space:]]for[[:space:]]Language[[:space:]]Models[[:space:]]via[[:space:]]an[[:space:]]Anthology[[:space:]]of[[:space:]]Backstories/510458f5-561c-4e99-8259-c4b547c29fbc_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Vision-Language[[:space:]]Model[[:space:]]Fine-Tuning[[:space:]]via[[:space:]]Simple[[:space:]]Parameter-Efficient[[:space:]]Modification/48625390-bfb2-4c3a-82f4-62af86aff7d4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Visual[[:space:]]Prompting[[:space:]]in[[:space:]]LLMs[[:space:]]for[[:space:]]Enhancing[[:space:]]Emotion[[:space:]]Recognition/ee83fe1d-9e45-445f-9c38-464febbe95b0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Visual[[:space:]]Text[[:space:]]Matters_[[:space:]]Improving[[:space:]]Text-KVQA[[:space:]]with[[:space:]]Visual[[:space:]]Text[[:space:]]Entity[[:space:]]Knowledge-aware[[:space:]]Large[[:space:]]Multimodal[[:space:]]Assistant/f1ce05bf-03aa-441a-aa86-9172ea2ec142_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Voices[[:space:]]Unheard_[[:space:]]NLP[[:space:]]Resources[[:space:]]and[[:space:]]Models[[:space:]]for[[:space:]]Yorùbá[[:space:]]Regional[[:space:]]Dialects/7f844077-1247-42e2-80dc-d49bf87b4b2f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Voices[[:space:]]in[[:space:]]a[[:space:]]Crowd_[[:space:]]Searching[[:space:]]for[[:space:]]clusters[[:space:]]of[[:space:]]unique[[:space:]]perspectives/a8cd7604-e232-4bd2-8dbf-cfb9b597ec3d_origin.pdf filter=lfs diff=lfs merge=lfs -text

2024/TraveLER_ A Modular Multi-LMM Agent Framework for Video Question-Answering/4933ce75-2f8b-4d94-99a0-a97b4270db9a_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:00991b4dfce974adcf8010237b72e6762dfdefdf3bc9a12f9a26ab92c36eaaf7
+size 34254489

2024/TraveLER_ A Modular Multi-LMM Agent Framework for Video Question-Answering/full.md

@@ -0,0 +1,1023 @@
+# TraveLER: A Modular Multi-LMM Agent Framework for Video Question-Answering
+
+Chuyi Shang\* Amos You\* Sanjay Subramanian Trevor Darrell Roei Herzig
+
+University of California, Berkeley
+
+# Abstract
+
+Recently, image-based Large Multimodal Models (LMMs) have made significant progress in video question-answering (VideoQA) using a frame-wise approach by leveraging large-scale pretraining in a zero-shot manner. Nevertheless, these models need to be capable of finding relevant information, extracting it, and answering the question simultaneously. Currently, existing methods perform all of these steps in a single pass without being able to adapt if insufficient or incorrect information is collected. To overcome this, we introduce a modular multi-LMM agent framework based on several agents with different roles, instructed by a Planner agent that updates its instructions using shared feedback from the other agents. Specifically, we propose TraveLER, a method that can create a plan to "Traverse" through the video, ask questions about individual frames to "Locate" and store key information, and then "Evaluate" if there is enough information to answer the question. Finally, if there is not enough information, our method is able to "Replan" based on its collected knowledge. Through extensive experiments, we find that the proposed TraveLER approach improves performance on several VideoQA benchmarks without the need to fine-tune on specific datasets. Our code is available at https://github.com/traveler-framework/TraveLER.
+
+# 1 Introduction
+
+Over the last few years, Large Multimodal Models (LMMs) have demonstrated tremendous progress in the area of video understanding, particularly for the video question-answering (VideoQA) domain (Fu et al., 2023; Wang et al., 2024). More recently, LMMs have been able to achieve impressive results through video-based models (Lin et al., 2023a; Sun et al., 2022; Ye et al., 2023; Li et al., 2023). However, video models require a high level
+
+of computational complexity to fine-tune, and annotations are difficult and expensive to collect. As a result, many recent approaches (Xue et al., 2023; Yu et al., 2023; Zhang et al., 2023a) operate on the frame level, leveraging large-scale image-based pretrained models in a zero-shot manner.
+
+Moreover, these models may need to do several tasks simultaneously in a single step. In particular, they should identify the correct events in videos by understanding what information is relevant and ignoring irrelevant information. Next, they would need to extract specific and question-relevant visual details and use them to answer the question. As such, one iteration might not be enough to collect all the necessary information. For example, many current approaches use simple captioning, which is often too general to extract specific details, or they might miss important events. In these cases, current approaches cannot revisit the video to find additional information. Decomposing this process into different components allows each component to adapt to newly collected information.
+
+To address this, we introduce a modular multi-LMM agent framework for VideoQA. A Planner agent instructs agents in charge of different tasks, such as navigating through the video, extracting visual information through a question-answering process, and reviewing this information to select an answer. Feedback from the agents is then passed on to each other and back to the Planner, who uses the feedback to update its instructions. In this way, we can revisit the video to narrow our focus or expand our search to more relevant information and extract specific details to answer the question.
+
+Consider the example in Figure 1. Suppose we are asked why the boy turned over in the middle of the video. In the first iteration, our method uses temporal cues from the question to skip to the middle of the video and asks questions to find the relevant frames. In the next iteration, we gather more information. Asking about what the boy is doing,
+
+![](images/0d323f3d66fa3b2bc4554f5e6f6246a86c27ad5535fe45262c766d00ffcda5fb.jpg)  
+Figure 1: A simplified overview of our TravelLER framework. Our proposed framework aims to answer the question by collecting relevant information from keyframes through interactive question-asking. To accomplish this, several agents (in colored boxes) with different roles interact (left-to-right in each row) over several iterations. TravelLER creates a plan (in blue) to "traverse" (in orange) through the video, asks questions regarding individual frames (in yellow) to "locate" and store key information and, "evaluates" whether there is sufficient information to answer the question (in green), and "replans" using past collected knowledge if there is not enough information. Click on the image to see the video.
+
+we learn that he is "standing up at the bottom of the slide" and is not looking at anything specific, which informs us that the boy is no longer "sitting down" (choice B) or "resting on the yellow object" (choice D). To eliminate these choices, we must confirm that the boy does not sit back down again by traveling to a timestamp near the end of the video. Finally, since we have collected enough information and followed the plans, we can select the right choice that the boy turns over to be on his stomach "to get down from slide" (Choice E).
+
+Our proposed approach - Traverse, Locate, Evaluate, and Replan (TraveLER), has four main stages. First, in the Traversal stage, an agent creates a plan to answer the question. In the Location stage, an agent uses the plan to decide which timestamp of the video to select. The corresponding frames are then sent to another agent, which asks ques
+
+tions and stores the answers in a memory bank for future iterations. Finally, in the Evaluation stage, an agent reviews all collected information and decides whether to answer or create a modified plan (Replan) to start the next iteration if necessary.
+
+To summarize, our main contributions are: (i) We introduce TraveLER, a modular multi-LMM agent framework for video question-answering. (ii) Our method shows improved performance on multiple difficult video question-answering benchmarks, such as NExT-QA, EgoSchema, Perception Test, and STAR. (iii) Our method is easy to employ with different LLMs and LMMs, highlighting the effectiveness of our modular approach.
+
+# 2 Related Work
+
+Video Question-Answering. VideoQA involves answering free-form or multiple-choice questions
+
+given an input video. Compared to image question answering, VideoQA poses unique challenges because it often requires strong temporal understanding and the ability to deal with long input sequences. Many recent works have focused on training end-to-end video-language models (Fu et al., 2023; Sun et al., 2022; Wang et al., 2024; Ye et al., 2023; Yu et al., 2022; Li et al., 2023), but doing so remains challenging due to computational constraints and difficulties in architecture scaling. As a result, many approaches adopt pretrained image models to the video domain by extracting information independently from each frame (Xue et al., 2023; Yu et al., 2023; Zhang et al., 2023a). Here, we design a framework that builds an adaptive plan to find and extract relevant information using a question-answering approach.
+
+LMMs for Video Understanding. LMMs have been shown to be extremely useful for VideoQA. Some methods use supervised or contrastive training to perform video-LMM pretraining (Zhao et al., 2023; Yang et al., 2023; Chen et al., 2024), while others adapt existing LMMs and use instruction tuning to adapt them to the video domain (Zhang et al., 2023b; Maaz et al., 2024; Lin et al., 2023a). However, recent improvements in LMM capabilities have allowed for many strong approaches for few-shot (Alayrac et al., 2022; Wang et al., 2022) and zero-shot VideoQA (Yang et al., 2022). In particular, zero-shot methods, such as LLoVi (Zhang et al., 2023a), use pre-trained LMMs to generate captions for each frame in the video. Nevertheless, uniformly sampling frames at random may miss important visual information and focus on unimportant frames (Wu et al., 2019; Lei et al., 2021b). Recent works like SeViLA (Yu et al., 2023) addressed this problem by performing parameter-efficient finetuning using captions to identify keyframes (Lu et al., 2022; Buch et al., 2022; Qian et al., 2023), but this requires fine-tuning on specific datasets. Unlike these works, which select all keyframes in a single pass, we introduce a novel iterative and modular approach instructed by a planner.
+
+LMM-based Agents for Videos. The strong reasoning abilities of LLMs (Brown et al., 2020; Chung et al., 2022) have made them effective in LLM-based agent approaches for videos, where an LLM performs much of the reasoning after collecting information from different modules (Chen et al., 2023; Lin et al., 2023b; Zhang et al., 2024; Zeng et al., 2023). For example, Socratic Models (Zeng et al., 2023) proposes a method to reason
+
+about videos based on generated audio transcriptions and CLIP frame similarity scores, while other works like VideoChatCaptioner (Chen et al., 2023) proposes a way to caption videos through chat dialogues between an LLM and a LMM. Recently, there have also been works that use program generation using an LLM to answer questions (Choudhury et al., 2023; Min et al., 2024). However, these works still use a single-pass approach and provide very general captions. In contrast, our work uses an iterative question-answering process to extract specific, relevant information in the frame.
+
+# 3 TravelER Framework
+
+We begin by describing some background on the LLM and LMM architectures (Section 3.1), then introduce each component of our framework (Section 3.2) and implementation details (Section 3.3). Our method is shown in Figure 2.
+
+# 3.1 Preliminaries
+
+Large language and multimodal models. LLMs are text-conditioned generative models. Given a prompt  $P$ , they encode it into a fixed language embedding  $l$  in an embedding space  $f(\cdot)$  and use this to produce text response  $R$ :  $R = f(l(P))$ . Similarly, Large Multimodal Models (LMMs) are adapted to jointly reason over vision and language. To map different modalities into the shared embedding space  $f(\cdot)$ , an image  $I$  is encoded using an encoder  $v$ , and the prompt  $P$  is encoded using a fixed language embedding  $l$ . The LMM outputs a textual response  $R$ :  $R = f(v(I), l(P))$ .
+
+Video question-answering. VideoQA involves viewing a video and answering questions. The model is usually evaluated through top-1 accuracy, which chooses the best answer out of a set of possible choices. Specifically, given a question  $Q$ , video input  $V$  consisting of a set of frames  $\{I_1, \dots, I_n\}$ , and a set of choices  $C = \{c_1, \dots, c_n\}$ , the model is asked to choose the best  $c_i$  to answer  $Q$ . Next, we introduce each component of our method.
+
+# 3.2 TravelER Components
+
+Traversal. In the Traversal stage, we create a plan for how to traverse through the video, which is a list of textual instructions that guide our approach to answering the question. To achieve this, we use the task prompt  $P_T$ , which is an instruction to create a plan for answering the question. We combine  $P_T$  with the question  $Q$ , and memory bank  $M$ , which
+
+![](images/669ac1d66dfbb406a47cf0ef9dcf2356dbda4608aeec6ab99f247b627ffc2cae.jpg)  
+Figure 2: TravelER framework. Our framework consists of four different modules, the Planner, Retriever, Extractor, and Evaluator. The Planner creates a plan and sends it to the Retriever. The Retriever uses the plan to select the next timestamp and sends this to the Extractor. The Extractor captions and generates questions about the timestamp, answers the questions, and saves the output in the memory bank. Finally, the Evaluator determines if there is enough information and if the plan has been followed. If yes, the Evaluator returns the answer, else the existing information is sent back to the Planner to begin a new iteration.
+
+is a dictionary of collected information keyed by timestamps and containing information from the corresponding frame, to receive the final prompt  $P_{\mathrm{T}}^{(1)}$ :  $P_{\mathrm{T}}^{(1)} = {}^{\prime \prime}[Q][M][P_T]$ .
+
+Our method uses a memory bank  $M$  to store collected information, which allows information to persist and to be updated as we proceed through different iterations. We initialize  $M$  with captions of 5 evenly sampled frames throughout the video. We find that this memory initialization gives the model good context for the general idea of the video, and performs better than starting with an empty memory  $M$ . After the first iteration, we add information iteratively using the Extractor module, which we discuss later in this section.
+
+Next, we input the prompt  $P_{\mathrm{T}}^{(1)}$  into  $\mathrm{LLM}_{\mathrm{planner}}$ , which returns response  $R_{T}$ , a step-by-step plan on how to traverse through the video and what information should be collected. This plan is revised in future iterations using collected information.
+
+$$
+R _ {T} = \mathrm {L L M} _ {\text {p l a n n e r}} (l \left(P _ {\mathrm {T}} ^ {(1)}\right))
+$$
+
+Our next step is to use the plan  $R_{T}$  in the Locator stage to locate relevant events and extract the information that we will use to answer the question.
+
+Locator. The Locator is a component that consists of two submodules, the Retriever and the Extractor. The Retriever selects the timestamps of the next frames to view, while the Extractor extracts relevant information from these frames using a question-answering process. Next, we discuss each component in more detail.
+
+(i) Retriever: The Retriever carries out the given plan  $R_{T}$  by selecting which frames to view next. The Retriever is an LLM-based submodule that finds the next timestamp  $t$  to view, given the plan  $R_{T}$ , collected information  $M$ , and video metadata (frame rate, length). The task prompt  $P_{R}$  is an instruction that contains information about the video length and asks which timestamp to view next. Thus, we insert the question  $Q$ , the plan  $R_{T}$ , and the collected information  $M$  into the task prompt  $P_{R}$  to create the new prompt  $P_{\mathrm{R}}^{(1)}$ :  $P_{\mathrm{R}}^{(1)} = [P_R][Q][R_T][M]$ .
+
+Given prompt  $P_{\mathrm{R}}^{(1)}$  , the LLM in the Retriever,  $\mathrm{LLM}_{\mathrm{retriever}}$  returns  $t$  , the next set of timestamps. Then, it retrieves frames  $I_{t}$  at timestamp  $t$
+
+(ii) Extractor: The Extractor is important because it allows us to capture more relevant and question-specific details from the visual input, unlike using only captions. We pass the frames selected by the retriever  $I_{t}$  into the Extractor submodule, which consists of two large models:  $\mathrm{LLM}_{\mathrm{extractor}}$ , to generate context-dependent questions about the frames  $I_{t}$ , and a different vision-language model  $\mathrm{LMM}_{\mathrm{extractor}}$ , which extracts the desired information from the same frames. We note that we use both an LLM and LMM since the LLM is better at reasoning about the plan and the collected information, while the LMM is able to collect visual information requested by the LLM.
+
+We first generate a general caption  $c_t$  for frame  $I_t$  using the LMMextractor. Then, we concatenate the caption  $c_t$ , plan  $R_T$ , and memory  $M$ , and the Extractor task prompt  $P_E$ , which is an instruction that asks to use available information to create 3 questions to ask about the current frame. This results in the new prompt  $P_{\mathrm{E}}^{(1)}: P_{\mathrm{E}}^{(1)} = [P_E][c_t][R_T][M]$ .
+
+Next, we input this new prompt  $P_{\mathrm{E}}^{(1)}$  into the LLM to get a set of questions  $\{q_1, q_2, \dots, q_n\}$  about each frame, where  $n$  is a parameter for how many questions to ask about each frame.
+
+$$
+\left\{q _ {1}, q _ {2}, \dots , q _ {n} \right\} = \mathrm {L L M} _ {\text {e x t r a c t o r}} (l \left(P _ {\mathrm {E}} ^ {(1)}\right))
+$$
+
+where  $l$  is the fixed language embedding.
+
+In this way, the generated questions take into account both the plan  $R_{T}$  and information from past and future frames of the video  $M$ . We then use the frame  $I_{t}$ , and the corresponding questions  $\{q_1,q_2,\dots ,q_n\}$  as input into  $\mathrm{LMM}_{\mathrm{extractor}}$ . The  $\mathrm{LMM}_{\mathrm{extractor}}$  then outputs a set of answers  $\{a_{1},a_{2},\dots ,a_{3}\}$ , where each answer  $a_{i}$  corresponds to the question  $q_{i}$ .
+
+$$
+\left\{a _ {1}, \dots , a _ {n} \right\} = \operatorname {L M M} _ {\text {e x t r a c t o r}} (v (I _ {t}), l (\left\{q _ {1}, \dots , q _ {n} \right\}))
+$$
+
+where  $v$  is the visual encoder.
+
+Finally, to use this collected information in future iterations, we update our memory bank  $M$ . To do this, we use the timestamp  $t$  of  $I_{t}$  as our key and the question-answer pair list as the value, and append this to our memory  $M$ . If the memory bank dictionary  $M$  is too long, we summarize it by using the memory bank as an input to another LLM and instruct it to make the memory bank entries more
+
+concise, while retaining the same keys and format. This output becomes our new memory bank.
+
+Evaluator. The Evaluator decides if there is enough information and determines if the plan has been followed. We concatenate the memory information  $M$ , the plan  $R_{T}$ , the question  $Q$ , and the choices  $C$  with the task prompt  $P_{V}$ . The task prompt  $P_{V}$  is an instruction to evaluate if there is enough information to answer the question and if the given plan  $R_{T}$  has been fulfilled. Thus, we get the new prompt  $P_{\mathrm{V}}^{(1)}: P_{\mathrm{V}}^{(1)} = [P_{\mathrm{V}}][Q][C][R_{T}][M]$ .
+
+We use this prompt  $P_{\mathrm{V}}^{(1)}$  as input into the LLM in the Evaluator,  $\mathrm{LLM}_{\mathrm{evaluator}}$ , which evaluates if there is enough information to answer the question and if the plan has been completely followed. If both are true,  $\mathrm{LLM}_{\mathrm{evaluator}}$  outputs the best choice  $c^*$  to answer the question  $Q$ . Otherwise, it provides an explanation  $E$  on why there is insufficient information and gives this explanation to the Planner to start a new iteration of the process.
+
+Re-planning. After each iteration, if the evaluator decides that there is not enough information to answer the question  $Q$  or if the plan  $P$  has not been completed, the existing memory  $M$  will be provided to the Planner in the next iteration, in addition to an explanation  $E$  for why an answer was not chosen. The Planner then outputs a new plan, restarting the process. We also set a limit on the number of iterations a question can take to prevent infinite loops. After reaching this limit, we force the Evaluator to choose the best choice.
+
+Summarizer. The Summarizer is an optional module used for some datasets to summarize the information. It is given the question  $Q$ , choices  $C$ , memory bank  $M$ , and task prompt  $P_{S}$ , which is an instruction to summarize the information in each timestamp of the memory bank  $M$ . The outputted summary is then used to replace the original memory bank  $M$ . This process helps reduce the amount of information passed to the LLM in future steps, which may struggle with very long inputs.
+
+# 3.3 Implementation Details
+
+Here, we discuss how we implement various components of our framework. More implementation details, such as prompts and dataset-specific details are in the Supplementary in Section B.
+
+Memory bank. We represent past collected information as a Python dictionary, with the timestamp of different frames as keys and a list of extracted
+
+Table 1: Results on Datasets. We show zero-shot results on different datasets. For fair comparisons, we gray out methods with fine-tuned components in their model.  
+
+<table><tr><td colspan="5">(a) NExT-QA</td><td colspan="2">(b) EgoSchema (Full)</td><td colspan="2">(c) STAR</td><td colspan="2">(d) Perception Test</td></tr><tr><td>Model</td><td>Cau.</td><td>Tem.</td><td>Des.</td><td>Avg.</td><td>Model</td><td>Acc.</td><td>Model</td><td>Avg.</td><td>Model</td><td>Acc.</td></tr><tr><td>SeViLA</td><td>61.3</td><td>61.5</td><td>75.6</td><td>63.6</td><td>mPLUG-Owl</td><td>31.1</td><td>SeViLA</td><td>44.6</td><td>SeViLA</td><td>46.2</td></tr><tr><td rowspan="2">MC-ViT-L</td><td rowspan="2">-</td><td rowspan="2">-</td><td rowspan="2">-</td><td rowspan="2">65.0</td><td>InternVideo</td><td>32.1</td><td>Flamingo-9B</td><td>41.8</td><td>MC-ViT-B</td><td>47.0</td></tr><tr><td>LongViViT</td><td>33.3</td><td>InternVideo</td><td>41.6</td><td>MC-ViT-L</td><td>48.1</td></tr><tr><td>InternVideo</td><td>43.4</td><td>48.0</td><td>65.1</td><td>49.1</td><td>Vamos</td><td>48.3</td><td>BLIP-2voting</td><td>40.3</td><td>Flamingo-3B</td><td>43.6</td></tr><tr><td>ViperGPT</td><td>-</td><td>-</td><td>-</td><td>60.0</td><td>LLoVi</td><td>50.3</td><td>BLIP-2concat</td><td>42.2</td><td>LongViViT</td><td>45.7</td></tr><tr><td>ProViQ</td><td>-</td><td>-</td><td>-</td><td>63.8</td><td rowspan="2">TravelLER</td><td rowspan="2">53.3 (+3.0)</td><td rowspan="2">TravelLER</td><td rowspan="2">44.9 (+2.7)</td><td rowspan="2">TravelLER</td><td rowspan="2">50.2 (+4.5)</td></tr><tr><td>LLoVi</td><td>69.5</td><td>61.0</td><td>75.6</td><td>67.7</td></tr><tr><td>TravelLER</td><td>70.0</td><td>60.5</td><td>78.2</td><td>68.2 (+0.5)</td><td></td><td></td><td></td><td></td><td></td><td></td></tr></table>
+
+information from the frame as the values. This extracted information consists of a brief caption of the frame and a list of question-answer pairs. To prevent the memory bank from becoming too large, we also implement a summarizer module that instructs an LLM to summarize the memory bank and return a more concise version in the same dictionary format as before.
+
+Agent model selection. Our modular approach has the benefit of allowing us to easily swap in different LLMs and LMMs (see Section 4.4). For our main experiments, we use LLaVA-1.6 (Liu et al., 2023) for  $\mathrm{LMM}_{\mathrm{extractor}}$  and GPT-3.5/GPT-4 (OpenAI et al., 2023) for  $\mathrm{LLM}_{\mathrm{planner}}$ ,  $\mathrm{LLM}_{\mathrm{retriever}}$ ,  $\mathrm{LLM}_{\mathrm{extractor}}$ , and  $\mathrm{LLM}_{\mathrm{evaluator}}$ .
+
+Multi-frame selection. We also allow the Retriever to select multiple frames instead of one. This helps to capture better events that happen quickly or require more context to recognize. For example, if we want to find the action of "a woman clapping her hands", a single frame selection may cause us to incorrectly assume the woman is not clapping if we view the frame where their hands are apart. We do this by creating an optional parameter called window size. The window size refers to the number of frames the Retriever extracts each time. When the window size is non-zero, the Retriever still specifies a single timestamp to go to, but when retrieving the frame at that timestamp, we also take the number of frames specified by the window size before and after the selected frame.
+
+# 4 Evaluation
+
+We evaluated our TraveLER framework on several benchmarks described in Section 4.1, and compared it to multiple baselines in Section 4.2. The results and ablations are in Section 4.3 and Sec
+
+tion 4.4. Additional results and ablations are in the Supplementary in Section A.
+
+# 4.1 Datasets
+
+We use the following datasets: (1) NExT-QA (Xiao et al., 2021) is a dataset that tests causal action reasoning and temporal understanding. Following the trend of works before us, we evaluate our method on the 5,000 questions in the NExT-QA validation set. (2) EgoSchema (Mangalam et al., 2023) is a challenging dataset that tests long-form video understanding. Viewers need to view 100 seconds of the video on average to answer the question correctly. (3) STAR (Wu et al., 2021) tests reasoning in real-world video situations. (4) Perception Test (Patraucean et al., 2023) is a challenging dataset that focuses on skills such as memory, abstraction, physics, and semantics and is intended to be approached in a few-shot or zero-shot manner.
+
+# 4.2 Baselines
+
+In our experiments, we compare our method to recent state-of-the-art zero-shot (ZS) methods, such as LLoVi (Zhang et al., 2023a), ProViQ (Choudhury et al., 2023), and other methods that are not necessarily ZS, such as SeViLA (Yu et al., 2023), and MC-ViT (Balazevic et al., 2024). We note that SeViLA uses fine-tuned components on QVHighlights (Lei et al., 2021a), while MC-ViT is fine-tuned on NExT-QA for Perception Test. Additional baselines are in Supplementary in Section A.2.
+
+# 4.3 Results
+
+Our results are shown in Table 1. We use GPT-4 for NExT-QA to ensure a fair comparison with LLoVi, which is the current state-of-the-art that uses GPT-4 to uniformly caption frames across the entire video. Interestingly, our method outperforms
+
+LLoVi despite viewing  $50\%$  fewer frames on average. Second, we also outperform SeViLA by  $+4.6\%$ , although SeViLA uses a keyframe selector that is fine-tuned on a video moment retrieval and grounding task while our method is fully ZS.
+
+For EgoSchema, we use GPT-4 with a retrained version of LaViLa that excludes overlapping Ego4D and EgoSchema videos to prevent data leakage. We show strong performance on long-form videos, where we outperform LLoVi by  $+3.0\%$ , while viewing  $95\%$  fewer clips on average.
+
+We use GPT-3.5 for Perception Test and STAR because it is cheaper than GPT-4, but results are likely to be improved even further with GPT-4. Nevertheless, we achieve higher accuracy than LongViViT on Perception Test by  $+4.5\%$  and MCViT by  $+2.1\%$ , although it was fine-tuned on NExTQA. We surpass both the best zero-shot approach by  $+2.7\%$  and the best fine-tuned result by  $+0.3\%$ .
+
+Finally, please refer to Section A for more experimental and ablation results, and Section D for additional visualizations. For example, in Figure 8, we see that question-answering is able to extract more relevant details in comparison to simple captioning. This may explain why our method significantly outperforms the descriptive split of NExT-QA. In Figure 7, we see our method is able to reason about vague references and correctly identify relevant objects through question-answering.
+
+# 4.4 Agent Ablations
+
+We perform ablations using 1000 randomly selected questions from the NExT-QA training set (see Table 2). Unless specified, we use GPT-3.5/LLaVA-1.6 as the LLM/LMM for all agents.
+
+Ablating the Planner. The Planner module outputs a plan, a list of instructions that guides the behavior of all other modules. We test the impact of removing it from our framework and find that it is worse. We hypothesize this is because the Planner provides many temporal cues that guide the Retriever module's search, such as "go to the middle of the video", and without these cues, the Retriever is not as good at selecting the next timestamp. Moreover, the Planner also helps the Evaluator better decide when to stop since in our iterative approach, the Evaluator uses the plan to determine when to stop. We also try removing re-planning by only running the Planner once in the beginning and keeping this plan fixed throughout. We find that this also reduces performance, showing the importance of adjusting plans to new information.
+
+Table 2: Ablations on the agents in our framework.  
+
+<table><tr><td>Agent</td><td>Ablation</td><td>Avg. (%)</td></tr><tr><td>——</td><td>Baseline</td><td>60.4</td></tr><tr><td rowspan="2">Planner</td><td>Removal</td><td>58.1</td></tr><tr><td>Fixed Plan</td><td>58.1</td></tr><tr><td>Retriever</td><td>Removal</td><td>56.9</td></tr><tr><td>Extractor</td><td>Caption Only</td><td>58.2</td></tr><tr><td>Evaluator</td><td>No Iteration</td><td>56.8</td></tr><tr><td>Summarizer</td><td>Removal</td><td>57.2</td></tr></table>
+
+Ablating the Retriever. The Retriever module determines the next timestamps to view, which helps focus our information collection. We ablate it by uniformly sampling frames from the video at 2-second intervals, similar to other methods like LLoVi, which performs worse. We believe this is because the Retriever allows us to capture frames that might have otherwise been skipped through uniform sampling and selects fewer unimportant frames that might mislead the model.
+
+Ablating the Extractor. Question asking is important as it allows us to capture more fine-grained and question-relevant information compared to simple caption generation, which produces a generic description. Thus, we ablate the Extractor by only allowing the LMM to caption frames. We find that this decreases performance by  $-2.2\%$ , suggesting that the ability to ask specific questions about a frame is important. We notice many generated captions capture the main idea of visual information in the frame, but are lacking in fine-grained details.
+
+Ablating the Evaluator. The Evaluator reflects on the collected information and decides if there is enough information to answer the question. To examine the impact of this reflection process, we make the Evaluator answer on the first iteration, finding it results in a  $-3.6\%$  performance decrease.
+
+Ablating the Summarizer. When collecting large amounts of information from videos, we use a Summarizer to condense the information, since long inputs can be challenging for LLMs. This has also been observed in recent work (Zhang et al., 2023a). To understand the impact of the Summarizer, we remove it. The results indicate that removing it degrades performance by  $-3.2\%$ , demonstrating the advantage of more concise information.
+
+# 4.5 Additional Experiments
+
+Substituting different LLMs/LMMs. To see how the choice of the LLM and LMM affects our framework's performance, we swap different LLMs and
+
+Table 3: Ablation Results. We perform ablations on 1000 randomly selected questions from the NExT-QA training set. We report (a), (b) replacing different LLMs and LMMs, (c) selecting different numbers of frames to view in the Retriever, and (d) changing the number of questions asked in the Extractor. We use GPT-3.5/LLaVA-1.6 as the LLM/LMM, 5 frames for the Retriever, and 3 questions for the Extractor, unless otherwise specified.  
+(a) Replacing Diff. LLMs  
+
+<table><tr><td>LLM</td><td>LMM</td><td>Accuracy</td></tr><tr><td>GPT-3.5</td><td></td><td>60.4</td></tr><tr><td>Llama 3</td><td rowspan="2">LLaVA-1.6</td><td>63.9</td></tr><tr><td>GPT-4</td><td>65.8</td></tr><tr><td>GPT-4o</td><td></td><td>68.0</td></tr></table>
+
+(b) Replacing Diff. LMMs  
+
+<table><tr><td>LLM</td><td>LMM</td><td>Accuracy</td></tr><tr><td>GPT-4</td><td>GPT-4V</td><td>64.7</td></tr><tr><td></td><td>BLIP-2</td><td>52.7</td></tr><tr><td>GPT-3.5</td><td>GPT-4V</td><td>59.5</td></tr><tr><td></td><td>LLaVA-1.6</td><td>60.4</td></tr></table>
+
+(c) # of Frames  
+
+<table><tr><td>#Frames</td><td>Accuracy</td></tr><tr><td>1</td><td>59.0</td></tr><tr><td>3</td><td>57.9</td></tr><tr><td>5</td><td>60.4</td></tr><tr><td>7</td><td>59.0</td></tr></table>
+
+(d) # of Questions  
+
+<table><tr><td># Questions</td><td>Accuracy</td></tr><tr><td>0</td><td>58.1</td></tr><tr><td>1</td><td>58.4</td></tr><tr><td>3</td><td>60.4</td></tr><tr><td>5</td><td>59.6</td></tr></table>
+
+LMMs into our framework (see Table 3a and Table 3b). We first try different LLMs while fixing the LMM to be LLaVA-1.6. While GPT-4 performs better than GPT-3.5 by a significant margin of  $+5.4\%$ , open-source model Llama 3 is very close  $(-1.9\%)$ , while incurring no additional cost. We also evaluate the newly released GPT-4o, which outperforms GPT-4 by  $+2.2\%$  while being  $61\%$  cheaper, showing that our method can leverage better future models. Second, we use different LMMs while fixing the LLM to be GPT-3.5. We find that LLaVA-1.6 does best, GPT-4V is slightly worse  $(-0.9\%)$ , and BLIP-2 is significantly worse  $(-7.7\%)$ . Finally, we run an experiment using GPT-4V as both the LLM and LMM, and find that this does worse than GPT-4 and LLaVA-1.6 by  $-1.1\%$ .
+
+GPT-4V Baseline. To get a baseline using GPT-4V, we use a method similar to LLoVi using a subset of 500 examples. We use GPT-4V to caption the video uniformly, then ask it to answer the question given the choices, which results in a performance change of  $-2.0\%$ . We note our motivation is to refrain from captioning every single frame, instead finding frames that help us answer the question.
+
+Retriever window size. We experiment with different window sizes in Table 3c which is the number of frames the Retriever extracts centered around the selected frame. This allows us to capture better actions that occur quickly or require more context to understand. We find that choosing 5 frames yields the best results and a  $+1.4\%$  increase when compared to selecting a single frame, but viewing more than 5 decreases performance. This suggests retrieving multiple frames can help the model better capture relevant information, but retrieving too many frames can lead to too much information.
+
+Memory Initialization. In order for the Planner to create effective plans, it is beneficial to initialize the memory bank properly. Memory initialization
+
+![](images/3cf55014e5fd03c64a502ebaadd95dce4307d88fda75b9413c3b6e2101d42441.jpg)  
+Figure 3: Comparison of different Memory Initialization (1, 3, 5 frames). 5 frames is optimal.
+
+allows the Planner to have a high-level overview of the video, and create a corresponding plan on how to traverse the video given the initial frames. We perform three different initializations with 1, 3, 5, 7, and 9 frames and display our results in Figure 3. We observe that initializing the memory bank with 5 frames uniformly sampled from the video (0, 0.25, 0.5, 0.75, 1 for beginning, quarter, middle, three-quarter, end) yields the best result. In contrast, we notice a decrease in accuracy of  $-2.9\%$  when using 9 frames,  $-3.1\%$  when using 7 frames,  $-0.6\%$  when initializing with 3 frames (0, 0.5, 1 for beginning, middle, end), and  $-2.5\%$  when initializing with 1 frame (0.5 for middle).
+
+Robustness to choosing incorrect frames. To see if our approach can recover from viewing incorrect frames, we choose random frames for the first 3 iterations (out of 5 total iterations) before using the Retriever. This leads to an accuracy drop of only  $2.8\%$ , showing our method is capable of recovering from viewing incorrect frames.
+
+The number of questions. Question answering allows us to extract more specific details from our visual inputs. However, we noticed that too many questions can yield irrelevant questions and false positives. As such, we experiment with modifying
+
+Table 4: Comparison of TraveLER with other few-shot and zero-shot keyframe localization methods. For fair comparisons, we gray out methods with fine-tuned components in their model. The best scores are in bold.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="4">NExT-QA (Random Subset)</td></tr><tr><td>Temporal</td><td>Causal</td><td>Descriptive</td><td>Average</td></tr><tr><td>SeViLA - Localizer (Yu et al., 2023)</td><td>48.8</td><td>61.2</td><td>68.3</td><td>58.2</td></tr><tr><td>Moment-DETR (Lei et al., 2021a)</td><td>45.3</td><td>55.8</td><td>70.8</td><td>54.6</td></tr><tr><td>SigLIP (Zhai et al., 2023)</td><td>48.4</td><td>61.5</td><td>73.8</td><td>59.1</td></tr><tr><td>TraveLER - Planner &amp; Retriever (ours)</td><td>50.9</td><td>62.7</td><td>72.4</td><td>60.3</td></tr></table>
+
+the number of questions asked for each frame by our extractor (see Table 3d). We record results for a 5-question, 3-question, and 1-question maximum. Note that 0-questions asked is equivalent to only allowing captions, which is discussed in the Extractor ablation. From our results, we find that a 3-question limit yields the best results compared to asking 1 or 5 questions  $(+2.0 / +0.8\%)$ . This suggests that asking questions helps in extracting relevant information, but too many questions can lead to false positives or too much irrelevant information.
+
+Memory bank initialization and formatting. Our memory bank  $M$  stores information that all modules rely on to make decisions. First, we experiment with different initializations as  $M$  must be initialized in the first iteration. We experiment with initialization of 1, 3, and 5 uniformly sampled captions. We find using 5 evenly spaced frames yields the best results, possibly because it starts the model with a general overview of the video before it starts to collect more relevant information. We also experiment with changing the format of the memory bank from JSON to markdown tables. We find this does worse by  $-2.9\%$ , possibly because LLMs are better able to understand JSON formats.
+
+Comparison with other keyframe selection methods. In Table 4, we compare our Planner and Retriever with other keyframe localization methods by replacing our Planner and Retriever with each of the other methods, and use our Extractor and Evaluator to perform the question answering. For all these methods, we use GPT-3.5 and LLaVA-1.6, and we evaluate these methods on a random subset of 1000 examples from the training set of NExTQA. Note that other methods find keyframes in one inference iteration, whereas our inference occurs over multiple iterations. Therefore, to ensure fair comparisons, we uniformly sample 32 frames and extract out 4 keyframes in the other methods,
+
+and we run 4 iterations of TravelLER using the Retriever with a window size of 2 to similarly find 4 keyframes among fewer than 32 viewed frames.
+
+We find that our Planner and Retriever surpasses other keyframe localization methods, despite considering fewer total frames (\(\sim\)25 total frames; we have 5 for memory initialization, and up to 5 frames each iteration). We would like to highlight that while our method is effective at finding keyframes, we do not need to find all keyframes to answer a question. Instead, we are often able to choose the correct answer with only a subset of the keyframes. Cost of inference. We test costs for GPT-3.5 and GPT-4, which have costs per inference of \(0.03\) and \)0.67$, respectively. We also test the open-source model Llama 3, which is free to run with only a minor performance decrease (-1.9%). Finally, we also try the new GPT-4o model, which is both better (+2.2%) and cheaper (61%) than GPT-4. We believe this trend will only continue as models become better and cheaper in the future.
+
+# 5 Conclusion
+
+We design a modular, multi-LMM agent framework for video-question answering based on several agents with different roles, instructed by a Planner agent that updates its instructions using shared feedback between the other agents. Our method creates a plan to "traverse" through the video, asking questions about individual frames to "locate" and store key information, and then "evaluate" if there is enough information to answer the question, "replanning" using new feedback if necessary. Through extensive experiments and ablations, we find that the proposed TravelER approach is not only easy to employ with different models but also improves performance on several video question-answering benchmarks without the need to fine-tune on specific datasets.
+
+# 6 Limitations
+
+In this work, we present a modular, zero-shot framework for video question answering (VideoQA) and demonstrate its effectiveness by improving on multiple state-of-the-art benchmarks. While TraveLER offers substantial benefits for VideoQA, it is important to recognize certain limitations that accompany our approach. Firstly, the effectiveness of our model relies heavily on the strength of the LLM and LMM. We notice that false-positives and incorrect statements from the LMM heavily impact performance. Our method's runtime also depends on the runtime of current existing methods, and with the modularity of our method we expect this to improve with faster and better models in the future. Finally, we do not anticipate negative impacts of this work, but, as with any Machine Learning method, we recommend exercising caution.
+
+# 7 Acknowledgements
+
+We would like to thank Suzie Petryk, Chancharik Mitra, Alon Mendelson, David Chan, Assaf Arbelle, and Leonid Karlinsky for helpful feedback and discussions. This project was supported in part by DoD, including PTG and/or LwLL programs, as well as BAIR's industrial alliance programs.
+
+# References
+
+Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. 2022. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems.  
+Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. 2016. Neural module networks. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 39-48.  
+Elad Ben Avraham, Roei Herzig, Karttikeya Mangalam, Amir Bar, Anna Rohrbach, Leonid Karlinsky, Trevor Darrell, and Amir Globerson. 2022. Bringing image scene structure to video via frame-clip consistency of object tokens. In Thirty-Sixth Conference on Neural Information Processing Systems.  
+Ivana Balazevic, Yuge Shi, Pinelopi Papalampidi, Rahma Chaabouni, Skanda Koppula, and Olivier J
+
+Henaff. 2024. Memory consolidation enables long-context video understanding. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pages 2527-2542. PMLR.  
+Amir Bar, Roei Herzig, Xiaolong Wang, Anna Rohrbach, Gal Chechik, Trevor Darrell, and A. Globerson. 2021. Compositional video synthesis with action graphs. In ICML.  
+Fabien Baradel, Natalia Neverova, Christian Wolf, Julien Mille, and Greg Mori. 2018. Object level visual reasoning in videos. In European Conference on Computer Vision, pages 105-121.  
+Peter Battaglia, Jessica Blake Chandler Hamrick, Victor Bapst, Alvaro Sanchez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner, Caglar Gulcehre, Francis Song, Andy Ballard, Justin Gilmer, George E. Dahl, Ashish Vaswani, Kelsey Allen, Charles Nash, Victoria Jayne Langston, Chris Dyer, Nicolas Heess, Daan Wierstra, Pushmeet Kohli, Matt Botvinick, Oriol Vinyals, Yujia Li, and Razvan Pascanu. 2018. Relational inductive biases, deep learning, and graph networks. arXiv.  
+Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems, volume 33, pages 1877-1901. Curran Associates, Inc.  
+Shyamal Buch, Cristobal Eyzaguirre, Adrien Gaidon, Jiajun Wu, Li Fei-Fei, and Juan Carlos Niebles. 2022. Revisiting the "Video" in Video-Language Understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR).  
+Jun Chen, Deyao Zhu, Kilichbek Haydarov, Xiang Li, and Mohamed Elhoseiny. 2023. Video chat captioner: Towards enriched spatiotemporal descriptions. Preprint, arXiv:2304.04227.  
+Sihan Chen, Handong Li, Qunbo Wang, Zijia Zhao, Mingzhen Sun, Xinxin Zhu, and Jing Liu. 2024. Vast: A vision-audio-subtitle-text omni-modality foundation model and dataset. Advances in Neural Information Processing Systems, 36.  
+Rohan Choudhury, Koichiro Niinuma, Kris M. Kitani, and László A. Jeni. 2023. Zero-shot video question answering with procedural programs. Preprint, arXiv:2312.00937.
+
+Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Alex Castro-Ros, Marie Pellat, Kevin Robinson, Dasha Valter, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei. 2022. Scaling instruction-finetuned language models. Preprint, arXiv:2210.11416.  
+Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. 2023. An empirical study of end-to-end videolanguage transformers with masked visual modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22898-22909.  
+Jiaxin Ge, Sanjay Subramanian, Baifeng Shi, Roei Herzig, and Trevor Darrell. 2023. Recursive visual programming. Preprint, arXiv:2312.02249.  
+Tanmay Gupta and Aniruddha Kembhavi. 2023. Visual programming: Compositional visual reasoning without training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14953-14962.  
+Roei Herzig, Ofir Abramovich, Elad Ben Avraham, Assaf Arbelle, Leonid Karlinsky, Ariel Shamir, Trevor Darrell, and Amir Globerson. 2024. Promptomyvit: Multi-task prompt learning improves video transformers using synthetic scene data. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 6803-6815.  
+Roei Herzig, Amir Bar, Huijuan Xu, Gal Chechik, Trevor Darrell, and Amir Globerson. 2020. Learning canonical representations for scene graph to image generation. In European Conference on Computer Vision.  
+Roei Herzig, Elad Ben-Avraham, Karttikeya Mangalam, Amir Bar, Gal Chechik, Anna Rohrbach, Trevor Darryll, and Amir Globerson. 2022. Object-region video transformers. In Conference on Computer Vision and Pattern Recognition (CVPR).  
+Roei Herzig, Elad Levi, Huijuan Xu, Hang Gao, Eli Brosh, Xiaolong Wang, Amir Globerson, and Trevor Darrell. 2019. Spatio-temporal action graph networks. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 0-0.  
+Roei Herzig, Alon Mendelson, Leonid Karlinsky, Assaf Arbelle, Rogerio Feris, Trevor Darrell, and Amir Globerson. 2023. Incorporating structured representations into pretrained vision  $\&$  language models using scene graphs. In The 2023 Conference on Empirical Methods in Natural Language Processing.
+
+Roei Herzig, Moshiko Raboh, Gal Chechik, Jonathan Berant, and Amir Globerson. 2018. Mapping images to scene graphs with permutation-invariant structured prediction. In Advances in Neural Information Processing Systems (NIPS).  
+Achiya Jerbi, Roei Herzig, Jonathan Berant, Gal Chechik, and Amir Globerson. 2020. Learning object detection from captions via textual scene attributes. ArXiv, abs/2009.14558.  
+Ranjay Krishna, Ines Chami, Michael S. Bernstein, and Li Fei-Fei. 2018. Referring relationships. European Conference on Computer Vision.  
+Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient memory management for large language model serving with pagedattention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles.  
+Jie Lei, Tamara Lee Berg, and Mohit Bansal. 2021a. Qvhighlights: Detecting moments and highlights in videos via natural language queries. In Advances in Neural Information Processing Systems.  
+Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L. Berg, Mohit Bansal, and Jingjing Liu. 2021b. Less is more: Clipbert for video-and-language learning via sparse sampling. In CVPR.  
+Jiangtong Li, Li Niu, and Liqing Zhang. 2022. From representation to reasoning: Towards both evidence and commonsense reasoning for video question-answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR).  
+Linjie Li, Zhe Gan, Kevin Lin, Chung-Ching Lin, Ce Liu, Zicheng Liu, and Lijuan Wang. 2023. Lavender: Unifying video-language understanding as masked language modeling. In Conference on Computer Vision and Pattern Recognition (CVPR).  
+Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. 2023a. Video-llava: Learning united visual representation by alignment before projection. Preprint, arXiv:2311.10122.  
+Kevin Lin, Faisal Ahmed, Linjie Li, Chung-Ching Lin, Ehsan Azarnasab, Zhengyuan Yang, Jianfeng Wang, Lin Liang, Zicheng Liu, Yumao Lu, Ce Liu, and Lijuan Wang. 2023b. Mm-vid: Advancing video understanding with gpt-4v(ison). Preprint, arXiv:2310.19773.  
+Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual instruction tuning. In Thirty-seventh Conference on Neural Information Processing Systems.  
+Haoyu Lu, Mingyu Ding, Nanyi Fei, Yuqi Huo, and Zhiwu Lu. 2022. LGDN: Language-guided denoising network for video-language modeling. In Advances in Neural Information Processing Systems.
+
+Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. 2024. Video-chatgpt: Towards detailed video understanding via large vision and language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024).  
+Tegan Maharaj, Nicolas Ballas, Anna Rohrbach, Aaron Courville, and Christopher Pal. 2017. A dataset and exploration of models for understanding video data through fill-in-the-blank question-answering. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 7359-7368.  
+Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. 2023. Egoschema: A diagnostic benchmark for very long-form video language understanding. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track.  
+Juhong Min, Shyamal Buch, Arsha Nagrani, Minsu Cho, and Cordelia Schmid. 2024. Morevqa: Exploring modular reasoning models for video question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR).  
+Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. 2024. Compositional chain of thought prompting for large multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR).  
+OpenAI, :, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mo Bavarian, Jeff Belgium, Irwan Bello, Jake Berdine, Gabriel Bernadett-Shapiro, Christopher Berner, Lenny Bogdonoff, Oleg Boiko, Madelaine Boyd, Anna-Luisa Brakman, Greg Brockman, Tim Brooks, Miles Brundage, Kevin Button, Trevor Cai, Rosie Campbell, Andrew Cann, Brittany Carey, Chelsea Carlson, Rory Carmichael, Brooke Chan, Che Chang, Fotis Chantzis, Derek Chen, Sully Chen, Ruby Chen, Jason Chen, Mark Chen, Ben Chess, Chester Cho, Casey Chu, Hyung Won Chung, Dave Cummings, Jeremiah Currier, Yunxing Dai, Cory Decareaux, Thomas Degry, Noah Deutsch, Damien Deville, Arka Dhar, David Dohan, Steve Dowling, Sheila Dunning, Adrien Ecoffet, Atty Eleti, Tyna Eloundou, David Farhi, Liam Fedus, Niko Felix, Simón Posada Fishman, Juston Forte, Isabella Fulford, Leo Gao, Elie Georges, Christian Gibson, Vik Goel, Tarun Gogineni, Gabriel Goh, Rapha Gontijo-Lopes, Jonathan Gordon, Morgan Grafstein, Scott Gray, Ryan Greene, Joshua Gross, Shixiang Shane Gu, Yufei Guo, Chris Hallacy, Jesse Han, Jeff Harris, Yuchen He, Mike Heaton, Johannes Heidecke, Chris Hesse, Alan Hickey, Wade Hickey, Peter Hoeschele, Brandon Houghton, Kenny Hsu, Shengli Hu, Xin Hu, Joost Huizinga, Shantanu Jain, Shawn Jain, Joanne Jang, Angela Jiang, Roger Jiang, Haozhun Jin, Denny Jin, Shino Jomoto, Billie
+
+Jonn, Heewoo Jun, Tomer Kaftan, Lukasz Kaiser, Ali Kamali, Ingmar Kanitscheider, Nitish Shirish Keskar, Tabarak Khan, Logan Kilpatrick, Jong Wook Kim, Christina Kim, Yongjik Kim, Hendrik Kirchner, Jamie Kiros, Matt Knight, Daniel Kokotajlo, Lukasz Kondraciuk, Andrew Kondrich, Aris Konstantinidis, Kyle Kosic, Gretchen Krueger, Vishal Kuo, Michael Lampe, Ikai Lan, Teddy Lee, Jan Leike, Jade Leung, Daniel Levy, Chak Ming Li, Rachel Lim, Molly Lin, Stephanie Lin, Mateusz Litwin, Theresa Lopez, Ryan Lowe, Patricia Lue, Anna Makanju, Kim Malfacini, Sam Manning, Todor Markov, Yaniv Markovski, Bianca Martin, Katie Mayer, Andrew Mayne, Bob McGrew, Scott Mayer McKinney, Christine McLeavey, Paul McMillan, Jake McNeil, David Medina, Aalok Mehta, Jacob Menick, Luke Metz, Andrey Mishchenko, Pamela Mishkin, Vinnie Monaco, Evan Morikawa, Daniel Mossing, Tong Mu, Mira Murati, Oleg Murk, David Mely, Ashvin Nair, Reiichiro Nakano, Rajeev Nayak, Arvind Neelakantan, Richard Ngo, Hyeonwoo Noh, Long Ouyang, Cullen O'Keefe, Jakub Pachocki, Alex Paino, Joe Palermo, Ashley Pantuliano, Giambattista Parascandolo, Joel Parish, Emy Parparita, Alex Passos, Mikhail Pavlov, Andrew Peng, Adam Perelman, Filipe de Avila Belbute Peres, Michael Petrov, Henrique Ponde de Oliveira Pinto, Michael Pokorny, Michelle Pokrass, Vitchyr Pong, Tolly Powell, Alethea Power, Boris Power, Elizabeth Proehl, Raul Puri, Alec Radford, Jack Rae, Aditya Ramesh, Cameron Raymond, Francis Real, Kendra Rimbach, Carl Ross, Bob Rotsted, Henri Roussez, Nick Ryder, Mario Saltarelli, Ted Sanders, Shibani Santurkar, Girish Sastry, Heather Schmidt, David Schnurr, John Schulman, Daniel Selsam, Kyla Sheppard, Toki Sherbakov, Jessica Shieh, Sarah Shoker, Pranav Shyam, Szymon Sidor, Eric Sigler, Maddie Simens, Jordan Sitkin, Katarina Slama, Ian Sohl, Benjamin Sokolowsky, Yang Song, Natalie Staudacher, Felipe Petroski Such, Natalie Summers, Ilya Sutskever, Jie Tang, Nikolas Tezak, Madeleine Thompson, Phil Tillet, Amin Tootoonchian, Elizabeth Tseng, Preston Tuggle, Nick Turley, Jerry Tworek, Juan Felipe Cerón Uribe, Andrea Vallone, Arun Vijayvergiya Chelsea Voss, Carroll Wainwright, Justin Jay Wang. Alvin Wang. Ben Wang. Jonathan Ward. Jason Wei CJ Weinmann Akila Welihinda Peter Welinder Jiayi WengLilian WengMatt Wiethoff Dave Willner Clemens Winter,Samuel WolrichHannah Wong Lauren Workman. Sherwin Wu. Jeff Wu. Michael Wu. Kai Xiao Tao Xu Sarah Yoo Kevin Yu Qiming Yuan Wojciech Zaremba Rowan Zellers Chong Zhang. Marvin Zhang. Shengjia Zhao. Tianhao Zheng. Juntang Zhuang. William Zhuk and Barret Zoph. 2023. Gpt-4 technical report. Preprint arXiv:2303.08774.
+
+Pinelopi Papalampidi, Skanda Koppula, Shreya Pathak, Justin Chiu, Joe Heyward, Viorica Patraucean, Jiajun Shen, Antoine Miech, Andrew Zisserman, and Aida Nematzdeh. 2024. A simple recipe for contrastively pre-training video-first encoders beyond 16 frames. In Proceedings of the IEEE/CVF Conference on Com
+
+puter Vision and Pattern Recognition (CVPR), pages 14386-14397.  
+Viorica Patraucean, Lucas Smaira, Ankush Gupta, Adria Recasens Continente, Larisa Markeeva, Dylan Sunil Banarse, Skanda Koppula, Joseph Heyward, Mateusz Malinowski, Yi Yang, Carl Doersch, Tatiana Matejovicova, Yury Sulsky, Antoine Miech, Alexandre Fréchette, Hanna Klimczak, Raphael Koster, Junlin Zhang, Stephanie Winkler, Yusuf Aytar, Simon Osindero, Dima Damen, Andrew Zisserman, and Joao Carreira. 2023. Perception test: A diagnostic benchmark for multimodal video models. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track.  
+Tianwen Qian, Ran Cui, Jingjing Chen, Pai Peng, Xiaowei Guo, and Yu-Gang Jiang. 2023. Locate before answering: Answer guided question localization for video question answering. Trans. Multi., 26:4554-4563.  
+Hung-Ting Su, Yulei Niu, Xudong Lin, Winston H. Hsu, and Shih-Fu Chang. 2023. Language models are causal knowledge extractors for zero-shot video question answering. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 4951-4960.  
+Sanjay Subramanian, Medhini Narasimhan, Kushal Khangaonkar, Kevin Yang, Arsha Nagrani, Cordelia Schmid, Andy Zeng, Trevor Darrell, and Dan Klein. 2023. Modular visual question answering via code generation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics, Toronto, Canada. Association for Computational Linguistics.  
+Yuchong Sun, Hongwei Xue, Ruihua Song, Bei Liu, Huan Yang, and Jianlong Fu. 2022. Long-form video-language pre-training with multimodal temporal contrastive learning. In Advances in Neural Information Processing Systems, volume 35, pages 38032-38045. Curran Associates, Inc.  
+Dídac Surís, Sachit Menon, and Carl Vondrick. 2023. Vipergpt: Visual inference via python execution for reasoning. Proceedings of IEEE International Conference on Computer Vision (ICCV).  
+Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, Ping Luo, Ziwei Liu, Yali Wang, Limin Wang, and Yu Qiao. 2024. Internvid: A large-scale video-text dataset for multimodal understanding and generation. In The Twelfth International Conference on Learning Representations.  
+Zhenhailong Wang, Manling Li, Ruochen Xu, Luowei Zhou, Jie Lei, Xudong Lin, Shuohang Wang, Ziyi Yang, Chenguang Zhu, Derek Hoiem, Shih-Fu Chang, Mohit Bansal, and Heng Ji. 2022. Language models with image descriptors are strong few-shot video-language learners. In Advances in Neural Information Processing Systems, volume 35, pages 8483-8497. Curran Associates, Inc.
+
+Bo Wu, Shoubin Yu, Zhenfang Chen, Joshua B Tenenbaum, and Chuang Gan. 2021. Star: A benchmark for situated reasoning in real-world videos. In Thirty-fifth Conference on Neural Information Processing Systems (NeurIPS).  
+Wenhao Wu, Dongliang He, Xiao Tan, Shifeng Chen, and Shilei Wen. 2019. Multi-agent reinforcement learning based frame sampling for effective untrimmed video recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV).  
+Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. 2021. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9777-9786.  
+Hongwei Xue, Yuchong Sun, Bei Liu, Jianlong Fu, Ruihua Song, Houqiang Li, and Jiebo Luo. 2023. CLIPvip: Adapting pre-trained image-text model to video-language alignment. In The Eleventh International Conference on Learning Representations.  
+Hongyang Xue, Zhou Zhao, and Deng Cai. 2017. Unifying the video and question attentions for open-ended video question answering. IEEE Transactions on Image Processing, 26(12):5656-5666.  
+Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. 2021. Just ask: Learning to answer questions from millions of narrated videos. In ICCV.  
+Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. 2022. Zero-shot video question answering via frozen bidirectional language models. In NeurIPS.  
+Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. 2023. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In CVPR.  
+Qinghao Ye, Guohai Xu, Ming Yan, Haiyang Xu, Qi Qian, Ji Zhang, and Fei Huang. 2023. Hitea: Hierarchical temporal-aware video-language pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15405-15416.  
+Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. 2022. Coca: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research.  
+Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. 2023. Self-chained image-language model for video localization and question answering. In NeurIPS.
+
+Andy Zeng, Maria Attarian, brian richter, Krzysztof Marcin Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael S Ryoo, Vikas Sindhwani, Johnny Lee, Vincent Vanhoucke, and Pete Florence. 2023. Socratic models: Composing zero-shot multimodal reasoning with language. In The Eleventh International Conference on Learning Representations.  
+Kuo-Hao Zeng, Tseng-Hung Chen, Ching-Yao Chuang, Yuan-Hong Liao, Juan Carlos Niebles, and Min Sun. 2017. Leveraging video descriptions to learn video question answering. Proceedings of the AAAI Conference on Artificial Intelligence, 31(1).  
+Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. 2023. Sigmoid loss for language image pre-training. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 11941-11952.  
+Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. 2023a. A simple llm framework for long-range video question-answering. Preprint, arXiv:2312.17235.  
+Chaoyi Zhang, Kevin Lin, Zhengyuan Yang, Jianfeng Wang, Linjie Li, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. 2024. Mm-narrator: Narrating long-form videos with multimodal in-context learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR).  
+Hang Zhang, Xin Li, and Lidong Bing. 2023b. VideoLLaMA: An instruction-tuned audio-visual language model for video understanding. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 543–553, Singapore. Association for Computational Linguistics.  
+Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. 2023. Learning video representations from large language models. In CVPR.  
+Lianmin Zheng, Liangsheng Yin, Zhiqiang Xie, Jeff Huang, Chuyue Sun, Cody Hao Yu, Shiyi Cao, Christos Kozyrakis, Ion Stoica, Joseph E. Gonzalez, Clark Barrett, and Ying Sheng. 2023. Efficiently programming large language models using sglang. Preprint, arXiv:2312.07104.  
+Linchao Zhu, Zhongwen Xu, Yi Yang, and Alexander G. Hauptmann. 2017. Uncovering the temporal context for video question answering. International Journal of Computer Vision, 124(3):409-421.
+
+# Supplementary Material for "TraveLER"
+
+Here we provide additional information about our experimental results, qualitative examples, implementation details, and datasets. Specifically, Section A provides more experiment results, Section B provides additional implementation details, Section C provides additional related work, and Section D provides qualitative visualizations to illustrate our approach.
+
+# A Additional Experiment Results
+
+We begin by presenting several additional ablations in Section A.1 that further demonstrate the benefits of our TraveLER approach. We also present additional results in Section A.2.
+
+# A.1 Additional Ablations
+
+In what follows, we provide additional ablations that further illustrate the benefits of TraveLER. For all ablations, we compare the ablated experiment with the corresponding best-performing TraveLER results on a random sample of 1000 examples from the training set of the NExT-QA dataset. We use GPT-3.5 as the LLM and LLaVA-1.6 as the LMM. LMM response length. The LMM in our framework is crucial because it allows us to capture more relevant and question-specific details from visual input. However, if the LMM's responses are too long, the memory bank will become too large, whereas if the LMM's responses are too short, insufficient information will be captured. Thus, we conduct an experiment to determine the optimal LMM response length, and display our results in Fig 4. We find that limiting the LMM response to 150 tokens yields the most optimal performance, while accuracy decreases by  $-2.2\%$  and  $-1.7\%$  if the response is limited to 75 tokens and 300 tokens respectively. This supports the fact that there is a tradeoff between not collecting enough information for short response lengths and collecting too much information as the LMM response size increases.
+
+Prompt analysis. In each module of our framework, we use a task prompt to provide instructions to our agents (LLMs or LMMs). The construction of these prompts plays a large role in how instructions are executed. Currently, we use the question  $Q$  as input into all prompts  $(P_{\mathrm{T}}^{(1)}), (P_{\mathrm{R}}^{(1)}), (P_{\mathrm{E}}^{(1)}), (P_{\mathrm{V}}^{(1)})$ . However, we use the choices  $C$  as input only for the Planner and Evaluator prompts since the Planner needs the choices to tailor its plan, and the Evaluator needs
+
+![](images/80d4b6fcf8c45291dee260bba6185bc16ac35afd515a6671c30dd2f64a0d8788.jpg)  
+Figure 4: Comparison of different LMM Response Length (75, 150, 300 max tokens). 150 is optimal.
+
+the choices to answer the question. We experiment with adding the choices  $C$  to the Retriever and Extractor prompts, and find that this degrades performance by  $-1.6\%$ . This may be because the incorrect choices mislead the Retriever into searching for non-existent events or the Extractor into asking irrelevant questions.
+
+# A.2 Additional Results
+
+Results on Causal-VidQA. We also report zero-shot results on Causal-VidQA (Li et al., 2022), a dataset designed to focus on causal related questions to facilitate deeper video understanding towards video reasoning. We see that our method outperforms the best zero-shot method by  $16.9\%$ .
+
+Additional Baselines. We also compare our approach with the concurrent work MoReVQA (Min et al., 2024). It can be seen that MoReVQA performs better on NExT-QA by  $1\%$ , while our method performs better on the more difficult long-form video understanding EgoSchema benchmark by  $1.6\%$ . We also note some key differences. Firstly, our method is able to re-plan based on feedback from previous iterations while MoReVQA uses a single forward pass through the stages and cannot modify its approach during the process. Secondly, we use an iterative approach while MoReVQA does not, which allows our method to adjust and improve using feedback from previous iterations.
+
+# B Additional Implementation Details
+
+To run our models on larger benchmarks, we use 8 NVIDIA RTX 6000 GPUs and split the dataset across multiple processes. In addition, we use the
+
+Table 5: Zero-shot (ZS) results on Causal-VidQA.  
+
+<table><tr><td>Model</td><td>Acc. (%)</td></tr><tr><td>Just-Ask (Yang et al., 2021)</td><td>27.1</td></tr><tr><td>CaKE + CoMem (Su et al., 2023)</td><td>30.0</td></tr><tr><td>CaKE + HGA (Su et al., 2023)</td><td>30.6</td></tr><tr><td>TraveLER (ours)</td><td>47.5 (+16.9)</td></tr></table>
+
+SGLang (Zheng et al., 2023) package, which provides a variety of performance optimizations for our LMMs and enables us to perform batched inference for models that do not natively support doing so. We serve our LMM on a single GPU and implement a queue that is shared across all runs. This allows individual runs to asynchronously call the LMM using an API request instead of creating a new instance of the LMM for each run. Typically, we have 4-5 processes sharing the same LMM.
+
+Next, we use all default parameters. For our LMMs, we experiment with modifying the maximum output token length as an ablation, but use all default parameters otherwise. For the Llama 3 ablation in 3a, we serve Llama 3 70B on 4 NVIDIA A100s using vLLM (Kwon et al., 2023). We report results from a single run for all experiments.
+
+# B.1 Prompts
+
+Our prompts are shown in Table 6, Table 7, Table 8, and 9. The black text is the base prompt template, and we replace the blue text with the corresponding information from the relevant video. The generated outputs are in the orange text.
+
+# B.2 NExT-QA
+
+Dataset. NExT-QA is a challenging dataset that tests causal action reasoning and temporal understanding. It contains 5,440 videos with an average length of 44s. Compared to earlier VideoQA benchmarks (Papalampidi et al., 2024; Xue et al., 2017; Zeng et al., 2017; Zhu et al., 2017), NExT-QA requires going beyond simple recognition of objects and actions to answer the questions correctly. Each question requires selecting the best option out of 5 choices, often with very similar degrees of plausibility. Additionally, each question is categorized into either a Temporal, Causal, or Descriptive type. Temporal questions often ask what happens during, before, or after an event or action, while causal questions require advanced reasoning and inference about why an event or action occurs. Following the trend of works before us, we evaluate our method
+
+on the 5,000 questions in the NExT-QA validation set, which consist of 500 different videos. This dataset is in English.
+
+Inference Details. For NExT-QA, we use LLaVA-1.6 (Vicuna 13B) as the LMM and GPT-4 (gpt-4-1106-preview) as the LLM. We used the longer, comprehensive prompts, with no answer choices included in the Extractor prompt. We also initialize the memory bank with 5 frames, and use the multi-frame Retriever with 5 frames.
+
+# B.3 Perception Test
+
+Dataset. In comparison with earlier VideoQA datasets (Maharaj et al., 2017; Zhu et al., 2017; Zeng et al., 2017) that focus on computational tasks such as classification, detection, or tracking, Perception Test is a dataset that focuses on skills such as memory, abstraction, physics, and semantics. Moreover, it is designed to test the transfer capabilities of different models and intended to be approached in a few-shot or zero-shot manner. The dataset consists of 11.6k real-world videos with an average length of 23 seconds, and 38K multiple choice QA questions. This dataset is in English.
+
+Inference Details. For Perception Test, we use LLaVA-1.6 (Vicuna 13B) as the LMM and GPT-3.5 (gpt-3.5-turbo-0125) as the LLM. We used shorter, simplified prompts, with no answer choices included in the Extractor prompt. We also initialize the memory bank with 5 frames, and use the multi-frame Retriever with 5 frames.
+
+# B.4 EgoSchema
+
+Dataset. EgoSchema is a dataset that tests long-form video reasoning and is intended to be answered in a few-shot or zero-shot manner. It introduces the idea of certificate lengths, which are the minimum number of seconds it takes to be able to answer the question correctly. It has 5k questions in the full set, and a 500 question subset is used as the validation set. We use the full test set for evaluation. This dataset is in English.
+
+Inference Details. For EgoSchema, we use GPT-4 (gpt-4-1106-preview) as the LLM and LaViLa as the LMM. To prevent data leakage, we use a retrained version of LaViLa that does not use overlapping videos between Ego4D and EgoSchema. For window size in the Retriever, we take the selected frame, along with the frame 1 second before and 1-2 seconds after the selected frame.
+
+# User:
+
+Create the best plan to gather information to answer the question.
+
+# QUESTION: QUESTION CHOICES: CHOICES
+
+You are provided information collected from individual frames of the video and is represented as a dictionary keyed by the timestamps of the frames.
+
+# INFORMATION: INFO
+
+You are also given an explanation for why you aren't able to definitively answer the question with the current information.
+
+# EXPLANATION: EXPLANATION
+
+# Follow these rules:
+
+1. You only have access to individual frames of the video, with no audio. You can go to a certain timestamp, search for actions or settings, and describe or ask questions about individual frames.  
+2. Make sure that you have viewed the relevant frames.
+
+Make your plan as simple and straightforward as possible, and no longer than 5 steps long. Return your plan as a numbered list, after PLAN. Do not include any other response or explanation. Let's think step-by-step.
+
+Assistant: Output: PLAN
+
+Table 6: Planner prompt  ${P}_{T}$  :
+
+# User:
+
+You are given the following information about a LENGTH second video, with information from individual frames at different timestamps.
+
+# INFORMATION: INFO
+
+# PLAN: PLAN
+
+Currently, you are viewing second CURR. Choose the timestamp, in seconds, of the next frame to view. When choosing the next frame to view, remember that you are trying to collect information to answer this multiple choice question: QUESTION
+
+Think of what information you need, and consider what information you already have. Use the temporal nature of the video and your past information to choose the next frame. Do not choose a frame you already have information about, and make sure that the frame you choose is at least WINDOW SIZE seconds apart from the second you are currently viewing.
+
+Return your answer as a single Python float representing the second you want to view. Don't provide any other response or explanation.
+
+Assistant: Output: TIMESTAMP
+
+Table 7: Retriever prompt  ${P}_{R}$  :
+
+# User:
+
+You are given the following information about a LENGTH second video, with information from individual frames at different timestamps.
+
+INFORMATION: INFO
+
+Currently, you are viewing second CURRENT TIMITAMP, which has the caption: FRAME CAPTION
+
+Form up to 3 questions about this frame to best help answer the multiple-choice question: QUESTION.
+
+Follow these rules:
+
+1. Use the given information to decide what further visual information you need to answer the question.  
+2. Since you are asking questions about a single frame, you cannot ask about other frames, reference past or future events, or ask about specific timestamps.
+
+Return your questions as a Python list of strings (in double quotes) and don't include any numbered lists, backsticks, or language hints. Follow Python syntax. Make sure you have followed the steps. Don't provide any other response or explanation.
+
+Assistant: Output: QUESTIONS
+
+Table 8: Extractor prompt  ${P}_{E}$  :
+
+# User:
+
+Evaluate if there is enough information to answer a multiple-choice question about a video and if the plan has been completed.
+
+If there is enough information to choose the correct answer with complete certainty and the plan has been followed, return the index of the choice after a brief explanation. Otherwise, return None after a brief explanation of why you can't narrow down to a single answer choice.
+
+Be strict and don't guess
+
+INFORMATION: INFO
+
+PLAN: PLAN
+
+QUESTION: QUESTION
+
+CHOICES: CHOICES
+
+Give a brief explanation. Then, include your final answer after the words "Final Answer:" in your response at the end. Do not include anything other than the answer as an integer or None after "Final Answer:"
+
+Let's think step by step.
+
+Assistant: Output: ANSWER
+
+Table 9:Evaluator prompt  $P_V$  ..
+
+# B.5 STAR
+
+Dataset. STAR is a dataset that tests reasoning in real-world video situations. It consists of 22K video clips, with 60K situated reasoning questions, with 4 possible choices each. Questions are broadly divided into 4 main categories: interaction, sequence, prediction, and feasibility. This dataset is in English.
+
+Inference Details. For STAR, we use LLaVA-1.6 (Vicuna 13B) as the LMM and GPT-3.5 (gpt-3.5-turbo-0125) as the LLM. We used shorter, simplified prompts, with no answer choices included in the Extractor prompt. We also initialize the memory bank with 5 frames, and use the multi-frame Retriever with 3 frames, since the videos are shorter.
+
+# C Additional Related Work
+
+Modular Vision Frameworks. There has been a long history of work (Andreas et al., 2016; Herzig et al., 2018; Krishna et al., 2018; Baradel et al., 2018; Battaglia et al., 2018; Herzig et al., 2020, 2022, 2019, 2023; Bar et al., 2021; Avraham et al., 2022; Jerbi et al., 2020; Herzig et al., 2024; Mitra et al., 2024) that attempts to combine deep neural networks with modularity. Recently, works like VisProg (Gupta and Kembhavi, 2023), CodeVQA (Subramanian et al., 2023), RVP (Ge et al., 2023), and ViperGPT (Surís et al., 2023) have leveraged the improved coding capabilities of LMMs to generate code to compose different submodules together to answer visual questions. In addition, ProViQ (Choudhury et al., 2023) extends ViperGPT's work in the video domain by adding more modules for VideoQA. Similarly, we leverage the strong power of LMMs in a modular approach. However, while these approaches have shown promising results, they are limited to single-shot planning when generating code, resulting in a fixed plan that cannot adapt. In contrast to these works, our approach has the advantage of being able to iteratively replan based on new information collected.
+
+# D Qualitative Visualizations
+
+We present further qualitative success and failure cases of our TraveLER framework. For each dataset, we display qualitative visualizations for 2 successes and 2 failure cases. For the success cases, we show expanded visualizations in Figures Table 5, Table 6, Table 7, and abridged versions in 8 that demonstrate the benefits of our question-answering approach compared to regular caption
+
+ing. For failures, we also present abridged versions for each dataset in Figures Table 9, 10, and 11. Finally, we present some additional success and failure cases in 12. For the visualizations, we display 3 iterations of the question answering process with 2 Extractor QA pairs each, compressing the output text by displaying the most important parts for brevity. For the abridged versions, we display the video on top, and the traversal order using the numbered orange circles. In the row beneath, we display the frames in the order they are selected, and display corresponding Extractor question-answer output in yellow and captions in gray.
+
+# E Licenses and Privacy
+
+The license, PII, and consent details of each dataset are in the respective papers. In addition, we wish to emphasize that the datasets we use do not contain any harmful or offensive content, as many other papers in the field also use them. Thus, we do not anticipate a specific negative impact, but, as with any Machine Learning method, we recommend to exercise caution.
+
+# Successes
+
+# NeXT-QA
+
+# Question and Choices
+
+Q: Why is there a ladder in the middle of the land?
+
+A: (A) for child to play in (B) supervising boy (C) tree trimming (D) cross over fence (E) to play
+
+![](images/3831137945b6aa6164fe899e809083bca18cefce34fae1bfc6a7566288c475d9.jpg)  
+Iteration 1
+
+![](images/2dfa109d1b80f7f9060b29fe4c5be541b3aec8c78c585d1930c6b8747908ab40.jpg)  
+Iteration 2
+
+![](images/db8e090825c053fcda526be166d4325a6010ce7f502c3cbd27171b89c2eb6af9.jpg)  
+Iteration 3  
+Current Frame  
+Figure 5: NeXT-QA Success Predictions. We can see that our framework can adapt to new information collected in past iterations. For example, in Iteration 3, our Planner module is able to use information about the wooden post from a previous iteration and ask further questions to identify the correct answer.
+
+# Successes
+
+# STAR
+
+# Question and Choices
+
+Q: What happened before the person took the blanket?
+
+A: (A) Opened the box. (B) Opened the door. (C) Ate the sandwich. (D) Threw the towel.
+
+![](images/22dc845e38c6f52fd7ab1a2123516446d1371975aa347a98273c5ecf33f21a9b.jpg)  
+Iteration 1  
+Iteration 3  
+Current Frame
+
+![](images/d73496ae909b0cdb4d44d55ba94b210f496b0be470a059a3a717dac44b9f1004.jpg)
+
+![](images/4f43af0dbbca1eee69520eed1205e7d0d37e8cc3b1902d6e695a23b1429e04e4.jpg)  
+Iteration 4  
+Current Frame  
+Figure 6: STAR Success Predictions. Here, we can see that our method does not require viewing frames sequentially. For example, we view the beginning of the video in Iteration 1, the middle of the video in Iterations 2 and 3, and return to the beginning in Iteration 4. Moreover, our method can collect information and double-check ambiguous information across different timestamps. For example, in Iteration 2, we are told the man is taking a blanket, and then we can view a different frame to confirm that he is indeed holding a towel in Iteration 3.
+
+# Successes
+
+# Perception Test
+
+# Question and Choices
+
+Q: What stops the motion of the object placed on the slanted plane after being released by the person the first time? A: (A) I don't know. (B) The person or collision with another object. (C) High friction with the surface.
+
+![](images/ca5e8f883843e9fa94e67b85bd8005d72a5f667c8c5114a775e0a6a1811d1295.jpg)  
+Current Frame
+
+![](images/490b19ef8043c383f0109d910f2e4b6c9fd04332be971cf32c9c26ed9f48e1da.jpg)  
+Iteration 2  
+Current Frame
+
+![](images/932c771670439521dd8b734a53ff8854e11c66600e69cb85544dcc563073d93f.jpg)  
+Iteration 3  
+Current Frame  
+Figure 7: Perception Test Success Predictions. We display some success cases for the challenging Perception Test dataset. Here, our method is able to infer which objects the question refers to through our question-asking approach, even though the question does not explicitly describe them.
+
+# Successes
+
+# NeXT-QA
+
+# Question and Choices
+
+Q: What is the color of the shirt of the lady cycling?
+
+A: (A) black (B) light blue (C) white and red (D) blue (E) purple
+
+![](images/fb3a8dd48631a0fcd633eb746691a8b4273cc6e7d0910ce9fe2f405710ed666b.jpg)
+
+# TravelER Output (Ours)
+
+![](images/f27c2ba67129184100cc8787ee76873f816a6fdcf4c75063a01849955023f024.jpg)
+
+# Caption Comparison (Others)
+
+C: A man is holding a keyboard instrument in a desert
+
+C: A man holds a keyboard and a woman is on a bike
+
+C: A man holds a keyboard and a woman is on a bike
+
+C: Two men are in a desert, possibly dancing together
+
+# STAR
+
+# Question and Choices
+
+Q: What object did the person throw before they took the book?
+
+A: (A) the towel (B) the laptop (C) the bag (D) the blanket
+
+![](images/54ae06e43c43562ff0e7f0c910e737b7a9486ddf165673771f33af40b9653243.jpg)
+
+# TravelER Output (Ours)
+
+![](images/93715490d4d588ecc3eb93b2f29607f3be0a68eb4a5bf2f6934aaf98b20529dc.jpg)
+
+# Caption Comparison (Others)
+
+![](images/7f00d19da144ac6ed9726c5c5f471ec405569465b68cccd1b98260202ad26aa3.jpg)
+
+# Perception Test
+
+# Question and Choices
+
+Q: Is the paper lined or plain?
+
+A: (A) lined (B) I don't know (C) plain
+
+![](images/62965fe0c98f8e491bbcd24be59a065d6641b5b39e52b3f928c1b179265ea4d4.jpg)  
+Figure 8: Comparison with Captioning Approaches. For each example, we display the videos on top, and the traversal order using the numbered orange circles. In the rows beneath, we display the frames in the order they are selected, and display corresponding Extractor question-answer output in yellow and captions in gray. We display compressed versions of GPT-4V-generated captions for a visual comparison. By asking specific questions, we can extract more detailed and relevant information than a general description.
+
+# TravelER Output (Ours)
+
+![](images/a33fac1bc108f338a0eb99a4151120fb0806a66ae3350801ab0a19bd17387fee.jpg)
+
+# Caption Comparison (Others)
+
+C: Person seated with paper and a blue pen
+
+C: Person writing with toys in the background
+
+C: Person flips the paper, showing what they wrote
+
+C: Person shows the paper with "MOON" written
+
+# Failures
+
+# NeXT-QA
+
+# Example 1
+
+Question and Choices
+
+Q: How many drums in total are there for the people to play?
+
+A: (A) four (B) two (C) six (D) three (E) eleven
+
+![](images/96a865e03560e5917c3401568de729b8a3caaa5a754d01db610a13dab4345ded.jpg)
+
+# TravelER Output (Ours)
+
+![](images/f2f1c1247ca764e824e20a4afe623572637beb3cad72b310aa63be4a74afa0b2.jpg)
+
+Q: How many drums are there?
+
+A:Four
+
+![](images/8f7a10627001bd329982772471ad0e0548cb2fd1a0d778934c3ff46e901bc522.jpg)
+
+Q: How many drums are there?
+
+A:Four
+
+![](images/02c4e599a71ff32550e1543cd19904c2a4a535eec9b534d9ce8d5a0559e7190e.jpg)
+
+Q: How many drums are there?
+
+A: None
+
+![](images/2e7efe0c408ccde4386d1b3d6c0ab6748a3cc89a138c4c65d5fc1a9676552d77.jpg)
+
+Q: How many drums are there?
+
+A: Three
+
+# Caption Comparison (Others)
+
+C: A child holds a drum set and looks forward
+
+C: A child is focused, while another person plays the drums in the foreground
+
+C: A woman appears to be instructing the people playing the drums
+
+C: A child is playing a drum with concentration, with another person behind them
+
+# Example 2
+
+Question and Choices
+
+Q: How is the atmosphere of this group discussion?
+
+A: (A) happy (B) serious (C) sad (D) angry (E) shocking
+
+![](images/1b89a2aa968d8e00594f76675ef3a4f36e064241f8db789cd4252deb9c6ccc61.jpg)
+
+# TravelER Output (Ours)
+
+![](images/475033c88fa894462e32dbb29dbe9a1f21be2fd54388e6ee841286194b05feaa.jpg)
+
+Q: What are the person's emotions?
+
+A: They are focused
+
+![](images/94ed72c06ac18463e721a75f5ae539d67974c5f50f5502c94f155ed9fa9b1ab9.jpg)
+
+Q: Is anyone smiling or laughing?
+
+A: Yes, the two women are smiling.
+
+![](images/f512eaf02289a9146f69004f2dde11fbe4acb17228959830baad4418dd6ed616.jpg)
+
+Q: Is the person angry or shouting?
+
+A: No
+
+![](images/516d43afd354e351c283cbf0389d0a141a095e6e7b360129720a78fbe2e2357b.jpg)
+
+Q: Are they surprised or shocked?
+
+A:No
+
+# Caption Comparison (Others)
+
+C: A person looking towards something off-camera
+
+C: Two people smiling and engaged in conversation
+
+C: A person looking towards something off-camera
+
+C: People gathered in a cozy living room
+
+Figure 9: NeXT-QA Failure Predictions. Here, we display some failure cases for NeXT-QA. Like before, we display the video on top, and the traversal order using the numbered orange circles. In the row beneath, we display the frames in the order they are selected, and display corresponding Extractor question-answer output in yellow and captions in gray. We can see that conflicting information or false positives can mislead our approach. We also observe that counting can be a challenge for certain LMMs, but this can be mitigated in the future by swapping in stronger LMMs.
+
+# Failures
+
+# STAR
+
+# Example 1
+
+# Answer
+
+Q: What did the person do to the window after holding the broom?
+
+A: (A) Closed. (B) Opened. (C) Washed. (D) Lied on.
+
+![](images/1099b83caa53dc600113da148103315880b61f5d8fb8bb308e2137625b6de488.jpg)
+
+# TravelER Output (Ours)
+
+![](images/00dc5d0d6c6c3a195e87d83600b19d9581639e7245ff5b6a542a4bd72910876d.jpg)
+
+Q: Is the person holding a broom?
+
+A: Yes
+
+![](images/cdf732a76f3cfab6d693a4c7e078cb84446417129bf90ac64fb191dd0d6e5b63.jpg)
+
+Q: Is the person interacting with a window?
+
+A: Yes
+
+![](images/e0363bd8e17fb049bdf7524007edddf496a84b195383b1d77515ba2bf0491a20.jpg)
+
+Q: Are they opening or closing the window
+
+A: Opening
+
+![](images/b9c0d40614c5e9939630b9b89aebb857f5591381fdaf50876d7b763260697383.jpg)
+
+Q: Are they opening or closing the window
+
+A: Opening
+
+# Caption Comparison (Others)
+
+C: Person standing near washing machine with object in hand
+
+C: Person moving towards the right, reaching upwards
+
+C: A person mid gesture, next to washing machines
+
+C: A person near a wall and washing machines
+
+# Example 2
+
+# Answer
+
+Q: What did the person do with the book?
+
+A: (A) Took. (B) Opened. (C) Threw. (D) Put down.
+
+![](images/b786bd395211bb0cf76cb29b3b25feeae8732c9004d9fe891cdb1c803a947e5d.jpg)  
+Figure 10: STAR Failure Cases. Here, we display some failure cases for the STAR dataset, using the same abridged representation described previously. We see that a limitation of a framewise approach is that it may be difficult to capture very temporal actions. For example, in Example 1, it is difficult to understand if the woman is opening or closing the window.
+
+# TravelER Output (Ours)
+
+![](images/c01a2741311cd989c4adcbe513131f9a556fb9492c5767399a98c4af6a303a1c.jpg)
+
+Q: What is the person doing?
+
+A: They are standing with a book
+
+![](images/d5e56687d110368715843ea01566062be25a2119dade6fe4c0a6f90478faae5c.jpg)
+
+Q: What is the person doing?
+
+A: They are reading a book
+
+![](images/629c16365079ac0bdb121f514f753af4762a7ac13a78ad2b3150b90b57eef2ca.jpg)
+
+Q: Is the person still reading?
+
+A: No, they are standing
+
+![](images/ee3036a04343894ad4826dc13b29849e150c5dcc3852dd7d215b08128bf8d8f9.jpg)
+
+Q: Does the person put down or throw the book?
+
+A: No
+
+# Caption Comparison (Others)
+
+C: A person sitting upright with a book or tablet in hand
+
+C: A person sitting in a living room reading a book
+
+C: A person standing up in a living room
+
+C: A person sitting in a living room reading a book
+
+# Failures
+
+# Perception Test
+
+# Example 1
+
+# Question and Choices
+
+Q: The person uses multiple similar objects to play an occlusion game. Where is the hidden object at the end of the game from the person's point of view?
+
+A: (A) Under the third object from the left. (B) Under the second object from the left. (C) Under the first object from the left.
+
+![](images/24eb0ca0d882b07d7a48e526ff45c5db556832552544c13ff55992fd2eb6e8a7.jpg)
+
+# TravelER Output (Ours)
+
+![](images/9dddb94f78faaf97c3b59110c08acb25a195887acca102cf5119079eb9a80eae.jpg)
+
+# Caption Comparison (Others)
+
+C: A hand is positioned above three red mugs on a table
+
+C: A hand touches one of the mugs, about to move it
+
+C: A hand holds a spoon above a cup
+
+C: A hand is hovering around the rightmost mug
+
+# Example 2
+
+# Question and Choices
+
+Q: Can the person close the door during the last attempt?
+
+A: (A) No, the person cannot close the door. (B) Yes, the person can close the door. (C) The person didn't attempt to close the door in this video.
+
+![](images/0243552ee0b8e70dd4b2c45d91d093c54d78cf075afd89417c8b1ac05ee1e2be.jpg)  
+Figure 11: Perception Test Failure Cases. We display some qualitative visualizations for Perception Test failure cases using the abridged representation discussed previously. We see that for some cases where objects of interest are occluded or not in the frame, our method might have difficulties extracting useful information.
+
+# TravelER Output (Ours)
+
+![](images/38cc6dad55035d0747c81ffc3cc54833f9e071dc2358d76a87e1686541552d8b.jpg)
+
+# Caption Comparison (Others)
+
+C: A finger points at a small red bottle on a tiled floor
+
+C: A finger is pushing the bottle, causing it to tip
+
+C: A bottle is falling, it is upside down
+
+C: A jar and a fallen bottle are on a red tile floor
+
+# Successes
+
+# NExT-QA
+
+Q: Why are the men wearing hats?
+
+A: (A) Sunny (B) Fashion (C) Costume of play (D) Uniform costume (E) Road repairing
+
+![](images/68a0f7ad713ff22d8c73930baa2a95ab42438e68dc1e7c8c8fa979ca8ccc872c.jpg)
+
+Q: What is the weather like?
+
+A: It is sunny
+
+![](images/772563287d66f1e16cb1342fd396cc697505a6d49ef8de66ec1b724f9752a92a.jpg)
+
+Q: Are the men wearing uniforms
+
+A: No, they wear outdoor gear
+
+C: A group of hikers, with one person resting an arm on another
+
+![](images/0b926718f933abf64c8f3e96551f3c48e1b2019cef39383246ac32ed43d6a22d.jpg)
+
+Q: What are the men doing?
+
+A: They are walking
+
+C: Friends hiking together, with one's arm on another
+
+![](images/88f4d181cc6947f3199882d7afc9be3c9f4ac7e2e50f7ae3c237da14df296aad.jpg)
+
+Q: Are there shadows that show sunny weather?
+
+A: Yes, there are
+
+C: A group of hikers, walking together along a road
+
+# Perception Test
+
+Q: Why can't the box be closed at the end?
+
+A: (A) The person put one or more objects that are too large to fit in the box, even if placed by themselves. (B) Each of the objects would fit in the box individually, but the person put too many objects. (C) The lid does not match the box.
+
+![](images/2e683345631ccc1ec807d624f086dd6242e83bc50370e8baa1bcf8c2fbbf4a36.jpg)
+
+Q: Why can't the box be closed?
+
+A: It is not clear
+
+C: A person standing behind a table, presenting a shoebox with a GoPro logo on it.
+
+![](images/8fe6e4f6514284cb6f8bdfb4b7765e864746c240c2e146c1fdb00c33ebdaf229.jpg)
+
+Q: What is in the box?
+
+A: A small bottle and container
+
+C: A person is unpacking it on a table. There are objects like a plant vase and boxes
+
+![](images/214d31e372ae7e3866e799fc2641aa4c889fcfaf936788743e8a4d12f3aa4228.jpg)
+
+Q: Are the objects smaller than the box?
+
+A: Yes, they are
+
+C: A person organizes items on a table with objects like a box with a yellow lid, and a potted plant.
+
+![](images/3a54c25eb990e396e1bfa925678aac7c6325756c24b02297238e01dd41d66c9e.jpg)
+
+Q: Are there too many objects?
+
+A: Yes, it is overflowing
+
+C: A person appears to be sorting through items on a table, with a stack of boxes and containers
+
+# Failures
+
+# NExT-QA
+
+Q: What does the lady do after taking the microphone from the man in red cape at the start?
+
+A: (A) put 2 fingers up (B) talk to the audience (C) turn behind her (D) eat (E) put hands together
+
+![](images/b455d210329e6d313322e1ba98708dae01a71e61433960022c6e6227c2f00ac2.jpg)
+
+Q: Does the lady have the mic?
+
+A: No, the man has it
+
+C: A karaoke night with participants on stage, one draped in a Swiss flag.
+
+![](images/de65ef6fd669af622436a394af35b6c55c2e1996042808919633511a23f62df6.jpg)
+
+Q: Does the lady have the mic?
+
+A: Yes
+
+C: An eclectic performance at a colorful music event.
+
+![](images/883ba7ffa35d4d0e4fcb728a02bdd78ed94d7999a024333ff619a32785f231a9.jpg)
+
+Q: What is the lady doing?
+
+A: She is holding it
+
+C: Band performing on stage with vocalist as other members look on.
+
+![](images/9cccf0224d2468ff891a43ca146fc25c96b358f095526ece60cd27e316f906a1.jpg)
+
+Q: What is she doing?
+
+A: She is high-fiving
+
+C: A high-spirited karaoke night where a performer gets a high-five on stage.
+
+# Perception Test
+
+Q: The person interacts with a lighting device among other objects. Is the lighting device on at any point?
+
+A: (A) I don't know (B) yes (C) no
+
+![](images/1a302c4de2e3947ed89614a827a25c046f30ebddf2150f6d6d1ae07035d72e8b.jpg)
+
+Q: Is there a lighting device?
+
+A: Yes, a lamp
+
+C: A person is reaching out towards an electronic device with wires and routers.
+
+![](images/78f9561d240eac934b728075831da71861d55034e259bf4bc837304684a0003d.jpg)
+
+Q: Is the lamp on?
+
+A: No, it is off
+
+C: A person is standing next to a shelf with household items.
+
+![](images/aeb56597b06060e01102b5bd57c477aec2570d2affda4e3b2efae20a897096a7.jpg)
+
+Q: Is the lamp on?
+
+A: No, is off
+
+C: An individual is standing by a shelf with electronic devices and a router.
+
+![](images/3a95f07b98b453204b5ab3f93eb748e6b9f63293d9e8c95db1064c5a6cfea0af.jpg)
+
+Q: Is the lamp on?
+
+A: No, it is off
+
+C: A person is standing next to a shelf with household items.
+
+Figure 12: An additional visualization of predictions. We show more qualitative visualizations of our method on NExT-QA and Perception Test using our abridged representation, with successes on top and failures on the bottom. We compare our generated question-answer pairs for each frame (in yellow) with captions (labeled C in gray) generated from the same frame. We see that our method is able to extract more fine-grained and relevant information compared to simple captioning.

2024/TraveLER_ A Modular Multi-LMM Agent Framework for Video Question-Answering/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:deb7b5324a88ff350cf73c15334ed19adf5ad1d69b9058a6905e4f187a1d0e78
+size 1920567

2024/Tree of Problems_ Improving structured problem solving with compositionality/418dcc40-3e0e-4ed6-aa5d-caf94c578b26_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:96682abd52c6bc61112f87befa3cad2b2232be3fabd8f0a256bc3a0ea2169d9f
+size 562431

2024/Tree of Problems_ Improving structured problem solving with compositionality/full.md

@@ -0,0 +1,466 @@
+# Tree of Problems: Improving structured problem solving with compositionality
+
+Armel Zebaze, Benoit Sagot, Rachel Bawden
+
+Inria, Paris, France
+
+{firstname这个名字}@inria.fr
+
+# Abstract
+
+Large Language Models (LLMs) have demonstrated remarkable performance across multiple tasks through in-context learning. For complex reasoning tasks that require step-by-step thinking, Chain-of-Thought (CoT) prompting has given impressive results, especially when combined with self-consistency. Nonetheless, some tasks remain particularly difficult for LLMs to solve. Tree of Thoughts (ToT) and Graph of Thoughts (GoT) emerged as alternatives, dividing the complex problem into paths of subproblems. In this paper, we propose Tree of Problems (ToP), a simpler version of ToT, which we hypothesise can work better for complex tasks that can be divided into identical subtasks. Our empirical results show that our approach outperforms ToT and GoT, and in addition performs better than CoT on complex reasoning tasks. All code for this paper is publicly available here: https://github.com/ArmelRandy/tree-of-problems.
+
+# 1 Introduction
+
+In-Context Learning (ICL) (Brown et al., 2020) is the ability of Large Language Models (LLMs) to perform a task with the help of a few demonstrations within their context. It is widely used to evaluate LLMs on various tasks. These models, whose number of parameters and training corpus size has increased massively over recent years, keep pushing the state of the art on a wide range of natural language tasks (Anil et al., 2023; Touvron et al., 2023; Gemma Team et al., 2024). However, they still struggle to perform complex tasks, notably those requiring multiple reasoning steps (Hendrycks et al., 2021a,b; Suzgun et al., 2023). Recently, Chain-of-Thought (CoT) prompting (Wei et al., 2022; Kojima et al., 2022) has greatly helped to enhance reasoning abilities of LLMs by helping them to mimic step-by-step reasoning. However, CoT implicitly requires the model to generalize beyond
+
+the cases seen in its prompt, which often leads to poor out-of-domain performance (Zhou et al., 2023). Applying CoT with self-consistency (Wang et al., 2023b) drives the model to explore multiple reasoning paths and to choose the most consistent answer, usually yielding better performance, but helping only marginally with out-of-distribution generalization. Moreover, solving complex problems involves understanding their underlying structure; this can help to avoid lengthy CoTs that are prone to reasoning errors.
+
+In this paper, we propose to tackle complex problem-solving and out-of-distribution generalization by dividing complex tasks into a series of simpler sub-tasks. We draw inspiration from techniques such as dynamic programming and divide and conquer in order to efficiently guide LLMs through complex problem solving. Such problems have previously been tackled using approaches adding structure to CoT, such as Tree of Thoughts (ToT) (Yao et al., 2023) and Graph of Thoughts (GoT) (Besta et al., 2024), which consist in sampling diverse reasoning paths (where path states represent subproblems) and finding the optimal path. We argue that for a subset of complex reasoning problems, where an instance can be decomposed into multiple analogous subinstances, ToT and GoT are overly complex, and the tasks can be better solved by a simpler approach. This simpler approach, which we name Tree of Problems (ToP) consists in building a tree structure, where each node represents a problem instance similar to the main instance. The deepest instances, which correspond to atomic problems, are solved first with CoT prompting and the internal nodes are recursively solved by merging their children's solutions. Figure 1 illustrates our method on the tasks of Last Letter Concatenation and Navigate from the BIG-Bench Hard benchmark (Suzgun et al., 2023).
+
+We conduct a comprehensive evaluation on several LLMs, including GPT-3.5, on multiple hard
+
+![](images/d2a688c5ba10215c715cf13b77bb49b8df410347608baf5bff04c69dd0f7ef94.jpg)  
+Figure 1: Overview of the Tree of Problems (ToP) framework for two tasks. On the left (a canonical task consisting of independent subproblems organised in a tree structure), the task is to concatenate the last letters of a list of names, accomplished by breaking the list in two, finding their solutions, and recombining them. On the right (an extension of the canonical structure to handle sequential tasks), the task is to determine the final position of an object after a series of steps. We first find its position after half of the steps, and then determine the final position by tracing the object through the remaining steps. See Section 3 for a description of ToP.
+
+![](images/66e293c58514481092af3cc00e8f0fc9062dd7696a98b23cc8146ec72742ff45.jpg)
+
+tasks. We find that ToP improves LLMs' problem solving abilities on structured tasks outperforming CoT, ToT and GoT by a large margin.
+
+# 2 Related Work
+
+CoT prompting was proposed to enhance reasoning by incorporating step-by-step logic into few-shot prompt demonstrations (Wei et al., 2022). It showed significant improvement over standard input-output (IO) prompting across various mathematical and symbolic reasoning benchmarks. Building on this, Kojima et al. (2022) and Wang et al. (2023a) inter alia demonstrated that zero-shot CoT could be achieved by using reasoning-inducing words at the end of the zero-shot prompt. Other works showed that wisely designing the CoT demonstrations could yield further improvements (Zhang et al., 2023; Fu et al., 2022). CoT Self-Consistency (CoT-SC; Wang et al. 2023b) improved on CoT by sampling diverse reasoning steps and selecting the most consistent answer after marginalizing over the reasoning paths. Our research also builds on the body of work addressing problem-solving through compositionality, which involves teaching LLMs to tackle complex problems by breaking them down into a series of subproblems and recursively solving them to derive
+
+the final answer, e.g. Least-to-Most (Zhou et al., 2023), decomposed (Khot et al., 2023) and successive (Dua et al., 2022) prompting. While these works align with our approach through their use of problem decomposition, we focus on breaking a main task into multiple similar subtasks, solvable using the same prompt. Moreover, our approach uses a tree structure that allows for greater flexibility and coverage in problem solving. The most closely related approaches are Tree of Thoughts (ToT) (Yao et al., 2023) and Graph of Thoughts (GoT) (Besta et al., 2024). ToT builds on the idea of sampling diverse reasoning paths but redefines problem solving as a search over a thought space, where states represent partial solutions. GoT extends ToT by including thought aggregation, which is analogous to our merge operation and by allowing refining (Madaan et al., 2023). While a "thought" represents a general reasoning step in their approach, we focus on reasoning through subproblems. We do not perform a search over a tree of thoughts, nor do we score or refine (improve) our tree nodes. Instead, each node in the tree of problems is directly relevant to solving the problem, and their bottom-up recombination produces the final solution. ToP is therefore a simpler and more cost-effective alternative to ToT and GoT.
+
+# 3 Our method
+
+Solving a complex problem often requires reasoning, partly explaining the success of CoT prompting for such problems. Reasoning involves understanding a problem's structure and design. This aspect is frequently overlooked in CoT because incorporating it can be challenging. Our method addresses this by constructing a tree of simpler, closely related subproblems to solve a more complex problem. We hypothesize that the capability of an LLM to solve simple instances can be extended to more complex ones. The ability of an LLM to solve a complex instance therefore lies in how accurately it can solve simpler ones and then combine their answers. The main class of problems we aim to tackle are complex problems that are divisible into independent subproblems resembling the initial one (we refer to these as canonical tasks). However, we also experiment with relaxing the independency constraint in order to tackle sequential tasks, which require finding the final state of a system after a series of independent processing steps (See the right of Figure 1). Our method relies on the following components:
+
+- A decomposer divides a problem instance into a series of smaller related instances, algorithmically or via few-shot prompting with a divide_prompt. We recursively build a tree of problems (nodes) considering 2 parameters: the breadth (the number of children of each internal node) and the depth of the tree, directly related to the granularity of the atomic subproblems. The root of the tree is the main problem. In this paper, ToP  $(b, d)$  refers to using breadth  $b$  and depth  $d$ .  
+- A solver is used to do the task of interest, namely the simplest instances obtained after decomposition (in our case an LLM with a task-specific solve_prompt).  
+- A merger receives the solved subproblems (problem statement and solution) at level  $k$  to build and solve the problem at level  $k - 1$ . It uses a specific merge_prompt to get the LLM to learn to combine the subproblems' solutions into the parent solution. As opposed to L2M, the prompt to get the solution of a problem at a level  $k$  only depends on the directly connected problems (at the level  $k + 1$ ).
+
+The workflow can be described as follows: The decomposer builds the tree of problems, the solver addresses the subproblems at the tree's leaves, and the merger recursively derives each node's solution
+
+by combining its children's solutions in a bottom-up approach. The total number of inference calls (omitting the cost of problem decomposition) is equal to the number of nodes in the tree structure.
+
+In addition to canonical tasks with a classic tree structure (see the left of Figure 1), ToP can also be used for sequential tasks, where a given subproblem needs the result of a previous subproblem as an input (see the right of Figure 1). Our standard ToP paradigm described above can be used to solve such problems by setting the breadth to 1. This has the effect that the problem is decomposed into a sequence of  $n$  subproblems organised as hierarchy of depth  $n$ . When solving the  $(k + 1)$ -th subproblem, the solver will have access to its child subproblem's result, i.e. the result of subproblem  $k$ , thereby accounting for the sequentiality of the decomposition. The LLM is no longer required to merge subproblems' solutions; it is directly fed with a new problem formulation automatically computed using the corresponding child's solution. The final solution is obtained by solving the last subproblem, and so the main problem instance (root node) does not influence the inference cost.
+
+For both tasks, all problems at the same level of the tree are solved in parallel to promote efficiency. We further detail the method with more examples in Appendix A.
+
+# 4 Experiments
+
+We first compare ToP to ToT and GoT to test our hypothesis that our simpler approach is more adapted to canonical tasks. We do this using the GoT tasks proposed by Besta et al. (2024). We then show that ToP is more effective in comparison to IO (direct input-output) and CoT prompting across a wider ranger of canonical tasks, namely Last Letter Concatenation (Wei et al., 2022) and 5 BIG-Bench Hard (Srivastava et al., 2023) tasks fitting the description. Finally, we test ToP on sequential tasks.
+
+# 4.1 Datasets
+
+GoT tasks. Besta et al. (2024) compared GoT to ToT, IO, and CoT prompting on three tasks (each with 100 examples): (i) Sorting, which involves arranging a list of 32 numbers ranging from 0 to 9 (both inclusive) in order, (ii) Set Intersection, which involves finding the common elements between two sets, each containing 32 elements and (iii) Keyword Counting, which involves identifying countries mentioned in a text and counting how
+
+many times each country appears.
+
+Symbolic Reasoning. We use two toy tasks introduced by Wei et al. (2022) (each with 500 examples): (i) Last Letter Concatenation, where the LLM is tasked with recovering the concatenation of the last letters from a list of names and (ii) Coin Flip, which evaluates if the LLM can deduce the final state of a coin (heads or tails) after people either flip it or not. During evaluation, we consider various list lengths (4, 8 and 16) for the first task, and different numbers of people involved (4, 8 and 16) for the second.
+
+BIG-Bench Hard (BBH). BBH consists of 23 BIG-Bench (Srivastava et al., 2023) tasks that have been shown to benefit from CoT (Suzgun et al., 2023). We use 8 tasks: Boolean Expressions, Hyperbaton, Multi-Step Arithmetic Two, Navigate, Object Counting, Tracking Shuffled Objects (3, 5, 7), Web of Lies and Word Sorting.
+
+# 4.2 Language models and prompts
+
+We experiment with gpt-3.5-turbo and gpt-3.5-turbo-instruct. For the solve_prompts, we use the CoT prompts of Suzgun et al. (2023) on BBH tasks, with minor changes. The CoT prompts for Symbolic Reasoning are inspired by those in (Wei et al., 2022), which contain 8 examples of 2-letters or 2-flips and those for GoT tasks are the same as in Besta et al. (2024). We report some implementation details in Appendix D and Appendix E.
+
+# 4.3 Main results
+
+GoT tasks. Table 1 compares our results on the GoT tasks with those obtained by rerunning the CoT, ToT and GoT approaches from (Besta et al., 2024). More precisely, we use the highest accuracy achieved with ToT and GoT on each task with gpt-3.5-turbo-0125. For Sorting, we intuitively choose  $b = 2$  as in merge sort and  $d = 2$  for performance. We use the same  $b$  for Keyword Counting, with  $d = 4$  to get simple atomic instances. In Set Intersection, we use  $b = 4$  because each set is divided into two disjoint subsets, resulting in four pairs of subsets (one pair per subproblem). Such a large breadth was sufficient to produce simple atomic problems, so we used  $d = 1$ . ToP outperforms
+
+ToT and GoT by a large margin on sorting with an absolute improvement of  $40\%$  over GoT. Similarly, ToP outperforms GoT by  $19\%$  and  $5\%$  respectively on Set Intersection and Keyword Counting.
+
+<table><tr><td rowspan="2">GoT Tasks</td><td colspan="4">gpt-3.5-turbo</td></tr><tr><td>CoT</td><td>ToT (best)</td><td>GoT (best)</td><td>ToP (ours)</td></tr><tr><td>Sorting</td><td>0.02</td><td>0.17</td><td>0.28</td><td>0.68</td></tr><tr><td>Set Intersection</td><td>0.07</td><td>0.25</td><td>0.46</td><td>0.65</td></tr><tr><td>Keyword Counting</td><td>0.00</td><td>0.00</td><td>0.26</td><td>0.31</td></tr></table>
+
+Last Letter Concatenation. We consider ToP (2, 1). Subproblems are obtained by dividing the main list into  $b = 2$  lists of equal length.
+
+Table 1: Results on 3 tasks from (Besta et al., 2024). In all results tables, best results are highlighted in bold.  
+
+<table><tr><td rowspan="2">Last Letter Concatenation</td><td colspan="3">gpt-3.5-turbo-instruct</td></tr><tr><td>IO</td><td>CoT</td><td>ToP (ours)</td></tr><tr><td>Four</td><td>0.032</td><td>0.900</td><td>0.990</td></tr><tr><td>Eight</td><td>0.000</td><td>0.662</td><td>0.854</td></tr><tr><td>Sixteen</td><td>0.000</td><td>0.252</td><td>0.444</td></tr></table>
+
+Table 2: Results on Symbolic Reasoning tasks.
+
+Comparison to Least-to-Most Prompting and CoT with Self-consistency. Least-to-most (L2M) prompting has also been successfully applied to Last Letter Concatenation (Zhou et al., 2023). Given a list of  $L$  names, L2M requires  $L - 1$  inference calls, the first to concatenate the first 2 last letters and the  $L - 2$  other to add the remaining last letters one after the other. Following Huang et al. (2024), we provide a fair comparison of L2M to ToP by adapting ToP's tree structure to require the same number of inference calls as L2M. This is done by using trees of breadth 2 and depth  $\log_2(L) - 1$  for lists of length  $L$ . We compare ToP to L2M as well as CoT self-consistency with  $L$  reasoning paths. The results (Table 3) show that for  $L = 4$  or  $L = 8$ , ToP (breadth = 2, depth = 1) achieves comparable performance to L2M while requiring half as many inference calls. When the number of inference calls is matched between the two methods, ToP demonstrates superior performance in all scenarios. CoT-SC lags behind both L2M and ToP.
+
+Moreover, since L2M is similar to ToP  $(1, L - 1)$ , we compare its accuracy to ToP (match) at each level of the tree. As illustrated in Figure 2, both methods start with a perfect score that gradually
+
+<table><tr><td rowspan="2">Last Letter Concatenation</td><td colspan="4">gpt-3.5-turbo-instruct</td></tr><tr><td>CoT-SC</td><td>L2M</td><td>ToP</td><td>ToP (match)</td></tr><tr><td>Four</td><td>0.908</td><td>0.988</td><td>0.990</td><td>0.990</td></tr><tr><td>Eight</td><td>0.574</td><td>0.870</td><td>0.854</td><td>0.932</td></tr><tr><td>Sixteen</td><td>0.116</td><td>0.742</td><td>0.444</td><td>0.858</td></tr></table>
+
+![](images/31650aef31947236e145ab8badc04f55276a217d01f547e2f98fceb9f0e7d09f.jpg)  
+Figure 2: Per-level accuracy of Least to Most prompting and ToP (match) for Last Letter Concatenation.
+
+decreases as they approach the task's resolution. ToP (match) consistently outperforms L2M at each step across all three settings.
+
+# 4.4 Complementary results
+
+We have successfully applied ToP to problems that can be divided into multiple independent instances. In this section, we report additional results for more such tasks and sequential tasks.
+
+# 4.4.1 Canonical BBH tasks
+
+BBH tasks such as Boolean Expressions, Hyperbaton, Multistep Arithmetic Two, Object Counting, and Word Sorting can be decomposed into multiple independent instances, whose solutions are later combined. They therefore correspond to canonical ToP tasks. We apply ToP (2, 1) to them and report results in Table 4. ToP yields an absolute improvement over CoT of  $21.2\%$  on Word Sorting and  $9.8\%$  on Hyperbaton. However, it is slightly worse than CoT on Boolean Expressions, Multistep Arithmetic Two and Object Counting with an average deterioration of  $3.6\%$  on the 3 tasks. We attribute this loss of accuracy to reasoning inconsistencies and we explore this in more detail in Appendix C.4.
+
+Table 3: Comparison of ToP to L2M and CoT-SC for Last Letter Concatenation. ToP (match) refers to ToP with the same number of inference calls as L2M.  
+
+<table><tr><td></td><td colspan="3">gpt-3.5-turbo-instruct</td></tr><tr><td></td><td>IO</td><td>CoT</td><td>ToP</td></tr><tr><td>Boolean Expressions</td><td>0.908</td><td>0.924</td><td>0.896</td></tr><tr><td>Hyperbaton</td><td>0.528</td><td>0.804</td><td>0.902</td></tr><tr><td>Multistep Arithmetic Two</td><td>0.032</td><td>0.780</td><td>0.736</td></tr><tr><td>Object Counting</td><td>0.412</td><td>0.928</td><td>0.892</td></tr><tr><td>Word Sorting</td><td>0.837</td><td>0.619</td><td>0.831</td></tr></table>
+
+# 4.4.2 Sequential tasks
+
+Coin Flip is an example of a sequential task. Using ToP (1, 2), the problem at the leaves is to find the state of the coin after going through the first half of the people. The final solution is obtained by determining how this state changes as the coin goes through the remaining half of the people. Navigate, Tracking Shuffled Objects, and Web of Lies can be modeled in a similar way. ToP outperforms CoT on all tasks, as shown in Table 5. ToP reaches near perfect accuracy on Coin Flip with 4 and 8 people. Moreover, it is more robust to the out-of-distribution setting than CoT, showing a lesser performance drop as the number of people increases. Compared to CoT, it yields an absolute improvement of  $5.2\%$  on Web of Lies,  $5.9\%$  on average on Tracking Shuffled Objects and  $2\%$  on Navigate.
+
+Table 4: Results on the canonical BBH tasks.  
+
+<table><tr><td></td><td colspan="3">gpt-3.5-turbo-instruct</td></tr><tr><td></td><td>IO</td><td>CoT</td><td>ToP</td></tr><tr><td>Coin Flip</td><td></td><td></td><td></td></tr><tr><td>Four</td><td>0.512</td><td>0.998</td><td>0.998</td></tr><tr><td>Eight</td><td>0.502</td><td>0.840</td><td>0.998</td></tr><tr><td>Sixteen</td><td>0.476</td><td>0.718</td><td>0.756</td></tr><tr><td>BIG-Bench Hard</td><td></td><td></td><td></td></tr><tr><td>Navigate</td><td>0.204</td><td>0.864</td><td>0.884</td></tr><tr><td>Tracking Shuffled Objects (3)</td><td>0.004</td><td>0.536</td><td>0.524</td></tr><tr><td>Tracking Shuffled Objects (5)</td><td>0.004</td><td>0.324</td><td>0.440</td></tr><tr><td>Tracking Shuffled Objects (7)</td><td>0.000</td><td>0.044</td><td>0.118</td></tr><tr><td>Web of Lies</td><td>0.528</td><td>0.920</td><td>0.972</td></tr></table>
+
+Table 5: Results on Coin Flip and sequential BBH tasks.
+
+# Conclusion
+
+We use compositionality to grant the ability to solve complex and structured problems to LLMs via the Tree of Problems (ToP) framework. ToP is a simplification of the ToT framework, and involves decomposing complex tasks into identical subtasks. Our experiments demonstrate that LLMs can benefit from ToP and solve certain complex problems better than ToT, GoT and L2M approaches and generalize better than with CoT.
+
+# Acknowledgements
+
+This work was partly funded by the last two authors' chairs in the PRAIRIE institute funded by the French national agency ANR as part of the "Investissements d'avvenir" programme under the reference ANR-19-P3IA-0001. The authors are grateful to the OPAL infrastructure from Université Côte d'Azur for providing resources and support. We thank Maciej Besta for answering our questions regarding Graph of Thoughts and Lydia Nishimwe for her helpful feedback.
+
+# Limitations
+
+Applicability of the Tree of Problems framework. Although ToP is a powerful prompting strategy that significantly widens the range of tasks that LLMs can handle accurately; it is limited to problems which have a structure (mathematical tasks, algorithmic tasks etc.) that can be decomposed into analogous subproblems. The founding hypothesis of ToP is the fact that LLMs can solve simple instances of a task and this ability can be efficiently translated to more complex instances.
+
+Reasoning consistency of LLMs. LLMs can surprisingly fail to be robust to minor changes in a problem formulation. They can fail to solve a problem closely related to another one that they are capable to solve. We note this as a typical failure case of ToP in Appendix C.4 on Object Counting and Multistep Arithmetic Two.
+
+# References
+
+Rohan Anil, Andrew M. Dai, Orhan Firat, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, Eric Chu, Jonathan H. Clark, Laurent El Shafey, Yanping Huang, Kathy Meier-Hellstern, Gaurav Mishra, Erica Moreira, Mark Omernick, Kevin Robinson, Sebastian Ruder, Yi Tay, Kefan Xiao, Yuanzhong Xu, Yujing Zhang, Gustavo Hernandez Abrego, Junwhan Ahn, Jacob Austin, Paul Barham, Jan Botha, James Bradbury, Siddhartha Brahma, Kevin Brooks, Michele Catasta, Yong Cheng, Colin Cherry, Christopher A. Choquette-Choo, Aakanksha Chowdhery, Clement Crepy, Shachi Dave, Mostafa Dehghani, Sunipa Dev, Jacob Devlin, Mark Diaz, Nan Du, Ethan Dyer, Vlad Feinberg, Fangxiaoyu Feng, Vlad Fienber, Markus Freitag, Xavier Garcia, Sebastian Gehrmann, Lucas Gonzalez, Guy Gur-Ari, Steven Hand, Hadi Hashemi, Le Hou, Joshua Howland, Andrea Hu, Jeffrey Hui, Jeremy Hurwitz, Michael Isard, Abe Ittycheriah, Matthew Jagielski, Wenhao Jia, Kathleen Kenealy, Maxim Krikun
+
+Sneha Kudugunta, Chang Lan, Katherine Lee, Benjamin Lee, Eric Li, Music Li, Wei Li, YaGuang Li, Jian Li, Hyeontaek Lim, Hanzhao Lin, Zhongtao Liu, Frederick Liu, Marcello Maggioni, Aroma Mahendru, Joshua Maynez, Vedant Misra, Maysam Moussalem, Zachary Nado, John Nham, Eric Ni, Andrew Nystrom, Alicia Parrish, Marie Pellat, Martin Polacek, Alex Polozov, Reiner Pope, Siyuan Qiao, Emily Reif, Bryan Richter, Parker Riley, Alex Castro Ros, Aurko Roy, Brennan Saeta, Rajkumar Samuel, Renee Shelby, Ambrose Slone, Daniel Smilkov, David R. So, Daniel Sohn, Simon Tokumine, Dasha Valter, Vijay Vasudevan, Kiran Vodrahalli, Xuezhi Wang, Pidong Wang, Zirui Wang, Tao Wang, John Wieting, Yuhuai Wu, Kelvin Xu, Yunhan Xu, Linting Xue, Pengcheng Yin, Jiahui Yu, Qiao Zhang, Steven Zheng, Ce Zheng, Weikang Zhou, Denny Zhou, Slav Petrov, and Yonghui Wu. 2023. Palm 2 technical report. Preprint, arXiv:2305.10403.
+
+Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Michal Podstawski, Lukas Gianinazzi, Joanna Gajda, Tomasz Lehmann, Hubert Niewiadomski, Piotr Nczyk, and Torsten Hoefler. 2024. Graph of thoughts: Solving elaborate problems with large language models. Proceedings of the AAAI Conference on Artificial Intelligence, 38(16):17682-17690.
+
+Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems, volume 33, pages 1877-1901. Curran Associates, Inc.
+
+Dheeru Dua, Shivanshu Gupta, Sameer Singh, and Matt Gardner. 2022. Successive prompting for decomposing complex questions. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 1251-1265, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.
+
+Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston Zhang, Aurelien Rodriguez, Austen Gregerson, Ava Spataru, Baptiste Roziere, Bethany Biron, Binh Tang, Bobbie Chern, Charlotte Caucheteux, Chaya Nayak, Chloe Bi, Chris Marra, Chris McConnell, Christian Keller, Christophe Touret, Chunyang Wu, Corinne Wong, Cristian Canton Ferrer, Cyrus Nikolaidis, Damien Alonsius, Daniel Song, Danielle Pintz, Danny Livshits, David Esiobu, Dhruv Choudhary, Dhruv Mahajan,
+
+Diego Garcia-Olano, Diego Perino, Dieuwke Hupkes, Egor Lakomkin, Ehab AlBadawy, Elina Lobanova, Emily Dinan, Eric Michael Smith, Filip Radenovic, Frank Zhang, Gabriel Synnaeve, Gabrielle Lee, Georgia Lewis Anderson, Graeme Nail, Gregoire Mialon, Guan Pang, Guillem Cucurell, Hailey Nguyen, Hannah Korevaar, Hu Xu, Hugo Touvron, Iliyan Zarov, Imanol Arrieta Ibarra, Isabel Kloumann, Ishan Misra, Ivan Evtimov, Jade Copet, Jaewon Lee, Jan Geffert, Jana Vranes, Jason Park, Jay Mahadeokar, Jeet Shah, Jelmer van der Linde, Jennifer Billock, Jenny Hong, Jenya Lee, Jeremy Fu, Jianfeng Chi, Jianyu Huang, Jiawen Liu, Jie Wang, Jiecao Yu, Joanna Bitton, Joe Spisak, Jongsoo Park, Joseph Rocca, Joshua Johnstun, Joshua Saxe, Junteng Jia, Kalyan Vasuden Alwala, Kartikeya Upasani, Kate Plawiak, Ke Li, Kenneth Heafield, Kevin Stone, Khalid El-Arini, Krithika Iyer, Kshitiz Malik, Kuenley Chiu, Kunal Bhalla, Lauren Rantala-Yeary, Laurens van der Maaten, Lawrence Chen, Liang Tan, Liz Jenkins, Louis Martin, Lovish Madaan, Lubo Malo, Lukas Blecher, Lukas Landzaat, Luke de Oliveira, Madeline Muzzi, Mahesh Pasupuleti, Mannat Singh, Manohar Paluri, Marcin Kardas, Mathew Oldham, Mathieu Rita, Maya Pavlova, Melanie Kambadur, Mike Lewis, Min Si, Mitesh Kumar Singh, Mona Hassan, Naman Goyal, Narjes Torabi, Nikolay Bashlykov, Nikolay Bogoychev, Niladri Chatterji, Olivier Duchenne, Onur Celebi, Patrick Alrassy, Pengchuan Zhang, Pengwei Li, Petar Vasic, Peter Weng, Prajjwal Bhargava, Pratik Dubal, Praveen Krishnan Punit Singh Koura, Puxin Xu, Qing He, Qingxiao Dong, Ragavan Srinivasan, Raj Ganapathy, Ramon Calderer, Ricardo Silveira Cabral, Robert Stojnic Roberta Raileanu Rohit Girdhar Rohit Patel,Romain SauvestreRonnie Polidoro Roshan Sumbaly Ross Taylor Ruan Silva Rui Hou Rui Wang Saghar Hosseini,Sahana Chennabasappa,Sanjay Singh Sean Bell Seohyun Sonia Kim,Sergey Edunov Shaoliang Nie, Sharan Narang, Sharath Raparthy Sheng Shen Shengye Wan Shruti BhosaleShun ZhangSimon Vandenhende Soumya Batra Spencer WhitmanSten Sootla Stephane CollotSuchin GururanganSydney BorodinskyTamar Herman,Tara Fowler Tarek Sheasha Thomas GeorgiouThomas Scialom Tobias Speckbacher,Todor Mihaylov Tong Xiao Ujjwal Karn,Vedanuj GoswamiVibhor Gupta,Vignesh Ramanathan,Viktor Kerkez,Vincent Gonguet,Virginie Do,Vish Vogeti,Vladan Petrovic Weiwei Chu,Wenhan Xiong,Wenyin Fu whitney Meers Xavier Martinet Xiaodong WangXiaqing Ellen Tan Xinfeng Xie Xuchao Jia Xuewei Wang,Yaelle Goldschlag,Yashesh Gaur,Yasmine Babaei Yi WenYiwen Song Yuchen Zhang Yue Li,Yuning Mao,Zacharie Delpierre Coudert Zheng YanZhengxing Chen,Zoe Papakipos,Aaditya Singh Aaron Grattafori Abha Jain Adam KelseyAdam Shajnfeld Adithya Gangidi Adolfo Victoria,Ahuva Goldstand Ajay Menon Ajay Sharma Alex BoesenbergAlex Vaughan Alexei Baevski Allie Feinstein Amanda Kallet Amit Sangani Anam Yunus Andrei Lupu Andres Alvarado Andrew Caples Andrew GuAndrew Ho Andrew Poulton Andrew Ryan Ankit Ramchandani Annie Franco Apara
+
+jita Saraf, Arkabandhu Chowdhury, Ashley Gabriel, Ashwin Bharambe, Assaf Eisenman, Azadeh Yazdan, Beau James, Ben Maurer, Benjamin Leonhardi, Bernie Huang, Beth Loyd, Beto De Paola, Bhargavi Paranjape, Bing Liu, Bo Wu, Boyu Ni, Braden Hancock, Bram Wasti, Brandon Spence, Brani Stojkovic, Brian Gamido, Britt Montalvo, Carl Parker, Carly Burton, Catalina Mejia, Changhan Wang, Changkyu Kim, Chao Zhou, Chester Hu, Ching-Hsiang Chu, Chris Cai, Chris Tindal, Christoph Feichtenhofer, Damon Civin, Dana Beaty, Daniel Kreymer, Daniel Li, Danny Wyatt, David Adkins, David Xu, Davide Testuggine, Delia David, Devi Parikh, Diana Liskovich, Didem Foss, Dingkang Wang, Duc Le, Dustin Holland, Edward Dowling, Eissa Jamil, Elaine Montgomery, Eleonora Presani, Emily Hahn, Emily Wood, Erik Brinkman, Esteban Arcaute, Evan Dunbar, Evan Smothers, Fei Sun, Felix Kreuk, Feng Tian, Firat Ozgenel, Francesco Caggioni, Francisco Guzmán, Frank Kanayet, Frank Seide, Gabriela Medina Florez, Gabriella Schwarz, Gada Badeer, Georgia Swee, Gil Halpern, Govind Thattai, Grant Herman, Grigory Sizov, Guangyi, Zhang, Guna Lakshminarayanan, Hamid Shojanazeri, Han Zou, Hannah Wang, Hanwen Zha, Haroun Habeeb, Harrison Rudolph, Helen Suk, Henry Aspegren, Hunter Goldman, Ibrahim Damlaj, Igor Molybog, Igor Tufanov, Irina-Elena Veliche, Itai Gat, Jake Weissman, James Geboski, James Kohli, Japhet Asher, Jean-Baptiste Gaya, Jeff Marcus, Jeff Tang, Jennifer Chan, Jenny Zhen Jeremy Reizenstein, Jeremy Teboul Jessica Zhong. Jian Jin,Jingyi Yang Joe Cummings Jon Carvill Jon Shepard Jonathan McPhie Jonathan Torres Josh Ginsburg Junjie Wang Kai Wu Kam Hou U,Karan Saxena,Karthik Prasad,Kartikay Khandealwal Katayoun Zand Kathy Matosich Kaushik Veeraraghavan Kelly Michelena Keqian Li Kun HuangKunal ChawlaKushal LakhotiaKyle Huang Lailin ChenLakshya Garg Lavender A. Leandro Silva,Lee Bell Lei Zhang Liangpeng Guo Licheng YuLiron MoshkovichLuca WehrstedtMadian Khabsa Manav Avalani Manish Bhatt Maria Tsimpoukelli,Martynas Mankus,Matan Hasson Matthew LennieMatthias Reso Maxim GroshevMaxim NaumovMaya Lathi Meghan Keneally Michael L. SeltzerMichal Valko Michelle RestrepoMihir Patel,Mik Vyatskov,Mikayel Samvelyan Mike Clark,Mike Macey,Mike Wang,Miquel Jubert Hermoso Mo Metanat Mohammad Rastegari Munish Bansal Nandhini Santhanam,Natascha Parks Natasha White,Navyata Bawa,Nayan Singhal,Nick EgeboNicolas Usunier,Nikolay Pavlovich Laptev Ning DongNing ZhangNorman ChengOleg ChernoguzOlivia Hart Omkar Salpekar Ozlem Kalinli Parkin Kent Parth ParekhPaul Saab,Pavan Balaji Pedro Rittner Philip Bontrager Pierre Roux Piotr DollarPolina ZvyaginaPrashant Ratanchandani,Pritish Yuvraj Qian Liang,Rachad Alao Rachel RodriguezRafi Ayub Ragtoham Murthy Raghu Nayani,Rahul MitraRaymond LiRebekkah Hogan Robin Battey,Rocky Wang Rohan Maheswari Russ Howes,Ruty Rinott Sai Jayesh Bondu Samyak Datta Sara Chugh Sara Hunt Sargun Dhillon,Sasha Sidorov,Satadru PanSaurabh Verma
+
+Seiji Yamamoto, Sharadh Ramaswamy, Shaun Lindsay, Shaun Lindsay, Sheng Feng, Shenghao Lin, Shengxin Cindy Zha, Shiva Shankar, Shuqiang Zhang, Shuqiang Zhang, Sinong Wang, Sneha Agarwal, Soji Sajuyigbe, Soumith Chintala, Stephanie Max, Stephen Chen, Steve Kehoe, Steve Satterfield, Sudarshan Govindaprasad, Sumit Gupta, Sungmin Cho, Sunny Virk, Suraj Subramanian, Sy Choudhury, Sydney Goldman, Tal Remez, Tamar Glaser, Tamara Best, Thilo Kohler, Thomas Robinson, Tianhe Li, Tianjun Zhang, Tim Matthews, Timothy Chou, Tzook Shaked, Varun Vontimitta, Victoria Ajayi, Victoria Montanez, Vijai Mohan, Vinay Satish Kumar, Vishal Mangla, Vitor Albiero, Vlad Ionescu, Vlad Poenaru, Vlad Tiberiu Mihailescu, Vladimir Ivanov, Wei Li, Wenchen Wang, Wenwen Jiang, Wes Bouaziz, Will Constable, Xiaocheng Tang, Xiaofang Wang, Xiaojian Wu, Xiaolan Wang, Xide Xia, Xilun Wu, Xinbo Gao, Yanjun Chen, Ye Hu, Ye Jia, Ye Qi, Yenda Li, Yilin Zhang, Ying Zhang, Yossi Adi, Youngjin Nam, Yu, Wang, Yuchen Hao, Yundi Qian, Yuzi He, Zach Rait, Zachary DeVito, Zef Rosnbrick, Zhaoduo Wen, Zhenyu Yang, and Zhiwei Zhao. 2024. The llama 3 herd of models. Preprint, arXiv:2407.21783.
+
+Yao Fu, Hao Peng, Ashish Sabharwal, Peter Clark, and Tushar Khot. 2022. Complexity-based prompting for multi-step reasoning. In *The Eleventh International Conference on Learning Representations*.
+
+Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021a. Measuring massive multitask language understanding. In International Conference on Learning Representations.
+
+Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. 2021b. Measuring mathematical problem solving with the math dataset. NeurIPS.
+
+Jie Huang, Xinyun Chen, Swaroop Mishra, Huaixiu Steven Zheng, Adams Wei Yu, Xinying Song, and Denny Zhou. 2024. Large language models cannot self-correct reasoning yet. In The Twelfth International Conference on Learning Representations.
+
+Tushar Khot, Harsh Trivedi, Matthew Finlayson, Yao Fu, Kyle Richardson, Peter Clark, and Ashish Sabharwal. 2023. Decomposed prompting: A modular approach for solving complex tasks. In *The Eleventh International Conference on Learning Representations*.
+
+Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. In Advances in Neural Information Processing Systems.
+
+Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, Shashank Gupta, Bodhisattwa Prasad Majumder,
+
+Katherine Hermann, Sean Welleck, Amir Yazdanbakhsh, and Peter Clark. 2023. Self-refine: Iterative refinement with self-feedback. In Thirty-seventh Conference on Neural Information Processing Systems.
+
+Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, Pouya Tafti, Léonard Hussenot, Pier Giuseppe Sessa, Aakanksha Chowdhery, Adam Roberts, Aditya Barua, Alex Botev, Alex Castro-Ros, Ambrose Slone, Amélie Héliou, Andrea Tacchetti, Anna Bulanova, Antonia Paterson, Beth Tsai, Bobak Shahriari, Charline Le Lan, Christopher A. Choquette-Choo, Clément Crepy, Daniel Cer, Daphne Ippolito, David Reid, Elena Buchatskaya, Eric Ni, Eric Noland, Geng Yan, George Tucker, George-Christian Muraru, Grigory Rozhdestvenskiy, Henryk Michalewski, Ian Tenney, Ivan Grishchenko, Jacob Austin, James Keeling, Jane Labanowski, Jean-Baptiste Lespiau, Jeff Stanway, Jenny Brennan, Jeremy Chen, Johan Ferret, Justin Chiu, Justin Mao-Jones, Katherine Lee, Kathy Yu, Katie Millican, Lars Lowe Sjoesund, Lisa Lee, Lucas Dixon, Rachel Reid, Maciej Mikula, Mateo Wirth, Michael Sharman, Nikolai Chinaev, Nithum Thain, Olivier Bachem, Oscar Chang, Oscar Wahltinez, Paige Bailey, Paul Michel, Petko Yotov, Rahma Chaabouni, Ramona Comanescu, Reena Jana, Rohan Anil, Ross McIlroy, Ruibo Liu, Ryan Mullins, Samuel L Smith, Sebastian Borgeaud, Sertan Girgin, Sholto Douglas, Shree Pandya, Siamak Shakeri, Soham De, Ted Klimenko, Tom Hennigan, Vlad Feinberg, Wojciech Stokowiec, Yu hui Chen, Zafarali Ahmed, Zhitao Gong, Tris Warkentin, Ludovic Peran, Minh Giang, Clément Farabet, Oriol Vinyals, Jeff Dean, Koray Kavukcuoglu, Demis Hassabis, Zoubin Ghahramani, Douglas Eck, Joelle Barral, Fernando Pereira, Eli Collins, Armand Joulin, Noah Fiedel, Evan Senter, Alek Andreev, and Kathleen Kenealy. 2024. Gemma: Open models based on gemini research and technology. Preprint, arXiv:2403.08295.
+
+Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam Fisch, Adam R. Brown, Adam Santoro, Aditya Gupta, Adrià Garriga-Alonso, Agnieszka Kluska, Aitor Lewkowycz, Akshit Agarwal, Alethea Power, Alex Ray, Alex Warstadt, Alexander W. Kocurek, Ali Safaya, Ali Tazarv, Alice Xiang, Alicia Parrish, Allen Nie, Aman Hussain, Amanda Askell, Amanda Dsouza, Ambrose Slone, Ameet Rahane, Anantharaman S. Iyer, Anders Johan Andreassen, Andrea Madotto, Andrea Santilli, Andreas Stuhlmüller, Andrew M. Dai, Andrew La, Andrew Lampinen, Andy Zou, Angela Jiang, Angelica Chen, Anh Vuong, Animesh Gupta, Anna Gottardi, Antonio Norelli, Anu Venkatesh, Arash Gholamidavoodi, Arfa Tabassum, Arul Menezes, Arun Kirubarajan, Asher Mullokandov, Ashish Sabharwal, Austin Herrick, Avia Efrat, Aykut Erdem, Ayla Karakas, B. Ryan Roberts, Bao Sheng Loe, Barret Zoph, Bartlomiej Bojanowski, Batuhan Özyurt, Behnam
+
+Hedayatnia, Behnam Neyshabur, Benjamin Inden, Benno Stein, Berk Ekmekci, Bill Yuchen Lin, Blake Howald, Bryan Orion, Cameron Diao, Cameron Dour, Catherine Stinson, Cedrick Argueta, Cesar Ferri, Chandan Singh, Charles Rathkopf, Chenlin Meng, Chitta Baral, Chiyu Wu, Chris Callison-Burch, Christopher Waites, Christian Voigt, Christopher D Manning, Christopher Potts, Cindy Ramirez, Clara E. Rivera, Clemencia Siro, Colin Raffel, Courtney Ashcraft, Cristina Garbacea, Damien Sileo, Dan Garrette, Dan Hendrycks, Dan Kilman, Dan Roth, C. Daniel Freeman, Daniel Khashabi, Daniel Levy, Daniel Mosegui Gonzalez, Danielle Perszyk, Danny Hernandez, Danqi Chen, Daphne Ippolito, Dar Gilboa, David Dohan, David Drakard, David Jurrgens, Debajyoti Datta, Deep Ganguli, Denis Emelin, Denis Kleyko, Deniz Yuret, Derek Chen, Derek Tam, Dieuwke Hupkes, Diganta Misra, Dilyar Buzan, Dimitri Coelho Mollo, Diyi Yang, Dong-Ho Lee, Dylan Schrader, Ekaterina Shutova, Ekin Dogus Cubuk, Elad Segal, Eleanor Hagerman, Elizabeth Barnes, Elizabeth Donoway, Ellie Pavlick, Emanuele Rodolà, Emma Lam, Eric Chu, Eric Tang, Erkut Erdem, Ernie Chang, Ethan A Chi, Ethan Dyer, Ethan Jerzak, Ethan Kim, Eunice Engefu Manyasi, Evgenii Zheltonozhskii, Fanyue Xia, Fatemeh Siar, Fernando Martínez-Plumed, Francesca Happé, François Chollet, Frieda Rong, Gaurav Mishra, Genta Indra Winata, Gerard de Melo, Germán Kruszewski Giambattista Parascandolo, Giorgio Mariani Gloria Xinyue Wang, Gonzalo Jaimovitch-Lopez, Gregor Betz, Guy Gur-Ari, Hana Galijasevic, Hannah Kim, Hannah Rashkin, Hannaneh Hajishirzi, Harsh Mehta, Hayden Bogar, Henry Francis Anthony Shevlin, Hinrich Schuetze Hiromu Yakura Hongming Zhang Hugh Mee Wong Ian Ng Isaac Noble Jaap Jumelet Jack Geissinger Jackson Kernion Jacob Hilton Jaehoon Lee Jaime Fernández Fisac James B Simon James Koppel James Zheng James Zou Jan Kocon Jana Thompson Janelle Wingfield Jared Kaplan Jarema Radom Jascha Sohl-Dickstein Jason Phang Jason Wei Jason Yosinski Jekaterina Novikova Jelle Bosscher Jennifer Marsh Jeremy Kim Jeroen Taal Jesse Engel Jesujoba Alabi Jiacheng Xu Jiaming Song Jillian Tang Joan Waweru John Burden John Miller John U. Balis Jonathan Batchelder Jonathan Berant Jörg Frohberg Jos Rozen Jose FernandezOrallo Joseph Boudeman Joseph Guerr Joseph Jones Joshua B. TenenbaumJoshua S. RuleJoyce Chua Kamil Kanclerz Karen Livescu Karl Krauth Karthik Gopalakrishnan Katerina Ignatyeva Katja Markert Kaustubh Dhole Kevin Gimpel Kevin Omondi Kory Wallace Mathewson Kristen Chiafullo Ksenia Shkaruta Kumar Shridhar Kyle McDonell Kyle Richardson Laria Reynolds Leo Gao Li Zhang Liam Dugan Lianhui Qin Lidia Contreras-Ochando Louis-Philippe Morency Luca Moschella Lucas Lam Lucy Noble Ludwig Schmidt Luheng He Luis Oliveros-Colón Luke Metz Lütfi Kerem Senel Maarten Bosma Maarten Sap Maartje Ter Hoeve Maheen Farooqi Manaal Faruqui Mantas Mazeika Marco Baturan Marco Marelli Marco Maru Maria Jose Ramirez-Quintana Marie Volkiehn Mario Giulianelli Martha Lewis Martin Pothast
+
+Matthew L Leavitt, Matthias Hagen, Mátías Schubert, Medina Orduna Baitemirova, Melody Arnaud, Melvin McElrath, Michael Andrew Yee, Michael Cohen, Michael Gu, Michael Ivanitskiy, Michael Starritt, Michael Strube, Michal Swedrowski, Michele Bevilacqua, Michihiro Yasunaga, Mihir Kale, Mike Cain, Minnie Xu, Mirac Suzgun, Mitch Walker, Mo Tiwari, Mohit Bansal, Moin Aminnaseri, Mor Geva, Mozhdeh Gheini, Mukund Varma T, Nanyun Peng, Nathan Andrew Chi, Nayeon Lee, Neta Gur-Ari Krakover, Nicholas Cameron, Nicholas Roberts, Nick Doiron, Nicole Martinez, Nikita Nangia, Niklas Deckers, Niklas Muennighoff, Nitish Shirish Keskar, Niveditha S. Iyer, Noah Constant, Noah Fiedel, Nuan Wen, Oliver Zhang, Omar Agha, Omar Elbaghdadi, Omer Levy, Owain Evans, Pablo Antonio Moreno Casares, Parth Doshi, Pascale Fung, Paul Pu Liang, Paul Vicol, Pegah Alipoformolabashi, Peiyuan Liao, Percy Liang, Peter W Chang, Peter Eckersley, Phu Mon Htut, Pinyu Hwang, Piotr Młkowski, Piyush Patil, Pouya Pezeshkpour, Priti Oli, Qiaozhu Mei, Qing Lyu, Qinlang Chen, Rabin Banjade, Rachel Etta Rudolph, Raefer Gabriel, Rahul Habacker, Ramon Risco, Raphael Milliere, Rhythm Garg, Richard Barnes, Rif A. Saurous, Riku Arakawa, Robbe Raymaekers, Robert Frank, Rohan Sikand, Roman Novak, Roman Sitelew, Ronan Le Bras, Rosanne Liu, Rowan Jacobs, Rui Zhang, Russ Salakhutdinov, Ryan Andrew Chi, Seungjae Ryan Lee, Ryan Stovall, Ryan Teehan, Ryan Yang, Sahib Singh, Saif M. Mohammad, Sajant Anand, Sam Dillavou, Sam Shleifer, Sam Wiseman, Samuel Gruetter, Samuel R. Bowman, Samuel Stern Schoenholz, Sanghyun Han, Sanjeev Kwatra, Sarah A. Rous, Sarik Ghazarian, Sayan Ghosh, Sean Casey, Sebastian Bischoff, Sebastian Gehrmann, Sebastian Schuster, Sepideh Sadeghi, Shadi Hamdan, Sharon Zhou, Shashank Srivastava, Sherry Shi, Shikhar Singh, Shima Asaadi, Shixiang Shane Gu, Shubh Pachchigar, Shubham Toshiwal, Shyam Upadhyay, Shyamolima Shammie Debnath, Siamak Shakeri, Simon Thormeyer, Simone Melzi, Siva Reddy, Sneha Priscilla Makini, Soo-Hwan Lee, Spencer Tirene, Sriharsha Hatwar, Stanislas Dehaene, Stefan Divic, Stefano Ermon, Stella Biderman, Stephanie Lin, Stephen Prasad Steven Piantadosi, Stuart Shieber, Summer Misherghi, Svetlana Kiritchenko, Swaroop Mishra Tal Linzen, Tal Schuster Tao Li Tao Yu Tariq Ali Tatsunori Hashimoto Te-Lin Wu Theo Desbordes Theodore Rothschild Thomas Phan Tianle Wang Tiberius Nkinyili Timo Schick Timofei Kornev Titus Tunduny Tobias Gerstenberg Trenton Chang Trishala Neeraj Tushar Khot Tyler Shultz Uri Shaham,Vedant Misra,Vera Demberg,Victoria Nyamai,Vikas Raunak,Vinay Venkatesh Ramasesh vinay uday prabhu,Vishakh Padmakumar Vivek Srikumar William Fedus William Saunders William Zhang,Wout Vossen Xiang Ren Xiaoyu TongXinran Zhao Xinyi WuXudong Shen Yadollah Yaghoobzadeh,Yair Lakretz,Yangqiu Song Yasaman Bahri,Yejin Choi,Yichi Yang,Yiding HaoYifu ChenYonatan BelinkovYu HouYu fang Hou Yuntao Bai Zachary Seid Zhuoye Zhao
+
+Zijian Wang, Zijie J. Wang, Zirui Wang, and Ziyi Wu. 2023. Beyond the imitation game: Quantifying and extrapolating the capabilities of language models. Transactions on Machine Learning Research.
+
+Mirac Suzgun, Nathan Scales, Nathanael Scharli, Sebastian Gehrmann, Yi Tay, Hyung Won Chung, Aakanksha Chowdhery, Quoc Le, Ed Chi, Denny Zhou, and Jason Wei. 2023. Challenging BIG-bench tasks and whether chain-of-thought can solve them. In Findings of the Association for Computational Linguistics: ACL 2023, pages 13003-13051, Toronto, Canada. Association for Computational Linguistics.
+
+Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinez, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. 2023. Llama 2: Open foundation and finetuned chat models. Preprint, arXiv:2307.09288.
+
+Lei Wang, Wanyu Xu, Yihuai Lan, Zhiqiang Hu, Yunshi Lan, Roy Ka-Wei Lee, and Ee-Peng Lim. 2023a. Plan-and-solve prompting: Improving zero-shot chain-of-thought reasoning by large language models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2609-2634, Toronto, Canada. Association for Computational Linguistics.
+
+Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H. Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023b. Self-consistency improves chain of thought reasoning in language models. In The Eleventh International Conference on Learning Representations.
+
+Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian richter, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou. 2022. Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems, volume 35, pages 24824-24837. Curran Associates, Inc.
+
+Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. Griffiths, Yuan Cao, and Karthik R Narasimhan. 2023. Tree of thoughts: Deliberate
+
+problem solving with large language models. In Thirty-seventh Conference on Neural Information Processing Systems.
+
+Zhuosheng Zhang, Aston Zhang, Mu Li, and Alex Smola. 2023. Automatic chain of thought prompting in large language models. In *The Eleventh International Conference on Learning Representations*.
+
+Denny Zhou, Nathanael Scharli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans, Claire Cui, Olivier Bousquet, Quoc V Le, and Ed H. Chi. 2023. Least-to-most prompting enables complex reasoning in large language models. In The Eleventh International Conference on Learning Representations.
+
+# A Clarifications
+
+# A.1 Canonical Tasks
+
+In Figure 1 we showed how to apply  $\text{ToP}(2,1)$  to an instance of Last Letter Concatenation. We illustrate how  $\text{ToP}(2,2)$  would look for concatenating the last letters of a list of 8 words in Figure 3. The decomposition is done on two levels, the leaves being solved first and the merge operation being recursively applied from the bottom to the top.
+
+# A.2 Sequential tasks
+
+Let us say that we have a system at state  $s_0$ , and we want to find its state after going through  $m$  processing steps  $(p_1, \ldots, p_m)$  in this order (i.e. a sequential task). Applying  $ToP(1, k)$  is equivalent to grouping the above steps into  $k$  groups  $G_1 = (p_1, \ldots, p_{\lceil \frac{m}{k} \rceil}), \ldots, G_k = (p_{m - \lfloor \frac{m}{k} \rfloor + 1}, \ldots, p_m)$ . We build a path graph from top to bottom, where the root is the main instance, and the leaf is the instance defined by  $s_0$  and  $G_1$ . Solving it yields a state  $s_1$  to which we apply the steps  $G_2$  and so on until we reach  $G_k$ . Tracking Shuffled Objects is an example of such a task. At the start,  $L$  people are assigned one object each. We are interested in recovering the assignment between people and objects after  $L$  swaps (transpositions). Figure 4 illustrates the application of  $ToP(1, 3)$  to an instance with 3 swaps. We first decompose the main instance into 3 subinstances; here, each instance corresponds to one swap. After decomposition, only the first instance has the correct initial assignment (grey part). For the remaining instances, placeholders are used, which will later be replaced by the solutions to the problems they depend on.
+
+![](images/5a639605737e1cd78430e47efcae19246e99ee44de4cf0e0e521f566ce29d6fd.jpg)  
+Figure 3: Overview of ToP (2, 2) for Last Letter Concatenation. The list of words is divided into 2 sublists which are recursively divided into two sublists. The problems at the leaves, which consist into concatenating the last letters of 2-word lists are solved first. The solutions are then merged in a bottom-up way until the main instance is solved.
+
+![](images/74e548d0899b5377d77c7bfdf874904453180bdffecc0ebf50dbed8c8c9e148f.jpg)  
+Figure 4: Overview of  $\text{ToP}(1,3)$  on an instance of Tracking Shuffled Objects (three objects).
+
+# A.3 Comparison with Least-to-Most Prompting
+
+Least-to-Most prompting also handles Last Letter Concatenation as a sequential task. In this regards, it is similar to  $\text{ToP}(1, L)$  on list with  $L$  words. As illustrated in Figure 5, L2M uses all couples
+
+instance-solution preceding an instance to build the prompt to solve it whereas ToP only uses the couples directly connected to it in the tree hierarchy.
+
+![](images/08917401017a7452ead65be12e0be06645abdcbfde6fa77f4b3aeed2df1f63bc.jpg)  
+Figure 5: Overview of L2M prompting on Last Letter Concatenation with 4 words.
+
+# B Additional Experiments
+
+# B.1 Scaling behaviour
+
+In this section, we study how ToP behaves as we vary the model scale. In Figure 6, we plot the performance of both IO and CoT prompting and ToP as a function of model scale for LLaMA 2 models (Touvron et al., 2023) and 3 BBH tasks. We use ToP (2, 1) for canonical tasks and ToP (1, 2) for sequential tasks. For all tasks, scaling up model size improved the performance of ToP beyond CoT prompting. LLaMA 2 70B achieves a  $98\%$  accuracy on Object Counting, an absolute improvement of  $18.8\%$  over CoT. ToP improves over random accuracy of IO and CoT on Web of Lies with LLaMA 2 7B, with an accuracy of  $72.8\%$ .
+
+We report IO prompting, CoT prompting and ToP performance on 8 BBH tasks in Table 6. ToP consistently yields an improvement of performance compared to IO and CoT prompting for most tasks and at all scales.
+
+# C Analysis
+
+We aim to gain a comprehensive understanding of the performance improvements offered by the ToP framework. We theoretically derive an upper bound of expected ToP's performance, then we study the impact of the tree structure on the results obtained. For the experiments in this section, we
+
+use LLaMA 3 8B (Dubey et al., 2024) unless stated otherwise.
+
+# C.1 Theoretical Analysis
+
+Let us consider a task with  $n$  problems. Each problem is further divided into  $k$  subproblems, resulting in a total of  $nk$  subproblems. If we evaluate an LLM on these  $nk$  subproblems and obtain  $m$  incorrect answers, we can infer the number of incorrect answers likely to occur when evaluating the original  $n$  problems. Assuming that an incorrect answer to a subproblem implies an incorrect answer to its corresponding main problem, we can analyze the outcomes in two scenarios. In the worst case, each of the  $m$  incorrect subproblems is associated with a distinct main problem and thus there would be  $m$  main problems with incorrect answers. The best case is when the  $m$  incorrect subproblems are distributed such that each affected main problem has  $k$  or  $m\% k$  incorrect subproblems. Consequently, the number of main problems with incorrect answers would be at most  $\lceil \frac{m}{k} \rceil$ . From this analysis, we can deduce that the accuracy at any level  $l$  of the problem hierarchy is constrained by the accuracy at level  $l - 1$ . Therefore, the accuracy for the overall task (the root of the hierarchy) is bounded by the accuracy observed at the most granular level (the leaves of the hierarchy). We validate this analysis by comparing the accuracy at level 1 to the accuracy at level 0 (main problem) for some of the
+
+![](images/9157697dbeab75988da497bfcdf08d5b65821083a3a9d11f4307324d190e192d.jpg)
+
+![](images/16bf5addfd0787582e982004b28152cc99076d2cf62fbfa671d0f99914cc441b.jpg)
+
+![](images/be97057dfb9eb549e94c53f2570d6f9a760c5c80d9c99326882c7bcd7ffabc31.jpg)
+
+Figure 6: Scaling behavior of ToP compared to IO and CoT with the LLaMA 2 family on 3 BBH tasks.  
+
+<table><tr><td rowspan="2">BBH Tasks</td><td colspan="3">LLaMA 2 7B</td><td colspan="3">LLaMA 2 13B</td><td colspan="3">LLaMA 2 70B</td></tr><tr><td>IO</td><td>CoT</td><td>ToP</td><td>IO</td><td>CoT</td><td>ToP</td><td>IO</td><td>CoT</td><td>ToP</td></tr><tr><td>Boolean Expressions</td><td>0.680</td><td>0.628</td><td>0.672</td><td>0.728</td><td>0.768</td><td>0.728</td><td>0.812</td><td>0.868</td><td>0.924</td></tr><tr><td>Hyperbaton</td><td>0.530</td><td>0.550</td><td>0.562</td><td>0.508</td><td>0.666</td><td>0.700</td><td>0.530</td><td>0.684</td><td>0.830</td></tr><tr><td>Multistep Arithmetic Two</td><td>0.008</td><td>0.004</td><td>0.012</td><td>0.012</td><td>0.024</td><td>0.044</td><td>0.016</td><td>0.196</td><td>0.216</td></tr><tr><td>Navigate</td><td>0.272</td><td>0.164</td><td>0.088</td><td>0.340</td><td>0.308</td><td>0.156</td><td>0.336</td><td>0.400</td><td>0.284</td></tr><tr><td>Object Counting</td><td>0.404</td><td>0.476</td><td>0.516</td><td>0.492</td><td>0.532</td><td>0.616</td><td>0.540</td><td>0.792</td><td>0.98</td></tr><tr><td colspan="10">Tracking Shuffled Objects</td></tr><tr><td>Three</td><td>0.156</td><td>0.156</td><td>0.136</td><td>0.076</td><td>0.184</td><td>0.132</td><td>0.056</td><td>0.584</td><td>0.568</td></tr><tr><td>Five</td><td>0.000</td><td>0.000</td><td>0.000</td><td>0.012</td><td>0.044</td><td>0.048</td><td>0.080</td><td>0.528</td><td>0.664</td></tr><tr><td>Seven</td><td>0.000</td><td>0.000</td><td>0.000</td><td>0.000</td><td>0.000</td><td>0.004</td><td>0.000</td><td>0.288</td><td>0.592</td></tr><tr><td>Web of Lies</td><td>0.488</td><td>0.528</td><td>0.728</td><td>0.552</td><td>0.984</td><td>1.000</td><td>0.488</td><td>0.996</td><td>0.984</td></tr><tr><td>Word Sorting</td><td>0.418</td><td>0.146</td><td>0.244</td><td>0.538</td><td>0.261</td><td>0.320</td><td>0.788</td><td>0.445</td><td>0.717</td></tr></table>
+
+Table 6: Few-shot prompting performance of the LLaMA 2 family on BIG-Bench Hard (BBH).
+
+![](images/2bfa94e294eb0915777dbb688e67fec72d698ee61078fa0a8d16671159c01313.jpg)  
+Figure 7: Comparison between CoT, ToP, ToP with an Oracle Merger and the leaves' accuracy on 3 BBH tasks.
+
+aforementioned BBH tasks. The results are summarized in Figure 7. The Oracle Merger represents the accuracy that would be achieved if the merger process were flawless.
+
+As expected, the accuracy at the leaves acts as an upper bound for ToP. Moreover, the Oracle Merger yields better performance than vanilla ToP. This suggests that there is a loss in accuracy when going from level  $k$  to level  $k - 1$ , which can prevent ToP from achieving an even higher performance. Interestingly, what happens with Multistep Arithmetic Two comes close to the worst case scenario that we depicted earlier. Despite the leaves' accuracy be
+
+ing  $55\%$ , ToP + Oracle Merger fails to outperform CoT's  $34\%$  accuracy, showing that the distribution of the correct leaves' instances inherently undermines ToP performance in this scenario.
+
+# C.2 Impact of the tree structure.
+
+GoT Tasks. We analyze the impact of the tree structure on ToP's results. As shown previously, there may be a loss in accuracy during the merge operation. A deeper tree means more of these losses, but it also means easier subproblems. For the three GoT tasks, we analyze the impact of the tree's depth when the breadth is set to two with LLaMA 3 70B Instruct (Dubey et al., 2024).
+
+As suggested by Figure 8, deeper trees led to a higher accuracy for all three tasks. This is because we observed very few errors during the merge operation performed by the LLM. Going deeper, even with a near perfect merger can negatively affect performance as observed with the Set Intersection task, which has an accuracy of  $47\%$  with  $d = 4$  but  $74\%$  with  $d = 3$  and  $62\%$  with  $d = 2$ . The small errors performed at the leaves being propa
+
+![](images/98cf94e0b12ebec4d940c60f194c26e378be35cd0b4e5e39daaf273129f63fcf.jpg)  
+Figure 8: Impact of the tree structure (depth) on the accuracy on the 3 GoT Tasks with LLaMA 3 70B Instruct. Depth = 0 represents CoT prompting.
+
+gated during the repetitive merge operations impact the overall accuracy of ToP. In terms of breadth, applying ToP (4, 1) to Set Intersection yields the same accuracy of  $62\%$  as ToP (2, 2). We observed ToP (4, 2) to have a  $49\%$  accuracy, comparable to ToP (2, 4)'s  $47\%$ .
+
+BBH Tasks. Tracking Shuffled Objects involves recovering the final assignment of  $L$  objects given to  $L$  people ( $L \in \{3, 5, 7\}$ ) after a series of  $L$  transpositions (pairwise swaps). Applying ToP ( $1, d$ ) to these tasks implies using  $d$  even subseries of swaps in a manner akin to Navigate (see Figure 1). We study the impact of various depths and we report the results in Figure 9.
+
+Across all settings, the task accuracy gradually increases with deeper trees and reaches its maximum when all the subproblems involve only one swap (depth  $= L - 1$ ). The trade-off between the number of merge operations and the accuracy of simple instances is not at play here.
+
+On Multistep Arithmetic Two, ToP (2, 1) and ToP (2, 2) respectively achieve  $30.8\%$  and  $57.2\%$  accuracy where CoT is at  $34\%$ . Similarly, ToP (1, 2) and ToP (1, 3) respectively achieve  $60\%$  and  $66.4\%$  where CoT is at  $60.4\%$  on Navigate. This suggests that the tree structure can greatly affect the quality of ToP.
+
+# C.3 Robustness to the solve prompt.
+
+Throughout our experiments, we used CoT prompting to solve the most granular subproblems (the tree's leaves). In this section, we examine the impact of using IO prompting to solve the leaves. We conduct experiments on Word Sorting, which did not benefit from CoT prompting as shown in
+
+![](images/0316a116f672ece1ebae39eca459d7e0ed3d85d680ad7016736b3f6f5725c72c.jpg)  
+Figure 9: Impact of the tree structure (depth) on the accuracy of ToP on Tracking Shuffled Objects  $(L\in \{3,5,7\})$ . Depth  $= 0$  and depth  $= 1$  represent CoT prompting.
+
+Table 4. Additionally, we include Tracking Shuffled Objects (3, 5), Boolean Expressions, Multistep Arithmetic Two, and Object Counting, where IO prompting produced much poorer results compared to CoT. The results are summarized in Table 7.  
+
+<table><tr><td rowspan="2">BBH tasks</td><td colspan="2">LLaMA 3 8B</td></tr><tr><td>IO</td><td>IO + ToP</td></tr><tr><td>Boolean Expressions</td><td>0.824</td><td>0.876</td></tr><tr><td>Multistep Arithmetic Two</td><td>0.008</td><td>0.036</td></tr><tr><td>Object Counting</td><td>0.492</td><td>0.552</td></tr><tr><td>Tracking Shuffled Objects</td><td></td><td></td></tr><tr><td>Three</td><td>0.132</td><td>0.196</td></tr><tr><td>Five</td><td>0.004</td><td>0.008</td></tr><tr><td>Web of Lies</td><td>0.528</td><td>0.484</td></tr><tr><td>Word Sorting</td><td>0.647</td><td>0.679</td></tr></table>
+
+Table 7: Comparison of IO prompting and IO + ToP on BBH tasks.
+
+We observe that  $\mathrm{IO} + \mathrm{ToP}$  consistently outperforms standalone IO prompting on all the BBH tasks except Web of Lies. This is because the LLM's performance is random (0.528) on that specific task and dividing it into simpler instances does not resolve the issue (leaves'  $\mathrm{EM} = 0.496$ ). Moreover, applying  $\mathrm{IO} + \mathrm{ToP}$  on Multistep Arithmetic Two with a depth of 2 instead of 1 gives a  $22\%$  accuracy. These results demonstrate that ToP is robust to low-quality solve_prompt but achieves the best results with a good solve_prompt.
+
+# C.4 Analysis of failure cases
+
+In Table 4, we showed that ToP induces considerable performance gains on various complex BBH tasks with gpt-3.5-turbo-instruct. However, ToP specifically failed to improve performance on Multistep Arithmetic Two and Object Counting. An analysis of the model's outputs suggests 2 main causes:
+
+1. The founding hypothesis of ToP is not always verified. The founding hypothesis of ToP is the ability of an LLM to perform better on simpler instances of a task compared to more complex ones. When this is not verified, ToP's performance tends to fall behind CoT prompting. We strikingly observe such a behaviour on Object Counting. For multiple instances, the LLM was able to count the relevant items in a given list but fails to do so in a subset of the same list, as illustrated in Table 8. This also frequently occurs with Multistep Arithmetic Two instances, where the  $73.6\%$  ToP's accuracy cannot be improved even with an Oracle merger.
+
+MAININSTANCE  
+
+<table><tr><td>Q: I have three oranges, a pig, a frog, a cow, three bananas, a nectarine, and a snail. How many animals do I have?</td></tr><tr><td>Let&#x27;s think step by step.We first identify the animals on the list and include their quantity in parentheses:- pig (1)- frog (1)- cow (1)- snail (1)Now, let&#x27;s add the numbers in parentheses: 1 + 1 + 1 + 1 = 4. So the answer is 4.</td></tr><tr><td>SUBSTANCE</td></tr><tr><td>Q: I have three bananas, a nectarine, and a snail. How many animals do I have?</td></tr><tr><td>Let&#x27;s think step by step.We first identify the animals on the list and include their quantity in parentheses:- bananas (3)- nectarine (1)- snail (1)Now, let&#x27;s add the numbers in parentheses: 3 + 1 + 1 = 5. So the answer is 5.</td></tr></table>
+
+Table 8: gpt-3.5-turbo-instruct is able to accurately count the number of animals in a given list but fails to count it in a sublist.
+
+2. The LLM fails to merge subproblems' solutions into their parent's solution. In Appendix C.1, we showed that the use of an oracle
+
+merger leads to even greater improvements in the ToP framework. Additionally, a bad merger can bring ToP below CoT performance. Table 9 shows an example on the Object Counting task, where the LLM accurately solve both subproblems but fail to combine their answers. We believe that such issues can be significantly alleviated with a better merge_prompt.
+
+SUBSTANCE 1  
+
+<table><tr><td>Q: I have a yam, a cauliflower, a bed, two cabbages, a garlic, and an oven. How many vegetables do I have?</td></tr><tr><td>Let&#x27;s think step by step.</td></tr><tr><td>We first identify the vegetables on the list and include their quantity in parentheses:</td></tr><tr><td>- yam (1)</td></tr><tr><td>- cauliflower (1)</td></tr><tr><td>- cabbage (2)</td></tr><tr><td>- garlic (1)</td></tr><tr><td>Now, let&#x27;s add the numbers in parentheses: 1 + 1 + 2 + 1 = 5. So the answer is 5.</td></tr></table>
+
+SUBSTANCE 2  
+
+<table><tr><td>Q: I have a carrot, a head of broccoli, a potato, a stalk of celery, a lettuce head, and a toaster. How many vegeta-
+bles do I have?</td></tr><tr><td>Let&#x27;s think step by step.
+We first identify the vegetables on the list and include their quantity in parentheses:
+- carrot (1)
+- broccoli (1)
+- potato (1)
+- celery (1)
+- lettuce (1)</td></tr><tr><td>Now, let&#x27;s add the numbers in parentheses: 1 + 1 + 1 + 1 + 1 = 5. So the answer is 5.</td></tr><tr><td>MAIN instance</td></tr><tr><td>Q: I have a yam, a cauliflower, a bed, two cabbages, a garlic, an oven, a carrot, a head of broccoli, a potato, a stalk of celery, a lettuce head, and a toaster. How many vegetables do I have?</td></tr><tr><td>Let&#x27;s think step by step.
+The list of words can be written as the concatenation of A and B where
+- A = a yam, a cauliflower, a bed, two cabbages, and a garlic.
+- B = an oven, a carrot, a head of broccoli, a potato, a stalk of celery, and a lettuce head.
+According to the premises, A contains 5 vegetables.
+According to the premises, B contains 6 vegetables.
+Then, the total number of vegetables is A + B = 5 + 6 = 11. So the answer is 11.</td></tr></table>
+
+Table 9: gpt-3.5-turbo-instruct is unable to use the information in its context to solve the given problem.
+
+# D Implementation details
+
+# D.1 Language Models and Datasets
+
+In Table 10, we list the links to the relevant resources used to build this work.
+
+# D.2 Sampling scheme and evaluation metrics
+
+For prompting strategies that require sampling, we use nucleus sampling with  $T = 0.7$  and top-p  $= 0.95$ . Otherwise, we use greedy decoding  $(T = 0.0)$ . In all experiments, we generate at most 2000 new tokens and use Exact String Match (EM) as the evaluation metric.
+
+# D.3 Decomposition
+
+All the decomposers are algorithmic. For sorting and set intersection, a list  $\mathsf{L}$  is divided into  $\mathsf{L}[\emptyset : \mathsf{len}(\mathsf{L}) // 2]$  and  $\mathsf{L}[\mathsf{len}(\mathsf{L}) // 2]$ . For keyword sorting, we decompose the input text based on its sentences with the help of a sentence splitter module.<sup>4</sup>
+
+# D.4 How to choose the breadth and the depth
+
+ToP depends on 2 parameters, the breadth and the depth of the tree structure. A quick analysis of the problem can lead to an informed guess about what a good breadth should be. This is typically the case of sorting problems when a breadth of 2 helps to mimic the merge sort algorithm. We mostly experimented with a breadth of 2 for canonical tasks and saw that it yielded very good results. When it comes to sequential problems, the breadth is 1 and the depth plays the role of the number of block of steps before reaching the final state (depth-wise decomposition). Using 2 blocks also gave good results, but deeper trees tend to always give better results for such problems.
+
+# E Prompts
+
+# E.1 GoT Tasks
+
+We provide the links to all the prompts used to solve the GoT tasks in Table 11.
+
+# E.2 BBH tasks
+
+We describe the modification applied to 3 BBH tasks: Hyperbaton, Navigate and Tracking Shuffled Objects. Instead of choosing which of two sentences has the right adjective ordering, we turned
+
+each instance of Hyperbaton into 2 separate instances of Yes/No questions relative to adjectives ordering. Table 12 shows an example instance before and after the modification (we only reported one separate instance for visibility). We modify Navigate to require to find the final position after a series of instructions instead of asking if they lead to the starting point. Table 13 shows an example instance before and after modification. In Tracking Shuffled Objects, we ask the LLM to recover the final object assignments after the pairwise swaps as exemplified in Table 14.
+
+We provide the links to all the prompts used to solve the BBH tasks in Table 15.
+
+# E.3 Symbolic Reasoning
+
+We provide the links to all the prompts used to solve Coin Flip and Last Letter Concatenation in Table 16.
+
+<table><tr><td colspan="2">Datasets for Comparisons</td></tr><tr><td>BIG-Bench Hard</td><td>https://huggingface.co/datasets/lukaemon/bbh</td></tr><tr><td colspan="2">Models for Evaluations</td></tr><tr><td>GPT-3.5-Turbo</td><td>gpt-3.5-turbo (gpt-3.5-turbo-0125)</td></tr><tr><td>GPT-3.5-Turbo-Instruct</td><td>gpt-3.5-turbo-instruct</td></tr><tr><td>LLaMA 3 8B</td><td>https://huggingface.co/meta-llama/Meta-Llama-3-8B</td></tr><tr><td>LLaMA 3 70B</td><td>https://huggingface.co/casperhausen/llama-3-70b-instruct-awq</td></tr><tr><td>LLaMA 2 7B</td><td>https://huggingface.co/meta-llama/Llama-2-7b-hf</td></tr><tr><td>LLaMA 2 13B</td><td>https://huggingface.co/meta-llama/Llama-2-13b-hf</td></tr><tr><td>LLaMA 2 70B</td><td>https://huggingface.co/TheBloke/Llama-2-70B-AWQ</td></tr></table>
+
+Table 10: Links to datasets, benchmarks and language models.  
+
+<table><tr><td colspan="2">CoT</td></tr><tr><td>Sorting</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/got/prompts/cot/sorting.txt</td></tr><tr><td>Set Intersection</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/got/prompts/cot/set Intersection.txt</td></tr><tr><td>Keyword Counting</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/got/prompts/cot/keyword_counting.txt</td></tr><tr><td colspan="2">Merge</td></tr><tr><td>Sorting</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/got/prompts/merge/sorting.txt</td></tr><tr><td>Set Intersection (2)</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/got/prompts/merge/set Intersection.txt</td></tr><tr><td>Set Intersection (4)</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/got/prompts/merge/set Intersection_4.txt</td></tr><tr><td>Keyword Counting</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/got/prompts/merge/keyword_counting.txt</td></tr></table>
+
+Table 11: Links to solve and merge prompts of the GoT Tasks.  
+
+<table><tr><td>BEFORE</td></tr><tr><td>Q: Which sentence has the correct adjective order:
+Options:
+(A) rubber terrible ship
+(B) terrible rubber ship</td></tr><tr><td>IO Answer</td></tr><tr><td>(B)</td></tr><tr><td>CoT Answer</td></tr><tr><td>A: Let&#x27;s think step by step.
+When there is more than one adjective before a noun, the adjectives need to respect the following order before a noun: &quot;[1. opinion] [2. size] [3. age] [4. shape] [5. color] [6. origin] [7. material] [8. purpose] noun&quot;.
+Option (A): &quot;rubber terrible ship&quot;. (1) rubber&quot; falls into the material category. (2) &quot;terrible&quot; falls into the opinion category.
+Option (A) has the following adjective order: [7. material] [1. opinion] (or, in numeric terms, 7 1). Because 7 &lt; 1 is not correct, (A) does not have the correct ordering.
+Option (B): &quot;terrible rubber ship&quot;. Option (B) has the following adjective order: [1. opinion] [7. material] (or, in numeric terms, 1 7). Because 1 &lt; 7 is correct, (B) has the correct ordering. So the answer is (B).</td></tr><tr><td>AFTER</td></tr><tr><td>Q: Answer with Yes or No. Does the following sentence have the correct adjective order?
+rubber terrible ship</td></tr><tr><td>IO Answer</td></tr><tr><td>No</td></tr><tr><td>CoT Answer</td></tr><tr><td>A: Let&#x27;s think step by step.
+When there is more than one adjective before a noun, the adjectives need to respect the following order before a noun: &quot;[1. opinion] [2. size] [3. age] [4. shape] [5. color] [6. origin] [7. material] [8. purpose] noun&quot;.
+The sentence is &quot;rubber terrible ship&quot;. (1) rubber&quot; falls into the material category. (2) &quot;terrible&quot; falls into the opinion category.
+The sentence has the following adjective order: [7. material] [1. opinion] (or, in numeric terms, 7 1). Because 7 &lt; 1 is not correct, the sentence does not have the correct ordering. So the answer is No.</td></tr></table>
+
+Table 12: Modification applied to Hyperbaton.
+
+<table><tr><td>BEFORE</td></tr><tr><td>Q: If you follow these instructions, do you return to the starting point? Turn left. Turn around. Turn left. Take 7 steps. Take 2 steps. Take 4 steps. Take 8 steps.</td></tr><tr><td>IO Answer</td></tr><tr><td>No</td></tr><tr><td>CoT Answer</td></tr><tr><td>A: Let&#x27;s think step by step. We start at the origin (0, 0), facing the positive y-axis.
+(1) Turn left: (0, 0), facing the negative x-axis.
+(2) Turn around: (0, 0), facing the positive x-axis.
+(3) Turn left: (0, 0), facing the positive y-axis.
+(4) Take 7 steps: (0, 7), facing the positive y-axis.
+(5) Take 2 steps: (0, 9), facing the positive y-axis.
+(6) Take 4 steps: (0, 13), facing the positive y-axis.
+(7) Take 8 steps: (0, 21), facing the positive y-axis.
+Since (0, 21) is not (0, 0), we are not where we started. So the answer is No.</td></tr><tr><td>AFTER</td></tr><tr><td>Q: If you follow these instructions, what are the coordinates of the end point if you start at the point (0, 0), facing the positive y-axis? Turn left. Turn around. Turn left. Take 7 steps. Take 2 steps. Take 4 steps. Take 8 steps.</td></tr><tr><td>IO Answer</td></tr><tr><td>(0, 21)</td></tr><tr><td>CoT Answer</td></tr><tr><td>A: Let&#x27;s think step by step. We start at the origin (0, 0), facing the positive y-axis.
+(1) Turn left: (0, 0), facing the negative x-axis.
+(2) Turn around: (0, 0), facing the positive x-axis.
+(3) Turn left: (0, 0), facing the positive y-axis.
+(4) Take 7 steps: (0. 7), facing the positive y-axis.
+(5) Take 2 steps: (0, 9), facing the positive y-axis.
+(6) Take 4 steps: (0, 13), facing the positive y-axis.
+(7) Take 8 steps: (0, 21), facing the positive y-axis.
+So the answer is (0, 21).</td></tr></table>
+
+Table 13: Modification applied to Navigate.
+
+# BEFORE
+
+Q: Alice, Bob, and Claire are friends and avid readers who occasionally trade books. At the start of the semester, they each buy one new book: Alice gets Ulysses, Bob gets Frankenstein, and Claire gets Lolita. As the semester proceeds, they start trading around the new books. First, Claire and Bob swap books. Then, Bob and Alice swap books. Finally, Claire and Bob swap books.
+
+At the end of the semester, Bob has
+
+Options:
+
+(A) Ulysses  
+(B) Frankenstein  
+(C) Lolita
+
+# IO Answer
+
+# (B)
+
+# CoT Answer
+
+A: Let's think step by step.
+
+(0) At the start: Alice: Ulysses, Bob: Frankenstein, Claire: Lolita.  
+(1) Claire and Bob swap books: Alice: Ulysses, Bob: Lolita, Claire: Frankenstein.  
+(2) Bob and Alice swap books: Alice: Lolita, Bob: Ulysses, Claire: Frankenstein.  
+(3) Claire and Bob swap books: Alice: Lolita, Bob: Frankenstein, Claire: Ulysses.
+
+At the end of the semester, Bob has Frankenstein. So the answer is (B).
+
+# AFTER
+
+Q: Alice, Bob, and Claire are friends and avid readers who occasionally trade books. At the start of the semester, they each buy one new book: Alice gets Ulysses, Bob gets Frankenstein, and Claire gets Lolita. As the semester proceeds, they start trading around the new books. First, Claire and Bob swap books. Then, Bob and Alice swap books. Finally, Claire and Bob swap books.
+
+At the end of the semester, what is the assignment of books?
+
+# IO Answer
+
+# Alice: Lolita, Bob: Frankenstein, Claire: Ulysses
+
+# CoT Answer
+
+A: Let's think step by step.
+
+(0) At the start: Alice: Ulysses, Bob: Frankenstein, Claire: Lolita.  
+(1) Claire and Bob swap books: Alice: Ulysses, Bob: Lolita, Claire: Frankenstein.  
+(2) Bob and Alice swap books: Alice: Lolita, Bob: Ulysses, Claire: Frankenstein.  
+(3) Claire and Bob swap books: Alice: Lolita, Bob: Frankenstein, Claire: Ulysses.  
+So the answer is Alice: Lolita, Bob: Frankenstein, Claire: Ulysses.
+
+Table 14: Modification applied to Tracking Shuffled Objects (Three objects).  
+
+<table><tr><td colspan="2">CoT</td></tr><tr><td>Boolean Expressions</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/bbh/prompts/cot/booleanExpressions.txt</td></tr><tr><td>Hyperbaton</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/bbh/prompts/cot/hyperbaton.txt</td></tr><tr><td>Multistep Arithmetic Two</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/bbh/prompts/cot/multistep_arithmetic_two.txt</td></tr><tr><td>Navigate</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/bbh/prompts/cot/navate.txt</td></tr><tr><td>Object Counting</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/bbh/prompts/cot/object_counting.txt</td></tr><tr><td>Tracking Shuffled Objects</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/bbh/prompts/cot/tracking_shuffled Objects.txt</td></tr><tr><td>Web of Lies</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/bbh/prompts/cot/web_of Lies.txt</td></tr><tr><td>Word Sorting</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/bbh/prompts/cot/word_sorting.txt</td></tr><tr><td colspan="2">IO</td></tr><tr><td>Boolean Expressions</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/bbh/prompts/standard/booleanExpressions.txt</td></tr><tr><td>Hyperbaton</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/bbh/prompts/standard/hyperbaton.txt</td></tr><tr><td>Multistep Arithmetic Two</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/bbh/prompts/standard/multistep_arithmetic_two.txt</td></tr><tr><td>Navigate</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/bbh/prompts/standard/navate.txt</td></tr><tr><td>Object Counting</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/bbh/prompts/standard/object_counting.txt</td></tr><tr><td>Tracking Shuffled Objects</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/bbh/prompts/standard/tracking_shuffled Objects.txt</td></tr><tr><td>Web of Lies</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/bbh/prompts/standard/web_of Lies.txt</td></tr><tr><td>Word Sorting</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/bbh/prompts/standard/word_sorting.txt</td></tr><tr><td colspan="2">Merge</td></tr><tr><td>Boolean Expressions</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/bbh/prompts/merge/booleanExpressions.txt</td></tr><tr><td>Hyperbaton</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/bbh/prompts/merge/hyperbaton.txt</td></tr><tr><td>Multistep Arithmetic Two</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/bbh/prompts/merge/multistep_arithmetic_two.txt</td></tr><tr><td>Navigate</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/bbh/prompts/merge/navate.txt</td></tr><tr><td>Object Counting</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/bbh/prompts/merge/object_counting.txt</td></tr><tr><td>Tracking Shuffled Objects</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/bbh/prompts/merge/tracking_shuffled Objects.txt</td></tr><tr><td>Web of Lies</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/bbh/prompts/merge/web_of Lies.txt</td></tr><tr><td>Word Sorting</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/bbh/prompts/merge/word_sorting.txt</td></tr></table>
+
+Table 15: Links to solve and merge prompts for the BBH tasks.
+
+<table><tr><td colspan="2">CoT</td></tr><tr><td>Coin</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/coin/prompts/cot/cot8.txt</td></tr><tr><td>Concatenation</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/concatenation/prompts/cot/cot8.txt</td></tr><tr><td colspan="2">IO</td></tr><tr><td>Coin</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/coin/prompts/standard/standard8.txt</td></tr><tr><td>Concatenation</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/concatenation/prompts/standard/standard8.txt</td></tr><tr><td colspan="2">Merge</td></tr><tr><td>Coin</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/coin/prompts/merge/merge.txt</td></tr><tr><td>Concatenation</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/concatenation/prompts/merge/merge.txt</td></tr><tr><td colspan="2">L2M</td></tr><tr><td>Coin</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/coin/prompts/merge/12m.txt</td></tr><tr><td>Concatenation</td><td>https://github.com/ArmelRandy/tree-of-problems/blob/master/top/concatenation/prompts/merge/12m.txt</td></tr></table>
+
+Table 16: Links to solve and merge prompts for Coin Flip and Last Letter Concatenation.

2024/Tree of Problems_ Improving structured problem solving with compositionality/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1bfd889775a93ca2006878ad52c1bee3629eaea29b979438998e17490d9c1122
+size 1759805

2024/Triad_ A Framework Leveraging a Multi-Role LLM-based Agent to Solve Knowledge Base Question Answering/2bfc46f7-f4ed-4d88-9bf5-1ead6a8c1d52_content_list.json

@@ -0,0 +1,2514 @@
+[
+    {
+        "type": "text",
+        "text": "Triad: A Framework Leveraging a Multi-Role LLM-based Agent to Solve Knowledge Base Question Answering",
+        "text_level": 1,
+        "bbox": [
+            117,
+            85,
+            878,
+            124
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Chang Zong $^{1}$ , Yuchen Yan $^{1}$ , Weiming Lu $^{1\\dagger}$ , Jian Shao $^{1}$   \nYongfeng Huang $^{2}$ , Heng Chang $^{3}$ , Yueting Zhuang $^{1\\dagger}$",
+        "bbox": [
+            258,
+            135,
+            739,
+            170
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "<sup>1</sup>College of Computer Science and Technology, Zhejiang University  \n<sup>2</sup>The Chinese University of Hong Kong  \n<sup>3</sup>Tsinghua University  \n{zongchang, luwm, yzhuang}@zju.edu.cn",
+        "bbox": [
+            226,
+            171,
+            776,
+            237
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            260,
+            339,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Recent progress with LLM-based agents has shown promising results across various tasks. However, their use in answering questions from knowledge bases remains largely unexplored. Implementing a KBQA system using traditional methods is challenging due to the shortage of task-specific training data and the complexity of creating task-focused model structures. In this paper, we present Triad, a unified framework that utilizes an LLM-based agent with multiple roles for KBQA tasks. The agent is assigned three roles to tackle different KBQA subtasks: agent as a generalist for mastering various subtasks, as a decision maker for the selection of candidates, and as an advisor for answering questions with knowledge. Our KBQA framework is executed in four phases, involving the collaboration of the agent's multiple roles. We evaluated the performance of our framework using three benchmark datasets, and the results show that our framework outperforms state-of-the-art systems on the LC-QuAD and YAGO-QA benchmarks, yielding F1 scores of  $11.8\\%$  and  $20.7\\%$ , respectively.",
+        "bbox": [
+            141,
+            288,
+            460,
+            631
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            645,
+            258,
+            659
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "A question-answering system is designed to extract information by converting a natural language question into a structured query that can retrieve precise information from an existing knowledge base (Omar et al., 2023a). The resolution of Knowledge Base Question Answering (KBQA) typically involves phases including question understanding, URI linking, and query execution. Traditional KBQA systems require the use of specialized models trained with domain datasets for question parsing and entity linking (Hu et al., 2018; Omar et al., 2023a; Hu et al., 2021). Large language models (LLMs), however, have shown promising competencies in in-context learning using task-specific",
+        "bbox": [
+            112,
+            671,
+            490,
+            896
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "demonstrations (Dong et al., 2022). LLMs have recently been employed as agents in the execution of complex problems. A framework that employs LLM-augmented agents can generate actions or coordinate multiple agents, thus improving the capacity to handle complex situations (Liu et al., 2023). Despite the remarkable performance of LLMs in various tasks as evidenced in previous studies, a comprehensive qualitative and quantitative evaluation of KBQA frameworks empowered with an LLM-based agent remains insufficiently explored.",
+        "bbox": [
+            507,
+            261,
+            884,
+            439
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/99ca4918bdf6bc83a47e697322f730e3148f67aa4f29ce2ff8375a19d1064a9d.jpg",
+        "image_caption": [
+            "Figure 1: A system with multiple roles who focus on sub-problems of each phase to solve a complex task."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            542,
+            453,
+            852,
+            646
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Studies on KBQA with LLMs has attracted considerable attention. Some works focus primarily on highlighting the inability of LLMs to generate complete factoid results (Hu et al., 2023b; Tan et al., 2023c) or demonstrating their potential efficacy in future research (Omar et al., 2023b; Tan et al., 2023b). Other works concentrates on generating answers by prompt learning and incorporating external knowledge bases (Baek et al., 2023; Tan et al., 2023a). Concurrently, LLMs can be deployed to address each phase within Text2SQL challenges(Li et al., 2023, 2024) or theorem proof tasks(Dong et al., 2023). However, each phase of",
+        "bbox": [
+            507,
+            712,
+            884,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "†Corresponding authors.",
+        "bbox": [
+            142,
+            906,
+            294,
+            920
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "1698",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 1698-1710 November 12-16, 2024 ©2024 Association for Computational Linguistics",
+        "bbox": [
+            159,
+            945,
+            836,
+            973
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "KBQA can be further decomposed into subtasks and completed through an agentic approach that provides feedback and cooperation. Additionally, decomposing the task reduces the complexity of cooperative working by allowing each role to concentrate on smaller sub-problems (Wang et al., 2020). As illustrated in Figure 1, three roles in an organization work together to provide the final answer for the overall task. The above observations spur our exploration into the following question: How does an LLM-based agent solve KBQA tasks by serving as multiple roles, and its performance is comparable to systems trained specifically?",
+        "bbox": [
+            112,
+            84,
+            490,
+            294
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this study, we introduce Triad, a unified framework that leverages an LLM-based agent with three roles to address KBQA tasks. Specifically, we implement the agent consisting of an LLM as the core, supplemented by various task-specific modules such as memory and executing functions. The agent is assigned three distinct roles: a generalist (G-Agent) adept at mastering numerous small tasks by the given examples, a decision maker (D-Agent) proficient at identifying options and selecting candidates, and an advisor (A-Agent) skilled at providing answers using internal and external knowledge. The cooperation of these agent roles composes a KBQA process containing four phases: question parsing, URI linking, query construction, and answer generation. We evaluate our framework on three benchmark datasets in various difficulties. The results show that our framework outperforms state-of-the-art systems, demonstrated by  $11.8\\%$  and  $20.7\\%$  F1 scores on the LC-QuAD and YAGO-QA benchmarks, respectively<sup>1</sup>.",
+        "bbox": [
+            115,
+            294,
+            490,
+            632
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The contributions of this study can be summarized as follows:",
+        "bbox": [
+            112,
+            633,
+            489,
+            664
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We propose Triad, the first framework that leverages an LLM-based agent to solve KBQA tasks in all its four phases, without specialized training models.",
+            "- We implement an LLM-based agent with various task-specific modules that can act as three roles, including a generalist, a decision maker, and an advisor, to collaboratively solve KBQA via focusing on subtasks.",
+            "- We evaluate the performance of Triad. The results show a competitive ability compared"
+        ],
+        "bbox": [
+            136,
+            678,
+            489,
+            883
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "to both state-of-the-art KBQA systems and pure LLM methods.",
+        "bbox": [
+            544,
+            84,
+            880,
+            116
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Preliminaries",
+        "text_level": 1,
+        "bbox": [
+            507,
+            128,
+            662,
+            143
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 Phases of KBQA",
+        "text_level": 1,
+        "bbox": [
+            507,
+            155,
+            687,
+            171
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "A typical KBQA system has a process that encompasses four phases:",
+        "bbox": [
+            507,
+            177,
+            884,
+            209
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Question parsing involves converting natural language questions into a structured format that incorporates references to entities and relations.",
+        "bbox": [
+            507,
+            218,
+            882,
+            266
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "URI linking entails associating and replacing these entity and relation mentions with their corresponding URIs within a knowledge base.",
+        "bbox": [
+            507,
+            275,
+            884,
+            325
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Query construction involves creating executable queries in a standard format to extract answers from knowledge bases.",
+        "bbox": [
+            507,
+            335,
+            884,
+            382
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Answer generation seeks to obtain the ultimate answers either by performing queries within knowledge bases or by directly querying an agent.",
+        "bbox": [
+            507,
+            392,
+            882,
+            441
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2 Roles of LLM-based Agent",
+        "text_level": 1,
+        "bbox": [
+            507,
+            451,
+            768,
+            467
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Drawing an analogy to a software development scenario, where coders complete small development tasks, with the process and plan being decided by the manager, and ultimately the outcome inspected by the leader, we assign the following three roles to an LLM-based agent to solve the KBQA task:",
+        "bbox": [
+            507,
+            473,
+            882,
+            569
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Agent as a generalist (G-Agent) is capable of mastering various small tasks by providing a few examples.",
+        "bbox": [
+            507,
+            579,
+            882,
+            627
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Agent as a decision-maker (D-Agent) adepts at analyzing options and providing candidate results as procedural feedback.",
+        "bbox": [
+            507,
+            637,
+            880,
+            684
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Agent as an advisor (A-Agent) is skilled in providing final answers with the aid of both external and its own knowledge.",
+        "bbox": [
+            507,
+            696,
+            882,
+            744
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.3 Task Formulation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            755,
+            695,
+            769
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "A KBQA task refers to the process of solving a set of subtasks  $S$ . Each subtask  $S_{t} \\in S$  contributes to one phase of the whole process. An LLM-based agent  $Agent_{r}$  with a role  $r$  can be used to resolve a type of subtasks by its task-specific components, including a language model  $LLM$ , a memory  $Mem_{t}$ , a function  $F_{t}$ , a prompt  $Pmt_{t}$  and a set of parameters  $\\theta_{t}$ , using the set of role-related hyperparameters  $\\sigma_{r}$ . The task can be formulated as follows:",
+        "bbox": [
+            507,
+            776,
+            884,
+            920
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{1}$ Code and data are available at https://github.com/ZJU-DCDLab/Triad.",
+        "bbox": [
+            112,
+            894,
+            487,
+            919
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "1699",
+        "bbox": [
+            482,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nf (K B Q A) = \\bigoplus_ {t = 1} ^ {T} f (S _ {t})\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            179,
+            110,
+            368,
+            152
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nf \\left(S _ {t}\\right) = A g e n t _ {r} \\left(L L M, M e m _ {t}, F _ {t}, P m t _ {t}, \\theta_ {t}, \\sigma_ {r}\\right) \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            115,
+            156,
+            487,
+            186
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": ", where  $T$  is the total number of subtasks,  $\\bigoplus$  is the way to coordinate subtasks to solve the whole.",
+        "bbox": [
+            112,
+            187,
+            485,
+            219
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Triad Framework",
+        "text_level": 1,
+        "bbox": [
+            112,
+            231,
+            304,
+            247
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The overall architecture of Triad is shown in Figure 2. Each role of the LLM-based agent, along with its associated subtasks, is illustrated as follows.",
+        "bbox": [
+            112,
+            258,
+            487,
+            306
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 G-Agent as a Generalized Solver",
+        "text_level": 1,
+        "bbox": [
+            112,
+            319,
+            416,
+            334
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "A generalized agent (G-Agent) proficiently manages numerous tasks by leveraging learning from limited examples through an LLM. In our framework, a G-Agent can perform question parsing, query template generation, or answer type classification as actions solely utilizing an LLM. These three subtasks are illustrated as follows:",
+        "bbox": [
+            112,
+            340,
+            487,
+            451
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Triplet mention extraction: The process of extracting triplet mentions in question parsing involves the conversion of a naturally phrased question, denoted as  $Q$ , into formatted triplets of entities and relations. This subtask is executed employing an LLM, which is guided by a prompt with a set of prerequisites and a selection of examples. This subtask can be represented as follows:",
+        "bbox": [
+            112,
+            463,
+            489,
+            590
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nf \\left(S _ {t r i}\\right) = A g e n t _ {g} \\left(L L M, P m t _ {t r i}, Q, \\mathcal {N}\\right) \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            136,
+            604,
+            487,
+            627
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nP m t _ {t r i} = \\left[ I n s _ {t r i}, S h o t _ {t r i}, C o T _ {t r i} \\right]\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            137,
+            626,
+            400,
+            640
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": ", where  $A g e n t_{g}$  is the agent as a generalist to perform the triplet extraction subtask with  $\\mathcal{N}$  examples.  $P m t_{tri}$  is the prompt to guide LLM to generate triplets from the question  $Q$ , which consists of instruction  $I n s_{tri}$ , examples  $S h o t_{tri}$ , and chain-of-thought prompt  $C o T_{tri}$  (Kojima et al., 2022).",
+        "bbox": [
+            112,
+            653,
+            489,
+            750
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "SPARQL template generation: The generation of SPARQL templates in query construction involves the use of an LLM to create a SPARQL template that articulates the question using standard SPARQL syntax, replacing URI identifiers with entity and relation variables. To derive precise and comprehensive answers from the knowledge base using SPARQL queries, there are two potential strategies. One approach involves the direct generation of an executable SPARQL using",
+        "bbox": [
+            112,
+            760,
+            489,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "an LLM, though this method may significantly increase LLM call times and error rates when numerous candidate queries are in play. Alternatively, a SPARQL template can initially be generated with entity and relation variables, which are subsequently replaced with linked URIs. For the sake of stability and efficiency, we opt for the second strategy. This subtask can be denoted as:",
+        "bbox": [
+            507,
+            84,
+            884,
+            212
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nf (S _ {q t}) = A g e n t _ {g} (L L M, P m t _ {q t}, \\theta_ {q t}, \\mathcal {N}),\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            529,
+            223,
+            838,
+            241
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nP m t _ {q t} = \\left[ I n s _ {q t}, S h o t _ {q t}, C o T _ {q t} \\right], \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            531,
+            243,
+            880,
+            260
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\theta_ {q t} = [ Q, f (S _ {t r i}) ]\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            559,
+            263,
+            695,
+            280
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": ", where  $A g e n t_{g}$  is the agent as a generalist to perform SPARQL template generation with  $\\mathcal{N}$  examples,  $f(S_{tri})$  is the triplets derived from the previous subtask,  $P m t_{q t}$  is the prompt for LLM to generate SPARQL template.",
+        "bbox": [
+            507,
+            288,
+            884,
+            369
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Answer type classification: In the phase of answer generation, the answer type classification subtask refers to the process of assigning a specific category to a response according to the question. This process serves as a guiding mechanism for the framework to generate comprehensive and accurate answers. This classification subtask is denoted as:",
+        "bbox": [
+            507,
+            378,
+            882,
+            489
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nf \\left(S _ {c l s}\\right) = A g e n t _ {g} \\left(L L M, P m t _ {c l s}, Q, \\mathcal {N}\\right), \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            529,
+            497,
+            882,
+            523
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nP m t _ {c l s} = \\left[ I n s _ {c l s}, S h o t _ {c l s}, C o T _ {c l s} \\right]\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            531,
+            518,
+            794,
+            533
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": ", where  $A g e n t_{g}$  is the agent as a generalist to perform type classification subtask with  $\\mathcal{N}$  examples,  $P m t_{c l s}$  is the prompt for LLM.",
+        "bbox": [
+            507,
+            544,
+            882,
+            592
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 D-Agent as a Decision-Maker",
+        "text_level": 1,
+        "bbox": [
+            507,
+            602,
+            786,
+            617
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "An agent as a decision maker (D-Agent) is capable of making candidate selections step by step through filtering and choosing from given options, harnessing the capabilities of an LLM and KB as memory. The D-Agent can effectively handle three subtasks, which are delineated as follows:",
+        "bbox": [
+            507,
+            623,
+            882,
+            718
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Candidate entity selection: The selection of candidate entities in URI linking is pivotal to the ultimate efficacy of KBQA. Prior research has focused primarily on developing a semantic similarity model to address this linking challenge. However, the linking task requires numerous iterations of searching within the knowledge base, which poses a compatibility issue for LLM-oriented methods. In our framework, an agent as a decision maker is utilized initially to filter all potential entity URIs from the knowledge base, subsequently deploying an LLM to select candidate URIs from a pool of",
+        "bbox": [
+            507,
+            728,
+            882,
+            920
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "1700",
+        "bbox": [
+            482,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/65f8a985484e660617f9c6e72ffacae12e9ff46eef3caef196387195c75dbcf4.jpg",
+        "image_caption": [
+            "Figure 2: Our Triad framework leverages an LLM-based agent with three different roles including a generalist, a decision-maker, and an advisor to cooperatively handle a series of subtasks in the four phases of a KBQA process."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            115,
+            83,
+            394,
+            312
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/19cd017a6f14bd1a81b57338bff9017ac9b677a59f7f97203cdb46895b63f89a.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            400,
+            83,
+            895,
+            312
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "potential identifiers. For each entity, our aim is to find the  $\\kappa$  most possible entity URIs which can be used to traverse over the KB to get the final answer. The entity selection subtask can be denoted as:",
+        "bbox": [
+            112,
+            384,
+            487,
+            447
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nf (S _ {e s}) = A g e n t _ {d} (L L M, M e m _ {e s}, F _ {e s},\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            132,
+            480,
+            423,
+            497
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\left. P m t _ {e s}, \\theta_ {e s}, \\mathcal {K}\\right), \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            205,
+            500,
+            485,
+            516
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nM e m _ {e s} = [ K B, L i s t _ {e s} ], \\theta_ {e s} = [ Q, f (S _ {t r i}) ]\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            127,
+            519,
+            455,
+            537
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": ", where  $Agent_d$  is the agent as a decision maker to perform the entity selection subtask with question  $Q$ , extracted triplets  $f(S_{tri})$  and memory  $Mem_{es}$ ,  $Mem_{es}$  is composed of a knowledge base  $KB$  and a list of entity URIs  $List_{es}$  filtered from  $KB$  using a text similarity matching function  $F_{es}$ ,  $Pmt_{es}$  is the prompt for LLM to perform the subtask,  $\\mathcal{K}$  is the hyperparameter of  $Agent_d$ , indicating the number of candidates selected by LLM.",
+        "bbox": [
+            112,
+            552,
+            487,
+            697
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Candidate relation selection: The task of candidate relation selection in URI linking presents considerable challenges due to the discrepancies between word forms and meanings. Nevertheless, the existence of reasoning paths in the knowledge base can be utilized to allow for a significant reduction of the search space in relation linking. In our framework, an agent as a decision maker endeavors to sieve through all potential relation URIs by navigating the knowledge base with candidate entity URIs generated from the previous subtask. Subsequently, an LLM is used to select the top  $\\kappa$  most probable relation URIs for output. The relation",
+        "bbox": [
+            112,
+            712,
+            489,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "selection subtask can be denoted as:",
+        "bbox": [
+            507,
+            384,
+            779,
+            399
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nf \\left(S _ {r s}\\right) = A g e n t _ {d} (L L M, M e m _ {r s}, F _ {r s},\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            529,
+            409,
+            818,
+            425
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\left. P m t _ {r s}, \\theta_ {r s}, \\mathcal {K}\\right), \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            601,
+            428,
+            880,
+            445
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nM e m _ {r s} = \\left[ K B, L i s t _ {r s} \\right], \\theta_ {r s} = \\left[ Q, f \\left(S _ {e s}\\right) \\right]\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            522,
+            448,
+            848,
+            464
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": ", where memory  $Mem_{rs}$  is composed of the knowledge base  $KB$  and a list of possible relation URIs  $List_{rs}$  filtered from  $KB$  using a one-order traversing function  $F_{rs}$ .  $Pmt_{rs}$  is the prompt for LLM to perform relation selection.  $\\kappa$  is the number of relation URIs selected by LLM.",
+        "bbox": [
+            507,
+            476,
+            884,
+            571
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Candidate SPARQL selection: The subtask of candidate SPARQL selection in query construction involves determining the appropriate SPARQL queries to obtain the final answers. Given a SPARQL template generated by the G-Agent, along with multiple candidate URIs selected from the D-Agent in previous subtasks, our D-Agent is targeted to identify the most plausible query. To further reduce the difficulty of selection, an executor function is applied to eliminate queries that cannot retrieve any results from the knowledge base. In conclusion, our aim in this subtask is to use D-Agent to construct executable SPARQLs and find the most possible one given a query candidate list with supported information. The SPARQL selection subtask can be denoted as:",
+        "bbox": [
+            507,
+            580,
+            884,
+            837
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} f (S _ {q s}) = A g e n t _ {d} (L L M, M e m _ {q s}, F _ {q s}, \\\\ \\left. \\begin{array}{l} P m t _ {q s}, \\theta_ {q s}, \\mathcal {K}), \\\\ [ ] \\end{array} \\right. \\tag {7} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            544,
+            847,
+            880,
+            891
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nM e m _ {q s} = \\left[ K B, L i s t _ {q s} \\right],\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            539,
+            887,
+            729,
+            904
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\theta_ {q s} = [ Q, f (S _ {e s}), f (S _ {r s}), f (S _ {q t}) ]\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            574,
+            906,
+            821,
+            923
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "1701",
+        "bbox": [
+            482,
+            927,
+            517,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": ", where memory  $Mem_{qs}$  is composed of a knowledge base  $KB$  and a list of possible SPARQLs  $List_{qs}$  constructed with SPARQL template  $f(S_{qt})$ , entity URIs  $f(S_{es})$ , and relation URIs  $f(S_{rs})$  by the function  $F_{qs}$ ,  $Pmt_{qs}$  is the prompt for LLM to perform query selection,  $\\mathcal{K} = 1$  is the number of queries selected by LLM.",
+        "bbox": [
+            112,
+            84,
+            489,
+            197
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3 A-Agent as a Comprehensive Advisor",
+        "text_level": 1,
+        "bbox": [
+            112,
+            218,
+            455,
+            233
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "An advisory agent (A-Agent) is capable of processing a question and a corresponding type of answer as input. Its response is generated by either extracting information from an external knowledge base or by utilizing its internal knowledge to provide a direct answer. This comprehensive answering subtask can be described as follows:",
+        "bbox": [
+            112,
+            244,
+            489,
+            356
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Comprehensive answering: The objective of comprehensive answering in the answer generation phase is to derive a definitive response based on an incoming question. Previous work (Omar et al., 2023b) has demonstrated that LLMs are more proficient in delivering single-fact responses and making Boolean judgments. Given this understanding, we implement an advisory agent that incorporates a simple policy to facilitate a comprehensive answering approach. Specifically, if a question yields a final SPARQL generated from the preceding steps, A-Agent extracts elements from the knowledge base to give the answer. Conversely, if the agent does not receive a feasible SPARQL, A-Agent provides a direct response with LLM's internal knowledge, following the prompt based on the type of the answer. Additionally, A-Agent will send a retry signal to previous phases if no result is generated. The subtask can be formulated as below:",
+        "bbox": [
+            112,
+            376,
+            489,
+            697
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nf (S _ {c a}) = A g e n t _ {a} (L L M, M e m _ {c a}, F _ {c a},\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            129,
+            715,
+            425,
+            733
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\left. P m t _ {c a}, \\theta_ {c a}, \\mathcal {T}\\right), \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            203,
+            736,
+            485,
+            753
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nM e m _ {c a} = [ K B ], \\theta_ {c a} = [ Q, f (S _ {q s}), f (S _ {c l s}) ]\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            124,
+            755,
+            457,
+            774
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": ", where  $A g e n t_{a}$  is the agent as an advisor to perform a comprehensive answering for the question  $Q$  with a memory  $M e m_{c a}$  of knowledge base,  $P m t_{c a}$  is the prompt for LLM to perform a direct response according to the type of the answer,  $f(S_{q s})$  is the final query and  $f(S_{c l s})$  is the answer type,  $\\mathcal{T}$  is the maximum times to retry for previous phases if no result is returned from  $K B$ .",
+        "bbox": [
+            112,
+            791,
+            489,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4 Performance Evaluation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            83,
+            754,
+            98
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1 Experimental Settings",
+        "text_level": 1,
+        "bbox": [
+            507,
+            109,
+            732,
+            124
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Indexed Knowledge Bases: The efficacy of our framework is assessed through the collection of two real knowledge bases, specifically DBpedia and YAGO. DBpedia (Auer et al., 2007) serves as an accessible knowledge base extracted from Wikipedia, while YAGO (Pellissier Tanon et al., 2020) is a large knowledge base that includes individuals, cities, nations, and organizations. We index the triples and the mentions of entities and relations in a Virtuoso endpoint and an Elasticsearch server, respectively.",
+        "bbox": [
+            507,
+            130,
+            884,
+            307
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "KBQA Benchmark Datasets: We evaluate our framework on datasets including YAGO-QA, LCQuAD 1.0, and QALD-9, which have various difficulties in interpreting the questions. These datasets contain questions in English, paired with their respective SPARQL queries, and accurate responses derived from a specific knowledge base. QALD-9 (Usbeck et al., 2018) and LC-QuAD 1.0 (Trivedi et al., 2017) are frequently used to evaluate QA systems with DBpedia. The recently published YAGO-QA in (Omar et al., 2023a), features questions accompanied by annotated SPARQL queries sourced from YAGO. The statistics for three benchmarks, along with their associated knowledge bases, are depicted in Table 1.",
+        "bbox": [
+            507,
+            316,
+            885,
+            557
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Baseline Methods: We evaluate Triad against traditional KBQA systems such as KGQAN (Omar et al., 2023a), EDGQA (Hu et al., 2021) and gAnswer (Hu et al., 2018). This comparison shows how our LLM-based agent framework can rival full-shot systems with just a few examples. Additionally, we contrast our framework with pure GPT models like GPT-3.5 Turbo and GPT-4 to exhibit Triad's architectural performance. We treat these foundation models as few-shot methods to answer the questions referring to some examples.",
+        "bbox": [
+            507,
+            565,
+            885,
+            744
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Implementation Details: Triad is implemented with Python 3.9. We incorporate LLM capabilities to our multi-role agent via OpenAI's API services. The names of entities and relations from knowledge bases are indexed in an ElasticSearch 7.5.2 server for text matching. All triples are imported into an SPARQL endpoint of Virtuoso 07.20.3237 for retrieval. Triad requires four hyperparameters: the number of examples G-Agent uses for subtask",
+        "bbox": [
+            507,
+            753,
+            885,
+            897
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{2}$ https://platform.openai.com/docs/models",
+        "bbox": [
+            529,
+            904,
+            791,
+            920
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "1702",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/f067ba9e32c5b22450e5b7df9e37764fac19483a3acaad5a39656811da784653.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Benchmarks</td><td colspan=\"5\">Benchmark Statistics</td></tr><tr><td>#Questions</td><td>KB</td><td>#Triples</td><td>Virtuoso Size</td><td>ES size</td></tr><tr><td>LC-QuAD 1.0</td><td>1000</td><td>DBpedia-04</td><td>397M</td><td>35.40G</td><td>1.56G</td></tr><tr><td>QALD-9</td><td>150</td><td>DBpedia-10</td><td>374M</td><td>36.89G</td><td>1.57G</td></tr><tr><td>YAGO-QA</td><td>100</td><td>YAGO-4</td><td>207M</td><td>24.85G</td><td>0.54G</td></tr></table>",
+        "bbox": [
+            181,
+            83,
+            816,
+            172
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/92754dd5a326d2ead11fd35555850845ce66281773f19eace5e83b1357fce48b.jpg",
+        "table_caption": [
+            "Table 1: The statistics of KBQA benchmarks, including the number of questions number, the number of triples, the size of index in Virtuoso and Elasticsearch."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Type</td><td rowspan=\"2\">Frameworks</td><td colspan=\"3\">LC-QuAD 1.0</td><td colspan=\"3\">QALD-9</td><td colspan=\"3\">YAGO-QA</td></tr><tr><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td></tr><tr><td rowspan=\"3\">full-shot</td><td>gAnswer</td><td>-</td><td>-</td><td>-</td><td>0.293</td><td>0.327</td><td>0.298</td><td>0.585</td><td>0.341</td><td>0.430</td></tr><tr><td>EDGQA</td><td>0.505</td><td>0.560</td><td>0.531</td><td>0.313</td><td>0.403</td><td>0.320</td><td>0.419</td><td>0.408</td><td>0.414</td></tr><tr><td>KGQAN</td><td>0.587</td><td>0.461</td><td>0.516</td><td>0.511</td><td>0.387</td><td>0.441</td><td>0.485</td><td>0.652</td><td>0.556</td></tr><tr><td rowspan=\"4\">few-shot</td><td>GPT-3.5</td><td>0.269</td><td>0.251</td><td>0.266</td><td>0.240</td><td>0.217</td><td>0.228</td><td>0.171</td><td>0.142</td><td>0.155</td></tr><tr><td>GPT-4</td><td>0.336</td><td>0.344</td><td>0.340</td><td>0.250</td><td>0.249</td><td>0.249</td><td>0.193</td><td>0.190</td><td>0.191</td></tr><tr><td>Triad-GPT3.5</td><td>0.490</td><td>0.519</td><td>0.504</td><td>0.293</td><td>0.302</td><td>0.297</td><td>0.660</td><td>0.639</td><td>0.649</td></tr><tr><td>Triad-GPT4</td><td>0.561</td><td>0.568</td><td>0.564</td><td>0.408</td><td>0.425</td><td>0.416</td><td>0.690</td><td>0.664</td><td>0.677</td></tr></table>",
+        "bbox": [
+            117,
+            225,
+            880,
+            379
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 2: The performance of our proposed Triad on three benchmarks, comparing with traditional KBQA systems (full-shot) and pure LLM (few-shot) baselines. The optimal and suboptimal scores are highlighted with bold and underlined text, respectively.",
+        "bbox": [
+            112,
+            388,
+            882,
+            432
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "learning, the number of candidates D-Agent selects for entity and relation linking, and the retry times for handling non-response SPARQLs. The optimal values for these parameters are 3, 2, 2, and 3, respectively. The framework and its variants are tested five times on each benchmark, with the average scores reported as the final results. For traditional systems, we report the results recorded in their papers. For pure LLM baselines, we write prompts to hire an LLM to answer questions directly referring to examples, and then link the mentions from the responses to the URIs in our indexed knowledge bases via built-in similarity search.",
+        "bbox": [
+            112,
+            457,
+            489,
+            667
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2 Performance Comparison",
+        "text_level": 1,
+        "bbox": [
+            112,
+            680,
+            364,
+            696
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The performance of Triad compared to traditional KBQA systems and pure LLM generation methods is shown in Table 2. Evaluation metrics precision(P), recall(R), and F1-score(F1) are reported. We can observe from the experimental results that:",
+        "bbox": [
+            112,
+            703,
+            489,
+            784
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Few-shot can be competitive with full-shot. Our multi-role LLM-based agent framework, though executing a few-shot prompt learning, exhibits competitive performance with cutting-edge full-shot KBQA systems.",
+        "bbox": [
+            112,
+            796,
+            489,
+            878
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Underlying capability matters. The use of GPT-4 as the core in an LLM-based agent significantly",
+        "bbox": [
+            112,
+            889,
+            489,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "outperforms GPT-3.5 on all benchmarks, demonstrating the importance of the underlying capabilities of an agent.",
+        "bbox": [
+            507,
+            457,
+            884,
+            505
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Explicit knowledge is necessary. Pure LLM models with GPT-3.5 and GPT-4 display deficiencies in generating accurate responses without an auxiliary knowledge base as a memory for intermediary steps such as URI linking.",
+        "bbox": [
+            507,
+            513,
+            884,
+            594
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Performance varies with complexity. Triad demonstrates superior results on the LC-QuAD and YAGO-QA benchmarks compared to QALD-9, due to an increasing failure in response to complex questions, which will be discussed later.",
+        "bbox": [
+            507,
+            601,
+            882,
+            681
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3 Study on Capabilities of Agent Roles",
+        "text_level": 1,
+        "bbox": [
+            507,
+            690,
+            843,
+            707
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We assess the efficacy of G-Agent with various other language models as the core. The framework without G-task uses the text-davinci-002, which is not as powerful as GPT-3.5 and GPT-4 in solving many tasks, and the one without G-chat uses text-davinci-003 to eliminate the chat and alignment abilities. We test the ability of D-Agent without D-uri and D-query by replacing the URI selection and query selection with URI matching and query generation, respectively. We evaluate the contribution of A-Agent eliminating A-llm and A-fact by responding to questions without using LLM's assistance or use an LLM to answer Boolean questions",
+        "bbox": [
+            507,
+            712,
+            884,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "1703",
+        "bbox": [
+            482,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "for auxiliary rather than single-fact questions. The F1 results of the role ablation experiments on two representative datasets are shown in Table 3. The results indicate that every component pertaining to each role contributes to the overall performance. More specifically, a G-Agent that employs a less powerful LLM as its core can drastically undermine performance. D-Agent assumes a more pivotal role during the linking phase compared to the query construction phase. A-Agent, on the other hand, proves to be an efficient solution for managing situations where SPARQL results are absent.",
+        "bbox": [
+            112,
+            84,
+            492,
+            277
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/346b10410304a8645f02c845b5225d26628bf1517956ef6cf0536a24af650bb4.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>G-task</td><td>G-chat</td><td>LC-QuAD 1.0</td><td>QALD-9</td></tr><tr><td>X</td><td>X</td><td>0.343</td><td>0.159</td></tr><tr><td>✓</td><td>X</td><td>0.443</td><td>0.248</td></tr><tr><td>✓</td><td>✓</td><td>0.564</td><td>0.416</td></tr><tr><td>D-uri</td><td>D-query</td><td>LC-QuAD 1.0</td><td>QALD-9</td></tr><tr><td>X</td><td>✓</td><td>0.274</td><td>0.210</td></tr><tr><td>✓</td><td>X</td><td>0.431</td><td>0.301</td></tr><tr><td>✓</td><td>✓</td><td>0.564</td><td>0.416</td></tr><tr><td>A-llm</td><td>A-fact</td><td>LC-QuAD 1.0</td><td>QALD-9</td></tr><tr><td>X</td><td>X</td><td>0.459</td><td>0.382</td></tr><tr><td>✓</td><td>X</td><td>0.473</td><td>0.385</td></tr><tr><td>✓</td><td>✓</td><td>0.564</td><td>0.416</td></tr></table>",
+        "bbox": [
+            115,
+            291,
+            500,
+            514
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.4 Analysis of Role Hyperparameters",
+        "text_level": 1,
+        "bbox": [
+            112,
+            586,
+            431,
+            602
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We concentrate on three hyperparameters of roles, including the number of examples  $(\\mathcal{N}\\in$ $\\{1,2,3\\})$  provided for G-Agent to learn subtasks, the number of URI candidates  $(\\mathcal{K}\\in$ $\\{(1,1),(1,2),(2,2),(2,3)\\})$  selected by D-Agent for query construction, and the number of retry times  $(\\mathcal{T}\\in \\{1,2,3\\})$  launched by A-Agent when there is no response. Table 4 presents the F1 results of Triad's performance, employing three hyperparameters on two benchmarks. We discover that:",
+        "bbox": [
+            112,
+            608,
+            489,
+            769
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Quality is more important than quantity. More examples provided to G-Agent do not always improve the performance. The efficacy of G-Agent is significantly influenced by the quality of examples.",
+        "bbox": [
+            112,
+            781,
+            489,
+            845
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "More options may harm the result. Choosing more candidate URIs for entities and relations could potentially disrupt subsequent query phases, thus affecting overall performance.",
+        "bbox": [
+            112,
+            857,
+            489,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/472bc0a6457be2ac74c5ba9f2054b1ee8eadd96e56c36024c036244c93a9f362.jpg",
+        "table_caption": [
+            "Table 3: Study on the roles of LLM-based agent by eliminating an element or downgrading the capability."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Triad Variants</td><td>LC-QuAD 1.0</td><td>QALD-9</td></tr><tr><td>Triad-1-Shot</td><td>0.556</td><td>0.376</td></tr><tr><td>Triad-2-Shot</td><td>0.511</td><td>0.402</td></tr><tr><td>Triad-3-Shot</td><td>0.564</td><td>0.416</td></tr><tr><td>Triad-Top1-1</td><td>0.528</td><td>0.281</td></tr><tr><td>Triad-Top1-2</td><td>0.562</td><td>0.375</td></tr><tr><td>Triad-Top2-2</td><td>0.564</td><td>0.416</td></tr><tr><td>Triad-Top2-3</td><td>0.558</td><td>0.384</td></tr><tr><td>Triad-1-Try</td><td>0.529</td><td>0.375</td></tr><tr><td>Triad-2-Tries</td><td>0.561</td><td>0.407</td></tr><tr><td>Triad-3-Tries</td><td>0.564</td><td>0.416</td></tr></table>",
+        "bbox": [
+            515,
+            83,
+            878,
+            269
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Performance evaluation on three hyperparameters that related to each role of an LLM-based agent.",
+        "bbox": [
+            507,
+            279,
+            884,
+            307
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "More chances benefits the framework. Persistently attempting to construct and execute SPARQL queries is an effective strategy that improves the probability of obtaining accurate answers. Considering the efficiency of overall execution, we set the maximum retry times as 3 in practice.",
+        "bbox": [
+            507,
+            332,
+            884,
+            430
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.5 Analysis of Linking Recall",
+        "text_level": 1,
+        "bbox": [
+            507,
+            441,
+            764,
+            456
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The process of linking is a relatively complex subtask in both the Text2SQL and the KBQA process (Li et al., 2024). Calculating the recall ratio of accurate URIs using D-Agent provides clarity on which step most adversely impacts performance. In the entity linking phase, considering all URIs of entities in the testing set as the ground truth of the linking results,  $80.75\\%$  of the correct URIs are contained from the output of the entity matching filter in D-Agent and  $70.50\\%$  of the correct URIs are retained from the entity selection performed by the LLM in D-Agent. Whereas, in the relation linking phase, only  $52.54\\%$  of the correct relation URIs survive from the selection of LLM, which indicates a greater difficulty in relation linking.",
+        "bbox": [
+            507,
+            462,
+            882,
+            703
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.6 Study on Complex Cases",
+        "text_level": 1,
+        "bbox": [
+            507,
+            714,
+            752,
+            730
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Despite the impressive performance of Triad in certain benchmarks, notable deficiencies remain in its ability to understand questions and generate queries for complex questions. A critical analysis of unsuccessful cases in QALD-9, which has the lowest F1 score, has revealed three primary reasons for this failure, as detailed below:",
+        "bbox": [
+            507,
+            734,
+            880,
+            847
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Complex Syntax signifies that advanced SPARQL queries incorporate keywords such as GROUP BY and HAVING. These terms augment the error propensity in the generation of SPARQL",
+        "bbox": [
+            507,
+            857,
+            882,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "1704",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/074382887f1027b7651a3c35651684bf409c981a5e197bb4363268df4d63760d.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Fail Reason</td><td>Ratio</td><td>Example</td></tr><tr><td>Complex Syntax</td><td>20%</td><td>Q42: Which countries have places with more than two caves?</td></tr><tr><td>Unexploited Semantics</td><td>17%</td><td>Q199: Give me all Argentine films.</td></tr><tr><td>Implicit Reasoning</td><td>5%</td><td>Q133: What are the names of the Teenage Mutant Ninja Turtles?</td></tr></table>",
+        "bbox": [
+            132,
+            83,
+            472,
+            269
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 5: The major reasons of complexity that result in failures, with their corresponding ratios of occurrence in failed cases.",
+        "bbox": [
+            112,
+            278,
+            489,
+            321
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "templates such as the example: Which frequent flyer program has the most airlines?",
+        "bbox": [
+            110,
+            353,
+            487,
+            387
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Unexploited Semantics indicates that semantics of an implicit entity should be comprehended in order to exclude irrelevant URIs. In the example Give me all Argentine films, the meaning of films should be used to narrow down the scope of potential entities in order to eliminate unrelated answers.",
+        "bbox": [
+            112,
+            404,
+            489,
+            501
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Implicit Reasoning presents a challenge that requires a deeper level of traversal by the framework to deduce accurate results from the posed question. For example, another failure question, How many grand-children did Jacques Cousteau have?, the term grand-children must be interpreted to son of son to ensure an accurate response.",
+        "bbox": [
+            112,
+            520,
+            490,
+            633
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.7 Cost Comparison and Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            653,
+            403,
+            669
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "According to our evaluation on the three datasets, the average cost of running a single case is 0.007 USD on average using Triad-GPT3.5 and 0.05 USD on average using Triad-GPT4. Specifically, most API calls occur in the phases of URL linking and comprehensive answering. Meanwhile, traditional KBQA baselines require a lot of training data and local training resources to achieve the SOTA performance, whereas Triad follows a zero- or few-shot manner to save computational cost locally. Furthermore, as shown in Section 4.4, in practice, adjusting the hyperparameters can make the cost as low as possible while preserving overall performance. As the cost of LLM services decreases, the value of Triad will increase accordingly.",
+        "bbox": [
+            112,
+            680,
+            489,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5 Related Work",
+        "text_level": 1,
+        "bbox": [
+            507,
+            83,
+            665,
+            98
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.1 SPARQL-based and LLM-based KBQA",
+        "text_level": 1,
+        "bbox": [
+            507,
+            109,
+            870,
+            124
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Traditional KBQA methods transform natural language queries into SPARQL requests for data extraction. Specific models are employed either for question understanding or URI linking, utilizing domain-based training datasets. Hu et al. (2018) introduces a semantic query graph to structurally represent the natural language query, thereby simplifying the task into a subgraph matching problem. Hu et al. (2021) proposes an entity description graph to represent natural language queries for question parsing and element linking. Omar et al. (2023a) restructures the question parsing task as a text generation issue using a sequence-to-sequence model. With the advent of LLMs, certain phases of KBQA can be enhanced with LLM-integrated methods. Baek et al. (2023) aims to augment LLM-based QA tasks with pertinent facts extracted from knowledge bases, offering a fully zero-shot architecture. Tan et al. (2023a) leverages the general applicability of LLMs to filter linking candidates by making selections via few-shot in-context learning. Omar et al. (2023b) provides a thorough comparison between LLMs and QA systems, recommending further studies to improve KBQA with LLM capabilities. However, apart from the above studies, our study proposes a complete framework incorporating both an LLM and few-shot learning across all KBQA phases from a systematic perspective.",
+        "bbox": [
+            505,
+            131,
+            885,
+            583
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.2 LLM-based Agents for Complex Tasks",
+        "text_level": 1,
+        "bbox": [
+            507,
+            594,
+            858,
+            609
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "LLMs have recently gained significant attention due to their ability to approximate human-level intelligence. This has led to numerous studies focusing on LLM-based agents. A recent survey (Wang et al., 2023) proposes a unified architecture for LLM-based agents, which consists of four modules that include profile, memory, plan, and action. CHATDB(Hu et al., 2023a) employs an LLM controller to generate SQL instructions, which allows for symbolic memory and complex multi-hop reasoning. ART(Paranjape et al., 2023) uses a frozen LLM to generate reasoning steps and further integrates tools for new tasks with minimal human intervention. Toolformer(Schick et al., 2024) takes a different approach by training an LLM to plan and execute tools for the next token prediction by learning API calls generation. ReAct(Yao et al., 2023) focuses on overcoming LLM hallucination by interacting with external knowledge bases,",
+        "bbox": [
+            505,
+            615,
+            885,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "1705",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "thus generating interpretable task-solving strategies. CodeAgent(Tang et al., 2024) designs a multiagent collaboration system across four phases in a code review process. Divergent from the aforementioned studies, our framework concentrates on the solving KBQA tasks by introducing a multirole LLM-based agent that specializes in various subtasks distributed across different phases.",
+        "bbox": [
+            112,
+            84,
+            492,
+            212
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            112,
+            224,
+            247,
+            239
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this study, we aim to bridge the gap between KBQA tasks and the investigation of LLM-based agents. We introduce Triad, a framework to address the KBQA task through an LLM-based agent acting as multiple roles, including a generalist capable of mastering diverse tasks given minimal examples, a decision-maker concentrating on option analysis and candidate selection, and an advisor skilled in answering questions with the aid of both external and internal knowledge. Triad achieves the best or competitive performance across three benchmark datasets compared to traditional KBQA systems and pure LLM models. In future research, we plan to broaden our framework to handle more intricate questions, such as multi-hop reasoning, and exploring the integration between our framework and retrieval-augmented generation.",
+        "bbox": [
+            112,
+            249,
+            490,
+            523
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            533,
+            220,
+            549
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "The limitation of our research lies in following aspects: (1) In terms of data, a broader range of QA datasets needs to be evaluated, encompassing datasets from different domains, languages, and difficulty levels. (2) In terms of model, more LLMs need to be evaluated, including open-source and commercial models from different organizations and on various scales. (3) In terms of framework, more types of agent collaboration methods can be explored to solve KBQA problems.",
+        "bbox": [
+            112,
+            558,
+            489,
+            720
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Ethics Considerations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            731,
+            309,
+            746
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "All datasets utilized in this study are publicly available and we have adhered to ethical considerations by not introducing additional information as input during LLM text generation.",
+        "bbox": [
+            112,
+            756,
+            490,
+            821
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            114,
+            832,
+            285,
+            848
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This work is supported by the \"Pioneer\" and \"Leading Goose\" R&D Program of Zhejiang (Grant No. 2023C01152 and No. 2024C01034).",
+        "bbox": [
+            112,
+            857,
+            489,
+            904
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            510,
+            84,
+            608,
+            98
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Sören Auer, Christian Bizer, Georgi Kobilarov, Jens Lehmann, Richard Cyganiak, and Zachary Ives. 2007. Dbpedia: A nucleus for a web of open data. In The Semantic Web, pages 722-735, Berlin, Heidelberg. Springer Berlin Heidelberg.",
+            "Jinheon Baek, AlhamFikri Aji, and Amir Saffari. 2023. Knowledge-augmented language model prompting for zero-shot knowledge graph question answering. arXiv preprint arXiv:2306.04136.",
+            "Qingxiu Dong, Li Dong, Ke Xu, Guangyan Zhou, Yaru Hao, Zhifang Sui, and Furu Wei. 2023. Large language model for science: A study on p vs. np. arXiv preprint arXiv:2309.05689.",
+            "Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Zhiyong Wu, Baobao Chang, Xu Sun, Jingjing Xu, and Zhifang Sui. 2022. A survey for in-context learning. arXiv preprint arXiv:2301.00234.",
+            "Chenxu Hu, Jie Fu, Chenzhuang Du, Simian Luo, Junbo Zhao, and Hang Zhao. 2023a. Chatdb: Augmenting llms with databases as their symbolic memory. arXiv preprint arXiv:2306.03901.",
+            "Nan Hu, Yike Wu, Guilin Qi, Dehai Min, Jiaoyan Chen, Jeff Z Pan, and Zafar Ali. 2023b. An empirical study of pre-trained language models in simple knowledge graph question answering. _World Wide Web_, pages 1-32.",
+            "Sen Hu, Lei Zou, Jeffrey Xu Yu, Haixun Wang, and Dongyan Zhao. 2018. Answering natural language questions by subgraph matching over knowledge graphs. IEEE Transactions on Knowledge and Data Engineering, page 824-837.",
+            "Xixin Hu, Yiheng Shu, Xiang Huang, and Yuzhong Qu. 2021. Edg-based question decomposition for complex question answering over knowledge bases. In The Semantic Web – ISWC 2021: 20th International Semantic Web Conference, ISWC 2021, Virtual Event, October 24–28, 2021, Proceedings, page 128–145, Berlin, Heidelberg. Springer-Verlag.",
+            "Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199-22213.",
+            "Jinyang Li, Binyuan Hui, Reynold Cheng, Bowen Qin, Chenhao Ma, Nan Huo, Fei Huang, Wenyu Du, Luo Si, and Yongbin Li. 2023. Graphix-t5: Mixing pre-trained transformers with graph-aware layers for text-to-sql parsing. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 13076-13084.",
+            "Jinyang Li, Binyuan Hui, Ge Qu, Jiaxi Yang, Binhua Li, Bowen Li, Bailin Wang, Bowen Qin, Ruiying Geng, Nan Huo, et al. 2024. Can llm already serve as a database interface? a big bench for large-scale"
+        ],
+        "bbox": [
+            510,
+            105,
+            885,
+            921
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "1706",
+        "bbox": [
+            482,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "database grounded text-to-sqls. Advances in Neural Information Processing Systems, 36.",
+            "Zhiwei Liu, Weiran Yao, Jianguo Zhang, Le Xue, Shelby Heinecke, Rithesh Murthy, Yihao Feng, Zeyuan Chen, Juan Carlos Niebles, Devansh Arpit, et al. 2023. Bolaa: Benchmarking and orchestrating llm-augmented autonomous agents. arXiv preprint arXiv:2308.05960.",
+            "Reham Omar, Ishika Dhall, Panos Kalnis, and Essam Mansour. 2023a. A universal question-answering platform for knowledge graphs. Proceedings of the ACM on Management of Data, 1(1):1-25.",
+            "Reham Omar, Omij Mangukiya, Panos Kalnis, and Essam Mansour. 2023b. Chatgpt versus traditional question answering for knowledge graphs: Current status and future directions towards knowledge graph chatbots. arXiv preprint arXiv:2302.06466.",
+            "Bhargavi Paranjape, Scott Lundberg, Sameer Singh, Hannaneh Hajishirzi, Luke Zettlemoyer, and MarcoTulio Ribeiro. 2023. Art: Automatic multi-step reasoning and tool-use for large language models. arXiv preprint arXiv:2303.09014.",
+            "Thomas Pellissier Tanon, Gerhard Weikum, and Fabian Suchanek. 2020. Yago 4: A reason-able knowledge base. In The Semantic Web, pages 583–596, Cham. Springer International Publishing.",
+            "Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettle-moyer, Nicola Cancedda, and Thomas Scialom. 2024. Toolformer: Language models can teach themselves to use tools. Advances in Neural Information Processing Systems, 36.",
+            "Chuanyuan Tan, Yuehe Chen, Wenbiao Shao, Wenliang Chen, Zhefeng Wang, Baoxing Huai, and Min Zhang. 2023a. Make a choice! knowledge base question answering with in-context learning. arXiv preprint arXiv:2305.13972.",
+            "Yiming Tan, Dehai Min, Yu Li, Wenbo Li, Nan Hu, Yongrui Chen, and Guilin Qi. 2023b. Can chatgpt replace traditional kbqa models? an in-depth analysis of the question answering performance of the gpt llm family. In International Semantic Web Conference, pages 348-367. Springer.",
+            "Yiming Tan, Dehai Min, Yu Li, Wenbo Li, Nan Hu, Yongrui Chen, and Guilin Qi. 2023c. Evaluation of chatgpt as a question answering system for answering complex questions. arXiv preprint arXiv:2303.07992.",
+            "Daniel Tang, Zhenghan Chen, Kisub Kim, Yewei Song, Haoye Tian, Saad Ezzini, Yongfeng Huang, and Jacques Klein Tegawende F Bissyande. 2024. Collaborative agents for software engineering. arXiv preprint arXiv:2402.02172."
+        ],
+        "bbox": [
+            115,
+            85,
+            485,
+            920
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Priyansh Trivedi, Gaurav Maheshwari, Mohnish Dubey, and Jens Lehmann. 2017. Lc-quad: A corpus for complex question answering over knowledge graphs. In The Semantic Web – ISWC 2017, pages 210–218, Cham. Springer International Publishing.",
+            "Ricardo Usbeck, Ria Hari Gusmita, Axel-Cyrille Ngonga Ngomo, and Muhammad Saleem. 2018. 9th challenge on question answering over linked data (qald-9) (invited paper). In Semdeep/NLIWoD@ISWC.",
+            "Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, et al. 2023. A survey on large language model based autonomous agents. arXiv preprint arXiv:2308.11432.",
+            "Tonghan Wang, Tarun Gupta, Anuj Mahajan, Bei Peng, Shimon Whiteson, and Chongjie Zhang. 2020. Rode: Learning roles to decompose multi-agent tasks. arXiv preprint arXiv:2010.01523.",
+            "Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023. React: Synergizing reasoning and acting in language models."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            426
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "1707",
+        "bbox": [
+            482,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "A Response Time Analysis",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            361,
+            99
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "We analyze various QA frameworks in response time to a question. The average latency of each phase including question parsing (QP), URI linking (UL), and answer generation (AG) for each knowledge base is reported. We randomly select 10 samples from each dataset for evaluation to obtain the average response times for Triad-1 and Triad-3, which represent retrying three times and generating an answer in one go, respectively, during the answer generation phase. The comparison between traditional QA systems and Triad is shown in Figure 3. Triad generally shows a competitive time-consuming performance to latest traditional QA systems. Specifically, compared to other phases, URL linking consumes more time due to the need to invoke LLM multiple times. Moreover, according to Section 4.4, with smaller retry times of A-Agent, Triad can significantly reduce time cost while only causing slight performance degradation, revealing the advantages of our framework in balancing performance and efficiency.",
+        "bbox": [
+            112,
+            108,
+            490,
+            447
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "image",
+        "img_path": "images/f49b3177223a6632d6f88a99553a955e45e71fb82605a396c8fdba2b7fe01725.jpg",
+        "image_caption": [
+            "Figure 3: Response time of traditional KBQA systems and Triad on three datasets. Each bar shows average response time of a particular phase of KBQA."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            457,
+            490,
+            615
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "B Role Performance on YAGO-QA",
+        "text_level": 1,
+        "bbox": [
+            114,
+            700,
+            431,
+            717
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "We choose LC-QuAD 1.0 and QALD-9 as our two representative datasets in Section 4.3, as the questions among them vary in difficulty, and the tasks in these two datasets are relatively more challenging than YAGO-QA. We provide the performance of agent roles on YAGO-QA in Table 6, which shows a consistent result with other datasets in Table 3.",
+        "bbox": [
+            112,
+            724,
+            487,
+            838
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "C Prompts Provided to LLMs of G-Agent for Solving Various Subtasks in KBQA",
+        "text_level": 1,
+        "bbox": [
+            114,
+            848,
+            487,
+            881
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "The prompt given to LLMs of  $Agent_{g}$  to perform triplet extraction from the question  $Q$  is as follows:",
+        "bbox": [
+            112,
+            889,
+            489,
+            921
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "table",
+        "img_path": "images/d5794f62b29a5ef9829e89ab726e9fe1b28602047520412768658df88e9e4ccd.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>G-task</td><td>G-chat</td><td>YAGO-QA</td></tr><tr><td>×</td><td>×</td><td>0.427</td></tr><tr><td>✓</td><td>×</td><td>0.553</td></tr><tr><td>✓</td><td>✓</td><td>0.677</td></tr><tr><td>D-uri</td><td>D-query</td><td>YAGO-QA</td></tr><tr><td>×</td><td>✓</td><td>0.346</td></tr><tr><td>✓</td><td>×</td><td>0.534</td></tr><tr><td>✓</td><td>✓</td><td>0.677</td></tr><tr><td>A-llm</td><td>A-fact</td><td>YAGO-QA</td></tr><tr><td>×</td><td>×</td><td>0.626</td></tr><tr><td>✓</td><td>×</td><td>0.647</td></tr><tr><td>✓</td><td>✓</td><td>0.677</td></tr></table>",
+        "bbox": [
+            563,
+            83,
+            831,
+            307
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Table 6: Performance of roles of LLM-based agent by eliminating an element or downgrading the capability.",
+        "bbox": [
+            507,
+            317,
+            882,
+            347
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "You are an assistant to identify triples within a provided sentence. Please adhere to the following guidelines:",
+        "text_level": 1,
+        "bbox": [
+            514,
+            393,
+            889,
+            441
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. Triples should be structured in the format <entity1, relation, entity2>.",
+            "2. The sentence must contain at least one triple, so you should provide at least one.",
+            "3. Entities should represent the smallest semantic units and should not contain descriptive details.",
+            "4. Entities can take the form of explicit or implicit references. Explicit entities refer to specific named resources, whereas implicit entities are less certain.",
+            "5. When an entity is implicit, utilize a variable format such as '?variable' to denote it, for example, '?location' or '?person'."
+        ],
+        "bbox": [
+            514,
+            457,
+            887,
+            650
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Here are some examples:",
+        "text_level": 1,
+        "bbox": [
+            515,
+            667,
+            714,
+            682
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Which city's founder is John Forbes? : <?city, foundeer, John Forbes>",
+        "bbox": [
+            514,
+            683,
+            887,
+            713
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "How many races have the horses bred by Jacques Van't Hart participated in? : <?horse, participated in, ?race> <?horse, breeder, Jacques Van't Hart>",
+        "bbox": [
+            514,
+            715,
+            887,
+            762
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Is camel of the chordate phylum? : <camel, phylum, chordate>",
+        "bbox": [
+            514,
+            764,
+            887,
+            794
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Sentence: <Question Sentence>",
+        "bbox": [
+            514,
+            812,
+            757,
+            827
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Output:",
+        "text_level": 1,
+        "bbox": [
+            515,
+            828,
+            584,
+            843
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "The prompt given to LLMs of  $A g e n t_{g}$  for SPARQL template generation is as follows:",
+        "bbox": [
+            507,
+            857,
+            882,
+            889
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "1708",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "You are an assistant to generate a SPARQL query to address a specific question. Here are the guidelines to follow:",
+        "bbox": [
+            119,
+            84,
+            494,
+            131
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. Ensure that the resulting SPARQL query is designed to answer the provided question.",
+            "2. Adhere to the commonly accepted SPARQL standards when generating the query.",
+            "3. Make an effort to leverage the information provided to assist in the creation of the SPARQL query.",
+            "4. Strive to keep the generated SPARQL query as straightforward as possible.",
+            "5. Avoid including 'PREFIX' or ':' in the SPARQL query.",
+            "6. Enclose condition entities and predicates within angle brackets, such as <entity> or <predicate>.",
+            "7. Maintain the original order of the given triples without altering their sequence."
+        ],
+        "bbox": [
+            119,
+            148,
+            494,
+            390
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Question: <question sentence>",
+        "bbox": [
+            119,
+            406,
+            357,
+            420
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Triplets: <extracted triplets>",
+        "bbox": [
+            119,
+            422,
+            339,
+            438
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Output:",
+        "bbox": [
+            119,
+            438,
+            188,
+            454
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "You are an assistant to select  $\\langle K \\rangle$  URIs from a provided list of possible URIs for a specified entity, following these guidelines:",
+        "bbox": [
+            514,
+            84,
+            889,
+            131
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. Identify the  $\\langle \\mathrm{K} \\rangle$  most appropriate URIs from the given list that best represent the entity in question.",
+            "2. Seek to understand the semantic information associated with the specified entity by examining the provided question.",
+            "3. The output should consist of  $\\langle \\mathrm{K} \\rangle$  URIs chosen from the provided list of possible URIs.",
+            "4. Simply output these  $\\langle \\mathrm{K} \\rangle$  target URIs, each on a separate line, without providing any additional explanations."
+        ],
+        "bbox": [
+            514,
+            148,
+            887,
+            325
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Sentence: <question sentence>",
+        "bbox": [
+            514,
+            341,
+            752,
+            357
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Entity: <entity mention>",
+        "bbox": [
+            514,
+            357,
+            707,
+            373
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Possible entity URIs: <Entity URI list>",
+        "bbox": [
+            514,
+            374,
+            818,
+            388
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Output:",
+        "bbox": [
+            514,
+            390,
+            584,
+            406
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "The prompt given to LLMs of  $Agent_d$  for candidate relation selection is as follows:",
+        "bbox": [
+            507,
+            418,
+            884,
+            448
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "The prompt given to LLMs of  $Agent_g$  for question type classification is as follows:",
+        "bbox": [
+            114,
+            470,
+            489,
+            502
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "You are an assistant to determine the specific type of a given question according to the following guidelines:",
+        "bbox": [
+            119,
+            538,
+            492,
+            586
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. You must determine the most probable question type for the input question.",
+            "2. The type of question should be enclosed within angle brackets, denoted as  $\\text{’} < \\text{’}$  and  $\\text{’} > \\text{’}$ .",
+            "3. Possible question types include: <count>, <select>, and <yes or no>."
+        ],
+        "bbox": [
+            119,
+            602,
+            494,
+            699
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Question: <question sentence>",
+        "bbox": [
+            119,
+            715,
+            357,
+            730
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Output:",
+        "bbox": [
+            119,
+            732,
+            188,
+            747
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "You are an assistant tasked with selecting the  $\\langle K\\rangle$  relation URIs between entities mentioned in a sentence. Here are the guidelines:",
+        "bbox": [
+            514,
+            476,
+            889,
+            526
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. The two entities are listed one after the other, without a specific order.",
+            "2. Use the provided sentence to discern the semantic meaning of these entities.",
+            "3. The potential relation URIs are listed one by one.",
+            "4. Your output should consist of a maximum of  $\\langle \\mathrm{K} \\rangle$  possible relation URIs, although you may also output fewer if appropriate.",
+            "5. Ensure that your output is organized, prioritizing the most likely relationship first.",
+            "6. Provide a list of no more than  $\\langle \\mathrm{K} \\rangle$  relation URIs (each on a separate line if there are multiple) without any additional descriptions."
+        ],
+        "bbox": [
+            514,
+            542,
+            887,
+            766
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Sentence: <question sentence>",
+        "bbox": [
+            514,
+            783,
+            752,
+            797
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Entities: <entity pair>",
+        "bbox": [
+            514,
+            799,
+            687,
+            815
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Possible relation URIs: <URI list>",
+        "bbox": [
+            514,
+            816,
+            784,
+            829
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Output:",
+        "bbox": [
+            514,
+            831,
+            584,
+            847
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "D Prompts Provided to LLMs of D-Agent for Solving Selection Subtasks in KBQA",
+        "bbox": [
+            114,
+            778,
+            487,
+            829
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "The prompt given to LLMs of  $Agent_d$  for candidate entities selection is as follows:",
+        "bbox": [
+            112,
+            851,
+            489,
+            883
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "The prompt given to LLMs of  $Agent_d$  for the final SPARQL selection is as follows:",
+        "bbox": [
+            509,
+            859,
+            882,
+            890
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "1709",
+        "bbox": [
+            482,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "You are an assistant to select an appropriate SPARQL query from the provided list in order to respond to a specific question. Please adhere to the following guidelines:",
+        "bbox": [
+            119,
+            84,
+            494,
+            149
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. Select the most suitable SPARQL query from the given query list to address the question.",
+            "2. Select a SPARQL query solely from the provided list; avoid crafting your own SPARQL query.",
+            "3. The selected SPARQL query must be applicable to answer the given question."
+        ],
+        "bbox": [
+            119,
+            164,
+            490,
+            275
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Sentence: <question sentence>",
+        "bbox": [
+            119,
+            294,
+            356,
+            307
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "SPARQL candidates: <SPARQLs to choose> Output:",
+        "bbox": [
+            119,
+            309,
+            468,
+            341
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "E Prompts Provided to LLMs of A-Agent for Solving Answering Subtask in KBQA",
+        "text_level": 1,
+        "bbox": [
+            114,
+            353,
+            487,
+            404
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "The prompt given to LLMs of  $Agent_{a}$  to generate a yes or no answer for the give question is as follows:",
+        "bbox": [
+            112,
+            413,
+            489,
+            445
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "You are an assistant to answer a yes-or-no question. Please adhere to the following guidelines:",
+        "bbox": [
+            119,
+            464,
+            494,
+            495
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. If you believe that the answer is yes, provide an output of 'True'. If not, provide an output of 'False'.",
+            "2. Please do not include additional information or explanations in your response."
+        ],
+        "bbox": [
+            119,
+            512,
+            492,
+            592
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Sentence: <question sentence>",
+        "bbox": [
+            119,
+            609,
+            356,
+            623
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Output:",
+        "text_level": 1,
+        "bbox": [
+            121,
+            625,
+            186,
+            640
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "The prompt given to LLMs of  $Agent_{a}$  to generate a single-fact answer for the give question is as follows:",
+        "bbox": [
+            114,
+            646,
+            489,
+            693
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "You are an assistant to answer a question. Please adhere to the following guidelines:",
+        "bbox": [
+            119,
+            713,
+            492,
+            745
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. The answer to the question is a single entity.",
+            "2. You should just output the full expression of the answer without any punctuation.",
+            "3. Do not output any other description."
+        ],
+        "bbox": [
+            119,
+            760,
+            489,
+            826
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Sentence: <question sentence>",
+        "bbox": [
+            119,
+            841,
+            356,
+            857
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Output:",
+        "text_level": 1,
+        "bbox": [
+            121,
+            859,
+            186,
+            873
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "1710",
+        "bbox": [
+            482,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 12
+    }
+]

2024/Triad_ A Framework Leveraging a Multi-Role LLM-based Agent to Solve Knowledge Base Question Answering/2bfc46f7-f4ed-4d88-9bf5-1ead6a8c1d52_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:772344d74851f1c20a1fa34994b7a8732ea429515b4528dfa408ba9a8e1660a9
+size 445917

2024/Triad_ A Framework Leveraging a Multi-Role LLM-based Agent to Solve Knowledge Base Question Answering/full.md

@@ -0,0 +1,493 @@
+# Triad: A Framework Leveraging a Multi-Role LLM-based Agent to Solve Knowledge Base Question Answering
+
+Chang Zong $^{1}$ , Yuchen Yan $^{1}$ , Weiming Lu $^{1\dagger}$ , Jian Shao $^{1}$   
+Yongfeng Huang $^{2}$ , Heng Chang $^{3}$ , Yueting Zhuang $^{1\dagger}$
+
+<sup>1</sup>College of Computer Science and Technology, Zhejiang University  
+<sup>2</sup>The Chinese University of Hong Kong  
+<sup>3</sup>Tsinghua University  
+{zongchang, luwm, yzhuang}@zju.edu.cn
+
+# Abstract
+
+Recent progress with LLM-based agents has shown promising results across various tasks. However, their use in answering questions from knowledge bases remains largely unexplored. Implementing a KBQA system using traditional methods is challenging due to the shortage of task-specific training data and the complexity of creating task-focused model structures. In this paper, we present Triad, a unified framework that utilizes an LLM-based agent with multiple roles for KBQA tasks. The agent is assigned three roles to tackle different KBQA subtasks: agent as a generalist for mastering various subtasks, as a decision maker for the selection of candidates, and as an advisor for answering questions with knowledge. Our KBQA framework is executed in four phases, involving the collaboration of the agent's multiple roles. We evaluated the performance of our framework using three benchmark datasets, and the results show that our framework outperforms state-of-the-art systems on the LC-QuAD and YAGO-QA benchmarks, yielding F1 scores of  $11.8\%$  and  $20.7\%$ , respectively.
+
+# 1 Introduction
+
+A question-answering system is designed to extract information by converting a natural language question into a structured query that can retrieve precise information from an existing knowledge base (Omar et al., 2023a). The resolution of Knowledge Base Question Answering (KBQA) typically involves phases including question understanding, URI linking, and query execution. Traditional KBQA systems require the use of specialized models trained with domain datasets for question parsing and entity linking (Hu et al., 2018; Omar et al., 2023a; Hu et al., 2021). Large language models (LLMs), however, have shown promising competencies in in-context learning using task-specific
+
+demonstrations (Dong et al., 2022). LLMs have recently been employed as agents in the execution of complex problems. A framework that employs LLM-augmented agents can generate actions or coordinate multiple agents, thus improving the capacity to handle complex situations (Liu et al., 2023). Despite the remarkable performance of LLMs in various tasks as evidenced in previous studies, a comprehensive qualitative and quantitative evaluation of KBQA frameworks empowered with an LLM-based agent remains insufficiently explored.
+
+![](images/99ca4918bdf6bc83a47e697322f730e3148f67aa4f29ce2ff8375a19d1064a9d.jpg)  
+Figure 1: A system with multiple roles who focus on sub-problems of each phase to solve a complex task.
+
+Studies on KBQA with LLMs has attracted considerable attention. Some works focus primarily on highlighting the inability of LLMs to generate complete factoid results (Hu et al., 2023b; Tan et al., 2023c) or demonstrating their potential efficacy in future research (Omar et al., 2023b; Tan et al., 2023b). Other works concentrates on generating answers by prompt learning and incorporating external knowledge bases (Baek et al., 2023; Tan et al., 2023a). Concurrently, LLMs can be deployed to address each phase within Text2SQL challenges(Li et al., 2023, 2024) or theorem proof tasks(Dong et al., 2023). However, each phase of
+
+KBQA can be further decomposed into subtasks and completed through an agentic approach that provides feedback and cooperation. Additionally, decomposing the task reduces the complexity of cooperative working by allowing each role to concentrate on smaller sub-problems (Wang et al., 2020). As illustrated in Figure 1, three roles in an organization work together to provide the final answer for the overall task. The above observations spur our exploration into the following question: How does an LLM-based agent solve KBQA tasks by serving as multiple roles, and its performance is comparable to systems trained specifically?
+
+In this study, we introduce Triad, a unified framework that leverages an LLM-based agent with three roles to address KBQA tasks. Specifically, we implement the agent consisting of an LLM as the core, supplemented by various task-specific modules such as memory and executing functions. The agent is assigned three distinct roles: a generalist (G-Agent) adept at mastering numerous small tasks by the given examples, a decision maker (D-Agent) proficient at identifying options and selecting candidates, and an advisor (A-Agent) skilled at providing answers using internal and external knowledge. The cooperation of these agent roles composes a KBQA process containing four phases: question parsing, URI linking, query construction, and answer generation. We evaluate our framework on three benchmark datasets in various difficulties. The results show that our framework outperforms state-of-the-art systems, demonstrated by  $11.8\%$  and  $20.7\%$  F1 scores on the LC-QuAD and YAGO-QA benchmarks, respectively<sup>1</sup>.
+
+The contributions of this study can be summarized as follows:
+
+- We propose Triad, the first framework that leverages an LLM-based agent to solve KBQA tasks in all its four phases, without specialized training models.  
+- We implement an LLM-based agent with various task-specific modules that can act as three roles, including a generalist, a decision maker, and an advisor, to collaboratively solve KBQA via focusing on subtasks.  
+- We evaluate the performance of Triad. The results show a competitive ability compared
+
+to both state-of-the-art KBQA systems and pure LLM methods.
+
+# 2 Preliminaries
+
+# 2.1 Phases of KBQA
+
+A typical KBQA system has a process that encompasses four phases:
+
+Question parsing involves converting natural language questions into a structured format that incorporates references to entities and relations.
+
+URI linking entails associating and replacing these entity and relation mentions with their corresponding URIs within a knowledge base.
+
+Query construction involves creating executable queries in a standard format to extract answers from knowledge bases.
+
+Answer generation seeks to obtain the ultimate answers either by performing queries within knowledge bases or by directly querying an agent.
+
+# 2.2 Roles of LLM-based Agent
+
+Drawing an analogy to a software development scenario, where coders complete small development tasks, with the process and plan being decided by the manager, and ultimately the outcome inspected by the leader, we assign the following three roles to an LLM-based agent to solve the KBQA task:
+
+Agent as a generalist (G-Agent) is capable of mastering various small tasks by providing a few examples.
+
+Agent as a decision-maker (D-Agent) adepts at analyzing options and providing candidate results as procedural feedback.
+
+Agent as an advisor (A-Agent) is skilled in providing final answers with the aid of both external and its own knowledge.
+
+# 2.3 Task Formulation
+
+A KBQA task refers to the process of solving a set of subtasks  $S$ . Each subtask  $S_{t} \in S$  contributes to one phase of the whole process. An LLM-based agent  $Agent_{r}$  with a role  $r$  can be used to resolve a type of subtasks by its task-specific components, including a language model  $LLM$ , a memory  $Mem_{t}$ , a function  $F_{t}$ , a prompt  $Pmt_{t}$  and a set of parameters  $\theta_{t}$ , using the set of role-related hyperparameters  $\sigma_{r}$ . The task can be formulated as follows:
+
+$$
+f (K B Q A) = \bigoplus_ {t = 1} ^ {T} f (S _ {t})
+$$
+
+$$
+f \left(S _ {t}\right) = A g e n t _ {r} \left(L L M, M e m _ {t}, F _ {t}, P m t _ {t}, \theta_ {t}, \sigma_ {r}\right) \tag {1}
+$$
+
+, where  $T$  is the total number of subtasks,  $\bigoplus$  is the way to coordinate subtasks to solve the whole.
+
+# 3 Triad Framework
+
+The overall architecture of Triad is shown in Figure 2. Each role of the LLM-based agent, along with its associated subtasks, is illustrated as follows.
+
+# 3.1 G-Agent as a Generalized Solver
+
+A generalized agent (G-Agent) proficiently manages numerous tasks by leveraging learning from limited examples through an LLM. In our framework, a G-Agent can perform question parsing, query template generation, or answer type classification as actions solely utilizing an LLM. These three subtasks are illustrated as follows:
+
+Triplet mention extraction: The process of extracting triplet mentions in question parsing involves the conversion of a naturally phrased question, denoted as  $Q$ , into formatted triplets of entities and relations. This subtask is executed employing an LLM, which is guided by a prompt with a set of prerequisites and a selection of examples. This subtask can be represented as follows:
+
+$$
+f \left(S _ {t r i}\right) = A g e n t _ {g} \left(L L M, P m t _ {t r i}, Q, \mathcal {N}\right) \tag {2}
+$$
+
+$$
+P m t _ {t r i} = \left[ I n s _ {t r i}, S h o t _ {t r i}, C o T _ {t r i} \right]
+$$
+
+, where  $A g e n t_{g}$  is the agent as a generalist to perform the triplet extraction subtask with  $\mathcal{N}$  examples.  $P m t_{tri}$  is the prompt to guide LLM to generate triplets from the question  $Q$ , which consists of instruction  $I n s_{tri}$ , examples  $S h o t_{tri}$ , and chain-of-thought prompt  $C o T_{tri}$  (Kojima et al., 2022).
+
+SPARQL template generation: The generation of SPARQL templates in query construction involves the use of an LLM to create a SPARQL template that articulates the question using standard SPARQL syntax, replacing URI identifiers with entity and relation variables. To derive precise and comprehensive answers from the knowledge base using SPARQL queries, there are two potential strategies. One approach involves the direct generation of an executable SPARQL using
+
+an LLM, though this method may significantly increase LLM call times and error rates when numerous candidate queries are in play. Alternatively, a SPARQL template can initially be generated with entity and relation variables, which are subsequently replaced with linked URIs. For the sake of stability and efficiency, we opt for the second strategy. This subtask can be denoted as:
+
+$$
+f (S _ {q t}) = A g e n t _ {g} (L L M, P m t _ {q t}, \theta_ {q t}, \mathcal {N}),
+$$
+
+$$
+P m t _ {q t} = \left[ I n s _ {q t}, S h o t _ {q t}, C o T _ {q t} \right], \tag {3}
+$$
+
+$$
+\theta_ {q t} = [ Q, f (S _ {t r i}) ]
+$$
+
+, where  $A g e n t_{g}$  is the agent as a generalist to perform SPARQL template generation with  $\mathcal{N}$  examples,  $f(S_{tri})$  is the triplets derived from the previous subtask,  $P m t_{q t}$  is the prompt for LLM to generate SPARQL template.
+
+Answer type classification: In the phase of answer generation, the answer type classification subtask refers to the process of assigning a specific category to a response according to the question. This process serves as a guiding mechanism for the framework to generate comprehensive and accurate answers. This classification subtask is denoted as:
+
+$$
+f \left(S _ {c l s}\right) = A g e n t _ {g} \left(L L M, P m t _ {c l s}, Q, \mathcal {N}\right), \tag {4}
+$$
+
+$$
+P m t _ {c l s} = \left[ I n s _ {c l s}, S h o t _ {c l s}, C o T _ {c l s} \right]
+$$
+
+, where  $A g e n t_{g}$  is the agent as a generalist to perform type classification subtask with  $\mathcal{N}$  examples,  $P m t_{c l s}$  is the prompt for LLM.
+
+# 3.2 D-Agent as a Decision-Maker
+
+An agent as a decision maker (D-Agent) is capable of making candidate selections step by step through filtering and choosing from given options, harnessing the capabilities of an LLM and KB as memory. The D-Agent can effectively handle three subtasks, which are delineated as follows:
+
+Candidate entity selection: The selection of candidate entities in URI linking is pivotal to the ultimate efficacy of KBQA. Prior research has focused primarily on developing a semantic similarity model to address this linking challenge. However, the linking task requires numerous iterations of searching within the knowledge base, which poses a compatibility issue for LLM-oriented methods. In our framework, an agent as a decision maker is utilized initially to filter all potential entity URIs from the knowledge base, subsequently deploying an LLM to select candidate URIs from a pool of
+
+![](images/65f8a985484e660617f9c6e72ffacae12e9ff46eef3caef196387195c75dbcf4.jpg)  
+Figure 2: Our Triad framework leverages an LLM-based agent with three different roles including a generalist, a decision-maker, and an advisor to cooperatively handle a series of subtasks in the four phases of a KBQA process.
+
+![](images/19cd017a6f14bd1a81b57338bff9017ac9b677a59f7f97203cdb46895b63f89a.jpg)
+
+potential identifiers. For each entity, our aim is to find the  $\kappa$  most possible entity URIs which can be used to traverse over the KB to get the final answer. The entity selection subtask can be denoted as:
+
+$$
+f (S _ {e s}) = A g e n t _ {d} (L L M, M e m _ {e s}, F _ {e s},
+$$
+
+$$
+\left. P m t _ {e s}, \theta_ {e s}, \mathcal {K}\right), \tag {5}
+$$
+
+$$
+M e m _ {e s} = [ K B, L i s t _ {e s} ], \theta_ {e s} = [ Q, f (S _ {t r i}) ]
+$$
+
+, where  $Agent_d$  is the agent as a decision maker to perform the entity selection subtask with question  $Q$ , extracted triplets  $f(S_{tri})$  and memory  $Mem_{es}$ ,  $Mem_{es}$  is composed of a knowledge base  $KB$  and a list of entity URIs  $List_{es}$  filtered from  $KB$  using a text similarity matching function  $F_{es}$ ,  $Pmt_{es}$  is the prompt for LLM to perform the subtask,  $\mathcal{K}$  is the hyperparameter of  $Agent_d$ , indicating the number of candidates selected by LLM.
+
+Candidate relation selection: The task of candidate relation selection in URI linking presents considerable challenges due to the discrepancies between word forms and meanings. Nevertheless, the existence of reasoning paths in the knowledge base can be utilized to allow for a significant reduction of the search space in relation linking. In our framework, an agent as a decision maker endeavors to sieve through all potential relation URIs by navigating the knowledge base with candidate entity URIs generated from the previous subtask. Subsequently, an LLM is used to select the top  $\kappa$  most probable relation URIs for output. The relation
+
+selection subtask can be denoted as:
+
+$$
+f \left(S _ {r s}\right) = A g e n t _ {d} (L L M, M e m _ {r s}, F _ {r s},
+$$
+
+$$
+\left. P m t _ {r s}, \theta_ {r s}, \mathcal {K}\right), \tag {6}
+$$
+
+$$
+M e m _ {r s} = \left[ K B, L i s t _ {r s} \right], \theta_ {r s} = \left[ Q, f \left(S _ {e s}\right) \right]
+$$
+
+, where memory  $Mem_{rs}$  is composed of the knowledge base  $KB$  and a list of possible relation URIs  $List_{rs}$  filtered from  $KB$  using a one-order traversing function  $F_{rs}$ .  $Pmt_{rs}$  is the prompt for LLM to perform relation selection.  $\kappa$  is the number of relation URIs selected by LLM.
+
+Candidate SPARQL selection: The subtask of candidate SPARQL selection in query construction involves determining the appropriate SPARQL queries to obtain the final answers. Given a SPARQL template generated by the G-Agent, along with multiple candidate URIs selected from the D-Agent in previous subtasks, our D-Agent is targeted to identify the most plausible query. To further reduce the difficulty of selection, an executor function is applied to eliminate queries that cannot retrieve any results from the knowledge base. In conclusion, our aim in this subtask is to use D-Agent to construct executable SPARQLs and find the most possible one given a query candidate list with supported information. The SPARQL selection subtask can be denoted as:
+
+$$
+\begin{array}{l} f (S _ {q s}) = A g e n t _ {d} (L L M, M e m _ {q s}, F _ {q s}, \\ \left. \begin{array}{l} P m t _ {q s}, \theta_ {q s}, \mathcal {K}), \\ [ ] \end{array} \right. \tag {7} \\ \end{array}
+$$
+
+$$
+M e m _ {q s} = \left[ K B, L i s t _ {q s} \right],
+$$
+
+$$
+\theta_ {q s} = [ Q, f (S _ {e s}), f (S _ {r s}), f (S _ {q t}) ]
+$$
+
+, where memory  $Mem_{qs}$  is composed of a knowledge base  $KB$  and a list of possible SPARQLs  $List_{qs}$  constructed with SPARQL template  $f(S_{qt})$ , entity URIs  $f(S_{es})$ , and relation URIs  $f(S_{rs})$  by the function  $F_{qs}$ ,  $Pmt_{qs}$  is the prompt for LLM to perform query selection,  $\mathcal{K} = 1$  is the number of queries selected by LLM.
+
+# 3.3 A-Agent as a Comprehensive Advisor
+
+An advisory agent (A-Agent) is capable of processing a question and a corresponding type of answer as input. Its response is generated by either extracting information from an external knowledge base or by utilizing its internal knowledge to provide a direct answer. This comprehensive answering subtask can be described as follows:
+
+Comprehensive answering: The objective of comprehensive answering in the answer generation phase is to derive a definitive response based on an incoming question. Previous work (Omar et al., 2023b) has demonstrated that LLMs are more proficient in delivering single-fact responses and making Boolean judgments. Given this understanding, we implement an advisory agent that incorporates a simple policy to facilitate a comprehensive answering approach. Specifically, if a question yields a final SPARQL generated from the preceding steps, A-Agent extracts elements from the knowledge base to give the answer. Conversely, if the agent does not receive a feasible SPARQL, A-Agent provides a direct response with LLM's internal knowledge, following the prompt based on the type of the answer. Additionally, A-Agent will send a retry signal to previous phases if no result is generated. The subtask can be formulated as below:
+
+$$
+f (S _ {c a}) = A g e n t _ {a} (L L M, M e m _ {c a}, F _ {c a},
+$$
+
+$$
+\left. P m t _ {c a}, \theta_ {c a}, \mathcal {T}\right), \tag {8}
+$$
+
+$$
+M e m _ {c a} = [ K B ], \theta_ {c a} = [ Q, f (S _ {q s}), f (S _ {c l s}) ]
+$$
+
+, where  $A g e n t_{a}$  is the agent as an advisor to perform a comprehensive answering for the question  $Q$  with a memory  $M e m_{c a}$  of knowledge base,  $P m t_{c a}$  is the prompt for LLM to perform a direct response according to the type of the answer,  $f(S_{q s})$  is the final query and  $f(S_{c l s})$  is the answer type,  $\mathcal{T}$  is the maximum times to retry for previous phases if no result is returned from  $K B$ .
+
+# 4 Performance Evaluation
+
+# 4.1 Experimental Settings
+
+Indexed Knowledge Bases: The efficacy of our framework is assessed through the collection of two real knowledge bases, specifically DBpedia and YAGO. DBpedia (Auer et al., 2007) serves as an accessible knowledge base extracted from Wikipedia, while YAGO (Pellissier Tanon et al., 2020) is a large knowledge base that includes individuals, cities, nations, and organizations. We index the triples and the mentions of entities and relations in a Virtuoso endpoint and an Elasticsearch server, respectively.
+
+KBQA Benchmark Datasets: We evaluate our framework on datasets including YAGO-QA, LCQuAD 1.0, and QALD-9, which have various difficulties in interpreting the questions. These datasets contain questions in English, paired with their respective SPARQL queries, and accurate responses derived from a specific knowledge base. QALD-9 (Usbeck et al., 2018) and LC-QuAD 1.0 (Trivedi et al., 2017) are frequently used to evaluate QA systems with DBpedia. The recently published YAGO-QA in (Omar et al., 2023a), features questions accompanied by annotated SPARQL queries sourced from YAGO. The statistics for three benchmarks, along with their associated knowledge bases, are depicted in Table 1.
+
+Baseline Methods: We evaluate Triad against traditional KBQA systems such as KGQAN (Omar et al., 2023a), EDGQA (Hu et al., 2021) and gAnswer (Hu et al., 2018). This comparison shows how our LLM-based agent framework can rival full-shot systems with just a few examples. Additionally, we contrast our framework with pure GPT models like GPT-3.5 Turbo and GPT-4 to exhibit Triad's architectural performance. We treat these foundation models as few-shot methods to answer the questions referring to some examples.
+
+Implementation Details: Triad is implemented with Python 3.9. We incorporate LLM capabilities to our multi-role agent via OpenAI's API services. The names of entities and relations from knowledge bases are indexed in an ElasticSearch 7.5.2 server for text matching. All triples are imported into an SPARQL endpoint of Virtuoso 07.20.3237 for retrieval. Triad requires four hyperparameters: the number of examples G-Agent uses for subtask
+
+<table><tr><td rowspan="2">Benchmarks</td><td colspan="5">Benchmark Statistics</td></tr><tr><td>#Questions</td><td>KB</td><td>#Triples</td><td>Virtuoso Size</td><td>ES size</td></tr><tr><td>LC-QuAD 1.0</td><td>1000</td><td>DBpedia-04</td><td>397M</td><td>35.40G</td><td>1.56G</td></tr><tr><td>QALD-9</td><td>150</td><td>DBpedia-10</td><td>374M</td><td>36.89G</td><td>1.57G</td></tr><tr><td>YAGO-QA</td><td>100</td><td>YAGO-4</td><td>207M</td><td>24.85G</td><td>0.54G</td></tr></table>
+
+Table 1: The statistics of KBQA benchmarks, including the number of questions number, the number of triples, the size of index in Virtuoso and Elasticsearch.  
+
+<table><tr><td rowspan="2">Type</td><td rowspan="2">Frameworks</td><td colspan="3">LC-QuAD 1.0</td><td colspan="3">QALD-9</td><td colspan="3">YAGO-QA</td></tr><tr><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td></tr><tr><td rowspan="3">full-shot</td><td>gAnswer</td><td>-</td><td>-</td><td>-</td><td>0.293</td><td>0.327</td><td>0.298</td><td>0.585</td><td>0.341</td><td>0.430</td></tr><tr><td>EDGQA</td><td>0.505</td><td>0.560</td><td>0.531</td><td>0.313</td><td>0.403</td><td>0.320</td><td>0.419</td><td>0.408</td><td>0.414</td></tr><tr><td>KGQAN</td><td>0.587</td><td>0.461</td><td>0.516</td><td>0.511</td><td>0.387</td><td>0.441</td><td>0.485</td><td>0.652</td><td>0.556</td></tr><tr><td rowspan="4">few-shot</td><td>GPT-3.5</td><td>0.269</td><td>0.251</td><td>0.266</td><td>0.240</td><td>0.217</td><td>0.228</td><td>0.171</td><td>0.142</td><td>0.155</td></tr><tr><td>GPT-4</td><td>0.336</td><td>0.344</td><td>0.340</td><td>0.250</td><td>0.249</td><td>0.249</td><td>0.193</td><td>0.190</td><td>0.191</td></tr><tr><td>Triad-GPT3.5</td><td>0.490</td><td>0.519</td><td>0.504</td><td>0.293</td><td>0.302</td><td>0.297</td><td>0.660</td><td>0.639</td><td>0.649</td></tr><tr><td>Triad-GPT4</td><td>0.561</td><td>0.568</td><td>0.564</td><td>0.408</td><td>0.425</td><td>0.416</td><td>0.690</td><td>0.664</td><td>0.677</td></tr></table>
+
+Table 2: The performance of our proposed Triad on three benchmarks, comparing with traditional KBQA systems (full-shot) and pure LLM (few-shot) baselines. The optimal and suboptimal scores are highlighted with bold and underlined text, respectively.
+
+learning, the number of candidates D-Agent selects for entity and relation linking, and the retry times for handling non-response SPARQLs. The optimal values for these parameters are 3, 2, 2, and 3, respectively. The framework and its variants are tested five times on each benchmark, with the average scores reported as the final results. For traditional systems, we report the results recorded in their papers. For pure LLM baselines, we write prompts to hire an LLM to answer questions directly referring to examples, and then link the mentions from the responses to the URIs in our indexed knowledge bases via built-in similarity search.
+
+# 4.2 Performance Comparison
+
+The performance of Triad compared to traditional KBQA systems and pure LLM generation methods is shown in Table 2. Evaluation metrics precision(P), recall(R), and F1-score(F1) are reported. We can observe from the experimental results that:
+
+Few-shot can be competitive with full-shot. Our multi-role LLM-based agent framework, though executing a few-shot prompt learning, exhibits competitive performance with cutting-edge full-shot KBQA systems.
+
+Underlying capability matters. The use of GPT-4 as the core in an LLM-based agent significantly
+
+outperforms GPT-3.5 on all benchmarks, demonstrating the importance of the underlying capabilities of an agent.
+
+Explicit knowledge is necessary. Pure LLM models with GPT-3.5 and GPT-4 display deficiencies in generating accurate responses without an auxiliary knowledge base as a memory for intermediary steps such as URI linking.
+
+Performance varies with complexity. Triad demonstrates superior results on the LC-QuAD and YAGO-QA benchmarks compared to QALD-9, due to an increasing failure in response to complex questions, which will be discussed later.
+
+# 4.3 Study on Capabilities of Agent Roles
+
+We assess the efficacy of G-Agent with various other language models as the core. The framework without G-task uses the text-davinci-002, which is not as powerful as GPT-3.5 and GPT-4 in solving many tasks, and the one without G-chat uses text-davinci-003 to eliminate the chat and alignment abilities. We test the ability of D-Agent without D-uri and D-query by replacing the URI selection and query selection with URI matching and query generation, respectively. We evaluate the contribution of A-Agent eliminating A-llm and A-fact by responding to questions without using LLM's assistance or use an LLM to answer Boolean questions
+
+for auxiliary rather than single-fact questions. The F1 results of the role ablation experiments on two representative datasets are shown in Table 3. The results indicate that every component pertaining to each role contributes to the overall performance. More specifically, a G-Agent that employs a less powerful LLM as its core can drastically undermine performance. D-Agent assumes a more pivotal role during the linking phase compared to the query construction phase. A-Agent, on the other hand, proves to be an efficient solution for managing situations where SPARQL results are absent.
+
+<table><tr><td>G-task</td><td>G-chat</td><td>LC-QuAD 1.0</td><td>QALD-9</td></tr><tr><td>X</td><td>X</td><td>0.343</td><td>0.159</td></tr><tr><td>✓</td><td>X</td><td>0.443</td><td>0.248</td></tr><tr><td>✓</td><td>✓</td><td>0.564</td><td>0.416</td></tr><tr><td>D-uri</td><td>D-query</td><td>LC-QuAD 1.0</td><td>QALD-9</td></tr><tr><td>X</td><td>✓</td><td>0.274</td><td>0.210</td></tr><tr><td>✓</td><td>X</td><td>0.431</td><td>0.301</td></tr><tr><td>✓</td><td>✓</td><td>0.564</td><td>0.416</td></tr><tr><td>A-llm</td><td>A-fact</td><td>LC-QuAD 1.0</td><td>QALD-9</td></tr><tr><td>X</td><td>X</td><td>0.459</td><td>0.382</td></tr><tr><td>✓</td><td>X</td><td>0.473</td><td>0.385</td></tr><tr><td>✓</td><td>✓</td><td>0.564</td><td>0.416</td></tr></table>
+
+# 4.4 Analysis of Role Hyperparameters
+
+We concentrate on three hyperparameters of roles, including the number of examples  $(\mathcal{N}\in$ $\{1,2,3\})$  provided for G-Agent to learn subtasks, the number of URI candidates  $(\mathcal{K}\in$ $\{(1,1),(1,2),(2,2),(2,3)\})$  selected by D-Agent for query construction, and the number of retry times  $(\mathcal{T}\in \{1,2,3\})$  launched by A-Agent when there is no response. Table 4 presents the F1 results of Triad's performance, employing three hyperparameters on two benchmarks. We discover that:
+
+Quality is more important than quantity. More examples provided to G-Agent do not always improve the performance. The efficacy of G-Agent is significantly influenced by the quality of examples.
+
+More options may harm the result. Choosing more candidate URIs for entities and relations could potentially disrupt subsequent query phases, thus affecting overall performance.
+
+Table 3: Study on the roles of LLM-based agent by eliminating an element or downgrading the capability.  
+
+<table><tr><td>Triad Variants</td><td>LC-QuAD 1.0</td><td>QALD-9</td></tr><tr><td>Triad-1-Shot</td><td>0.556</td><td>0.376</td></tr><tr><td>Triad-2-Shot</td><td>0.511</td><td>0.402</td></tr><tr><td>Triad-3-Shot</td><td>0.564</td><td>0.416</td></tr><tr><td>Triad-Top1-1</td><td>0.528</td><td>0.281</td></tr><tr><td>Triad-Top1-2</td><td>0.562</td><td>0.375</td></tr><tr><td>Triad-Top2-2</td><td>0.564</td><td>0.416</td></tr><tr><td>Triad-Top2-3</td><td>0.558</td><td>0.384</td></tr><tr><td>Triad-1-Try</td><td>0.529</td><td>0.375</td></tr><tr><td>Triad-2-Tries</td><td>0.561</td><td>0.407</td></tr><tr><td>Triad-3-Tries</td><td>0.564</td><td>0.416</td></tr></table>
+
+Table 4: Performance evaluation on three hyperparameters that related to each role of an LLM-based agent.
+
+More chances benefits the framework. Persistently attempting to construct and execute SPARQL queries is an effective strategy that improves the probability of obtaining accurate answers. Considering the efficiency of overall execution, we set the maximum retry times as 3 in practice.
+
+# 4.5 Analysis of Linking Recall
+
+The process of linking is a relatively complex subtask in both the Text2SQL and the KBQA process (Li et al., 2024). Calculating the recall ratio of accurate URIs using D-Agent provides clarity on which step most adversely impacts performance. In the entity linking phase, considering all URIs of entities in the testing set as the ground truth of the linking results,  $80.75\%$  of the correct URIs are contained from the output of the entity matching filter in D-Agent and  $70.50\%$  of the correct URIs are retained from the entity selection performed by the LLM in D-Agent. Whereas, in the relation linking phase, only  $52.54\%$  of the correct relation URIs survive from the selection of LLM, which indicates a greater difficulty in relation linking.
+
+# 4.6 Study on Complex Cases
+
+Despite the impressive performance of Triad in certain benchmarks, notable deficiencies remain in its ability to understand questions and generate queries for complex questions. A critical analysis of unsuccessful cases in QALD-9, which has the lowest F1 score, has revealed three primary reasons for this failure, as detailed below:
+
+Complex Syntax signifies that advanced SPARQL queries incorporate keywords such as GROUP BY and HAVING. These terms augment the error propensity in the generation of SPARQL
+
+<table><tr><td>Fail Reason</td><td>Ratio</td><td>Example</td></tr><tr><td>Complex Syntax</td><td>20%</td><td>Q42: Which countries have places with more than two caves?</td></tr><tr><td>Unexploited Semantics</td><td>17%</td><td>Q199: Give me all Argentine films.</td></tr><tr><td>Implicit Reasoning</td><td>5%</td><td>Q133: What are the names of the Teenage Mutant Ninja Turtles?</td></tr></table>
+
+Table 5: The major reasons of complexity that result in failures, with their corresponding ratios of occurrence in failed cases.
+
+templates such as the example: Which frequent flyer program has the most airlines?
+
+Unexploited Semantics indicates that semantics of an implicit entity should be comprehended in order to exclude irrelevant URIs. In the example Give me all Argentine films, the meaning of films should be used to narrow down the scope of potential entities in order to eliminate unrelated answers.
+
+Implicit Reasoning presents a challenge that requires a deeper level of traversal by the framework to deduce accurate results from the posed question. For example, another failure question, How many grand-children did Jacques Cousteau have?, the term grand-children must be interpreted to son of son to ensure an accurate response.
+
+# 4.7 Cost Comparison and Analysis
+
+According to our evaluation on the three datasets, the average cost of running a single case is 0.007 USD on average using Triad-GPT3.5 and 0.05 USD on average using Triad-GPT4. Specifically, most API calls occur in the phases of URL linking and comprehensive answering. Meanwhile, traditional KBQA baselines require a lot of training data and local training resources to achieve the SOTA performance, whereas Triad follows a zero- or few-shot manner to save computational cost locally. Furthermore, as shown in Section 4.4, in practice, adjusting the hyperparameters can make the cost as low as possible while preserving overall performance. As the cost of LLM services decreases, the value of Triad will increase accordingly.
+
+# 5 Related Work
+
+# 5.1 SPARQL-based and LLM-based KBQA
+
+Traditional KBQA methods transform natural language queries into SPARQL requests for data extraction. Specific models are employed either for question understanding or URI linking, utilizing domain-based training datasets. Hu et al. (2018) introduces a semantic query graph to structurally represent the natural language query, thereby simplifying the task into a subgraph matching problem. Hu et al. (2021) proposes an entity description graph to represent natural language queries for question parsing and element linking. Omar et al. (2023a) restructures the question parsing task as a text generation issue using a sequence-to-sequence model. With the advent of LLMs, certain phases of KBQA can be enhanced with LLM-integrated methods. Baek et al. (2023) aims to augment LLM-based QA tasks with pertinent facts extracted from knowledge bases, offering a fully zero-shot architecture. Tan et al. (2023a) leverages the general applicability of LLMs to filter linking candidates by making selections via few-shot in-context learning. Omar et al. (2023b) provides a thorough comparison between LLMs and QA systems, recommending further studies to improve KBQA with LLM capabilities. However, apart from the above studies, our study proposes a complete framework incorporating both an LLM and few-shot learning across all KBQA phases from a systematic perspective.
+
+# 5.2 LLM-based Agents for Complex Tasks
+
+LLMs have recently gained significant attention due to their ability to approximate human-level intelligence. This has led to numerous studies focusing on LLM-based agents. A recent survey (Wang et al., 2023) proposes a unified architecture for LLM-based agents, which consists of four modules that include profile, memory, plan, and action. CHATDB(Hu et al., 2023a) employs an LLM controller to generate SQL instructions, which allows for symbolic memory and complex multi-hop reasoning. ART(Paranjape et al., 2023) uses a frozen LLM to generate reasoning steps and further integrates tools for new tasks with minimal human intervention. Toolformer(Schick et al., 2024) takes a different approach by training an LLM to plan and execute tools for the next token prediction by learning API calls generation. ReAct(Yao et al., 2023) focuses on overcoming LLM hallucination by interacting with external knowledge bases,
+
+thus generating interpretable task-solving strategies. CodeAgent(Tang et al., 2024) designs a multiagent collaboration system across four phases in a code review process. Divergent from the aforementioned studies, our framework concentrates on the solving KBQA tasks by introducing a multirole LLM-based agent that specializes in various subtasks distributed across different phases.
+
+# 6 Conclusion
+
+In this study, we aim to bridge the gap between KBQA tasks and the investigation of LLM-based agents. We introduce Triad, a framework to address the KBQA task through an LLM-based agent acting as multiple roles, including a generalist capable of mastering diverse tasks given minimal examples, a decision-maker concentrating on option analysis and candidate selection, and an advisor skilled in answering questions with the aid of both external and internal knowledge. Triad achieves the best or competitive performance across three benchmark datasets compared to traditional KBQA systems and pure LLM models. In future research, we plan to broaden our framework to handle more intricate questions, such as multi-hop reasoning, and exploring the integration between our framework and retrieval-augmented generation.
+
+# Limitations
+
+The limitation of our research lies in following aspects: (1) In terms of data, a broader range of QA datasets needs to be evaluated, encompassing datasets from different domains, languages, and difficulty levels. (2) In terms of model, more LLMs need to be evaluated, including open-source and commercial models from different organizations and on various scales. (3) In terms of framework, more types of agent collaboration methods can be explored to solve KBQA problems.
+
+# Ethics Considerations
+
+All datasets utilized in this study are publicly available and we have adhered to ethical considerations by not introducing additional information as input during LLM text generation.
+
+# Acknowledgements
+
+This work is supported by the "Pioneer" and "Leading Goose" R&D Program of Zhejiang (Grant No. 2023C01152 and No. 2024C01034).
+
+# References
+
+Sören Auer, Christian Bizer, Georgi Kobilarov, Jens Lehmann, Richard Cyganiak, and Zachary Ives. 2007. Dbpedia: A nucleus for a web of open data. In The Semantic Web, pages 722-735, Berlin, Heidelberg. Springer Berlin Heidelberg.  
+Jinheon Baek, AlhamFikri Aji, and Amir Saffari. 2023. Knowledge-augmented language model prompting for zero-shot knowledge graph question answering. arXiv preprint arXiv:2306.04136.  
+Qingxiu Dong, Li Dong, Ke Xu, Guangyan Zhou, Yaru Hao, Zhifang Sui, and Furu Wei. 2023. Large language model for science: A study on p vs. np. arXiv preprint arXiv:2309.05689.  
+Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Zhiyong Wu, Baobao Chang, Xu Sun, Jingjing Xu, and Zhifang Sui. 2022. A survey for in-context learning. arXiv preprint arXiv:2301.00234.  
+Chenxu Hu, Jie Fu, Chenzhuang Du, Simian Luo, Junbo Zhao, and Hang Zhao. 2023a. Chatdb: Augmenting llms with databases as their symbolic memory. arXiv preprint arXiv:2306.03901.  
+Nan Hu, Yike Wu, Guilin Qi, Dehai Min, Jiaoyan Chen, Jeff Z Pan, and Zafar Ali. 2023b. An empirical study of pre-trained language models in simple knowledge graph question answering. _World Wide Web_, pages 1-32.  
+Sen Hu, Lei Zou, Jeffrey Xu Yu, Haixun Wang, and Dongyan Zhao. 2018. Answering natural language questions by subgraph matching over knowledge graphs. IEEE Transactions on Knowledge and Data Engineering, page 824-837.  
+Xixin Hu, Yiheng Shu, Xiang Huang, and Yuzhong Qu. 2021. Edg-based question decomposition for complex question answering over knowledge bases. In The Semantic Web – ISWC 2021: 20th International Semantic Web Conference, ISWC 2021, Virtual Event, October 24–28, 2021, Proceedings, page 128–145, Berlin, Heidelberg. Springer-Verlag.  
+Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199-22213.  
+Jinyang Li, Binyuan Hui, Reynold Cheng, Bowen Qin, Chenhao Ma, Nan Huo, Fei Huang, Wenyu Du, Luo Si, and Yongbin Li. 2023. Graphix-t5: Mixing pre-trained transformers with graph-aware layers for text-to-sql parsing. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 13076-13084.  
+Jinyang Li, Binyuan Hui, Ge Qu, Jiaxi Yang, Binhua Li, Bowen Li, Bailin Wang, Bowen Qin, Ruiying Geng, Nan Huo, et al. 2024. Can llm already serve as a database interface? a big bench for large-scale
+
+database grounded text-to-sqls. Advances in Neural Information Processing Systems, 36.  
+Zhiwei Liu, Weiran Yao, Jianguo Zhang, Le Xue, Shelby Heinecke, Rithesh Murthy, Yihao Feng, Zeyuan Chen, Juan Carlos Niebles, Devansh Arpit, et al. 2023. Bolaa: Benchmarking and orchestrating llm-augmented autonomous agents. arXiv preprint arXiv:2308.05960.  
+Reham Omar, Ishika Dhall, Panos Kalnis, and Essam Mansour. 2023a. A universal question-answering platform for knowledge graphs. Proceedings of the ACM on Management of Data, 1(1):1-25.  
+Reham Omar, Omij Mangukiya, Panos Kalnis, and Essam Mansour. 2023b. Chatgpt versus traditional question answering for knowledge graphs: Current status and future directions towards knowledge graph chatbots. arXiv preprint arXiv:2302.06466.  
+Bhargavi Paranjape, Scott Lundberg, Sameer Singh, Hannaneh Hajishirzi, Luke Zettlemoyer, and MarcoTulio Ribeiro. 2023. Art: Automatic multi-step reasoning and tool-use for large language models. arXiv preprint arXiv:2303.09014.  
+Thomas Pellissier Tanon, Gerhard Weikum, and Fabian Suchanek. 2020. Yago 4: A reason-able knowledge base. In The Semantic Web, pages 583–596, Cham. Springer International Publishing.  
+Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettle-moyer, Nicola Cancedda, and Thomas Scialom. 2024. Toolformer: Language models can teach themselves to use tools. Advances in Neural Information Processing Systems, 36.  
+Chuanyuan Tan, Yuehe Chen, Wenbiao Shao, Wenliang Chen, Zhefeng Wang, Baoxing Huai, and Min Zhang. 2023a. Make a choice! knowledge base question answering with in-context learning. arXiv preprint arXiv:2305.13972.  
+Yiming Tan, Dehai Min, Yu Li, Wenbo Li, Nan Hu, Yongrui Chen, and Guilin Qi. 2023b. Can chatgpt replace traditional kbqa models? an in-depth analysis of the question answering performance of the gpt llm family. In International Semantic Web Conference, pages 348-367. Springer.  
+Yiming Tan, Dehai Min, Yu Li, Wenbo Li, Nan Hu, Yongrui Chen, and Guilin Qi. 2023c. Evaluation of chatgpt as a question answering system for answering complex questions. arXiv preprint arXiv:2303.07992.  
+Daniel Tang, Zhenghan Chen, Kisub Kim, Yewei Song, Haoye Tian, Saad Ezzini, Yongfeng Huang, and Jacques Klein Tegawende F Bissyande. 2024. Collaborative agents for software engineering. arXiv preprint arXiv:2402.02172.
+
+Priyansh Trivedi, Gaurav Maheshwari, Mohnish Dubey, and Jens Lehmann. 2017. Lc-quad: A corpus for complex question answering over knowledge graphs. In The Semantic Web – ISWC 2017, pages 210–218, Cham. Springer International Publishing.  
+Ricardo Usbeck, Ria Hari Gusmita, Axel-Cyrille Ngonga Ngomo, and Muhammad Saleem. 2018. 9th challenge on question answering over linked data (qald-9) (invited paper). In Semdeep/NLIWoD@ISWC.  
+Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, et al. 2023. A survey on large language model based autonomous agents. arXiv preprint arXiv:2308.11432.  
+Tonghan Wang, Tarun Gupta, Anuj Mahajan, Bei Peng, Shimon Whiteson, and Chongjie Zhang. 2020. Rode: Learning roles to decompose multi-agent tasks. arXiv preprint arXiv:2010.01523.  
+Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023. React: Synergizing reasoning and acting in language models.
+
+# A Response Time Analysis
+
+We analyze various QA frameworks in response time to a question. The average latency of each phase including question parsing (QP), URI linking (UL), and answer generation (AG) for each knowledge base is reported. We randomly select 10 samples from each dataset for evaluation to obtain the average response times for Triad-1 and Triad-3, which represent retrying three times and generating an answer in one go, respectively, during the answer generation phase. The comparison between traditional QA systems and Triad is shown in Figure 3. Triad generally shows a competitive time-consuming performance to latest traditional QA systems. Specifically, compared to other phases, URL linking consumes more time due to the need to invoke LLM multiple times. Moreover, according to Section 4.4, with smaller retry times of A-Agent, Triad can significantly reduce time cost while only causing slight performance degradation, revealing the advantages of our framework in balancing performance and efficiency.
+
+![](images/f49b3177223a6632d6f88a99553a955e45e71fb82605a396c8fdba2b7fe01725.jpg)  
+Figure 3: Response time of traditional KBQA systems and Triad on three datasets. Each bar shows average response time of a particular phase of KBQA.
+
+# B Role Performance on YAGO-QA
+
+We choose LC-QuAD 1.0 and QALD-9 as our two representative datasets in Section 4.3, as the questions among them vary in difficulty, and the tasks in these two datasets are relatively more challenging than YAGO-QA. We provide the performance of agent roles on YAGO-QA in Table 6, which shows a consistent result with other datasets in Table 3.
+
+# C Prompts Provided to LLMs of G-Agent for Solving Various Subtasks in KBQA
+
+The prompt given to LLMs of  $Agent_{g}$  to perform triplet extraction from the question  $Q$  is as follows:
+
+<table><tr><td>G-task</td><td>G-chat</td><td>YAGO-QA</td></tr><tr><td>×</td><td>×</td><td>0.427</td></tr><tr><td>✓</td><td>×</td><td>0.553</td></tr><tr><td>✓</td><td>✓</td><td>0.677</td></tr><tr><td>D-uri</td><td>D-query</td><td>YAGO-QA</td></tr><tr><td>×</td><td>✓</td><td>0.346</td></tr><tr><td>✓</td><td>×</td><td>0.534</td></tr><tr><td>✓</td><td>✓</td><td>0.677</td></tr><tr><td>A-llm</td><td>A-fact</td><td>YAGO-QA</td></tr><tr><td>×</td><td>×</td><td>0.626</td></tr><tr><td>✓</td><td>×</td><td>0.647</td></tr><tr><td>✓</td><td>✓</td><td>0.677</td></tr></table>
+
+Table 6: Performance of roles of LLM-based agent by eliminating an element or downgrading the capability.
+
+# You are an assistant to identify triples within a provided sentence. Please adhere to the following guidelines:
+
+1. Triples should be structured in the format <entity1, relation, entity2>.  
+2. The sentence must contain at least one triple, so you should provide at least one.  
+3. Entities should represent the smallest semantic units and should not contain descriptive details.  
+4. Entities can take the form of explicit or implicit references. Explicit entities refer to specific named resources, whereas implicit entities are less certain.  
+5. When an entity is implicit, utilize a variable format such as '?variable' to denote it, for example, '?location' or '?person'.
+
+# Here are some examples:
+
+Which city's founder is John Forbes? : <?city, foundeer, John Forbes>
+
+How many races have the horses bred by Jacques Van't Hart participated in? : <?horse, participated in, ?race> <?horse, breeder, Jacques Van't Hart>
+
+Is camel of the chordate phylum? : <camel, phylum, chordate>
+
+Sentence: <Question Sentence>
+
+# Output:
+
+The prompt given to LLMs of  $A g e n t_{g}$  for SPARQL template generation is as follows:
+
+You are an assistant to generate a SPARQL query to address a specific question. Here are the guidelines to follow:
+
+1. Ensure that the resulting SPARQL query is designed to answer the provided question.  
+2. Adhere to the commonly accepted SPARQL standards when generating the query.  
+3. Make an effort to leverage the information provided to assist in the creation of the SPARQL query.  
+4. Strive to keep the generated SPARQL query as straightforward as possible.  
+5. Avoid including 'PREFIX' or ':' in the SPARQL query.  
+6. Enclose condition entities and predicates within angle brackets, such as <entity> or <predicate>.  
+7. Maintain the original order of the given triples without altering their sequence.
+
+Question: <question sentence>
+
+Triplets: <extracted triplets>
+
+Output:
+
+You are an assistant to select  $\langle K \rangle$  URIs from a provided list of possible URIs for a specified entity, following these guidelines:
+
+1. Identify the  $\langle \mathrm{K} \rangle$  most appropriate URIs from the given list that best represent the entity in question.  
+2. Seek to understand the semantic information associated with the specified entity by examining the provided question.  
+3. The output should consist of  $\langle \mathrm{K} \rangle$  URIs chosen from the provided list of possible URIs.  
+4. Simply output these  $\langle \mathrm{K} \rangle$  target URIs, each on a separate line, without providing any additional explanations.
+
+Sentence: <question sentence>
+
+Entity: <entity mention>
+
+Possible entity URIs: <Entity URI list>
+
+Output:
+
+The prompt given to LLMs of  $Agent_d$  for candidate relation selection is as follows:
+
+The prompt given to LLMs of  $Agent_g$  for question type classification is as follows:
+
+You are an assistant to determine the specific type of a given question according to the following guidelines:
+
+1. You must determine the most probable question type for the input question.  
+2. The type of question should be enclosed within angle brackets, denoted as  $\text{’} < \text{’}$  and  $\text{’} > \text{’}$ .  
+3. Possible question types include: <count>, <select>, and <yes or no>.
+
+Question: <question sentence>
+
+Output:
+
+You are an assistant tasked with selecting the  $\langle K\rangle$  relation URIs between entities mentioned in a sentence. Here are the guidelines:
+
+1. The two entities are listed one after the other, without a specific order.  
+2. Use the provided sentence to discern the semantic meaning of these entities.  
+3. The potential relation URIs are listed one by one.  
+4. Your output should consist of a maximum of  $\langle \mathrm{K} \rangle$  possible relation URIs, although you may also output fewer if appropriate.  
+5. Ensure that your output is organized, prioritizing the most likely relationship first.  
+6. Provide a list of no more than  $\langle \mathrm{K} \rangle$  relation URIs (each on a separate line if there are multiple) without any additional descriptions.
+
+Sentence: <question sentence>
+
+Entities: <entity pair>
+
+Possible relation URIs: <URI list>
+
+Output:
+
+D Prompts Provided to LLMs of D-Agent for Solving Selection Subtasks in KBQA
+
+The prompt given to LLMs of  $Agent_d$  for candidate entities selection is as follows:
+
+The prompt given to LLMs of  $Agent_d$  for the final SPARQL selection is as follows:
+
+You are an assistant to select an appropriate SPARQL query from the provided list in order to respond to a specific question. Please adhere to the following guidelines:
+
+1. Select the most suitable SPARQL query from the given query list to address the question.  
+2. Select a SPARQL query solely from the provided list; avoid crafting your own SPARQL query.  
+3. The selected SPARQL query must be applicable to answer the given question.
+
+Sentence: <question sentence>
+
+SPARQL candidates: <SPARQLs to choose> Output:
+
+# E Prompts Provided to LLMs of A-Agent for Solving Answering Subtask in KBQA
+
+The prompt given to LLMs of  $Agent_{a}$  to generate a yes or no answer for the give question is as follows:
+
+You are an assistant to answer a yes-or-no question. Please adhere to the following guidelines:
+
+1. If you believe that the answer is yes, provide an output of 'True'. If not, provide an output of 'False'.  
+2. Please do not include additional information or explanations in your response.
+
+Sentence: <question sentence>
+
+# Output:
+
+The prompt given to LLMs of  $Agent_{a}$  to generate a single-fact answer for the give question is as follows:
+
+You are an assistant to answer a question. Please adhere to the following guidelines:
+
+1. The answer to the question is a single entity.  
+2. You should just output the full expression of the answer without any punctuation.  
+3. Do not output any other description.
+
+Sentence: <question sentence>
+
+# Output:

2024/Triad_ A Framework Leveraging a Multi-Role LLM-based Agent to Solve Knowledge Base Question Answering/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e2aa5ef7ac6fe968f60cfc5b76a1bcb549a0ee3e64aa53e973d8c2d3097d6f00
+size 551387

2024/TroL_ Traversal of Layers for Large Language and Vision Models/d3959391-dc37-4afa-81df-706ce45ce8e9_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2f56736f7d011a91ae4844282af853ce9a7e0d68c2b992741c592a6d729652fd
+size 6466745

2024/TroL_ Traversal of Layers for Large Language and Vision Models/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:797e620160e3ef2335aea452f7e269143688991275aca1908c7833c5e8158615
+size 1613573

2024/Turn Waste into Worth_ Rectifying Top-k Router of MoE/78b71285-cba8-4c67-9072-33097abe467c_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5b9eaf7b0d3c839c8cd0d340c8e00c86e0dce5a6c4a13b85b6df1e369fdcdc0b
+size 553993

2024/Turn Waste into Worth_ Rectifying Top-k Router of MoE/full.md

@@ -0,0 +1,367 @@
+# Turn Waste into Worth: Rectifying Top- $k$  Router of MoE
+
+Zhiyuan Zeng $^{1}$ , Qipeng Guo $^{2}$ , Zhaoye Fei $^{1,2}$ , Zhangyue Yin $^{1}$ , Yunhua Zhou $^{2}$ , Linyang Li $^{2}$ , Tianxiang Sun $^{1}$ , Hang Yan $^{2}$ , Dahua Lin $^{2}$ , Xipeng Qiu $^{1\dagger}$
+
+$^{1}$ School of Computer Science, Fudan University, Shanghai, China
+
+$^{2}$ Shanghai AI Laboratory
+
+cengzy23@m.fudan.edu.cn; xpqiu@fudan.edu.cn
+
+# Abstract
+
+Sparse Mixture of Experts (MoE) models are popular for training large language models due to their computational efficiency. However, the commonly used top- $k$  routing mechanism suffers from redundancy computation and memory costs due to the unbalanced routing. Some experts are overflow, where the exceeding tokens are dropped. While some experts are empty, which are padded with zeros, negatively impacting model performance. To address the dropped tokens and padding, we propose the Rectify-Router, comprising the Intra-GPU Rectification and the Fill-in Rectification. The Intra-GPU Rectification handles dropped tokens, efficiently routing them to experts within the GPU where they are located to avoid inter-GPU communication. The Fill-in Rectification addresses padding by replacing padding tokens with the tokens that have high routing scores. Our experimental results demonstrate that the Intra-GPU Rectification and the Fill-in Rectification effectively handle dropped tokens and padding, respectively. Furthermore, the combination of them achieves superior performance, surpassing the accuracy of the vanilla top-1 router by  $4.7\%$ .
+
+# 1 Introduction
+
+Sparse Mixture of Experts (MoE) is gaining popularity as a model architecture for training large language models (Fedus et al., 2022; Du et al., 2022; Zoph et al., 2022; Jiang et al., 2024; Dai et al., 2024) owing to its computational efficiency. In a sparse MoE model, each token is assigned to one or more experts based on a routing mechanism. The top- $k$  router is currently the most widely used routing mechanism, where tokens are directed to the experts with the top- $k$  scores.
+
+However, top-  $k$  router is unbalanced, where the number of tokens routed to different GPUs is not
+
+![](images/afb77ad1bb95cfb6d9ff4204dc27cb19b8c39e33145713f4e6a3eef99d81945b.jpg)  
+Figure 1: The illustration of dropped token and padding in top- $k$  router of MoE. Queue  $i$  represents the queue of tokens to be sent to expert  $i$ . The capacity of each expert is fixed to 3.
+
+the same. In order to achieve a balanced workload across GPUs, the top- $k$  routing imposes a maximum limit on the number of tokens that each expert can process, resulting in any tokens that exceed this limit being dropped and empty experts being filled with zeros, which negatively affects overall model performance (Gale et al., 2022).
+
+Previous studies have attempted to address the balance issue in routing by introducing auxiliary loss mechanisms (Shazeer et al., 2017; Lepikhin et al., 2021; Zoph et al., 2022). But there are drawbacks to the way, the performance degradation due to dropped tokens is still significant (Zhou et al., 2022; Gale et al., 2022). Even some methods have made improvements to propose absolutely balanced routers, but they have been found to underperform the original top- $k$  routing methodology (Yu et al., 2022).
+
+Rather than focusing on improving the balance of the top-  $k$  router, We propose an alternative approach called the Rectify-Router, which rectifies top-  $k$  router by post-processing the dropped tokens and padding from the top-  $k$  router. We propose two
+
+Rectify-Routers: the Intra-GPU Rectification and the Fill-in Rectification. The Intra-GPU Rectification is designed to handle the dropped tokens, while the Fill-in Rectification specifically addresses the padding issue.
+
+Post-processing the dropped tokens with another router may bring expensive communication cost. Therefore, we propose the Intra-GPU Rectification which routes the dropped tokens to the experts within the GPU where they are located, eliminating the need for inter-GPU communication. Our empirical experiments have demonstrated that the Intra-GPU Rectification effectively handles the post-processing of dropped tokens and is more efficient than the commonly used routers, in terms of communication.
+
+To address the padding issue, we present the Fill-in Rectification, which replace padding tokens with the tokens that have high routing scores. Fill-in Rectification first identifies the optimal expert for each token based on the routing scores and subsequently selects the tokens with the highest routing score to replace the padding for each expert. By employing Fill-in Rectification, tokens with the higher routing scores receive more computational allocation.
+
+The Intra-GPU Rectification and Fill-in Rectification are orthogonal approaches that can be seamlessly combined. Our experiments have demonstrated their effectiveness in handling dropped tokens and padding. Furthermore, combing the Intra-GPU Rectification and Fill-in Rectification yield improved performance compared to using them individually.
+
+Contributions The contributions of our work can be summarized as follows:
+
+1. We introduce the concept of Rectify-Router to handle the dropped tokens and padding in MoE models. Specifically, the dropped tokens are efficiently processed using the Intra-GPU Rectification, while the padding tokens are optimally managed using the Fill-in Rectification.  
+2. Our experiments validate that both the Intra-GPU Rectification and the Fill-in Rectification significantly improve the performance of the top- $k$  routing, even without additional training.  
+3. Experiments present that our methods are robust to various settings of expert capacity and
+
+that Intra-GPU Rectification can be used for accelerating MoE by reducing expert capacities.
+
+# 2 Related Works
+
+The routing of MoE can be classified into two categories: balanced and unbalanced. The balanced routing assigns the same number of tokens to each expert, while the unbalanced routing does not make sure that the number of tokens received by each expert is the same.
+
+Unbalanced Routing Top- $k$  routing was the most commonly used unbalanced routing proposed by Shazeer et al. (2017), which greedily assigns tokens to experts, according to the token-expert assignment scores. Numerous MoE models have adopted top- $k$  routing, including Switch Transformer (Fedus et al., 2022), Glam (Du et al., 2022), ST-MoE (Zoph et al., 2022), Flan-MoE (Shen et al., 2023), and NLLB (Koishekenov et al., 2022), to name just a few.
+
+It is worth noting that many unbalanced routing methods are variations or derivatives of top- $k$  routing. For example, Switch Transformer (Fedus et al., 2022) argues in favor of using top-1 routing instead of top-2 routing for improved efficiency. ST-MoE (Zoph et al., 2022) and LIMoE (Mustafa et al., 2022) propose auxiliary loss functions to enhance the stability of MoE during training. Additionally, SCoMoE (Zeng and Xiong, 2023) and Gating-Dropout (Liu et al., 2022) improve the efficiency of top- $k$  routing by designing hierarchical routing systems based on the hierarchical structure of the communication topology.
+
+The routing method proposed in this paper is also a variation of top- $k$  routing. However, unlike the aforementioned approaches, our objective is to address the issues of dropped tokens and padding that arise from unbalanced routing. Switch Transformer (Fedus et al., 2022) tackles the problem of dropped tokens by increasing the capacity of experts, allowing each expert to handle more tokens. While this approach reduces the number of dropped tokens, it introduces additional overhead in terms of both speed and memory. On the other hand, Megablocks (Gale et al., 2022) addresses the challenges of padding and dropped tokens by gathering all experts onto the same GPU and employing model parallelism rather than expert parallelism. However, the model parallelism is shown to be more expensive than the expert parallelism
+
+![](images/9dfee9ab902c08feffc6458ffaa37bebfe3fe49cec27c34542e2ace9dd911725.jpg)  
+Figure 2: Left: Post-processing of dropped tokens at GPU 0 with Intra-GPU Rectification. Right: Post-processing of padding at GPU 0 with Fill-in Rectification.
+
+by Tutel (Hwang et al., 2022).
+
+Balanced Routing In response to the imbalance issue inherent in top- $k$  routing, several balanced routing methods have been proposed. For instance, the Base Layer approach (Lewis et al., 2021) employs a balanced assignment algorithm to evenly distribute tokens among experts. However, their assumption that tokens within the same batch can be evenly clustered may not hold true in all cases, which can potentially result in poorer performance (Yu et al., 2022). Another alternative to balanced routing is random routing (Zuo et al., 2022), which assigns tokens to experts in a random manner. While random routing achieves balance and efficiency, it lacks any specialization or optimization in the routing process. Another approach called expert choices (Zhou et al., 2022) allows each expert to select a fixed number of tokens, rather than relying on tokens to determine their target experts. This approach helps to avoid padding issues but still results in dropped tokens. Soft routing (Puigcerver et al., 2023) is a method that compresses tokens by applying a linear transformation to generate fixed-size hidden states for each expert. However, this method is only suitable for encoder models with fixed input lengths and may not be applicable to autoregressive decoder models.
+
+# 3 Preliminary
+
+In this section, we will introduce expert parallelism, top-  $k$  routing, and two prevalent challenges that emerge while employing top-  $k$  routing: padding and dropped tokens.
+
+Expert Parallelism and Top- $k$  Routing In expert parallelism, experts are distributed across GPUs uniformly. If there are  $n$  experts and  $k$  GPUs, each GPU contains  $k/n$  experts. The process of transmitting tokens to the respective experts entails inter-GPU communication.
+
+Top-  $k$  routing greedily assigns tokens to experts according to the routing score:
+
+$$
+\mathbb {R} _ {i} = \operatorname {a r g t o p k} _ {j \in [ m ]} \left\{a _ {i j} \mid a _ {i j} = w _ {j} ^ {T} x _ {i} \right\} \tag {1}
+$$
+
+where  $a_{ij}$  is the score of assigning the  $i$ th token to the  $j$ th expert,  $w_{j}$  denotes the embedding vector of the  $j$ th expert,  $x_{i}$  corresponds to the hidden states of the  $i$  token. The index set  $\mathbb{R}_i$  signifies the target experts of the  $i$ th token. Given the scores of assigning token  $x_{i}$  to  $m$  experts, denoted as  $a_{i0}, a_{i1}, \ldots, a_{im}, \mathbb{R}_i$  contains the indices of experts with top- $k$  scores.
+
+Since each token undergoes processing by multiple experts, the outputs of these experts for the same token are consolidated through linear combination. The combining weights are determined by
+
+the normalized routing scores, as defined in Eq. (1):
+
+$$
+o _ {i} = \sum_ {j \in \mathbb {R} _ {i}} \frac {e ^ {a _ {i j}}}{\sum_ {j \in \mathbb {R} _ {i}} e ^ {a _ {i j}}} E _ {j} \left(x _ {i}\right). \tag {2}
+$$
+
+Here,  $o_i$  represents the combined result of token  $x_i$ . The term  $\frac{e^{a_{ij}}}{\sum_j^k e^{a_{ij}}}$  denotes the normalized routing scores, while  $E_j(x_i)$  refers to the outputs of the  $j$ th expert with token  $x_i$  as its input.
+
+The top- $k$  routing approach exhibits an inherent imbalance, wherein the distribution of tokens among different experts is not uniform. However, the current distributed framework exclusively supports balanced computation across GPUs. Consequently, there exists a limitation on the maximum number of tokens that each expert can receive, which is referred to as the capacity. The capacity is determined by the capacity factor, which is typically set to  $k$  for top- $k$  routing (Lepikhin et al., 2021; Rajbhandari et al., 2022). Mathematically, the capacity can be expressed as:
+
+capacity = capacity factor ×  $\frac{\text{number of tokens}}{\text{number of experts}}$ .
+
+Dropped Tokens and Padding The issue of dropped tokens and padding arises naturally when dealing with the expert capacity setting, as depicted in Figure 1. With a fixed expert capacity, overflow experts are compelled to drop tokens with the lowest routing scores and directly pass them to the next layer through residual connections, as highlighted in red in Figure 1. Consequently, due to the dropped tokens, the set  $\mathbb{R}_i$  defined in Eq. (1) only includes the successfully routed experts, i.e.,  $|\mathbb{R}_i| \leq k$ .
+
+Conversely, certain experts may receive fewer tokens than the capacity limitation, leading to redundant computation in the form of padding. These padding instances are illustrated in yellow in Figure 1.
+
+If the capacity factor for top-  $k$  routing is set to  $k$ , the number of dropped tokens and padding tokens will be equal. However, this equality does not hold if we modify the capacity factor. Increasing the capacity factor results in fewer dropped tokens but more padding. Conversely, reducing the capacity factor reduces padding tokens but increases the number of dropped tokens.
+
+# 4 Method
+
+In this paper, we introduce a novel approach to address both the dropped tokens and padding as
+
+sociated with top- $k$  routing by utilizing Rectify-Routers. Specifically, we propose two Rectify-Routers: the Intra-GPU Rectification and the Fill-in Rectification, which are visualized in Figure 2. The Intra-GPU Rectification is designed to efficiently post-process the dropped tokens, while the Fill-in Rectification is dedicated to addressing the padding problem.
+
+# 4.1 Rectify-Router for Dropped Tokens: Intra-GPU Rectification
+
+We expect to post-process the dropped tokens by evenly routing them across GPUs. But sending tokens among GPUs requires expensive communication cost. Furthermore, the dropped tokens have the lower routing scores than the other tokens routed to the same expert, which may be less important. Therefore, we propose an efficient Rectify-Router for the dropped tokens: Intra-GPU Rectification, which dispatch the dropped tokens to the experts inside GPU, which does not require any communication among GPUs. This process is visualized in the left part of Figure 2, where the dropped tokens from GPU 0 are routed to the expert 0 or expert 1 at GPU 0.
+
+Given the input token  $x_{i}$ , the Intra-GPU Rectification greedily assigns token  $x_{i}$  to the optimal expert within the same GPU according to the routing scores. The Intra-GPU Rectification can be seen as a variant of the top-  $k$  routing. If all experts are distributed in the same GPU, then the Intra-GPU Rectification is exactly the top-1 routing.
+
+In top- $k$  routing, the same token may be dropped by multiple times. Take the top-2 routing as an example, if a token  $x_{i}$  is dropped at both the first and second routing, it should be sent to two experts at Intra-GPU Rectification. To simplify the problem, we only send  $x_{i}$  to one expert, although it is dropped twice. In another example, the token  $x_{i}$  is dropped only at the second routing, while the first routing is successful. In this case, we have to combine the results of top- $k$  routing and Intra-GPU Rectification. We combine them linearly according to the routing scores:
+
+$$
+o _ {i} = \frac {\sum_ {j \in \mathbb {R} _ {i}} e ^ {a _ {i j}} E _ {j} \left(x _ {i}\right) + (k - \left| \mathbb {R} _ {i} \right|) e ^ {a _ {i h}} E _ {h} \left(x _ {i}\right)}{\left(\sum_ {j \in \mathbb {R} _ {i}} e ^ {a _ {i j}}\right) + (k - \left| \mathbb {R} _ {i} \right|) e ^ {a _ {i h}}}, \tag {3}
+$$
+
+where  $E_{j}(x_{i})$  represents the expert outputs obtained through top-  $k$  routing, while  $E_{h}(x_{i})$  denotes the expert outputs from Intra-GPU Rectification. We normalize the routing scores  $a_{ij}$  and  $a_{ih}$  as
+
+the combining weights of  $E_{j}(x_{i})$  and  $E_{h}(x_{i})$  respectively. Specifically, we scale the combining weights of  $E_{h}(x_{i})$  with a constant factor  $(k - |\mathbb{R}_i|)$ , because a token is dropped  $(k - |\mathbb{R}_i|)$  times but only processed by one expert in the Intra-GPU Rectification.
+
+Similar to the top- $k$  router, the Intra-GPU Rectification also exhibits imbalance. However, this imbalance does not affect the computational fairness among GPUs. Although the balance of routing determines the number of dropped tokens within a batch, it does not influence the variation in the number of dropped tokens across different batches. The number of dropped tokens in different batches is similar due to their identical data distribution. Since batches on different GPUs are sampled independently and identically distributed, the number of dropped tokens across GPUs is comparable.
+
+# 4.2 Rectify-Router for Padding: Fill-in Rectification
+
+Fill-in Rectification aims to replace the unnecessary padding with the tokens that have high routing scores, which is visualized in the right part of Figure 2. This process is divided into two separate stages. Firstly, we identify the most suitable expert for each token, and subsequently, we select the optimal tokens for each expert.
+
+During the initial stage, each token will choose the expert ranked as the  $k + 1$ th highest score as the optimal expert. This decision is based on the fact that the top-  $k$  experts have already been assigned, and the  $k + 1$ th expert is considered the most suitable among the remaining experts. Furthermore, each token is only allowed to select one expert, which avoids the same token being processed by multiple experts during the second stage.
+
+Upon completion of the first stage, we transition to the second stage. It is worth noting that multiple tokens may select the same expert. Consequently, it is possible that the number of tokens choosing a particular expert surpasses the number of padding tokens of that expert. In such scenarios, we prioritize the tokens with higher routing scores for replacing the padding tokens.
+
+Indeed, implementing this algorithm can be achieved by extending the top-  $k$  router to a top $k + 1$  router while ensuring the expert capacity remains unchanged. As the expert capacity is fixed, introducing the Fill-in Rectification incurs minimal additional overhead. Alternatively, we can view
+
+this approach as reducing the capacity factor of the top-  $k + 1$  routing from  $k + 1$  to  $k$  to avoid the padding.
+
+The Fill-in Rectification has a potential issue related to the normalization of routing scores, where the gradient of routing scores may vanish due to the invalid normalization. We address this issue in Appendix A with straight-through trick (Bengio et al., 2013).
+
+# 5 Experiments
+
+# 5.1 Experiment Settings
+
+Model We follow previous work (Komatsuzaki et al., 2023) to train MoE models from a pretrained dense model. We initialize all experts in the same layer of MoE as the FFN parameters of the corresponding layer in the Dense model. We use the LLama2-7b (Touvron et al., 2023) to initialize MoE models. In most of our experiments, we employ eight experts per layer in the MoE models. But in Appendix B, we explore the extension of the number of experts to 32. Our experiments are conducted using the MoE implementation of DeepSpeed (Rajbhandari et al., 2022) and the training framework of gpt-neox (Aandonian et al., 2021).
+
+For simplicity, we denote our Intra-GPU Rectification as IR, and the Fill-in Rectification as FR. The top-  $k$  router, depending on whether it uses the Intra-GPU Rectification or the Fill-in Rectification, will be denoted as Top-k +IR or Top-k+FR, respectively.
+
+Training During the training phase, we utilize the OpenOrca dataset (Lian et al., 2023) with 1.78B tokens, which is an open-source reimplementation of Orca dataset (Mukherjee et al., 2023). It augments the instructions from flan data (Longpre et al., 2023) by adding complex system prompts and generate the step-by-step reasoning or explanation using chatgpt (OpenAI et al., 2023).
+
+We conduct our model training on a cluster of 32 GPUs (80GB). The training process consists of 10k steps with a global batch size of 256 and a micro batch size of 8. Following Mukherjee et al. (2023), we construct training examples by concatenating instructions with their corresponding responses: "[instruct][response]". However, only the tokens in the response are utilized for the next-token-prediction loss. For optimization, we use the Adam optimizer (Kingma and Ba, 2015) with an initial learning rate of 1e-5, which is decayed to 1e
+
+<table><tr><td>Model</td><td>Router</td><td>CF</td><td>Train Speed</td><td>MMLU</td><td>SuperGLUE</td><td>TruthfulQA</td><td>LogiQA</td><td>Avg</td></tr><tr><td>LLama2-raw</td><td>-</td><td>-</td><td>3.2k</td><td>25.85</td><td>59.06</td><td>25.21</td><td>25.03</td><td>33.78</td></tr><tr><td>LLama2</td><td>-</td><td>-</td><td>3.2k</td><td>35.01</td><td>63.74</td><td>30.23</td><td>27.64</td><td>39.15</td></tr><tr><td rowspan="4">LLama-MoE (Top-1)</td><td>Top-1</td><td>1.0</td><td>2.4k</td><td>33.05</td><td>64.34</td><td>29.49</td><td>28.11</td><td>38.74</td></tr><tr><td>Top-1+IR</td><td>1.0</td><td>2.3k</td><td>36.27</td><td>64.52</td><td>30.35</td><td>30.56</td><td>40.42</td></tr><tr><td>Top-1+FR</td><td>1.0</td><td>2.3k</td><td>34.66</td><td>63.97</td><td>28.51</td><td>29.18</td><td>39.08</td></tr><tr><td>Top-1+FR+IR</td><td>1.0</td><td>2.2k</td><td>35.81</td><td>65.08</td><td>30.84</td><td>30.56</td><td>40.57</td></tr><tr><td rowspan="4">LLama-MoE (Top-2)</td><td>Top-2</td><td>2.0</td><td>1.7k</td><td>35.39</td><td>64.58</td><td>29.98</td><td>29.33</td><td>39.82</td></tr><tr><td>Top-2+IR</td><td>2.0</td><td>1.6k</td><td>35.92</td><td>65.11</td><td>29.98</td><td>29.03</td><td>40.01</td></tr><tr><td>Top-2 + FR</td><td>2.0</td><td>1.6k</td><td>35.90</td><td>64.35</td><td>31.08</td><td>29.80</td><td>40.28</td></tr><tr><td>Top-2 +FR+IR</td><td>2.0</td><td>1.5k</td><td>36.01</td><td>65.60</td><td>30.72</td><td>29.95</td><td>40.57</td></tr></table>
+
+Table 1: The performance of LLama2-7b and MoE models on MMLU, SuperGLUE, TruthfulQA and LogiQA. CF denotes the capacity factor defined in Eq. (3). Avg represents the average accuracy. The training speed is measured as the number of tokens that each GPU can process per second. All models were trained on OpenOrca except for LLama2-raw. Top- $k+$ FR and Top- $k+$ IR represents the top- $k$  router using Fill-in Rectification and Intra-GPU Rectification respectively. Top- $k+$ FR+IR combines both the Fill-in Rectification and the Intra-GPU Rectification.
+
+![](images/43963c23aa9acce84a4eb60411d014927d3cdbb301ad3f399d0c90694b0b46cc.jpg)  
+Figure 3: The proportion of dropped/padding tokens varied by training steps. Both the MoE models of 8 experts and 32 experts are trained on OpenOrca based on LLama2-7b. The proportion of the dropped tokens is the same as that of the padding tokens.
+
+6 using a cosine learning rate scheduler. Regarding the load-balance loss for the top-  $k$  router, we set the weights to 1e-2, following Fedus et al. (2022).
+
+Evaluation We evaluated our models on multiple benchmarks, including MMLU (Li et al., 2023), SuperGLUE (Wang et al., 2019), TruthfulQA (Lin et al., 2022) and LogiQA (Liu et al., 2020), which covers the evaluation in knowledge, natural language understanding, safety, and logical reasoning respectively. All evaluations were conducted in a zero-shot setting. Our evaluation metric was accuracy, and we utilized the lm-evaluation-harness (Gao et al., 2023) framework for conducting the evaluations.
+
+# 5.2 Proportion of Dropped/Padding Tokens
+
+In Figure 3, we show the proportion of dropped/padding tokens. Initially, the proportion of dropped tokens decreases and stabilizes after 2000 steps. This reduction is due to the load-balance loss proposed by Lepikhin et al. (2021). However, even with this loss, the proportion of dropped tokens remains high— $15\%$  for 32 experts. Additionally, the number of dropped tokens is significantly larger for 32 experts compared to 8 experts. This is expected, as balancing the routing for more experts is more challenging. With only one expert, there is no imbalance, but as the number of experts increases, the issue of dropped tokens becomes more pronounced.
+
+# 5.3 Main Results
+
+We trained both LLama2-7b and LLama-based MoE on OpenOrca and evaluated them on MMLU (knowledge), SuperGLUE (NLU), TruthfulQA (Safety) and LogiQA (Reasoning), the results of which are shown in Table 1. Comparing the performance of LLama2-raw (pretrained) and LLama2 (trained on OpenOrca), we observed that the LLama2 outperforms LLama2-raw substantially, which demonstrates the effectiveness of finetuning on openorca. To evaluate the effectiveness of our methods, we applied our Intra-GPU Rectification (IR) and Fill-in Rectification (FR) to both the top-1 router and top-2 router. These configurations are grouped as LLama-MoE (Top-1) and LLama-MoE (Top-2) in Table 1.
+
+<table><tr><td>Train Router</td><td>Test Router</td><td>Test CF</td><td>Test Speed</td><td>MMLU</td><td>SuperGLUE</td><td>TruthfulQA</td><td>LogiQA</td><td>Avg</td></tr><tr><td rowspan="6">Top-1</td><td>Top-1</td><td></td><td>9.4k</td><td>33.05</td><td>64.34</td><td>29.49</td><td>28.11</td><td>38.74</td></tr><tr><td>Expert Choices</td><td></td><td>9.4k</td><td>30.27</td><td>64.50</td><td>27.66</td><td>28.87</td><td>37.82</td></tr><tr><td>Megablocks</td><td>1.0</td><td>-1</td><td>36.39</td><td>64.09</td><td>29.98</td><td>30.41</td><td>40.21</td></tr><tr><td>Top-1+IR</td><td></td><td>9.2k</td><td>36.21</td><td>64.64</td><td>29.86</td><td>29.18</td><td>39.97</td></tr><tr><td>Top-1+FR</td><td></td><td>8.9k</td><td>33.28</td><td>62.76</td><td>28.51</td><td>29.49</td><td>38.51</td></tr><tr><td>Top-1+FR+IR</td><td></td><td>8.6k</td><td>36.40</td><td>63.94</td><td>29.98</td><td>29.80</td><td>40.03</td></tr><tr><td rowspan="6">Top-2</td><td>Top-2</td><td></td><td>6.2k</td><td>35.39</td><td>64.58</td><td>29.98</td><td>29.33</td><td>39.82</td></tr><tr><td>Expert Choices</td><td></td><td>6.2k</td><td>32.80</td><td>61.76</td><td>26.80</td><td>28.57</td><td>37.48</td></tr><tr><td>Megablocks</td><td>2.0</td><td>-</td><td>35.95</td><td>64.59</td><td>30.47</td><td>30.26</td><td>40.31</td></tr><tr><td>Top-2+IR</td><td></td><td>6.0k</td><td>35.70</td><td>64.40</td><td>30.47</td><td>29.95</td><td>40.13</td></tr><tr><td>Top-2+FR</td><td></td><td>5.8k</td><td>35.96</td><td>65.37</td><td>30.35</td><td>30.26</td><td>40.48</td></tr><tr><td>Top-2+FR+IR</td><td></td><td>5.5k</td><td>36.14</td><td>65.16</td><td>30.35</td><td>31.49</td><td>40.78</td></tr></table>
+
+Table 2: The performance of applying Intra-GPU Rectification and Fill-in Rectification only at inference. All models are trained with the vanilla top-1 router and top-2 router (referred to as the train router), but they were evaluated with Intra-GPU Rectification or Fill-in Rectification at inference (referred to as the test router). Test CF denotes the capacity factor set during inference. Test speed represents the number of tokens processed per second on each GPU during inference.
+
+LLama-MoE (Top-1) We conducted 4 top-1 based MoE models (Top-1, Top-1+FR, Top-1+IR, Top-1+FR+IR). The performance of the vanilla top-1 router is subpar, and it is even inferior to the dense model (LLama2-FT) on both MMLU and TruthfulQA. But after incorporating our proposed Intra-GPU Rectification (Top-1+IR), the performance of the top-1 router are significantly improved on all benchmarks, especially on MMLU and LogiQA. This indicates that the dropped tokens have a substantial impact on model performance, and the Intra-GPU Rectification effectively handles these dropped tokens. Our Fill-in Rectification (Top-1+FR) also significantly improves the performance of the model on MMLU and LogiQA tasks. But it is worth noting that the performance of the model declined on the other two benchmarks. Therefore, it can be concluded that the primary issue with top-1 routing lies in dropped tokens rather than padding. Combing the Intra-GPU Rectification and Fill-in Rectification resulted in the best top-1-based router (Top-1+FR+IR), which outperforms the vanilla top-1 router by 1.83 (4.7%) in terms of the average accuracy across benchmarks.
+
+LLama-MoE (Top-2) Top-2 based routers also encompass 4 routers (Top-2, Top-2-FR, Top-2-IR, Top-2-FR+IR). Both the Intra-GPU Rectification and Fill-in Rectification significantly enhance the performance of Top-2 router on at least 2 benchmarks, which demonstrate that our methods are effective for the top-2 router as well. Just as we observed with the top-1 routing results, combining
+
+the Intra-GPU Rectification and the Fill-in Rectification in the top-2 router yielded the best performance on all benchmarks. Specifically, the Top-2+FR+IR outperformed the vanilla top-2 router by a margin of 0.75 (1.8%) in terms of the average accuracy across benchmarks.
+
+Interestingly, we observed that the top-1 router outperformed the top-2 router in some benchmarks. For example, Top-1+FR+IR outperforms Top-1+FR+IR on both TruthfulQA and LogiQA, which raises concerns about potential overfitting in the top-2 router. Finally, it is important to note that that both the Intra-GPU Rectification and Fill-in Rectification do not alter the capacity of experts, hence they do not significantly influence the training speed.
+
+# 5.4 Improve Top- $k$  Routing at Inference
+
+In this experiment, we conducted a study to evaluate the effectiveness of applying Rectify-Routers at the inference stage of MoE models. The results are presented in Table 2. We found that both the Intra-GPU Rectification and Fill-in Rectification can improve the performance of top-1 and top-2 routers at inference, even they are not applied at training. Similar to the results in Table 1, combining Intra-GPU Rectification and Fill-in Rectification yielded better results than using either method alone. Moreover, both the Intra-GPU Rectification and Fill-in Rectification only slightly slows down  $(< 10\%)$  the inference speed of top- $k$  routers. For top-2 based models, the application of Rectify-Routers solely during the inference stage proves to be sufficient, as
+
+it demonstrates comparable performance to using them during both training and inference.
+
+We also compared our methods with Expert Choices and Megablocks. Expert Choices addressed the issue of padding but still suffers from the problem of dropped tokens. According to the results presented in Table 2, incorporating expert choices during the model inference phase reduces the model's performance. This suggests that expert choices need to be trained to perform well, whereas our method can be applied directly during the inference phase of MoE trained with a Top-k router. Megablocks overcomes both the issues of dropped tokens and padding by switching from expert parallelism to model parallelism. Although the performance of Megablocks is comparable to our method, the communication complexity of Megablocks  $(O(C_{g} \cdot W))$  is much higher than that of expert parallelism  $(O(C_{g}))$  (Hwang et al., 2022), where  $C_{g}$  denotes the token capacity per GPU and  $W$  denotes the world size of communication. As Tutel (Hwang et al., 2022) suggests, it is better to combine model parallelism with expert parallelism for greater efficiency. Therefore, our method and Megablocks are complementary.
+
+# 5.5 Capacity Factor Variation
+
+In the previous experiments, we maintained a fixed capacity factor of  $k$  for top-  $k$  routing. However, there are instances where it may be beneficial to adjust the capacity factor for improved efficiency or performance. Therefore, in this section, we examine the performance of our Rectify-Routers under different capacity factors.
+
+To minimize training costs, we train MoE models using the vanilla top- $k$  router with a capacity factor of  $k$ , and evaluate models with different capacity factors. We only present the average accuracy of models in Table 3 and Table 4. The complete results are shown in Appendix D
+
+Post Routing with Low Capacity From Table 3, we can see that decreasing the capacity factor improves the efficiency of both top-1 and top-2 based models. However, It also leads to noticeable decrease in the model performance on benchmarks. It is interesting that the top-2 router is more robust to the decrease in capacity factor. Specifically, reducing the capacity factor of the vanilla top-2 router
+
+<table><tr><td>Train Router</td><td>Test Router</td><td>Test CF</td><td>Test Speed</td><td>Avg</td></tr><tr><td rowspan="6">Top-1</td><td>Top-1</td><td rowspan="2">1.0</td><td>9.4k</td><td>38.74</td></tr><tr><td>Top-1+IR</td><td>9.2k</td><td>39.97</td></tr><tr><td>Top-1</td><td rowspan="2">0.75</td><td>12.1k</td><td>37.83</td></tr><tr><td>Top-1+IR</td><td>9.9k</td><td>40.06</td></tr><tr><td>Top-1</td><td rowspan="2">0.5</td><td>16k</td><td>34.84</td></tr><tr><td>Top-1+IR</td><td>10.6k</td><td>40.40</td></tr><tr><td rowspan="6">Top-2</td><td>Top-2</td><td rowspan="2">2.0</td><td>6.2k</td><td>39.82</td></tr><tr><td>Top-2+IR</td><td>6.0k</td><td>40.13</td></tr><tr><td>Top-2</td><td rowspan="2">1.5</td><td>7.4k</td><td>39.50</td></tr><tr><td>Top-2+IR</td><td>6.6k</td><td>39.60</td></tr><tr><td>Top-2</td><td rowspan="2">1.0</td><td>8.9k</td><td>38.51</td></tr><tr><td>Top-2+IR</td><td>7.3k</td><td>40.01</td></tr></table>
+
+Table 3: Performance of top-  $k$  routers and their variants with low capacity factors  $\left( { <  = k}\right)$  .
+
+from 2 to 1.5 only results in a slight performance decline (0.32).
+
+In contrast to the vanilla top-1 or top-2 routers, the MoE models incorporating our Intra-GPU Rectification (Top-1+IR and Top-2+IR) are robust to the decrease of capacity factor. We even observed that the lower capacity factor leads to a better performance for both Top-1+IR and Top-2+IR, which suggests that the Intra-GPU Rectification acts as a form of regularization for the MoE models by constraining the choices made by the experts. The similar results are also observed in Zeng and Xiong (2023); Liu et al. (2022). By setting the capacity factor of Top-1+IR to 0.5 and that of Top-2-IR to 1.0, we observed that they are faster than the vanilla top-1 (1.13x) and top-2 routers (1.18x) respectively, while maintaining comparable or superior performance.
+
+Fill-in Rectification with High Capacity Increasing the capacity factor of MoE models has been widely suggested in previous research studies (Fedus et al., 2022; Zoph et al., 2022). In alignment with these findings, we have also observed the benefits of increasing the capacity factor in terms of improving model performance, as demonstrated in Table 4. Notably, we have found that increasing the capacity factor of the top-1 router leads to a more substantial improvement in model performance than that of the top-2 router.
+
+Our Fill-in Rectification introduces a more significant and consistent improvement with the increase in capacity factor. Top-1+FR and Top-2+FR consistently outperform Top-1 and Top-2, respectively, across various capacity factor settings.
+
+<table><tr><td>Train Router</td><td>Test Router</td><td>Test CF</td><td>Test Speed</td><td>Avg</td></tr><tr><td rowspan="6">Top-1</td><td>Top-1</td><td rowspan="2">1.0</td><td>9.4k</td><td>38.74</td></tr><tr><td>Top-1+FR</td><td>8.9k</td><td>38.51</td></tr><tr><td>Top-1</td><td rowspan="2">1.25</td><td>8.6k</td><td>39.59</td></tr><tr><td>Top-1+FR</td><td>8.1k</td><td>40.10</td></tr><tr><td>Top-1</td><td rowspan="2">1.5</td><td>7.9k</td><td>39.86</td></tr><tr><td>Top-1+FR</td><td>7.3k</td><td>40.33</td></tr><tr><td rowspan="6">Top-2</td><td>Top-2</td><td rowspan="2">2.0</td><td>6.2k</td><td>39.82</td></tr><tr><td>Top-2+FR</td><td>5.8k</td><td>40.48</td></tr><tr><td>Top-2</td><td rowspan="2">2.5</td><td>5.4k</td><td>39.89</td></tr><tr><td>Top-2+FR</td><td>5.1k</td><td>40.51</td></tr><tr><td>Top-2</td><td rowspan="2">3.0</td><td>4.9k</td><td>40.03</td></tr><tr><td>Top-2+FR</td><td>4.5k</td><td>40.44</td></tr></table>
+
+Table 4: Performance of top-  $k$  routers and their variants with high capacity factors  $\left( { >  = k}\right)$  .  
+
+<table><tr><td>Router</td><td>Experts/GPU</td><td>Avg</td></tr><tr><td rowspan="3">Top-1+IR</td><td>1</td><td>39.97</td></tr><tr><td>2</td><td>39.95</td></tr><tr><td>4</td><td>39.83</td></tr><tr><td rowspan="3">Top-2+IR</td><td>1</td><td>40.13</td></tr><tr><td>2</td><td>40.01</td></tr><tr><td>4</td><td>40.17</td></tr></table>
+
+# 5.6 Impact of Expert Distribution
+
+Our Intra-GPU Rectification is a variant of the top-1 router, where tokens are assigned to the top-1 expert within GPU. When all experts are situated in the same GPU, the Intra-GPU Rectification essentially functions as the top-1 router. Therefore, the distribution of experts across GPUs can potentially influence the performance of the Intra-GPU Rectification. We conducted an investigation to explore this aspect and present the results in Table 5.
+
+Interestingly, we found that increasing the number of experts per GPU did not yield significant improvements for either the top-1 router or the top-2 router. This suggests that the Intra-GPU Rectification demonstrates robustness to variations in the number of experts per GPU.
+
+# 5.7 Impact of Load-Balance Loss
+
+The Rectify-Routers proposed in this paper were designed to address the issues of dropped tokens and padding resulting from unbalanced routing.
+
+Table 5: The performance of Intra-GPU Rectification evaluated under various settings of the number of experts per GPU.  
+
+<table><tr><td>Model</td><td>Aux-loss</td><td>Avg</td></tr><tr><td>Top-1</td><td>yes</td><td>38.74</td></tr><tr><td>Top-1</td><td>no</td><td>38.21</td></tr><tr><td>Top-1+FR+IR</td><td>yes</td><td>40.03</td></tr><tr><td>Top-1+FR+IR</td><td>no</td><td>39.98</td></tr></table>
+
+Table 6: The performance comparison of using vs. not using load-balance loss. Aux-loss represents whether load-balance loss is used.
+
+In our previous experiments, we utilized the loadbalanced loss introduced by Lepikhin et al. (2021) to enhance the balance of routing for all models, including those utilizing the Rectify-Routers. However, it is intriguing to investigate whether the Rectify-Routers remain effective in the absence of the load-balance loss. The results of this exploration are presented in Table 6.
+
+Upon analyzing the results in Table 6, we observed a notable disparity in the performance of the vanilla top-1 router with and without the load-balance loss, particularly in the case of SuperGLUE and TruthfulQA. This discrepancy suggests that the load-balance loss plays a crucial role in improving the performance of the vanilla top-1 router. However, when considering our Rectify-Routers  $(\mathrm{Top - 1 + FR + IR})$ , removing the load-balance loss does not result in a significant loss of performance. This finding indicates that our Rectify-Routers enhance the resilience of the top-1 router against the load-balance loss.
+
+Other Experiments 1) We scale the number of experts from 8 to 32 in Appendix B; 2) We validate the importance of straight-through trick in Appendix C.1;
+
+# 6 Conclusion
+
+In this paper, we present the Rectify-Router, a method to tackle dropped tokens and padding in MoE models. By introducing the Intra-GPU Rectification and the Fill-in Rectification, we effectively handle the issues of dropped tokens and padding, respectively. Experimental results demonstrate the individual effectiveness of both techniques and the synergistic performance improvement when they are combined. Furthermore, our methods prove to be effective in diverse settings, including varying numbers of experts, different expert capacities, and even without the load-balance loss.
+
+# 7 Limitation
+
+1. The MoE models were initialized from a dense model (LLama2-7b). Due to the high costs, we have not validated our methods by training from scratch. But our experimental results in Table 1 demonstrate that finetuning the pre-trained LLama2-7b into an MoE model can bring significant performance improvements.  
+2. Our experiments were conducted using LLama2-7b, while other configurations, such as LLama2-70B, were not explored due to high costs. But we have validated the scalability of our method by increasing the number of experts, which is presented in Appendix B.  
+These limitations highlight potential areas for future research and expansion of our work.
+
+# References
+
+Alex Andonian, Quentin Anthony, Stella Biderman, Sid Black, Preetham Gali, Leo Gao, Eric Hallahan, Josh Levy-Kramer, Connor Leahy, Lucas Nestler, Kip Parker, Michael Pieler, Shivanshu Purohit, Tri Songz, Wang Phil, and Samuel Weinbach. 2021. GPT-NeoX: Large Scale Autoregressive Language Modeling in PyTorch.  
+Yoshua Bengio, Nicholas Léonard, and Aaron C. Courville. 2013. Estimating or propagating gradients through stochastic neurons for conditional computation. CoRR, abs/1308.3432.  
+Damai Dai, Chengqi Deng, Chenggang Zhao, R. X. Xu, Huazuo Gao, Deli Chen, Jiashi Li, Wangding Zeng, Xingkai Yu, Y. Wu, Zhenda Xie, Y. K. Li, Panpan Huang, Fuli Luo, Chong Ruan, Zhifang Sui, and Wenfeng Liang. 2024. Deepseekmoe: Towards ultimate expert specialization in mixture-of-experts language models. CoRR, abs/2401.06066.  
+Nan Du, Yanping Huang, Andrew M. Dai, Simon Tong, Dmitry Lepikhin, Yuanzhong Xu, Maxim Krikun, Yanqi Zhou, Adams Wei Yu, Orhan First, Barret Zoph, Liam Fedus, Maarten P. Bosma, Zongwei Zhou, Tao Wang, Yu Emma Wang, Kellie Webster, Marie Pellat, Kevin Robinson, Kathleen S. Meier-Hellstern, Toju Duke, Lucas Dixon, Kun Zhang, Quoc V. Le, Yonghui Wu, Zhifeng Chen, and Claire Cui. 2022. Glam: Efficient scaling of language models with mixture-of-experts. In International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA, volume 162 of Proceedings of Machine Learning Research, pages 5547-5569. PMLR.  
+William Fedus, Barret Zoph, and Noam Shazeer. 2022. Switch transformers: Scaling to trillion parameter
+
+models with simple and efficient sparsity. J. Mach. Learn. Res., 23:120:1-120:39.  
+Trevor Gale, Deepak Narayanan, Cliff Young, and Matei Zaharia. 2022. Megablocks: Efficient sparse training with mixture-of-experts. CoRR, abs/2211.15841.  
+Leo Gao, Jonathan Tow, Baber Abbasi, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Alain Le Noac'h, Haonan Li, Kyle McDonell, Niklas Muennighoff, Chris Ociepa, Jason Phang, Laria Reynolds, Hailey Schoelkopf, Aviya Skowron, Lintang Sutawika, Eric Tang, Anish Thite, Ben Wang, Kevin Wang, and Andy Zou. 2023. A framework for few-shot language model evaluation.  
+Changho Hwang, Wei Cui, Yifan Xiong, Ziyue Yang, Ze Liu, Han Hu, Zilong Wang, Rafael Salas, Jithin Jose, Prabhat Ram, Joe Chau, Peng Cheng, Fan Yang, Mao Yang, and Yongqiang Xiong. 2022. Tutel: Adaptive mixture-of-experts at scale. CoRR, abs/2206.03382.  
+Albert Q. Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de Las Casas, Emma Bou Hanna, Florian Bressand, Gianna Lengyel, Guillaume Bour, Guillaume Lample, Lélio Renard Lavaud, Lucile Saulnier, Marie-Anne Lachaux, Pierre Stock, Sandeep Subramanian, Sophia Yang, Szymon Antoniak, Teven Le Scao, Théophile Gervet, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. 2024. Mixtral of experts. CoRR, abs/2401.04088.  
+Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings.  
+Yeskendir Koishekenov, Vassilina Nikoulina, and Alexandre Berard. 2022. Memory-efficient NLLB-200: language-specific expert pruning of a massively multilingual machine translation model. CoRR, abs/2212.09811.  
+Aran Komatsuzaki, Joan Puigcerver, James Lee-Thorp, Carlos Riquelme Ruiz, Basil Mustafa, Joshua Ainslie, Yi Tay, Mostafa Dehghani, and Neil Houlsby. 2023. Sparse upcycling: Training mixture-of-experts from dense checkpoints. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net.  
+Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan First, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. 2021. Gshard: Scaling giant models with conditional computation and automatic sharding. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net.
+
+Mike Lewis, Shruti Bhosale, Tim Dettmers, Naman Goyal, and Luke Zettlemoyer. 2021. BASE layers: Simplifying training of large, sparse models. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 6265-6274. PMLR.  
+Haonan Li, Yixuan Zhang, Fajri Koto, Yifei Yang, Hai Zhao, Yeyun Gong, Nan Duan, and Timothy Baldwin. 2023. CMMLU: measuring massive multitask language understanding in chinese. CoRR, abs/2306.09212.  
+Wing Lian, Bleys Goodson, Eugene Pentland, Austin Cook, Chanvichet Vong, and "Teknium". 2023. Openorca: An open dataset of gpt augmented flan reasoning traces. https://https://huggingface.co/Open-Orca/OpenOrca.  
+Stephanie Lin, Jacob Hilton, and Owain Evans. 2022. Truthfulqa: Measuring how models mimic human falsehoods. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2022, Dublin, Ireland, May 22-27, 2022, pages 3214-3252. Association for Computational Linguistics.  
+Jian Liu, Leyang Cui, Hanmeng Liu, Dandan Huang, Yile Wang, and Yue Zhang. 2020. Logiqa: A challenge dataset for machine reading comprehension with logical reasoning. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI 2020, pages 3622-3628. ijcai.org.  
+Rui Liu, Young Jin Kim, Alexandre Muzio, and Hany Hassan. 2022. Gating dropout: Communication-efficient regularization for sparsely activated transformers. In International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA, volume 162 of Proceedings of Machine Learning Research, pages 13782-13792. PMLR.  
+Shayne Longpre, Le Hou, Tu Vu, Albert Webson, Hyung Won Chung, Yi Tay, Denny Zhou, Quoc V. Le, Barret Zoph, Jason Wei, and Adam Roberts. 2023. The flan collection: Designing data and methods for effective instruction tuning.  
+Subhabrata Mukherjee, Arindam Mitra, Ganesh Jawahar, Sahaj Agarwal, Hamid Palangi, and Ahmed Awadallah. 2023. Orca: Progressive learning from complex explanation traces of GPT-4. CoRR, abs/2306.02707.  
+Basil Mustafa, Carlos Riquelme, Joan Puigcerver, Rodolphe Jenatton, and Neil Houlsby. 2022. Multimodal contrastive learning with limoe: the language-image mixture of experts. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022.
+
+OpenAI, , Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mo Bavarian, Jeff Belgium, Irwan Bello, Jake Berdine, Gabriel Bernadett-Shapiro, Christopher Berner, Lenny Bogdonoff, Oleg Boiko, Madelaine Boyd, Anna-Luisa Brakman, Greg Brockman, Tim Brooks, Miles Brundage, Kevin Button, Trevor Cai, Rosie Campbell, Andrew Cann, Brittany Carey, Chelsea Carlson, Rory Carmichael, Brooke Chan, Che Chang, Fotis Chantzis, Derek Chen, Sully Chen, Ruby Chen, Jason Chen, Mark Chen, Ben Chess, Chester Cho, Casey Chu, Hyung Won Chung, Dave Cummings, Jeremiah Currier, Yunxing Dai, Cory Decareaux, Thomas Degry, Noah Deutsch, Damien Deville, Arka Dhar, David Dohan, Steve Dowling, Sheila Dunning, Adrien Ecoffet, Atty Eleti, Tyna Eloundou, David Farhi, Liam Fedus, Niko Felix, Simón Posada Fishman, Juston Forte, Isabella Fulford, Leo Gao, Elie Georges, Christian Gibson, Vik Goel, Tarun Gogineni, Gabriel Goh, Rapha Gontijo-Lopes, Jonathan Gordon, Morgan Grafstein, Scott Gray, Ryan Greene, Joshua Gross, Shixiang Shane Gu, Yufei Guo, Chris Hallacy, Jesse Han, Jeff Harris, Yuchen He, Mike Heaton, Johannes Heidecke, Chris Hesse, Alan Hickey, Wade Hickey, Peter Hoeschele, Brandon Houghton, Kenny Hsu, Shengli Hu, Xin Hu, Joost Huizinga, Shantanu Jain, Shawn Jain, Joanne Jang, Angela Jiang, Roger Jiang, Haozhun Jin, Denny Jin, Shino Jomoto, Billie Jonn, Heewoo Jun, Tomer Kaftan, Lukasz Kaiser, Ali Kamali, Ingmar Kanitscheider, Nitish Shirish Keskar, Tabarak Khan, Logan Kilpatrick, Jong Wook Kim, Christina Kim, Yongjik Kim, Hendrik Kirchner, Jamie Kiros, Matt Knight, Daniel Kokotajlo, Lukasz Kondraciuk, Andrew Kondrich, Aris Konstantinidis, Kyle Kosic, Gretchen Krueger, Vishal Kuo, Michael Lampe, Ikai Lan, Teddy Lee, Jan Leike, Jade Leung, Daniel Levy, Chak Ming Li. Rachel Lim Molly Lin Stephanie Lin Mateusz Litwin Theresa Lopez Ryan Lowe Patricia Lue. Anna Makanju Kim Malfacini Sam Manning Todor Markov Yaniv Markovski Bianca Martin Katie Mayer. Andrew Mayne Bob McGrew Scott Mayer McKinney Christine McLeavey Paul McMillan. Jake McNeil David Medina Aalok Mehta Jacob Menick Luke Metz. Andrey Mishchenko Pamela Mishkin Vinnie Monaco Evan Morikawa Daniel Mossing Tong Mu Mira Murati Oleg Murk David Mély. Ashvin Nair. Reiichiro Nakano Rajeev Nayak. Arvind Neelakantan Richard Ngo. Hyeonwoo Noh Long Ouyang Cullen O'Keefe. Jakub Pachocki Alex Paino Joe Palermo Ashley Pantuliano Giambattista Parascandolo Joel Parish. Emy Parparita Alex Passos. Mikhail Pavlov. Andrew Peng Adam Perelman Filipe de Avila Belbute Peres Michael Petrov. Henrique Ponde de Oliveira Pinto Michael Pokorny. Michelle Pokrass. Vitchyr Pong. Tolly Powell. Alethea Power. Boris Power. Elizabeth Proehl. Raul Puri. Alec Radford. Jack Rae. Aditya Ramesh. Cameron Raymond. Francis Real. Kendra Rimbach
+
+Carl Ross, Bob Rotsted, Henri Roussez, Nick Ryder, Mario Saltarelli, Ted Sanders, Shibani Santurkar, Girish Sastry, Heather Schmidt, David Schnurr, John Schulman, Daniel Selsam, Kyla Sheppard, Toki Sherbakov, Jessica Shieh, Sarah Shoker, Pranav Shyam, Szymon Sidor, Eric Sigler, Maddie Simens, Jordan Sitkin, Katarina Slama, Ian Sohl, Benjamin Sokolowsky, Yang Song, Natalie Staudacher, Felipe Petroski Such, Natalie Summers, Ilya Sutskever, Jie Tang, Nikolas Tezak, Madeleine Thompson, Phil Tillet, Amin Tootoonchian, Elizabeth Tseng, Preston Tuggle, Nick Turley, Jerry Tworek, Juan Felipe Cerón Uribe, Andrea Vallone, Arun Vijayvergiya, Chelsea Voss, Carroll Wainwright, Justin Jay Wang, Alvin Wang, Ben Wang, Jonathan Ward, Jason Wei, CJ Weinmann, Akila Welihinda, Peter Welinder, Jiayi Weng, Lilian Weng, Matt Wiethoff, Dave Willner, Clemens Winter, Samuel Wolrich, Hannah Wong, Lauren Workman, Sherwin Wu, Jeff Wu, Michael Wu, Kai Xiao, Tao Xu, Sarah Yoo, Kevin Yu, Qiming Yuan, Wojciech Zaremba, Rowan Zellers, Chong Zhang, Marvin Zhang, Shengjia Zhao, Tianhao Zheng, Juntang Zhuang, William Zhuk, and Barret Zoph. 2023. Gpt-4 technical report.
+
+Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier, and Michael Auli. 2019. *fairoseq: A fast, extensible toolkit for sequence modeling*. In *Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies*, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Demonstrations, pages 48-53. Association for Computational Linguistics.
+
+Joan Puigcerver, Carlos Riquelme, Basil Mustafa, and Neil Houlsby. 2023. From sparse to soft mixtures of experts. CoRR, abs/2308.00951.
+
+Samyam Rajbhandari, Conglong Li, Zhewei Yao, Minjia Zhang, Reza Yazdani Aminabadi, Ammar Ahmad Awan, Jeff Rasley, and Yuxiong He. 2022. Deepspeed-moe: Advancing mixture-of-experts inference and training to power next-generation AI scale. In International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA, volume 162 of Proceedings of Machine Learning Research, pages 18332-18346. PMLR.
+
+Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc V. Le, Geoffrey E. Hinton, and Jeff Dean. 2017. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net.
+
+Sheng Shen, Le Hou, Yanqi Zhou, Nan Du, Shayne Longpre, Jason Wei, Hyung Won Chung, Barret Zoph, William Fedus, Xinyun Chen, Tu Vu, Yuexin Wu, Wuyang Chen, Albert Webson, Yunxuan Li, Vincent Zhao, Hongkun Yu, Kurt Keutzer, Trevor Darrell, and Denny Zhou. 2023. Mixture-of-experts meets
+
+instruction tuning:a winning combination for large language models.
+
+Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton-Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. 2023. Llama 2: Open foundation and finetuned chat models. CoRR, abs/2307.09288.
+
+Alex Wang, Yada Pruksachatkun, Nikita Nangia, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. 2019. Superglue: A stickier benchmark for general-purpose language understanding systems. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 3261-3275.
+
+Ping Yu, Mikel Artetxe, Myle Ott, Sam Shleifer, Hongyu Gong, Ves Stoyanov, and Xian Li. 2022. Efficient language modeling with sparse all-mlp. CoRR, abs/2203.06850.
+
+Zhiyuan Zeng and Deyi Xiong. 2023. Scomoe: Efficient mixtures of experts with structured communication. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net.
+
+Yanqi Zhou, Tao Lei, Hanxiao Liu, Nan Du, Yanping Huang, Vincent Zhao, Andrew M. Dai, Zhifeng Chen, Quoc V. Le, and James Laudon. 2022. Mixture-of-experts with expert choice routing. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022.
+
+Barret Zoph, Irwan Bello, Sameer Kumar, Nan Du, Yanping Huang, Jeff Dean, Noam Shazeer, and William Fedus. 2022. St-moe: Designing stable and transferable sparse expert models.
+
+Simiao Zuo, Xiaodong Liu, Jian Jiao, Young Jin Kim, Hany Hassan, Ruofei Zhang, Jianfeng Gao, and Tuo Zhao. 2022. Taming sparsely activated transformer
+
+# A Gradient Issues in Fill-in Rectification
+
+There is a potential issue in the Fill-in Rectification, which stems from the implementation of the top-  $k$  routing. According to Eq. (1), the routing scores of top-  $k$  routing are normalized on the selected experts  $\mathbb{R}_i$ , rather than considering all expert choices. Several implementations like deepspeed-moe (Rajbhandari et al., 2022) and fairseq-moe (Ott et al., 2019) first normalize the routing scores on all experts and then re-normalize the scores specifically for the selected experts:
+
+$$
+\begin{array}{l} g _ {i j} = \frac {e ^ {a _ {i j}}}{\sum_ {j} ^ {m} e ^ {a _ {i j}}} \tag {4} \\ o _ {i} = \sum_ {j \in \mathbb {R} _ {i}} \frac {g _ {i j}}{\sum_ {j} g _ {i j}} E _ {j} (x _ {i}), \\ \end{array}
+$$
+
+where  $g_{ij}$  represents the routing scores that are initially normalized across all experts and then further normalized specifically on the selected experts  $(\mathbb{R}_i)$ . However, their implementation is equivalent to directly normalizing the routing scores on  $\mathbb{R}_i$ .
+
+There are two potential issues of normalizing routing scores on  $\mathbb{R}_i$ : 1) the routing scores of activated experts can not influence those of inactivated experts. For example, the increase of  $a_{ij}(j \in \mathbb{R}_i)$  does not lead to the decrease of  $a_{il}(l \notin \mathbb{R}_i)$ . 2) In the case of top-2 routing, if the first routing of  $x_i$  is successful while the second routing fails due to the expert overflow, the gradients of all routing scores of  $x_i$  will be zero ( $\frac{\partial L}{\partial a_{ij}} = 0$ ). This is because that there is only one available expert choice for  $x_i$  ( $|\mathbb{R}_i| = 1$ ). Normalizing on  $|\mathbb{R}_i|$  would always yield a value of 1, regardless of the actual value of  $a_{ij}$ , leading to invalid gradients.
+
+This problem is more prominent for the Fill-in Rectification, since it brings more dropped tokens, i.e., more unsuccessful routing. To address this problem, we utilize the straight-through trick to stop the gradient of normalization item in Eq. (4), which ensures that the gradient of routing scores remain valid:
+
+$$
+\frac {\partial L}{\partial a _ {i j}} \equiv \frac {\partial L}{\sum_ {j} g _ {i j} \partial \frac {g _ {i j}}{\sum_ {j} g _ {i j}}} \frac {\partial g _ {i j}}{\partial a _ {i j}} \tag {5}
+$$
+
+No modifications have been made to the forward stage. But at the backward stage, the gradient of the routing score  $\frac{\partial L}{\partial g_{ij}}$  is calculated as  $\frac{\partial L}{\sum_{j} g_{ij} \partial \frac{g_{ij}}{\sum_{j} g_{ij}}}$  rather than 0, where the normalization item  $\sum_{j} g_{ij}$  is taken as a constant number without gradient.
+
+![](images/eb66606c183136e69cf2cb64f0822c41b26f8f6381bff75a175584efb6a888a5.jpg)  
+(a) The performance of 8-experts and 32-experts MoEs on MMLU
+
+![](images/26b2c5a9ab6ec7fc770b1be708d734f6e9cd7a07dc26faaa746fdca2d38609ee.jpg)  
+(b) The performance of 8-experts and 32-experts MoEs on SuperGLUE
+
+![](images/c96a7f11ff09452e06e21f482626e0998ca27b8e6ecba725c9ba4f92fdf91778.jpg)  
+(c) The performance of 8-experts and 32-experts MoEs on TruthfulQA
+
+![](images/70027715bb83521a34e8c1e271b9f89fca45ac27c4a7d85bdcc5926dd9ff8369.jpg)  
+(d) The performance of 8-experts and 32-experts MoEs on LogiQA  
+Figure 4: The performance of 8-experts and 32-experts MoEs on MMLU, SuperGLUE, TruthfulQA and LogiQA.
+
+# B Scaling to 32 Experts
+
+In this experiment, we aimed to investigate the effectiveness of our methods when applied to a larger number of experts. We expanded the number of experts from 8 to 32. To reduce training costs, we only applied the Rectify-Routers (Intra-GPU Rectification and Fill-in Rectification) during evaluation. The results of this experiment are presented in Figure 4.
+
+Interestingly, our findings indicate that increasing the number of experts from 8 to 32 does not necessarily result in improved model performance. In fact, in certain benchmarks, such as SuperGLUE, the performance of the model even declined. This observation aligns with previous research (Komatsuzaki et al., 2023), suggesting that increasing the number of experts can potentially be detrimental. One plausible explanation for this phenomenon is
+
+that a larger number of experts may lead to overfitting of the model. We believe that increasing the number of experts is helpful with enough training data. Notably, scaling from 8 to 32 experts only yielded notable benefits in the case of TruthfulQA.
+
+Despite the lack of consistent improvement when increasing the number of experts, our methods (Intra-GPU Rectification and Fill-in Rectification) still demonstrated significant enhancements compared to the vanilla top- $k$  routing approach in the context of 32 experts. For instance, while the vanilla top-1 and top-2 routers with 32 experts underperformed those with 8 experts on MMLU, our methods (Top-2+FR+IR) enabled the 32-expert models to outperform their 8-expert counterparts.
+
+<table><tr><td>Model</td><td>ST</td><td>MMLU</td><td colspan="2">SuperGLUETruthfulQA</td><td>LogiQA</td><td>Avg</td></tr><tr><td>Top-1+FR</td><td>Yes</td><td>34.66</td><td>63.97</td><td>28.51</td><td>29.18</td><td>39.08</td></tr><tr><td>Top-1+FR</td><td>No</td><td>33.96</td><td>62.75</td><td>29.25</td><td>28.57</td><td>38.63</td></tr><tr><td>Top-2</td><td>Yes</td><td>35.39</td><td>64.58</td><td>29.98</td><td>29.33</td><td>39.82</td></tr><tr><td>Top-2</td><td>No</td><td>35.86</td><td>64.73</td><td>29.98</td><td>28.26</td><td>39.70</td></tr></table>
+
+Table 7: The performance of Top-1+FR and Top-2 router with and without straight-through trick. The second column (ST) denotes whether the straight-through trick is used.
+
+# C Analysis
+
+# C.1 Impact of Straight-through Trick
+
+In Appendix A, we propose a solution to address the gradient issue associated with the Fill-in Rectification by utilizing the straight-through trick. To evaluate the effectiveness of this technique, we conducted an experiment comparing the performance of the Fill-in Rectification with versus without the straight-through trick. The results of this comparison are presented in Table 7.
+
+Our findings indicate that the straight-through trick proves to be beneficial in improving the performance of the Fill-in Rectification (Top-1+FR). This suggests that the straight-through trick is necessary for the Fill-in Rectification to achieve optimal results. However, the application of the straight-through trick does not yield a significant improvement in the performance of the top-2 router. This can be attributed to the fact that the proportion of unsuccessful routing is relatively small  $(5\%)$  for the top-2 router, while it is considerably large  $(50\%)$  when employing the Fill-in Rectification.
+
+# D Complete Results of Capacity Factor Variation
+
+In Section 5.5, we have discussed the performance of MoE models across various capacity factor settings. However, it is worth noting that only the average accuracy are reported in Table 3 and Table 4. For a comprehensive overview, we present the complete results in Table 8 and Table 9, which encompass the evaluation outcomes across all benchmarks.
+
+<table><tr><td>Train Router</td><td>Test Router</td><td>Test CF</td><td>Test Speed</td><td>MMLU</td><td>SuperGLUE</td><td>TruthfulQA</td><td>LogiQA</td><td>Avg</td></tr><tr><td rowspan="2">Top-1</td><td>Top-1</td><td rowspan="2">1.0</td><td>9.4k</td><td>33.05</td><td>64.34</td><td>29.49</td><td>28.11</td><td>38.74</td></tr><tr><td>Top-1+IR</td><td>9.2k</td><td>36.21</td><td>64.64</td><td>29.86</td><td>29.18</td><td>39.97</td></tr><tr><td rowspan="2">Top-1</td><td>Top-1</td><td rowspan="2">0.75</td><td>12.1k</td><td>30.88</td><td>62.68</td><td>29.37</td><td>28.41</td><td>37.83</td></tr><tr><td>Top-1+IR</td><td>9.9k</td><td>36.12</td><td>64.61</td><td>29.74</td><td>29.80</td><td>40.06</td></tr><tr><td rowspan="2">Top-1</td><td>Top-1</td><td rowspan="2">0.5</td><td>16k</td><td>26.43</td><td>59.71</td><td>26.68</td><td>26.57</td><td>34.84</td></tr><tr><td>Top-1+IR</td><td>10.6k</td><td>36.32</td><td>65.23</td><td>29.98</td><td>30.10</td><td>40.40</td></tr><tr><td rowspan="2">Top-2</td><td>Top-2</td><td rowspan="2">2.0</td><td>6.2k</td><td>35.39</td><td>64.58</td><td>29.98</td><td>29.33</td><td>39.82</td></tr><tr><td>Top-2+IR</td><td>6.0k</td><td>35.70</td><td>64.40</td><td>30.47</td><td>29.95</td><td>40.13</td></tr><tr><td rowspan="2">Top-2</td><td>Top-2</td><td rowspan="2">1.5</td><td>7.4k</td><td>35.22</td><td>65.00</td><td>30.47</td><td>27.34</td><td>39.50</td></tr><tr><td>Top-2+IR</td><td>6.6k</td><td>35.71</td><td>64.47</td><td>29.98</td><td>28.26</td><td>39.60</td></tr><tr><td rowspan="2">Top-2</td><td>Top-2</td><td rowspan="2">1.0</td><td>8.9k</td><td>33.28</td><td>62.76</td><td>28.51</td><td>29.49</td><td>38.51</td></tr><tr><td>Top-2+IR</td><td>7.3k</td><td>35.93</td><td>64.40</td><td>29.62</td><td>30.10</td><td>40.01</td></tr></table>
+
+Table 8: Performance of top-  $k$  routers and their variants with low capacity factors  $\left( { <  = k}\right)$  . The difference between this table and Table 3 is that the evaluation results on all benchmarks are reported in this table, but only the average accuracy is reported in Table 3  
+
+<table><tr><td>Train Router</td><td>Test Router</td><td>Test CF</td><td>Test Speed</td><td>MMLU</td><td>SuperGLUE</td><td>TruthfulQA</td><td>LogiQA</td><td>Avg</td></tr><tr><td rowspan="6">Top-1</td><td>Top-1</td><td rowspan="2">1.0</td><td>9.4k</td><td>33.05</td><td>64.34</td><td>29.49</td><td>28.11</td><td>38.74</td></tr><tr><td>Top-1+FR</td><td>8.9k</td><td>33.28</td><td>62.76</td><td>28.51</td><td>29.49</td><td>38.51</td></tr><tr><td>Top-1</td><td rowspan="2">1.25</td><td>8.6k</td><td>34.51</td><td>62.94</td><td>29.13</td><td>31.79</td><td>39.59</td></tr><tr><td>Top-1+FR</td><td>8.1k</td><td>35.01</td><td>65.22</td><td>30.23</td><td>29.95</td><td>40.10</td></tr><tr><td>Top-1</td><td rowspan="2">1.5</td><td>7.9k</td><td>36.40</td><td>64.21</td><td>28.88</td><td>29.95</td><td>39.86</td></tr><tr><td>Top-1+FR</td><td>7.3k</td><td>36.10</td><td>64.89</td><td>30.23</td><td>30.10</td><td>40.33</td></tr><tr><td rowspan="6">Top-2</td><td>Top-2</td><td rowspan="2">2.0</td><td>6.2k</td><td>35.39</td><td>64.58</td><td>29.98</td><td>29.33</td><td>39.82</td></tr><tr><td>Top-2+FR</td><td>5.8k</td><td>35.96</td><td>65.37</td><td>30.35</td><td>30.26</td><td>40.48</td></tr><tr><td>Top-2</td><td rowspan="2">2.5</td><td>5.4k</td><td>35.96</td><td>64.53</td><td>30.23</td><td>28.87</td><td>39.89</td></tr><tr><td>Top-2+FR</td><td>5.1k</td><td>36.00</td><td>64.92</td><td>30.59</td><td>30.56</td><td>40.51</td></tr><tr><td>Top-2</td><td rowspan="2">3.0</td><td>4.9k</td><td>35.72</td><td>64.65</td><td>30.59</td><td>29.18</td><td>40.03</td></tr><tr><td>Top-2+FR</td><td>4.5k</td><td>35.98</td><td>65.31</td><td>30.23</td><td>30.26</td><td>40.44</td></tr></table>
+
+Table 9: Performance of top-  $k$  routers and their variants with high capacity factors  $\left( { >  = k}\right)$  . The difference between this table and Table 4 is that the evaluation results on all benchmarks are reported in this table, but only the average accuracy is reported in Table 4

2024/Turn Waste into Worth_ Rectifying Top-k Router of MoE/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f033e8df73a456b89ea1c7ca6e47e66c9756151dec4a2b6f80ddf9ff465cf5d5
+size 659833

2024/Twists, Humps, and Pebbles_ Multilingual Speech Recognition Models Exhibit Gender Performance Gaps/75d52286-5b60-4f2d-833b-11b4a49fb28d_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f12c4705263676fb230a23fe7f688e09d7c7e22702d6d40a86a8edd90da33a50
+size 2307010

2024/Twists, Humps, and Pebbles_ Multilingual Speech Recognition Models Exhibit Gender Performance Gaps/full.md

@@ -0,0 +1,632 @@
+# Twists, Humps, and Pebbles: Multilingual Speech Recognition Models Exhibit Gender Performance Gaps
+
+Giuseppe Attanasio $^{\text{♥}}$ , Beatrice Savoldi $^{\text{♣}}$ , Dennis Fucci $^{\text{◇}\text{♣}}$ , Dirk Hovy $^{\text{♣}}$
+
+$\Leftrightarrow$  Instituto de Telecomunicações, Lisbon, Portugal
+
+$\diamond$  University of Trento, Trento, Italy
+
+Fondazione Bruno Kessler, Trento, Italy
+
+$\clubsuit$  Bocconi University, Milan, Italy
+
+giuseppe.attenasio@lx.it.pt
+
+# Abstract
+
+Current automatic speech recognition (ASR) models are designed to be used across many languages and tasks without substantial changes. However, this broad language coverage hides performance gaps within languages, for example, across genders. Our study systematically evaluates the performance of two widely used multilingual ASR models on three datasets, encompassing 19 languages from eight language families and two speaking conditions. Our findings reveal clear gender disparities, with the advantaged group varying across languages and models. Surprisingly, those gaps are not explained by acoustic or lexical properties. However, probing internal model states reveals a correlation with gendered performance gap. That is, the easier it is to distinguish speaker gender in a language using probes, the more the gap reduces, favoring female speakers. Our results show that gender disparities persist even in state-of-the-art models. Our findings have implications for the improvement of multilingual ASR systems, underscoring the importance of accessibility to training data and nuanced evaluation to predict and mitigate gender gaps. We release all code and artifacts at https://github.com/g8a9/multilingual-asr-gender-gap.
+
+# 1 Introduction
+
+A new class of multi-task, multilingual neural networks (Radford et al., 2022; Communication et al., 2023; Chu et al., 2023) has recently pushed the boundaries of several speech-related tasks, including automatic speech recognition (ASR). As these models offer support to an increasing number of languages at no cost, they have found widespread adoption and have been integrated into applications for the general public, such as real-time voice tran
+
+![](images/d92c105397e3fd16ae5b0e31c0ca88a375193ab48388375b2fb66434f87ca3b6.jpg)  
+Figure 1: Study overview. We transcribe three speech datasets across 19 languages and compute gender performance gaps. Next, we investigate the models for possible causes of these gaps.
+
+scription. $^{1}$  However, while the user base expands, one question has yet to be answered: will usability be the same for everyone? In this paper, we seek an answer to this question by studying how systems understand the voices of different genders and across multiple languages. These two axes of analysis are motivated by complementary interests.
+
+On the one hand, voice and language production are among the strongest identity traits. They vary across individuals, sociodemographic groups (Labov, 1964; Wolfram, 2004; Alim, 2004), and, crucially, across genders. Gendered differences in the voice—rooted in physiological factors linked to biological sex (e.g., vocal tract length), as well as in sociocultural ones (e.g., prescribed registers) (Zimman, 2017, 2020)—have been extensively studied in sociolinguistics (e.g., Coleman, 1976; Busby and Plant, 1995; Hillenbrand and Clark, 2009). Also,
+
+empirical studies have shown that gender affects the performance of traditional ASR systems in English (Garnerin et al., 2019). On the other hand, voice and speech can change drastically across languages, cultures, and in function of gender. For example, males' and females' pitch—a measure of the highness of the voice—is much closer in Japanese- than in English-speaking people (Loveday, 1981; Yuasa, 2008).
+
+Despite a large literature on gender, voice, and ASR, to our knowledge, no studies have tested whether multi-task, multilingual ASR models serve genders to a comparable level across languages. Arguably, overlooking potential disparities can result in unequal service quality for already socially disadvantaged individuals (Mengesha et al., 2021; Tatman, 2017).
+
+To fill this gap, we systematically study gendered performance gaps in massively multilingual ASR models. We set out to answer two research questions. (Q1) Do multi-task, multilingual ASR models perform equally well across speakers identifying as women, men, or neither of the two? If so, (Q2), can we relate gender gaps to acoustic and lexical variation in data or model internal states?
+
+To answer Q1, we evaluate two state-of-the-art multilingual open-weight ASR models on three datasets, covering 19 languages and two speech conditions (i.e. read and spontaneous). The results of our extensive evaluation show that models systematically exhibit gender performance gaps. However, whether models favor male or female speakers depends on the dataset and language. Results on one dataset also highlight subpar performance for speakers who do not identify with either gender compared to males.
+
+While studying acoustic and lexical phenomena in test data (Q2), we found no significant correlations of specific features with performance gaps between male and female speakers. This objective finding underscores the complexity of the issue and the need for further research. Instead, interpretability analyses suggest that when presented with speech from men and women, models build different internal representations that can serve as a proxy for gender disparities.
+
+Contributions. We conduct the first extensive evaluation of two widely used multilingual ASR models for gender performance differences. We document significant gender gaps by inspecting model internal states. We release code, data, and
+
+all artifacts we produce for future research.
+
+Bias Statement. When using gender as a variable, we rely on speakers' declared identity (see §8). We evaluate whether speech from individuals identifying as female, male, or neither of the two is processed equally well by multilingual ASR models. We measure disparities of minority groups versus the socially advantaged group, i.e., men. Performance parity is the ideal outcome. We define a system to be biased if it risks further contributing allocative (i.e., quality of service, technology less accessible) and representational harms impacting the minority groups, e.g., feeding into stereotypes about the inadequacy of women and speech technology or "shrill," "incorrect" voice (Tallon, 2019).
+
+# 2 Background
+
+Multi-task Multilingual Speech Models. Contemporary speech recognition systems process audio and text separately and use common cross-modal interactions—e.g., OpenAI's Whisper (Radford et al., 2022) is loosely inspired by the standard Transformer (Vaswani et al., 2017). Crucially, multitasking and multilinguality come at virtually no cost for the user. With Whisper, for example, special "task" and "language" tokens can be prepended to the decoder input to change its functions. This strategy allows multi-tasking and multilinguality without architectural changes or fine-tuning (see Figure 1, top-right, in Radford et al. (2022)). Similarly, Meta's SeamlessM4T (Communication et al., 2023) uses a speech encoder for audio and an encoder-decoder transformer for text (Costa-jussa et al., 2022, NLLB).
+
+Gender and Speech (Technologies). Gendered aspects of the voice are one of the most salient individual traits (Kreiman and Sirdis, 2011; Azul, 2015; Zimman, 2021) and have long been studied in linguistics (Zimman, 2020). The anatomy and makeup of the vocal tract do play a role, as it determines pitch range and formant variations, often regarded as the most distinctive vocal features of cisgender men and women. However, gender variation in the voice has been shown to arise from several aspects besides physical ones (Oates and
+
+Dacakis, 2015; Zimman, 2018; Becker et al., 2022). As studies including transgender individuals show (Zimman, 2017), sociocultural factors influence vocal use, including which parts of the pitch range are used (Loveday, 1981; Yuasa, 2008), and articulatory practices for sibilant consonants and vowels that are perceived as gender characterizing (Pharaoh et al., 2014; Podesva et al., 2016; Li, 2017).
+
+For speech technologies, sociodemographic variation has posed challenges to ASR systems (Sawalha and Abu Shariah, 2013; Liu et al., 2022; Rajan et al., 2022; Fucci et al., 2023). Various works have found ASR models for English and French to recognize better male speech and voices (Tatman, 2017; Garnerin et al., 2019, 2021). This effect is often a result of the under-representation of women in the training data (Meyer et al., 2020; Gaido et al., 2020; Garnerin et al., 2020). However, various works have found the reverse to be true. Several studies found comparatively better performance for women in more spontaneous, conversational data in Dutch (Feng et al., 2022) and English (Adda-Decker and Lamel, 2005; Koenecke et al., 2020). They attribute these findings to sociolinguistic factors, like the higher incidence of disfluencies and informal speech in men.
+
+Thus, the current literature provides a fragmented picture of gender disparities in speech technologies. However, most studies focus primarily on a single language, usually English, albeit across multiple datasets and models. We broaden this research to include large-scale multilingual speech models. In addition, we evaluate performance in a third gender category, which includes individuals who do not identify as male or female.
+
+# 3 Methodology
+
+This section describes the experimental design to answer Q1, i.e., whether there are gendered performance differences in multi-task multilingual ASR models. Our results (§4) confirm it.
+
+Models. We experiment with OpenAI's Whisper (Radford et al., 2022, WHISPER) and Meta's SeamlessM4T (Communication et al., 2023, SEAMLESS), two widely used, state-of-the-art multilingual ASR models (details in Appendix A.1).
+
+Datasets. Among other multilingual datasets available for ASR (Gales et al., 2014; Black, 2019; Pratap et al., 2020; Iranzo-Sanchez et al., 2022; Valk and Alumae, 2021), we use Mozilla Com
+
+mon Voice (CV, Ardila et al., 2020),  $^{4}$  Fleurs (Connieu et al., 2023) and VoxPopuli (Wang et al., 2021). Mainly, dataset selection is bound to the availability of reliable speakers' gender information. We use the gender labels that come with each dataset. $^{5}$  Moreover, these datasets cover two distinct recording and speech conditions:  $i$  read, where speech is typically well-articulated and based on the reading of pre-defined texts (CV and Fleurs), and  $ii$  spontaneous conditions elicited from public speeches (VoxPopuli), which allow for more speaker-dependent articulations and variation in word usage (Gabler et al., 2023), and thus represent a testbed closer to real-world use cases of ASR technologies. We base our analysis on the concatenation of validation and test splits of each dataset to avoid unreliable results due to training data contamination. See Appendix A.5 for a discussion on data contamination and how transcription performance varies across splits in our setup.
+
+Languages. We include 19 languages from CV and Fleurs, and a subset of those—11 in total, due to data availability—from VoxPopuli. They represent eight diverse language families and data availability conditions (i.e., high-low resource). The dataset choice depended on whether sufficient utterances stratified across genders were available for meaningful comparisons. See Tables 5-7, Appendix A.6 for all the statistics.
+
+# 3.1 Evaluation
+
+Quality Metrics. We use standard Word- and Character-Error Rate (WER and CER, respectively) to evaluate transcription quality. Following Radford et al. (2022), we  $i$  ) report WER for all languages but Yoruba and Japanese—where variability in orthography and ambiguous word units may affect evaluation (Rowlands, 1954; Matsuoka et al., 1997)—for which we use CER,  $ii$  ) transliterate all Russian and Serbian references and hypotheses into Cyrillic, and  $iii$  ) apply the official normalization routines to texts when evaluating Whisper. Moreover, we trim audio to the initial 30 seconds and
+
+filter out dataset noise by removing records with an empty reference and silence in the snippet.8
+
+When evaluated in terms of overall quality, both WHISPER and SEAMLESS showcase competitive results, respectively, on Fleurs (12.68 and 19.87 - avg. 19 langs); CV (17.51 and 16.42 - avg. 19 langs); VP (13.59 and 12.74 - avg. 11 langs). For disaggregated error rate (ER) scores, see Figure 6 in Appendix A.1.
+
+Gap Metrics. Measuring the ASR gender gap is equivalent to seeking what is commonly known in the field of machine learning as group fairness (Chouldechova and Roth, 2020), or, more specifically, demographic parity (Dwork et al., 2012). Conceptually, parity is achieved when a given statistical measure is equal across different groups.
+
+We operationalize demographic parity using the notation from Czarnowska et al. (2021). We introduce a Pairwise Comparison Metric, loosely inspired by the Disparity Score in Gaut et al. (2020), defined as follows:
+
+$$
+E \left(r _ {A}, r _ {B}\right) = 1 0 0 \cdot \left(\phi \left(r _ {A}\right) - \phi \left(r _ {B}\right)\right) / \phi \left(r _ {B}\right) \tag {1}
+$$
+
+where  $r_A$  and  $r_B$  are the set of audio snippets belonging to given groups (e.g., either "male", "female", or "other" in CV), and  $\phi(\cdot)$  is one of the quality metrics defined above (i.e., WER or CER). The ideal score is 0—i.e., the model performs perfectly equally on the two groups.
+
+Significance test. We use a bootstrapped approach (Koehn, 2004; Søgaard et al., 2014) to estimate  $\phi(\cdot)$ . We sample  $40\%$  of the smallest group and the same number of records from the largest group for  $n = 1000$  iterations. We sample stratifying on speakers to avoid skewing the speaker distribution. Stratified, gender-balanced sampling ensures a reliable comparison. We compute  $E(\cdot)$  on the arithmetic means of  $\phi(\cdot)$  across the  $n$  runs and use a two-sided Student's t-test to compute the statistical difference between the means. See Appendix A.5 for more details on comparing sampling performance vs. complete sets.
+
+Note that sampling requires attributing each record to an individual speaker—information not
+
+![](images/2ef2eb47ff20a767bfcc8dd8e3819557d375a56cab9d8577d54d873e8c721be1.jpg)
+
+![](images/e01aa1b2cb2e1f9c7e92495d38a4988eeba8ef77c0a11ea7541fd7c78a8b4ceb.jpg)  
+(a) VoxPopuli
+
+![](images/9f36d5d889ad77d7afa832c005ce89b863372869591acffdc86c0f8eb21a119f.jpg)  
+(b) Fleurs  
+(c) CV  
+Figure 2: Model error rate gap (Eq. 1). Positive values indicate better performance on men, negative ones on women.  $p < 0.05$ .
+
+available in Fleurs. For consistency, we attribute a speaker ID to each record automatically. Specifically, we  $i$  use state-of-the-art speaker verification embedding models to encode each recording, ii) cluster them using HDBSCAN (Campello et al., 2013), and iii) assign a speaker ID to each cluster. See Appendix A.4 for details on the pipeline.
+
+# 4 Gender Performance Gaps
+
+In the following, we report the results of evaluating WHISPER and SEAMLESS, focusing on gender performance disparities. We first investigate whether models are equally able to recognize female and male speakers (§4.1). Then, limited to CV, we repeat the analysis for speakers identifying with neither of the two against males (§4.2).
+
+# 4.1 "Female" - "Male" Gap
+
+Figure 2 reports  $E(r_{F}, r_{M})$ , i.e., the gender gap results for WHISPER and SEAMLESS on each dataset. Concerning our Q1, a broad overview reveals that these models do not perform equally across female and male speakers, often showing a preference for one gender over the other. This gender disparity in multilingual models is a key point of our investigation. Our analysis does not consistently reveal models disadvantageing the feminine group—see values below 0. This finding stands in stark contrast with mounting evidence of a strong masculine bias affecting a wide range of Natural Language Processing (NLP) tasks (e.g., Sun et al., 2019; Stanczak et al., 2022). It is especially noteworthy given the well-known under-representation of women in current resources used for model training (Garnerin et al., 2019; Zanon Boito et al., 2022; Sun and Peng, 2021). We observe that both speech models—albeit to varying degrees—better recognize female speech on average in two out of the three considered datasets (i.e., F1eurs and VP).
+
+Upon closer examination, we also observe substantially different behaviors across datasets. First, despite the higher degree of spoken language variation to be expected in spontaneous recordings from the VP dataset (Figure 2a), results show a comparatively reduced gender gap, with overlapping values between WHISPER and SEAMLESS for most languages (i.e., hu, es, fr, nl, cs, de). Conversely, it is on read resource that we attest to a higher degree of variability across models as well as languages, with occasional wide disparities toward either females or males in CV (2c). Notably, SEAMLESS remains closer to performance equality for most instances on Fleurs (2b), i.e., values between 11 and -10.
+
+As such, in line with previous research (see §2), analysis of (binary) gender disparities for the ASR reveals a complex picture, which we further analyze and unpack in the following sections.
+
+# 4.2 "Other" - "Male" Gap
+
+Restricting our analysis to validation and test sets leaves us with few records for speakers who do not identify as either male or female. As such, we report results for five high-resource languages, Catalan, German, English, Spanish, and French.[10]
+
+Table 1 reports  $E(r_{O}, r_{M})$  for both models. Results suggest that models penalize non-male speech and that gaps are larger than disparities
+
+<table><tr><td></td><td>ca</td><td>de</td><td>en</td><td>es</td><td>fr</td></tr><tr><td>WHISPER</td><td>4.68</td><td>37.20</td><td>28.50</td><td>-2.00</td><td>5.21</td></tr><tr><td>SEAMLESS</td><td>52.76</td><td>17.19</td><td>38.70</td><td>11.37</td><td>-0.50</td></tr></table>
+
+Table 1: Model error rate gap (Eq. 1). Positive values indicate better performance on men, negative ones on "Other."  $p < {0.05}$  .
+
+observed for the female/male groups. Given the limited sample size, we deem this analysis only exploratory. However, it is a valuable first overview of the model's behaviors for gender-non-conforming voices. We underscore the need for more representation of diverse voices and sociodemographic groups in current resources, and for which monitoring fairness in existing speech models remains out of reach. See §7 for an expanded discussion.
+
+# 5 Acoustic and Lexical Analysis
+
+Gendered performance gaps vary across languages and datasets. Here, we dive into a focused analysis on acoustic aspects or lexical phenomena present in the test data (Q2). Due to data availability, we conduct this analysis on the "Female" - "Male" setup.
+
+# 5.1 Acoustic Analysis
+
+Motivated by sociophonetic evidence for gendered differences in speech (§2), we examine voice and language production across speakers of different genders. Specifically, we measure three acoustic features in our evaluation records, and explore whether potential acoustic differences across gender groups relate to gender performance gaps.
+
+We include: i) pitch, measured as the mean of the fundamental frequency values (Hirst and de Looze, 2021), known to vary between biological sexes (Coleman, 1976; Hillenbrand and Clark, 2009) and languages (Loveday, 1981; Yuasa, 2008); ii) intensity, measured as the mean of fundamental frequencies (Pausewang Gelfer and Young, 1997), subject to recording conditions (Maryn and Zarowski, 2015); iii) speaking rate, measured as the number of tokens per minute (Künzel, 2013), subject to language and gender variation (Borsel and Maesschalck, 2008; Coupé et al., 2019) as well as read vs. spontaneous speech (Nakamura et al., 2008).<sup>11</sup>
+
+Findings. We started by exploring whether pitch, intensity, and speaking rate differ between genders in each dataset-language pair. Independent-sample T-tests revealed significant differences for most setups  $(p < 0.05)$ , consistent with previous literature. For example, pitch consistently showed statistically significant differences (see Appendix B.1).
+
+Motivated by these observations, we computed the mean acoustic values of each gender group and correlated it with performance gaps, i.e.,  $E(r_{F}, r_{M})$ , aggregating by dataset and model. No clear trends emerged overall. Sporadically, we found a strong and significant linear correlation, e.g., for SEAMLESS and VoxPopuli where Pearson's rho for pitch is  $-0.83$  (see Appendix B.1 for complete details). Fitting an ordinary least squares (OLS) regressor to predict  $E(r_{F}, r_{M})$  led to similar conclusions, as low  $\mathbb{R}^2$  scores suggest (max: 0.42, average $_{\sigma}$ : 0.24 $_{\pm 0.11}$ ). For a more fine-grained analysis, we repeated the analysis fitting an OLS model to predict sentence-level error rates (i.e.,  $r_{F}, r_{M}$ ) but found low  $\mathbb{R}^2$  scores (max: 0.20, average $_{\sigma}$ : 0.03 $_{\pm 0.04}$ ).
+
+In summary, despite pitch, intensity, and speaking rate vary significantly across gender groups within the same dataset and language, such variation seldom correlates with performance differences. These findings suggest that performance gaps are a complex phenomenon, and our analysis should expand elsewhere to explain it.
+
+# 5.2 Lexical Analysis
+
+In §5.1, we explored acoustic aspects of speech. However, what is uttered can be as crucial as how it is uttered. Indeed, prior work has found that overall speech perception can be impacted by aspects related to lexical and syntactic complexities (van Kni-jff et al., 2018; Carroll and Ruigendijk, 2012). Besides, certain lexical phenomena such as named entities represent a well-known challenge for speech models, especially in multilingual contexts (Gaido et al., 2021). Thus, we focus on speech content and study whether lexical phenomena—as measured in the reference transcripts—explain ASR disparities.
+
+For each record, we counted  $i$  ) the occurrences of part-of-speech tags and  $ii$  ) named entities, and computed  $iii$  ) lexical density (Halliday, 1989). Similarly to acoustic features, we contrast distribution between gender groups (details in Appendix B.2).
+
+Findings. Comparing the distribution of lexical features between female and male speakers, we
+
+found mixed results. Several dataset-language setups have significantly different distributions  $(p < 0.05)$ , but not all. Different distributions are also present in read datasets, suggesting that lexical variability is not controlled when collecting data. We proceeded similarly to the acoustic analysis to verify whether such differences explain error rates. Fitting an OLS model to predict error rates from lexical features only yields low  $\mathbf{R}^2$  scores (max: 0.35, average $_\sigma$ :  $0.09_{\pm 0.08}$ , across all datasets and languages). Moreover, we found no significant (linear) correlation between the difference of group means and  $E(r_F, r_M)$ .
+
+This finding echoes those from acoustic analysis: Lexical phenomena do not explain gender performance gaps in our data. We must research other aspects beyond the lexical content of utterances.
+
+# 6 Probing Gender in ASR Models
+
+The field of natural language processing (NLP) has now established that transformer language models encode syntactic (Hewitt and Manning, 2019), semantic (Tenney et al., 2019), and factual (Petroni et al., 2019; Meng et al., 2022) information in hidden representations. As such, recent work has focused on extracting this information through probes, i.e., supervised classifiers trained on the model's embeddings (Alain and Bengio, 2016; Belinkov and Glass, 2019). Compellingly, some have related extractability of sensitive attributes, e.g., gender, to bias in downstream application (Orgad et al., 2022). Probing is primarily motivated by the risk that models entangle protected attributes and predictions in sensitive use cases (Zhao et al., 2018; Ravfogel et al., 2020).
+
+Motivated by this line of research, we measure gender extractability in ASR models and ask (Q2) whether and to what extent it explains gender-based performance gaps. To our knowledge, ours is the first study of this kind.[12]
+
+Experimental design. We focus on one dataset-model configuration, namely WHISPER and CV. We attach our probes to the model encoder's last layer embeddings and train one distinct probe for every position. We use Logistic Regression and Mini
+
+![](images/ba82401d51c363fe34b65da00e06c672383218871b3eefb46561804b8c6b6683.jpg)  
+Figure 3: F1 Macro (top) and measured code length (bottom) for Logistic Regression and MDL probes, respectively. English, F-M setup, first 10 seconds of context. Actual score (blue pale line), exponential moving average (solid line,  $n = 3$ ), and average length of test snippets (dashed line).
+
+mum Description Length (MDL) (Voita and Titov, 2020) to probe gender in the female-male binary setup (see Appendix B.3 for details).
+
+Findings. Figure 3 shows gender extractability on the English CV split for the F-M setup. Trends are similar across languages (full results in Figure 9a). First, F1 scores suggest that gender extractability is relatively easy. Whisper produces representations of female and male speech that can be easily separated. Second, extractability is not constant over time. It starts high in the first milliseconds, drops during the actual signal (i.e., the speaker is talking), and finally plateaus around the initial value. This finding recalls the "attention sink" theory, implicating that transformers prioritize initial positions for specific tasks (Xiao et al., 2023). Third, the trends observed in MDL probes—where lower codelength indicates an easier task and higher extractability—suggest that simple logistic regression is a robust method, countering the issues identified in NLP (Voita and Titov, 2020).<sup>13</sup>
+
+Correlation with Error Rate. Do gender probing scores tell us something about ASR quality and gender bias (Q2)? To answer this question, we compare F1 scores from logistic regression scores and ASR error rates. We measure  $\mathrm{F1}_S$ , i.e., the average F1 score achieved by probes at the posi
+
+![](images/caf83d798db402a17d79fa602edfbbe2c2785b6129a97e3a5fd2a497ac126017.jpg)  
+Figure 4: Logistic probe  $\mathbf{F1}_S$  vs.  $E(r_{F},r_{M})$ . CV. The purple dashed line is a linear interpolation.
+
+tions corresponding to actual speech, $^{14}$ $\phi(r_{M})$  and  $\phi(r_{F})$  (error rates on male and female segments), and the disparity score  $E(r_{F}, r_{M})$ .
+
+$\mathrm{F1}_S$  has a weak linear correlation with  $\phi (r_M)$  and  $\phi (r_F)$  (Pearson's  $\rho$  is 0.13 and -0.003, respectively), hinting that  $\mathrm{F1}_S$  cannot explain per-group quality. However,  $\rho$  between  $\mathrm{F1}_S$  and  $E(r_{F},r_{M})$  is -0.65 ( $n = 19$ , see Figure 4). In other words, linear correlation suggests that the better probes can extract gender from WHISPER's hidden states, the lower the F-M error rate gap is. Note that, as per its definition,  $E(r_{F},r_{F})$  can be negative (see points below the zero line in Figure 4). Negative values indicate performance favoring the minority group (here, women). This finding suggests that WHISPER does encode recordings from speakers that identify as men and women differently. This aspect can serve as a proxy for measuring and mitigating gender disparities, e.g. by reducing gender extractability from hidden representations (Krishnan et al., 2024). However, we caution the reader from attributing high extractability to gendered voice and discourage using probes on people's voice to predict gender (see Ethical considerations, §8).
+
+# 7 Discussion
+
+Twists, humps, and pebbles—the way to equitable multilingual speech recognition models is no straight line. We have discovered that gender disparities vary across spoken setups and languages and that consistency is also weak between models. Perhaps of greater interest, the advantaged group is sometimes women, sometimes men, and never
+
+![](images/c4d3d09b5181d2328bccab47a2d3c802e4e00180072bb4aeba3b3b0aab6e0ffc.jpg)  
+Figure 5: Share of records (y axis, %) having a WER greater than the value on the x axis. WHISPER, Fleurs nl (top) and ar (bottom).
+
+anyone who does not identify with either.
+
+Such a high variability prevents us from foreseeing gender disparities and, in turn, the actual impact on system users. We discuss here how to  $i)$  reduce such variability and  $ii)$  better use the metrics and test data at our disposal.
+
+Free! the training data. While our acoustic and lexical inquiries on test data were inconsequential, experiments with probes on the model's internals showed that models learned and encoded properties to differentiate genders. Intuitively, a follow-up study would focus on what shaped those model's properties in the first place: training data. Models encode facts and information from training data, and some studies looked into how different sociodemographic groups are represented (Muller et al., 2023; Elazar et al., 2024). Studying gender distributions in pretraining multilingual speech data would be extremely valuable, as we could establish how gendered gaps and underrepresentation correlate. However, this analysis was impossible, as neither Whisper (Radford et al., 2022) nor SeamlessM4T (Communication et al., 2023) released such information. We call for more transparency.
+
+Go beyond group-wise metrics. A crucial choice in our design was measuring differences between two groups of speakers. We opted for  $E(\cdot)$ , measured as the relative difference between the group means. While the metric was primarily meant to make results across datasets, languages, and models comparable, it does not provide a complete picture of gender disparity's concrete impact. For instance, it hides absolute gaps: a  $E(\cdot) = 10$
+
+gap corresponds to a 0.2 WER gap if the model averages a WER of 2 on male speakers and to a 2 WER gap if the average is 20. These gaps may or may not be significant given many contextual factors (e.g., what the model is being used for, who is using it, and which ramifications such gaps can have). This thinking echoes that of recent critiques on measuring social bias in NLP, advocating for explicit statements about what system behavior must be considered good or bad and under which social values (Blodgett et al., 2020), and how to build contextualized metrics (Lum et al., 2024).
+
+We would like a metric to indicate a more precise estimation of ASR model failures and their severity for users. A step in this direction would be going beyond group-wise metrics and studying sentence-level errors. Drawing on the concepts of wealth distribution and inequality (O'sullivan et al., 2003), we can inspect error distributions between speakers and groups and address the question "assuming  $X$  is an acceptable error rate, what fraction of records from each group will be served to a satisfactory level?" Following Koenecke et al. (2020), we report in Figure 5 the proportion of records having an error rate at least that large. Due to space constraints, we focus on WHISPER's transcriptions of Fleur's Arabic and Dutch data. This visualization lets us identify where performance between genders starts to diverge. If, given our use case, we hypothesize that transcriptions become unusable for  $\mathrm{WER} > 20$ , men and women would be equally affected in Dutch. However, it would not be the case in Arabic where  $40\%$  of records from female speakers, but half of those from males, would not be acceptable. Parity is reached around  $\mathrm{WER} = 40$ .
+
+This concrete analysis highlighted gender disparities that group-wise metrics could not capture otherwise. It also served as an example of a broader argumentation: Future inquiries can benefit from using metrics and observations more closely related to real-world scenarios.
+
+Improve sociodemographic representation. Empirical studies like ours rely on the quality and quantity of evaluation data. To test speech recognition, we need representative splits from diverse speakers. However, current datasets are gender-skewed, with men often overrepresented in terms of records and speakers. While downsampling can help at evaluation time, it does not address the lack of speaker diversity. Skewed distributions are even more evident in
+
+low-resource languages when building standard train/Validation/test sets is required—e.g., in CV, Yoruba counts only 14 and 20 unique males and females, respectively. Catalan counts only 8 within "Other."
+
+Therefore, we call for much needed considerations when comparing gender groups in high-resource languages, and new collective efforts to collect more representative data for low-resource languages and gender identities beyond the binary.
+
+# 8 Conclusion
+
+We conducted the first extensive evaluation on gender-based performance gaps of Whisper (Radford et al., 2022) and SeamlessM4T (Communication et al., 2023) for ASR. These models consistently exhibit gender bias across 19 languages from eight language families. Depending on the language, disparities can favor men or women but rarely favor speakers who identify as neither. We locate a potential source of these gaps using gender probes from interpretability approaches. Our results show that probes can be a proxy for gender gaps and that group fairness in multi-task multilingual models remains unsolved.
+
+# Ethical Considerations
+
+The use of gender as a variable in this paper warrants ethical reflections.
+
+As one of the most salient perceptual traits of one's identity (Kreiman and Sirdis, 2011; Azul, 2015; Zimman, 2021) gendered differences represent a linchpin of much (socio)phonetic research (Zimman, 2020), which has unpacked several physical and sociocultural factors contributing to such gendered characteristics in the voice. Based on this evidence, our work does not intend to be normative nor assumes the existence of a single, unidimensional "female" or "male" voice. Instead, we question whether different gendered groups are equally recognized by current multilingual ASR models and incorporate sociophonetic knowledge in our analysis and discussions to isolate why that might not be the case. To do so, we do not make any inference about the gender of the speakers in the employed data. Instead, we rely on the declared gender of the speakers in the employed speech resources. In this regard, part of our analysis of the Mozilla Common Voice dataset uses a third category, "Other", which potentially aggregates diverse identities (e.g., transgender, non-binary, and other
+
+marginalized individuals) $^{15}$  under one single umbrella term. While this third category also allows us to include genders non-conforming to the binary in our study, we recognize that this label and category might be an oversimplification and risk erasing the experiences and representativity of many gender identities.
+
+Finally, gender probes represent a methodological approach to studying how models encode different audios from input signals. It is, hence, essential to remember the inherent limitations and ethical implications of relying on probes for gender classification. While these tools may offer a convenient means of categorization and can be suitable for exploring the models' behavior, they often overlook the nuanced and multifaceted nature of gender identity.
+
+# Limitations
+
+Our paper comes with a series of limitations. We divide them into two categories: data and methods.
+
+Data. Gender is the driving variable of our analysis. However, it is widely recognized that gender interacts with other sociocultural factors, e.g., dialect (Wolfram, 2004) or sexual orientation (Zimman, 2013), in voice production. By limiting to self-identified gender provided with the datasets, our analysis can only provide a partial view. We thus advocate for speech dataset releases that include a principled set of such factors.
+
+The scarcity of data limits the generalizability of our results in two more setups in our analysis. First, the category "Other" counts a deficient number of speakers and records in most of our setups (see §7). Second, the number of speakers we extracted from Fleurs is compared to CV and VoxPopuli (see 3.1). These factors hamper cross-dataset comparison and overall generalizability.
+
+We did not control for data quality on references. It might be that references are noisy (e.g., single words, empty) and can lead to over- or underestimation of our measurements.
+
+Finally, data contamination. Although we chose validation and test sets as our analysis targets, we cannot exclude that models were trained on part or all of them. We conducted a side analysis (Appendix A.5) that suggested that this might not be the case.
+
+Methods. Since we miss reliable multilingual tools for acoustic analyses, we estimated the speaking rate using Whisper's pre-trained tokenizer. We acknowledge that pretrained tokenizers yield fewer tokens for high-resource languages. Hence, our measurement of low-resource might be overestimating the phenomenon.
+
+If framed in the context of bias evaluation paradigms—which we are currently not doing—probing can be seen as an intrinsic bias paradigm. We correlate probing performance to a downstream task (and potential harm), i.e., quality gaps in ASR. However, studies in the field of NLP recognize that intrinsic and extrinsic (downstream) bias metrics do not necessarily correlate (Goldfarb-Tarrant et al., 2021; Kaneko et al., 2022). To our knowledge, ours is the first study to inspect such aspects in transformers for speech, and findings may not hold across modalities. We leave this research question to future work.
+
+# Acknowledgments
+
+We thank the reviewers, the members of the SARDINE and MilaNLP research groups, and Marco Gaido for the insightful comments. Giuseppe Attanasio was supported by the Portuguese Recovery and Resilience Plan through project C645008882-00000055 (Center for Responsible AI) and by Fundação para a Ciência e Tecnologia through contract UIDB/50008/2020. He conducted part of the work as a member of the MilaNLP group at Bocconi University, Milan. Dirk Hovy was supported by the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation program (grant agreement No. 949944, INTEGRATOR) and a MUR FARE 2020 initiative under grant agreement Prot. R20YSMBZ8S (INDOMITA). He is director of the Data and Marketing Insights Unit of the Bocconi Institute for Data Science and Analysis (BIDSA). Beatrice Savoldi was supported by the PNRR project FAIR - Future AI Research (PE00000013), under the NRRP MUR program funded by the NextGenerationEU.
+
+# References
+
+Martine Adda-Decker and Lori Lamel. 2005. Do speech recognizers prefer female speakers? In Proc. Interspeech 2005, pages 2205-2208.  
+Guillaume Alain and Yoshua Bengio. 2016. Understanding intermediate layers using linear classifier probes. arXiv preprint arXiv:1610.01644.
+
+H Samy Alim. 2004. You know my steez: An ethnographic and sociolinguistic study of styleshifting in a Black American speech community. Stanford University.  
+Rosana Ardila, Megan Branson, Kelly Davis, Michael Kohler, Josh Meyer, Michael Henrtty, Reuben Morais, Lindsay Saunders, Francis Tyers, and Gregor Weber. 2020. Common voice: A massively-multilingual speech corpus. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 4218-4222.  
+David Azul. 2015. On the varied and complex factors affecting gender diverse people's vocal situations: Implications for clinical practice. Perspectives on Voice and Voice Disorders, 25(2):75-86.  
+Kara Becker, Sameer ud Dowla Khan, and Lal Zimman. 2022. Beyond binary gender: creaky voice, gender, and the variationist enterprise. Language Variation and Change, 34(2):215-238.  
+Yonatan Belinkov and James Glass. 2019. Analysis methods in neural language processing: A survey. Transactions of the Association for Computational Linguistics, 7:49-72.  
+Michael Beukman and Manuel Fokam. 2023. Analysing cross-lingual transfer in low-resourced African named entity recognition. In Proceedings of the 13th International Joint Conference on Natural Language Processing and the 3rd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 199–224, Nusa Dua, Bali. Association for Computational Linguistics.  
+Alan W Black. 2019. Cmu wilderness multilingual speech dataset. In ICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5971-5975.  
+Su Lin Blodgett, Solon Barocas, Hal Daumé III, and Hanna Wallach. 2020. Language (technology) is power: A critical survey of "bias" in NLP. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5454-5476, Online. Association for Computational Linguistics.  
+Paul Boersma and David Weenink. 2001. Praat, a system for doing phonetics by computer. *Glot International*, 5(9/10):341-345.  
+John Van Borsel and Dorothy De Maesschalck. 2008. Speech rate in males, females, and male-to-female transsexuals. Clinical Linguistics & Phonetics, 22(9):679-685. PMID: 18608249.  
+Peter A Busby and Geoff L Plant. 1995. Formant frequency values of vowels produced by preadolescent boys and girls. The Journal of the Acoustical Society of America, 97(4):2603-2606.
+
+Ricardo JGB Campello, Davoud Moulavi, and Jörg Sander. 2013. Density-based clustering based on hierarchical density estimates. In Pacific-Asia conference on knowledge discovery and data mining, pages 160-172. Springer.  
+Rebecca Carroll and Esther Ruigendijk. 2012. The effects of syntactic complexity on processing sentences in noise. Journal of Psycholinguistic Research, 42:139 - 159.  
+Alexandra Chouldechova and Aaron Roth. 2020. A snapshot of the frontiers of fairness in machine learning. Communications of the ACM, 63(5):82-89.  
+Yunfei Chu, Jin Xu, Xiaohuan Zhou, Qian Yang, Shiliang Zhang, Zhijie Yan, Chang Zhou, and Jingren Zhou. 2023. Qwen-audio: Advancing universal audio understanding via unified large-scale audiolanguage models. arXiv preprint arXiv:2311.07919.  
+Ralph O. Coleman. 1976. A comparison of the contributions of two voice quality characteristics to the perception of maleness and femaleness in the voice. Journal of Speech & Hearing Research, 19(1):168-180.  
+Seamless Communication, Loic Barrault, Yu-An Chung, Mariano Coria Meglioli, David Dale, Ning Dong, Mark Duppenthaler, Paul-Ambroise Duquenne, Brian Ellis, Hady Elsahar, Justin Haaheim, John Hoffman, Min-Jae Hwang, Hirofumi Inaguma, Christopher Klaiber, Ilia Kulikov, Pengwei Li, Daniel Licht, Jean Maillard, Ruslan Mavlyutov, Alice Rakotoarison, Kaushik Ram Sadagopan, Abinesh Ramakrishnan, Tuan Tran, Guillaume Wenzek, Yilin Yang, Ethan Ye, Ivan Evtimov, Pierre Fernandez, Cynthia Gao, Prangthip Hansanti, Elahe Kalbassi, Amanda Kallet, Artyom Kozhevnikov, Gabriel Mejia Gonzalez, Robin San Roman, Christophe Touret, Corinne Wong, Carleigh Wood, Bokai Yu, Pierre Andrews, Can Balioglu, Peng-Jen Chen, Marta R. Costa-jussa, Maha Elbayad, Hongyu Gong, Francisco Guzmán, Kevin Heffernan, Somya Jain, Justine Kao, Ann Lee, Xutai Ma, Alex Mourachko, Benjamin Peloquin, Juan Pino, Sravya Popuri, Christophe Ropers, Safiyyah Saleem, Holger Schwenk, Anna Sun, Paden Tomasello, Changhan Wang, Jeff Wang, Skyler Wang and Mary Williamson. 2023. Seamless: Multilingual Expressive and Streaming Speech Translation.  
+Alexis Conneau, Min Ma, Simran Khanuja, Yu Zhang, Vera Axelrod, Siddharth Dalmia, Jason Riesa, Clara Rivera, and Ankur Bapna. 2023. Fleurs: Few-shot learning evaluation of universal representations of speech. In 2022 IEEE Spoken Language Technology Workshop (SLT), pages 798-805. IEEE.  
+Marta R Costa-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, et al. 2022. No language left behind: Scaling human-centered machine translation. arXiv preprint arXiv:2207.04672.
+
+Christophe Coupe, Yoon Mi Oh, Dan Dediu, and François Pellegrino. 2019. Different languages, similar encoding efficiency: Comparable information rates across the human communicative niche. Science Advances, 5(9):eaaw2594.  
+Paula Czarnowska, Yogarshi Vyas, and Kashif Shah. 2021. Quantifying social biases in NLP: A generalization and empirical comparison of extrinsic fairness metrics. Transactions of the Association for Computational Linguistics, 9:1249-1267.  
+Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Reingold, and Richard Zemel. 2012. Fairness through awareness. In Proceedings of the 3rd innovations in theoretical computer science conference, pages 214-226.  
+Yanai Elazar, Akshita Bhagia, Ian Helgi Magnusson, Abhilasha Ravichander, Dustin Schwenk, Alane Suhr, Evan Pete Walsh, Dirk Groeneveld, Luca Soldaini, Sameer Singh, Hannaneh Hajishirzi, Noah A. Smith, and Jesse Dodge. 2024. What's in my big data? In The Twelfth International Conference on Learning Representations.  
+Rui Feng, Chen Luo, Qingyu Yin, Bing Yin, Tuo Zhao, and Chao Zhang. 2022. CERES: Pretraining of graph-conditioned transformer for semi-structured session data. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 219–230, Seattle, United States. Association for Computational Linguistics.  
+Dennis Fucci, Marco Gaido, Matteo Negri, Mauro Cettolo, and Luisa Bentivogli. 2023. No pitch left behind: Addressing gender unbalance in automatic speech recognition through pitch manipulation. arXiv preprint arXiv:2310.06590.  
+Susanne Fuchs and Martine Toda. 2010. Do differences in male versus female/s/reflect biological or sociophonetic factors. *Turbulent sounds: An interdisciplinary guide*, 21:281-302.  
+Philipp Gabler, Bernhard C Geiger, Barbara Schuppler, and Roman Kern. 2023. Reconsidering Read and Spontaneous Speech: Causal Perspectives on the Generation of Training Data for Automatic Speech Recognition. Information, 14(2):137.  
+Marco Gaido, Susana Rodríguez, Matteo Negri, Luisa Bentivogli, and Marco Turchi. 2021. Is "moby dick" a whale or a bird? named entities and terminology in speech translation. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1707-1716.  
+Marco Gaido, Beatrice Savoldi, Luisa Bentivogli, Matteo Negri, and Marco Turchi. 2020. Breeding gender-aware direct speech translation systems. In Proceedings of the 28th International Conference on Computational Linguistics, pages 3951-3964, Barcelona, Spain (Online). International Committee on Computational Linguistics.
+
+Mark JF Gales, Kate M Knill, Anton Ragni, and Shakti P Rath. 2014. Speech recognition and keyword spotting for low-resource languages: Babel project research at cued. In Fourth International workshop on spoken language technologies for under-resourced languages (SLTU-2014), pages 16-23. International Speech Communication Association (ISCA).  
+Mahault Garnerin, Solange Rossato, and Laurent Be-sacier. 2019. Gender Representation in French Broadcast Corpora and Its Impact on ASR Performance. In 1st International Workshop on AI for Smart TV Content Production, Access and Delivery, page 3-9.  
+Mahault Garnerin, Solange Rossato, and Laurent Besacier. 2020. Gender representation in open source speech resources. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 6599-6605, Marseille, France. European Language Resources Association.  
+Mahault Garnerin, Solange Rossato, and Laurent Be-sacier. 2021. Investigating the Impact of Gender Representation in ASR Training Data: a Case Study on Librispeech. In 3rd Workshop on Gender Bias in Natural Language Processing, pages 86-92.  
+Andrew Gaut, Tony Sun, Shirlyn Tang, Yuxin Huang, Jing Qian, Mai ElSherief, Jieyu Zhao, Diba Mirza, Elizabeth Belding, Kai-Wei Chang, and William Yang Wang. 2020. Towards understanding gender bias in relation extraction. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2943-2953, Online. Association for Computational Linguistics.  
+Seraphina Goldfarb-Tarrant, Rebecca Marchant, Ricardo Muñoz Sánchez, Mugdha Pandya, and Adam Lopez. 2021. Intrinsic bias metrics do not correlate with application bias. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1926-1940, Online. Association for Computational Linguistics.  
+Michael A.K. Halliday. 1989. Spoken and written language.  
+John Hewitt and Christopher D. Manning. 2019. A structural probe for finding syntax in word representations. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4129-4138, Minneapolis, Minnesota. Association for Computational Linguistics.  
+James M. Hillenbrand and Michael J. Clark. 2009. The role of f0 and formant frequencies in distinguishing the voices of men and women. Attention Perception & Psychophysics, 71(5):1150-1166.  
+Daniel J. Hirst and Céline de Looze. 2021. Measuring Speech. Fundamental frequency and pitch. In
+
+Rachael-Anne Knight and Jane Setter, editors, Cambridge Handbook of Phonetics, 1, pages 336-361. Cambridge University Press.  
+Aliakbar Imani and Hadina Habil. 2017. Lexical features of academic writing.  
+Javier Iranzo-Sanchez, Javier Jorge Cano, Alejandro Pérez-González-de Martos, Adrián Giménez Pastor, Gonçal Garcés Díaz-Munio, Pau Baquero-Arnal, Joan Albert Silvestre-Cerdà, Jorge Civera Saiz, Albert Sanchis, and Alfons Juan. 2022. MLLP-VRAIN UPV systems for the IWSLT 2022 simultaneous speech translation and speech-to-speech translation tasks. In Proceedings of the 19th International Conference on Spoken Language Translation (IWSLT 2022), pages 255-264, Dublin, Ireland (in-person and online). Association for Computational Linguistics.  
+Masahiro Kaneko, Danushka Bollegala, and Naoaki Okazaki. 2022. Debiasing isn't enough! - on the effectiveness of debiasing MLMs and their social biases in downstream tasks. In Proceedings of the 29th International Conference on Computational Linguistics, pages 1299-1310, Gyeongju, Republic of Korea. International Committee on Computational Linguistics.  
+Philipp Koehn. 2004. Statistical significance tests for machine translation evaluation. In Proceedings of the 2004 Conference on Empirical Methods in Natural Language Processing, pages 388-395, Barcelona, Spain. Association for Computational Linguistics.  
+Allison Koenecke, Andrew Nam, Emily Lake, Joe Nudell, Minnie Quartey, Zion Mengesha, Connor Toups, John R. Rickford, Dan Jurafsky, and Sharad Goel. 2020. Racial disparities in automated speech recognition. Proceedings of the National Academy of Sciences, 117(14):7684-7689.  
+Jody Kreiman and Diana Sirdis. 2011. Foundations of voice studies: An interdisciplinary approach to voice production and perception. John Wiley & Sons.  
+Aravind Krishnan, Badr M. Abdullah, and Dietrich Klakow. 2024. On the encoding of gender in transformer-based asr representations.  
+Hermann J. Künzel. 2013. Some general phonetic and forensic aspects of speaking tempo. International Journal of Speech Language and The Law, 4:48-83.  
+William Labov. 1964. The social stratification of English in New York city. Ph.D. thesis, Columbia University.  
+Fangfang Li. 2017. The development of gender-specific patterns in the production of voiceless sibilant fricatives in Mandarin Chinese. Linguistics, 55(5):1021-1044.  
+Chunxi Liu, Michael Picheny, Leda Sarı, Pooja Chitkara, Alex Xiao, Xiaohui Zhang, Mark Chou, Andres Alvarado, Caner Hazirbas, and Yatharth Saraf. 2022.
+
+Towards measuring fairness in speech recognition: Casual conversations dataset transcriptions. In ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6162-6166.  
+Leo Loveday. 1981. Pitch, politeness and sexual role: An exploratory investigation into the pitch correlates of English and Japanese politeness formulae. Language and Speech, 24(1):71-89.  
+Kristian Lum, Macy Reese Anthis, Chirag Nagpal, and Alexander D'Amour. 2024. Bias in language models: Beyond trick tests and toward ruted evaluation. arXiv preprint arXiv:2402.12649.  
+Youri Maryn and Andrzej Zarowski. 2015. Calibration of clinical audio recording and analysis systems for sound intensity measurement. American Journal of Speech-Language Pathology, 24(4):608-618.  
+Tatsuo Matsuoka, Katsutoshi Ohtsuki, Takeshi Mori, Kotaro Yoshida, Sadaoki Furui, and Katsuhiko Shirai. 1997. Japanese large-vocabulary continuous-speech recognition using a business-newspaper corpus. In 1997 IEEE International Conference on Acoustics, Speech, and Signal Processing, volume 3, pages 1803-1806. IEEE.  
+Leland McInnes, John Healy, and James Melville. 2018. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426.  
+Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022. Locating and editing factual associations in gpt. Advances in Neural Information Processing Systems, 35:17359-17372.  
+Zion Mengesha, Courtney Heldreth, Michal Lahav, Juliana Sublewski, and Elyse Tuennerman. 2021. "I don't Think These Devices are Very Culturally Sensitive."—Impact of Automated Speech Recognition Errors on African Americans. Frontiers in Artificial Intelligence, 4.  
+Josh Meyer, Lindy Rauchenstein, Joshua D Eisenberg, and Nicholas Howell. 2020. Artie bias corpus: An open dataset for detecting demographic bias in speech applications. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 6462-6468.  
+Hosein Mohebbi, Grzegorz Chrupa, Willem Zuidema, and Afra Alishahi. 2023. Homophone disambiguation reveals patterns of context mixing in speech transformers. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 8249-8260, Singapore. Association for Computational Linguistics.  
+Benjamin Muller, Belen Alastruey, Prangthip Hansanti, Elahe Kalbassi, Christophe Ropers, Eric Smith, Adina Williams, Luke Zettlemoyer, Pierre Andrews, and Marta R. Costa-jussa. 2023. The gender-GAP
+
+pipeline: A gender-aware polyglot pipeline for gender characterisation in 55 languages. In Proceedings of the Eighth Conference on Machine Translation, pages 536-550, Singapore. Association for Computational Linguistics.  
+Masanobu Nakamura, Koji Iwano, and Sadaoki Furui. 2008. Differences between acoustic characteristics of spontaneous and read speech and their effects on speech recognition performance. Computer Speech & Language, 22(2):171-184.  
+Jennifer Oates and Georgia Dacakis. 2015. Transgender voice and communication: Research evidence underpinning voice intervention for male-to-female transsexual women. Perspectives on Voice and Voice Disorders, 25(2):48-58.  
+Hadas Orgad, Seraphina Goldfarb-Tarrant, and Yonatan Belinkov. 2022. How gender debiasing affects internal model representations, and why it matters. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2602-2628, Seattle, United States. Association for Computational Linguistics.  
+Arthur O'sullivan, Steven M Sheffrin, and Kathy Swan. 2003. Economics: Principles in action.  
+Eliana Pastor, Algis Koudounas, Giuseppe Attanasio, Dirk Hovy, and Elena Baralis. 2023. Explaining speech classification models via word-level audio segments and paralinguistic features. arXiv preprint arXiv:2309.07733.  
+Marylou Pausewang Gelfer and Shannon Ryan Young. 1997. Comparisons of intensity measures and their stability in male and female sneakers. Journal of Voice, 11(2):178-186.  
+Fabio Petroni, Tim Rocttäschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. 2019. Language models as knowledge bases? In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2463-2473, Hong Kong, China. Association for Computational Linguistics.  
+Nicolai Pharaoh, Marie Maegaard, Janus Spindler Møller, and Tore Kristiansen. 2014. Indexical meanings of [s] among Copenhagen youth: Social perception of a phonetic variant in different prosodic contexts. Language in Society, 43(1):1-31.  
+Robert J. Podesva, Janneke Van Hofwegen, Erez Levon, and Ronald Beline Mendes. 2016. S/exuality in small-town California: Gender normativity and the acoustic realization of/s. Language, sexuality, and power: Studies in intersectional linguistics, pages 16-88.
+
+Vineel Pratap, Qiantong Xu, Anuroop Sriram, Gabriel Synnaeve, and Ronan Collobert. 2020. Mls: A large-scale multilingual dataset for speech research. *ArXiv*, abs/2012.03411.  
+Peng Qi, Yuhao Zhang, Yuhui Zhang, Jason Bolton, and Christopher D. Manning. 2020. Stanza: A python natural language processing toolkit for many human languages. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 101-108, Online. Association for Computational Linguistics.  
+Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2022. Robust Speech Recognition via Large-Scale Weak Supervision. ArXiv:2212.04356 [cs, eess].  
+Sai Sathiesh Rajan, Sakshi Udeshi, and Sudipta Chattopadhyay. 2022. Aquevox: Automated fairness testing of speech recognition systems. In International Conference on Fundamental Approaches to Software Engineering, pages 245-267. Springer International Publishing Cham.  
+Mirco Ravanelli, Titouan Parcollet, Peter Plantinga, Aku Rouhe, Samuele Cornell, Loren Lugosch, Cem Subakan, Nauman Dawalatabad, Abdelwahab Heba, Jianyuan Zhong, Ju-Chieh Chou, Sung-Lin Yeh, Szu-Wei Fu, Chien-Feng Liao, Elena Rastorgueva, François Grondin, William Aris, Hwidong Na, Yan Gao, Renato De Mori, and Yoshua Bengio. 2021. SpeechBrain: A general-purpose speech toolkit. ArXiv:2106.04624.  
+Shauli Ravfogel, Yanai Elazar, Hila Gonen, Michael Twiton, and Yoav Goldberg. 2020. Null it out: Guarding protected attributes by iterative nullspace projection. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7237-7256, Online. Association for Computational Linguistics.  
+Evan C Rowlands. 1954. Types of word junction in yoruba. Bulletin of the School of Oriental and African Studies, 16(2):376-388.  
+Majdi Sawalha and Mohammad Abu Shariah. 2013. The effects of speakers' gender, age, and region on overall performance of Arabic automatic speech recognition systems using the phonetically rich and balanced Modern Standard Arabic speech corpus. In Proceedings of the 2nd Workshop of Arabic Corpus Linguistics WACL-2. Leeds.  
+Adrian P. Simpson. 2009. Phonetic differences between male and female speech. Language and linguistics compass, 3(2):621-640.  
+Anders Søgaard, Anders Johannsen, Barbara Plank, Dirk Hovy, and Hector Martínez Alonso. 2014. What's in a p-value in NLP? In Proceedings of the Eighteenth Conference on Computational Natural Language Learning, pages 1-10, Ann Arbor, Michigan. Association for Computational Linguistics.
+
+Karolina Stanczak, Edoardo Ponti, Lucas Torroba Hennigen, Ryan Cotterell, and Isabelle Augenstein. 2022. Same neurons, different languages: Probing morphosyntax in multilingual pre-trained models. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1589-1598, Seattle, United States. Association for Computational Linguistics.  
+Jiao Sun and Nanyun Peng. 2021. Men are elected, women are married: Events gender bias on wikipedia. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 350-360.  
+Tony Sun, Andrew Gaut, Shirlyn Tang, Yuxin Huang, Mai ElSherief, Jieyu Zhao, Diba Mirza, Elizabeth Belding, Kai-Wei Chang, and William Yang Wang. 2019. Mitigating gender bias in natural language processing: Literature review. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1630–1640, Florence, Italy. Association for Computational Linguistics.  
+Tina Tallon. 2019. A Century of "Shrill": How Bias in Technology Has Hurt Women's Voices. The New Yorker.  
+Rachael Tatman. 2017. Gender and dialect bias in YouTube's automatic captions. In Proceedings of the First ACL Workshop on Ethics in Natural Language Processing, pages 53-59, Valencia, Spain. Association for Computational Linguistics.  
+Ian Tenney, Dipanjan Das, and Ellie Pavlick. 2019. BERT rediscovers the classical NLP pipeline. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4593-4601, Florence, Italy. Association for Computational Linguistics.  
+Jörgen Valk and Tanel Alumäe. 2021. Voxlingua107: a dataset for spoken language recognition. In 2021 IEEE Spoken Language Technology Workshop (SLT), pages 652-658. IEEE.  
+Eline C van Knijff, Martine Coene, and Paul J. Govaerts. 2018. Speech understanding in noise in elderly adults: the effect of inhibitory control and syntactic complexity. International journal of language & communication disorders, 53 3:628-642.  
+Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30.  
+Elena Voita and Ivan Titov. 2020. Information-theoretic probing with minimum description length. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 183-196, Online. Association for Computational Linguistics.
+
+Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, and Emmanuel Dupoux. 2021. VoxPopuli: A large-scale multilingual speech corpus for representation learning, semi-supervised learning and interpretation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 993-1003, Online. Association for Computational Linguistics.
+
+Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics.
+
+Walt Wolfram. 2004. Urban african american vernacular english: Morphology and syntax. A handbook of varieties of English, 1:319-340.
+
+Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. 2023. Efficient streaming language models with attention sinks. arXiv preprint arXiv:2309.17453.
+
+Ikuko Patricia Yuasa. 2008. Culture and gender of voice pitch: A sociophonetic comparison of the Japanese and Americans. Equinox Publishing.
+
+Marcely Zanon Boito, Fethi Bouguares, Florentin Barbier, Souhir Gahbiche, Loic Barrault, Mickael Rouvier, and Yannick Esteve. 2022. Speech resources in the Tamasheq language. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 2066-2071, Marseille, France. European Language Resources Association.
+
+Jieyu Zhao, Yichao Zhou, Zeyu Li, Wei Wang, and Kai-Wei Chang. 2018. Learning gender-neutral word embeddings. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 4847-4853, Brussels, Belgium. Association for Computational Linguistics.
+
+Lal Zimman. 2013. Hegemonic masculinity and the variability of gay-sounding speech: The perceived sexuality of transgender men. Journal of Language and Sexuality, 2(1):1-39.
+
+Lal Zimman. 2017. Gender as stylistic bricolage: Transmasculine voices and the relationship between fundamental frequency and/s. Language in Society, 46(3):339-370.
+
+Lal Zimman. 2018. Transgender voices: Insights on identity, embodiment, and the gender of the voice. Language and Linguistics Compass, 12(8):e12284.
+
+Lal Zimman. 2020. Sociophonetics. The International Encyclopedia of Linguistic Anthropology, pages 1-5.
+
+Lal Zimman. 2021. Gender diversity and the voice. In *The Routledge handbook of language, gender, and sexuality*, pages 69–90. Routledge.
+
+# A Experimental Details
+
+# A.1 Multilingual ASR Models
+
+For both Whisper and SeamlessM4T, we used code and model checkpoints in transformers (Wolf et al., 2020). The Hub's model IDs are openai/whisper-large-v3 and facebook/seamless-m4t-v2-large, respectively. Both checkpoints correspond to the latest and best-performing versions available at the time of writing, February 2024.
+
+We used each models's standard decoding configurations to transcribe each audio snippet.
+
+In figure 6, we report WER results for SEAMLESS and WHISPER on the languages and datasets used in our experiments.
+
+# A.2 Languages
+
+We studied 19 languages in CV and Fleurs and 11 in VoxPopuli. Languages cover eight distinct language families. Due to data availability, we limited the comparison between speaker identifying with "Male" and "Other" in CV to Catalan, German, English, Spanish, and French. Table 4 reports an overview of languages and language families.
+
+# A.3 Voice Activity Detection
+
+While conducting our experiments, we noticed that other than records with an empty reference, datasets can contain also empty audio snippets. More precisely, these recordings do have a signal but it is mostly silence. We found this phenomenon primarily in Fleurs-es, with silence mostly coming from snippets attributed to women.
+
+Therefore, we used pyannote.audio's pretrained neural models and code for voice activity detection.16 After counting the number of segments where voice was detected, we filter out all snippets were no segment was detected. We release the the counts and IDs of silence records in our repository.
+
+# A.4 Speaker ID Attribution in Fleurs
+
+Speaker IDs are crucial in our pipeline to avoid over- or under-estimating performance gaps due to overly present speakers. Since F1eurs does not provide this piece of information, we attribute it automatically.
+
+![](images/3c6f204319c4d7862f4e301c42a49b2051d2f96119525fb00351d74231342a5e.jpg)  
+(a) VoxPopuli
+
+![](images/ac373ebdfb5bd46af486f7b427ae55ea97bc867f581cd5e89ada44c961124bd1.jpg)  
+(b) Fleurs
+
+![](images/bbbc8f24c29c8943530bbd8301db0671f9e32d0f72b51677ceda50cfc3fe52a5.jpg)  
+(c) CV
+
+![](images/35b9bb5d2f47c0b447ce9ee374e36acd00d408da0b28042224c25782d64171ec.jpg)  
+Figure 6: SEAMLESS and WHISPER transcription quality. Error rate results are computed on test splits.  
+Figure 7: UMAP projections of snippets in Fleur-en. We found 13 male speakers (left) and 16 female speakers (right). Color indicates cluster ID assigned by HDBSCAN.
+
+<table><tr><td colspan="2"></td><td>test</td><td>train</td><td>val</td><td>gender set</td><td>gender subset</td></tr><tr><td rowspan="2">Fleurs</td><td>Whisper</td><td>12.68±.11.3</td><td>12.64±.11.4</td><td>12.49±.11</td><td>12.60±.11.2</td><td>12.62±.11.2</td></tr><tr><td>S M4T</td><td>19.87±.6.4</td><td>19.72±.6.3</td><td>18.75±.6.2</td><td>19.65±.6.3</td><td>19.67±.6.3</td></tr><tr><td rowspan="2">CV</td><td>Whisper</td><td>17.51±.17.5</td><td>16.65±.19.8</td><td>17.02±.18.7</td><td>16.88±.19.2</td><td>18.45±.19.2</td></tr><tr><td>S M4T</td><td>16.42±.9.5</td><td>12.68±9.9</td><td>14.7±.8.9</td><td>13.06±.9.6</td><td>15.58±.9.3</td></tr><tr><td rowspan="2">VP</td><td>Whisper</td><td>13.59±.4.45</td><td>12.63±.3.79</td><td>12.99±.4.22</td><td>12.69±.3.8</td><td>13.32±.4.2</td></tr><tr><td>S M4T</td><td>12.74±.3.27</td><td>11.27±.2.63</td><td>12.33±.2.68</td><td>11.39±.2.6</td><td>12.54±.2.9</td></tr></table>
+
+Table 2: Error rate distribution across datasets splits (test/train/Validation), all pre-sampled female/male records, and on the sampled gender subset we used. Results averaged over 19 languages in CV and Fleurs, and 11 for VoxPopuli.  
+
+<table><tr><td colspan="2"></td><td>test</td><td>train</td><td>val</td><td>presample_all</td><td>overall</td></tr><tr><td rowspan="2">Fleurs</td><td>Whisper</td><td>-1.18±.2.9</td><td>-0.659±.1.8</td><td>-0.255±.2.6</td><td>-0.893±.1.7</td><td>-0.997±.1.9</td></tr><tr><td>SM4T</td><td>0.153±.1.9</td><td>-0.880±.1</td><td>-0.257±.2.7</td><td>-0.735±.0.72</td><td>-0.809±.0.72</td></tr><tr><td rowspan="2">CV</td><td>Whisper</td><td>1.17±.5.7</td><td>-0.65±.2.7</td><td>0.88±.5.6</td><td>-0.14±.3.5</td><td>1.06±.5.3</td></tr><tr><td>SM4T</td><td>2.01±.5.4</td><td>-1.22±.5.9</td><td>0.40±.2.7</td><td>0.59±.2.9</td><td>1.05±.3.3</td></tr><tr><td rowspan="2">VP</td><td>Whisper</td><td>0.31±.3.5</td><td>-0.80±.0.97</td><td>-0.58±.1.9</td><td>-0.73±.0.9</td><td>-0.104±.2.2</td></tr><tr><td>SM4T</td><td>0.16±.2.8</td><td>-0.77±.0.54</td><td>-0.71±.2.1</td><td>-0.75±.0.6</td><td>-0.268±.2.2</td></tr></table>
+
+Table 3: Average F-M ER absolute difference across datasets splits (test/train/validation), all pre-sampled female/male records, and on the sampled gender subset used in our experiments. The results are averaged over 19 languages for CV and 11 for VoxPopuli. For Fleurs, we average results over the only 9 languages comprising female/male speakers in all splits.  
+
+<table><tr><td>Language</td><td>ISO</td><td>Family</td></tr><tr><td>Japanese</td><td>ja</td><td>Japonic</td></tr><tr><td>Dutch</td><td>nl</td><td>Germanic</td></tr><tr><td>English</td><td>en</td><td></td></tr><tr><td>German</td><td>de</td><td></td></tr><tr><td>Swahili</td><td>sw</td><td>NC / Bantu</td></tr><tr><td>Yoruba</td><td>yo</td><td>NC / Volta-Niger</td></tr><tr><td>Catalan</td><td>ca</td><td>Romance</td></tr><tr><td>French</td><td>fr</td><td></td></tr><tr><td>Italian</td><td>it</td><td></td></tr><tr><td>Portuguese</td><td>pt</td><td></td></tr><tr><td>Romanian</td><td>ro</td><td></td></tr><tr><td>Spanish</td><td>es</td><td></td></tr><tr><td>Arab</td><td>ar</td><td>Semitic</td></tr><tr><td>Czech</td><td>cs</td><td>Slavic</td></tr><tr><td>Russian</td><td>ru</td><td></td></tr><tr><td>Serbian</td><td>sr</td><td></td></tr><tr><td>Slovak</td><td>sk</td><td></td></tr><tr><td>Finnish</td><td>fi</td><td>Uralic</td></tr><tr><td>Hungarian</td><td>hu</td><td></td></tr></table>
+
+Table 4: Languages, their ISO code, and family studied in this paper.
+
+Specifically, we encode each snippet using a pre
+
+trained ECAPA-TDNN model for speaker verification. We use code from SpeechBrain (Ravanelli et al., 2021) and the model checkpoint with Hub ID speechbrain/spkrec-ecapa-voxceleb. Next, we cluster embeddings separately per gender using HDBSCAN (Campello et al., 2013), with code from scikit-learn. We use set min_cluster_size to 2 and use cosine similarity. Finally, we assign a numerical ID to each found cluster. Figure 7 shows an example of clustered snippets projected with UMAP (McInnes et al., 2018). We release per dataset and language IDs in our repository.
+
+# A.5 Subset selection and data splits
+
+To ensure the soundness of our experimental framework and of the gender subsample selection described in Section 3, we conducted preliminary analyses by comparing WER transcription quality scores (Table 2) and gender F-M gap results (Figure 3) across:  $i$  ) different datasets splits (test/train/validation), ii) all female/male records available from any split, and iii) the sampled female/male subset used in our main experiments.
+
+Since no precise information concerning the data
+
+used to train WHISPER and SEAMLESS is publicly available, this evaluation is carried out with the goal of spotting potential data contamination issues. Namely, to ensure that we are not testing models on data samples comprised in their underlying training, thus posing the risk of obtaining unreliable results, which cannot be compared across datasets and models.
+
+Table 2 shows no notable WER variations for Fleurs. Instead, both models perform better on the train splits of CV and VP, especially for the former dataset. The influence of the split accounted in CV and VP can thus be reflective of the F-M WER differences reported in Figure 3. Informed by such results, the gender subset used in our experiments is sampled from the test and validation split only for CV and VP, whereas for Fleurs, we leverage gender data from the whole corpus.
+
+# A.6 Dataset Statistics
+
+Table 5, 7, and 6 reports all collected statistics on CV, Fleurs, and VoxPopuli, respectively.
+
+# A.7 Energy Statement
+
+Experiments were conducted using a private infrastructure on hardware of type A100 PCIe 80GB (TDP of 250W) with carbon efficiency of 0.233 kgCO2eq/kWh. $^{17}$  Consumption is mostly due to inference to generate transcripts. We estimate a total emission of  $3.6\mathrm{kgCO2eq}$  for experiments in the main body (transcripts of test and validation sets). Transcripts on training sets account for additional  $104.24\mathrm{kgCO2eq}$ . No emission was directly offset. Total emissions are equivalent to driving an internal combustion engine car for  $980\mathrm{km}$ . $^{18}$  Estimate was conducted using codecarbon $^{19}$  but does not account for energy for cooling the infrastructure.
+
+# B Details on the exploratory analysis
+
+# B.1 Acoustic Analysis
+
+We provide further details about the acoustic analysis carried out in §5, comparing male and female speakers. Specifically, Table 9 reports the statistics of the T-tests conducted between acoustic features comparing genders across different languages and datasets, and Table 8 presents statistics on the Pearson correlation between differences in acous
+
+tic features and error rate gaps, separately for each (dataset, model) configuration.
+
+We discuss the results of two representative languages only: Italian and Slovak. Italian exemplifies languages with comparable performance across genders, whereas Slovak is biased towards men and belongs to the set of languages with larger gender gap trends (see §4 for details). For both languages, acoustic features are, in most cases, statistically different. As shown in Figure 8, the most significant differences are found in mean pitch values, confirming the well-known differences between males and females. The distributions align with ranges suggested by the literature (Simpson, 2009) for both languages and all datasets. In contrast, differences in speaking rate are much less evident but still statistically significant according to the T-test (see Table 9). Intensity shows the most variability across all dimensions, even within gender groups, particularly noticeable in the Fleur's dataset.[20] In Italian (VoxPopuli), there is no difference in intensity between the two gender groups ( $p = 0.356$ ), and similarly in Slovak (CV) ( $p = 0.376$ ). However, for both languages, fitting an OLS regression at the sentence level using acoustic features to predict sentence-level error rates ( $r_F$ ,  $r_M$ ) showed no significant contribution, similar to other languages ( $\mathbf{R}^2$  with max: 0.20 and average $_\sigma$ : 0.03 $_{\pm 0.04}$ )
+
+# B.2 Lexical Analysis
+
+To capture lexical phenomena, we extracted from the reference transcript of each record two sets of features. Part-of-speech tags include NOUN, PROPN, VERB, ADJ, ADV, PRON, AUX, CCONJ, DET, PART. Named entities tags include F_LOC, F_ORG, F_PER, and MISC. For each language, we used the corresponding SpaCy (https://spacy.io/models) "medium" text analyzer (e.g., for English en_core_web_sm and de_core_news_sm for German) when available. For hu an ar we used models available in Stanza (Qi et al., 2020). For yo and sw we limited the analysis to POS tags due to NE parses unavailability. We used mbeukman/xlm-roberta-base-finetuned-ner-y orub and mbeukman/xlm-roberta-base-finetuned-swahili-finetuned-ner-swahili for yo and sw, respectively (Beukman and Fokam, 2023). We left sr, sk, and cs out of this analysis since no taggers were available. We normalized
+
+<table><tr><td>Lang</td><td># records</td><td>Seconds</td><td># Tokens</td><td># M</td><td>Gini (M)</td><td># F</td><td>Gini (F)</td><td># O</td><td>Gini (O)</td></tr><tr><td>ar</td><td>14013</td><td>4.37±1.51</td><td>19.92±17.55</td><td>319</td><td>0.66</td><td>102</td><td>0.68</td><td></td><td></td></tr><tr><td>ca</td><td>1645</td><td>6.12±1.73</td><td>17.05±6.22</td><td>265</td><td>0.22</td><td>241</td><td>0.22</td><td>8</td><td>0.26</td></tr><tr><td>cs</td><td>13746</td><td>4.60±1.38</td><td>20.77±8.71</td><td>293</td><td>0.64</td><td>39</td><td>0.61</td><td></td><td></td></tr><tr><td>de</td><td>3641</td><td>6.10±1.75</td><td>16.74±6.08</td><td>807</td><td>0.22</td><td>129</td><td>0.24</td><td>17</td><td>0.24</td></tr><tr><td>en</td><td>4938</td><td>5.96±2.28</td><td>11.08±3.99</td><td>2262</td><td>0.25</td><td>476</td><td>0.25</td><td>33</td><td>0.21</td></tr><tr><td>es</td><td>6404</td><td>6.09±1.60</td><td>15.61±5.01</td><td>1476</td><td>0.29</td><td>517</td><td>0.30</td><td>42</td><td>0.34</td></tr><tr><td>fi</td><td>1695</td><td>4.74±1.66</td><td>18.22±8.93</td><td>52</td><td>0.57</td><td>17</td><td>0.57</td><td></td><td></td></tr><tr><td>fr</td><td>4214</td><td>5.81±1.67</td><td>16.40±5.56</td><td>876</td><td>0.19</td><td>213</td><td>0.19</td><td>19</td><td>0.23</td></tr><tr><td>hu</td><td>12245</td><td>5.45±1.52</td><td>22.84±9.23</td><td>173</td><td>0.57</td><td>277</td><td>0.53</td><td></td><td></td></tr><tr><td>it</td><td>7088</td><td>6.14±1.69</td><td>17.52±6.44</td><td>888</td><td>0.29</td><td>203</td><td>0.28</td><td></td><td></td></tr><tr><td>ja</td><td>7981</td><td>5.02±2.29</td><td>22.64±15.07</td><td>1033</td><td>0.44</td><td>632</td><td>0.41</td><td></td><td></td></tr><tr><td>nl</td><td>12808</td><td>4.83±1.38</td><td>16.29±5.57</td><td>487</td><td>0.62</td><td>130</td><td>0.63</td><td></td><td></td></tr><tr><td>pt</td><td>7319</td><td>4.61±1.63</td><td>10.66±5.44</td><td>620</td><td>0.41</td><td>82</td><td>0.37</td><td></td><td></td></tr><tr><td>ro</td><td>6489</td><td>4.05±0.90</td><td>15.70±4.13</td><td>167</td><td>0.56</td><td>39</td><td>0.57</td><td></td><td></td></tr><tr><td>ru</td><td>10466</td><td>5.51±1.85</td><td>18.87±9.18</td><td>702</td><td>0.45</td><td>248</td><td>0.43</td><td></td><td></td></tr><tr><td>sk</td><td>3008</td><td>4.34±1.74</td><td>13.66±9.64</td><td>58</td><td>0.64</td><td>13</td><td>0.65</td><td></td><td></td></tr><tr><td>sr</td><td>1895</td><td>2.86±0.89</td><td>8.63±5.97</td><td>39</td><td>0.57</td><td>11</td><td>0.64</td><td></td><td></td></tr><tr><td>sw</td><td>13730</td><td>5.61±1.80</td><td>21.04±8.13</td><td>248</td><td>0.62</td><td>251</td><td>0.58</td><td></td><td></td></tr><tr><td>yo</td><td>1226</td><td>5.77±1.40</td><td>41.97±9.89</td><td>14</td><td>0.48</td><td>20</td><td>0.64</td><td></td><td></td></tr></table>
+
+Table 5: Statistic for CV. We used validation and test sets. Total number of records, average length in seconds, number of tokens as tokenized by Whisper's pretrained tokenizer, number of unique speakers and, gini index of the snippets-per-speaker dispersion for male (M), female (F), and “other” (O) subgroups.  
+
+<table><tr><td>Lang</td><td># records</td><td>Seconds</td><td># Tokens</td><td># M</td><td>Gini (M)</td><td># F</td><td>Gini (F)</td></tr><tr><td>cs</td><td>2208</td><td>9.93±5.38</td><td>60.04±31.51</td><td>45</td><td>0.58</td><td>16</td><td>0.47</td></tr><tr><td>de</td><td>4064</td><td>8.74±5.87</td><td>32.34±20.82</td><td>147</td><td>0.55</td><td>64</td><td>0.57</td></tr><tr><td>en</td><td>3485</td><td>9.92±6.17</td><td>24.48±15.05</td><td>274</td><td>0.52</td><td>104</td><td>0.52</td></tr><tr><td>es</td><td>3130</td><td>11.56±6.78</td><td>42.89±25.15</td><td>95</td><td>0.57</td><td>48</td><td>0.48</td></tr><tr><td>fi</td><td>1141</td><td>9.99±5.44</td><td>46.65±26.61</td><td>25</td><td>0.63</td><td>16</td><td>0.45</td></tr><tr><td>fr</td><td>3341</td><td>10.23±6.15</td><td>38.19±23.92</td><td>151</td><td>0.52</td><td>84</td><td>0.54</td></tr><tr><td>hu</td><td>2145</td><td>10.47±6.05</td><td>62.48±35.59</td><td>45</td><td>0.48</td><td>21</td><td>0.50</td></tr><tr><td>it</td><td>2419</td><td>13.00±7.19</td><td>52.78±28.99</td><td>87</td><td>0.47</td><td>37</td><td>0.53</td></tr><tr><td>nl</td><td>2351</td><td>8.16±5.23</td><td>36.63±22.88</td><td>69</td><td>0.54</td><td>42</td><td>0.54</td></tr><tr><td>ro</td><td>2701</td><td>11.29±5.75</td><td>57.31±31.54</td><td>51</td><td>0.52</td><td>19</td><td>0.57</td></tr><tr><td>sk</td><td>1266</td><td>10.33±5.55</td><td>62.52±35.16</td><td>26</td><td>0.57</td><td>10</td><td>0.47</td></tr></table>
+
+Table 6: Statistic for VoxPopuli. We used validation and test sets. Total number of records, average length in seconds, number of tokens as tokenized by Whisper's pretrained tokenizer, number of unique speakers and, gini index of the snippets-per-speaker dispersion for male (M) and female (F) subgroups.
+
+tags from different modes accordingly (e.g., we consider Stanza's PERSON and SpaCy F_PER the same feature). Moreover, we computed lexical complexity as specified in Imani and Habil (2017), i.e., the ratio between the sum of [NOUN, PROPN, VERB, ADJ, ADV] and the sum of [PRON, AUX, CCONJ, DET, PART].
+
+# B.3 Probing Analysis
+
+Training and test set for probing experiments are sampled from the concatenation of each language's validation and test sets for consistency with error rate experiments. We balanced the samples on gender, stratifying on the speaker distribution.
+
+We used standard logistic regression probes as per scikit-learn's implementation and standard parameters but a larger allowance of steps for con
+
+vergence (n=1000). For MDL probes, we used logistic regression as the backbone classifier for comparability and operationalized the probe via the online code configuration (Voita and Titov, 2020). Following previous work (Orgad et al., 2022), we set the dataset slices to (percentage): [0.2, 0.4, 0.8, 1.6, 3.2, 6.25, 12.5, 25, 50, 100].
+
+We probe a total of 1000 distinct positions that correspond roughly to the first tens seconds of each recording (Whisper encodes audio with 25 milliseconds-long rolling frame and stride of 10 milliseconds) (Radford et al., 2022).
+
+# C Release Statement
+
+The following is a list of artifacts we produced in this work that we release to facilitate future research.
+
+<table><tr><td>Lang</td><td># records</td><td>Seconds</td><td># Tokens</td><td># M</td><td>Gini (M)</td><td># F</td><td>Gini (F)</td></tr><tr><td>ar</td><td>836</td><td>10.83±4.08</td><td>57.52±21.70</td><td>9</td><td>0.31</td><td>38</td><td>0.51</td></tr><tr><td>ca</td><td>2914</td><td>11.87±4.00</td><td>39.93±14.87</td><td>5</td><td>0.15</td><td>5</td><td>0.05</td></tr><tr><td>cs</td><td>3772</td><td>12.05±4.45</td><td>53.38±19.93</td><td>7</td><td>0.36</td><td>5</td><td>0.21</td></tr><tr><td>de</td><td>3326</td><td>12.98±6.32</td><td>37.19±13.30</td><td>4</td><td>0.07</td><td>4</td><td>0.47</td></tr><tr><td>en</td><td>2832</td><td>9.75±3.58</td><td>24.55±8.54</td><td>13</td><td>0.33</td><td>15</td><td>0.28</td></tr><tr><td>es</td><td>3082</td><td>12.16±3.80</td><td>37.88±13.67</td><td>6</td><td>0.38</td><td>7</td><td>0.35</td></tr><tr><td>fi</td><td>3482</td><td>12.74±4.43</td><td>47.38±18.27</td><td>12</td><td>0.34</td><td>18</td><td>0.47</td></tr><tr><td>fr</td><td>3194</td><td>10.25±3.58</td><td>38.80±15.06</td><td>4</td><td>0.06</td><td>6</td><td>0.15</td></tr><tr><td>hu</td><td>4374</td><td>12.09±4.12</td><td>54.90±21.68</td><td>5</td><td>0.03</td><td>5</td><td>0.05</td></tr><tr><td>it</td><td>3528</td><td>14.53±5.34</td><td>40.59±14.30</td><td>6</td><td>0.33</td><td>6</td><td>0.31</td></tr><tr><td>ja</td><td>2348</td><td>12.99±3.77</td><td>49.50±18.25</td><td>6</td><td>0.50</td><td>16</td><td>0.66</td></tr><tr><td>nl</td><td>2594</td><td>9.55±3.73</td><td>40.59±14.63</td><td>8</td><td>0.53</td><td>6</td><td>0.23</td></tr><tr><td>pt</td><td>3944</td><td>12.50±4.24</td><td>35.32±12.88</td><td>5</td><td>0.08</td><td>4</td><td>0.19</td></tr><tr><td>ro</td><td>4022</td><td>10.22±3.63</td><td>48.98±18.03</td><td>5</td><td>0.08</td><td>7</td><td>0.35</td></tr><tr><td>ru</td><td>3496</td><td>11.39±3.97</td><td>42.57±16.14</td><td>13</td><td>0.36</td><td>11</td><td>0.26</td></tr><tr><td>sk</td><td>2600</td><td>11.65±3.81</td><td>53.15±19.93</td><td>8</td><td>0.41</td><td>10</td><td>0.57</td></tr><tr><td>sr</td><td>2820</td><td>10.76±3.40</td><td>59.46±21.59</td><td>4</td><td>0.07</td><td>6</td><td>0.08</td></tr><tr><td>sw</td><td>3588</td><td>14.08±4.52</td><td>49.60±17.25</td><td>18</td><td>0.42</td><td>17</td><td>0.39</td></tr><tr><td>yo</td><td>2850</td><td>16.32±5.66</td><td>73.91±34.05</td><td>7</td><td>0.36</td><td>6</td><td>0.50</td></tr></table>
+
+Table 7: Statistic for Fleurs. We used validation and test sets. Total number of records, average length in seconds, number of tokens as tokenized by Whisper's pretrained tokenizer, number of unique speakers and, gini index of the snippets-per-speaker dispersion for male (M) and female (F) subgroups. Speakers are extracted automatically.  
+
+<table><tr><td>Dataset</td><td>Model</td><td>Feature</td><td>rho</td><td>p</td></tr><tr><td rowspan="6">VoxPopuli</td><td rowspan="4">SEAMLESS</td><td>Sp. rate</td><td>-0.10</td><td>0.76</td></tr><tr><td>Intensity</td><td>-0.43</td><td>0.18</td></tr><tr><td>Pitch</td><td>-0.83</td><td>0.00</td></tr><tr><td>Sp. rate</td><td>-0.33</td><td>0.32</td></tr><tr><td rowspan="2">WHISPER</td><td>Intensity</td><td>0.23</td><td>0.50</td></tr><tr><td>Pitch</td><td>-0.30</td><td>0.36</td></tr><tr><td rowspan="6">Fleurs</td><td rowspan="4">SEAMLESS</td><td>Sp. rate</td><td>0.62</td><td>0.00</td></tr><tr><td>Intensity</td><td>0.11</td><td>0.66</td></tr><tr><td>Pitch</td><td>-0.06</td><td>0.82</td></tr><tr><td>Sp. rate</td><td>0.20</td><td>0.42</td></tr><tr><td rowspan="2">WHISPER</td><td>Intensity</td><td>-0.26</td><td>0.29</td></tr><tr><td>Pitch</td><td>-0.53</td><td>0.02</td></tr><tr><td rowspan="6">CV</td><td rowspan="4">SEAMLESS</td><td>Sp. rate</td><td>-0.51</td><td>0.02</td></tr><tr><td>Intensity</td><td>-0.25</td><td>0.31</td></tr><tr><td>Pitch</td><td>0.20</td><td>0.41</td></tr><tr><td>Sp. rate</td><td>-0.39</td><td>0.10</td></tr><tr><td rowspan="2">WHISPER</td><td>Intensity</td><td>-0.10</td><td>0.69</td></tr><tr><td>Pitch</td><td>0.38</td><td>0.11</td></tr></table>
+
+Table 8: Correlation between Acoustic Features and Error Rate Gaps. Pearson correlation between the difference of the mean of subgroups an and  $E(r_{F}, r_{M})$ . In bold the statistically relevant correlation ( $p < 0.05$ ).
+
+- Transcriptions obtained with Whisper and SeamlessM4T of the full datasets Mozilla Common Voice, Google Fleurs, and Meta Vox-Populi.  
+- Segment-level annotations of segments with voice activity extracted from the voice activity detection pipeline (§A.3).  
+- Segment-level acoustic features from our acoustic analysis ( $\S 5$ ).
+
+- Segment-level embeddings and cluster IDs extracted on Fleurs with the speaker identification pipeline (§A.4).  
+- Extensive statistics on speaker, gender, and record distributions on the three datasets—which we could not find anywhere else online.  
+- Dataset samples on which we computed the quality and gap metrics for reproducibility purposes ( $\S 3$ ).
+
+<table><tr><td></td><td>ar</td><td>ca</td><td>cs</td><td>de</td><td>en</td><td>es</td><td>fi</td><td>fr</td><td>hu</td><td>it</td><td>ja</td><td>nl</td><td>pt</td><td>ro</td><td>ru</td><td>sk</td><td>sr</td><td>sw</td><td>yo</td></tr><tr><td colspan="20">Fleurs</td></tr><tr><td>Intensity</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td></tr><tr><td>Pitch</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td></tr><tr><td>Speaking rate</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✘</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✘</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td></tr><tr><td colspan="20">VoxPopuli</td></tr><tr><td>Intensity</td><td></td><td>✓</td><td>✓</td><td>✘</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✘</td><td></td><td>✓</td><td></td><td>✓</td><td></td><td>✓</td><td></td><td></td><td></td></tr><tr><td>Pitch</td><td></td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td></td><td>✓</td><td></td><td>✓</td><td></td><td>✓</td><td></td><td></td><td></td></tr><tr><td>Speaking rate</td><td></td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✘</td><td>✘</td><td>✓</td><td>✓</td><td></td><td>✘</td><td></td><td>✓</td><td></td><td>✓</td><td></td><td></td><td></td><td></td></tr><tr><td colspan="20">CV</td></tr><tr><td>Intensity</td><td>✘</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✘</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✘</td><td>✓</td><td>✓</td><td>✓</td></tr><tr><td>Pitch</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td></tr><tr><td>Speaking rate</td><td>✓</td><td>✘</td><td>✓</td><td>✓</td><td>✘</td><td>✘</td><td>✓</td><td>✘</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✘</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td></tr></table>
+
+Table 9: Statistical differences of acoustic features between genders, separately by language and dataset.  
+√ indicates statistical difference, × indicates no difference (independent-sample Student's T-test,  $p < 0.05$ ).
+
+![](images/e59ae76e9f89264bfa961663b83ba8b355697db0db5efa76856138c23163266c.jpg)
+
+![](images/19d7f544a8c0d0a62b0ef0fbb73c73fca50329fb7a37fc2e4bcda399dc217d71.jpg)
+
+![](images/7e251abb4f8c0deeef2a5743797d857c4705a0971dfbc08bd495978850dbbbc0.jpg)
+
+![](images/60bd773dc6e7abf6db442d3779a9815affb40610aed20be1a681185d419ba11c.jpg)
+
+![](images/b3f0ffb439eec3757ed876c90aef04baeb504f82111b9d91bad4aeb3d3a80fd3.jpg)
+
+![](images/37203d0b4cdc3f4ae51624f3fbe448ea25e69f93c30321081fb989cc945df3dc.jpg)  
+Figure 8: Acoustic features. Distributions of mean intensity, mean pitch, and speaking rate for Italian and Slovak on our three datasets. "Male" (green, solid line), "Female" (red, dashed line).
+
+![](images/ad06d388cab28d300e7d08c165061a419a56185c3755e3921eb9a4a042f10fef.jpg)
+
+![](images/e9624f82ff80385fdb440ae36ef519098640e5d5b741eca7fc76a6d5f21b5ba3.jpg)  
+ru
+
+![](images/a75538a26ee21c84d33c2dde59e17fece314a883a0cff0c7dbd5ec13568d7319.jpg)  
+CS
+
+![](images/212091e020c11f1f9992208470b88a4f8e1afc9e8e64b6cd90ef6a551144cc74.jpg)  
+es
+
+![](images/f86b4e8dc5aab8e422af722396a4f055ee56c4e6ee512af0245dc6cb2a2e5baf.jpg)  
+ro
+
+![](images/3c55918b7357af91927d426b8899be3e63432233cae4d7f613110d35f987c6db.jpg)  
+ja
+
+![](images/49bca4bf1bf54eb15c0fdb98d995c7fcad6ba5a662ff0525d7991b4f72d42e37.jpg)  
+ar  
+Embedding position in time (s)
+
+![](images/712e087cbd4a35b6bda94b7b5622a5900d5caefb7fff8f8b24faf31482014818.jpg)
+
+![](images/35157376a517024b3ad761f35441ac80c2923d2c188d516638335dde17db5d96.jpg)  
+sr
+
+![](images/365503ca3517448b207e9ebf6faee367038b3da5c5c397acdce0e47c6c515402.jpg)  
+it
+
+![](images/a44b2359aee7cbb35cc658ed0212d487289ad8d629f4ebec5b17bf8b237c9978.jpg)  
+ca
+
+![](images/a3a1f70676013a097a6453d64f1e47503621ce121be4ab05689a6cf9d9d8e2d4.jpg)  
+SW
+
+![](images/9358524cd9fc45e17d0678358e3acb909a61b9679bd248caf1b5220b0ebcbf88.jpg)  
+hu  
+Embedding position in time (s)
+
+![](images/1f32bd5392d926ec476f65911145d0d0e8333b69a5d595ea0ca45b9102de12aa.jpg)  
+nI
+
+![](images/cb54b249f77aade39896acde1842536998ef8b0b92b5858bfb898f4128448ac0.jpg)  
+sk
+
+![](images/24616e071d0cd3771f5fafa680c52fd4ee831d0ccc314445f5774cfc19dd30dd.jpg)  
+fr
+
+![](images/16dc6c23482378afc6e74d3d53cac9eadb9562b297a4926099728b7d66076672.jpg)  
+pt
+
+![](images/630333871695ca95844d172eae75604d62a654f5544572f4c2d54dd691afd80d.jpg)  
+yo
+
+![](images/4f056e85210686c1070a36fb77d4ca076387d5246a32e841aeb53840d679ecea.jpg)  
+fi  
+Embedding position in time (s)
+
+![](images/ab8a930e1f0bf6c23032934d4b3e27d516287b3d0712db3dda60b544d0c3c2e8.jpg)  
+(a) Probes trained on the original labels.
+
+![](images/e2c6fcbc4be071b3648e6cd81e1af5a43174fbdeee416447f120f87c1c3ce8ab.jpg)  
+ru
+
+![](images/b5ac441e897875de806ffc8ebc2f925a88040a40aeb626a907643d0f165ce77a.jpg)  
+CS
+
+![](images/a6fc5c403422aa606fd6c0a509dc368a202a4c0bf95a6b8db9e4b7faa3938f56.jpg)  
+es
+
+![](images/ec178fcbea83b19dfcabc72797500cd37310d59cdc4b9dbea1ffe65ca13424b7.jpg)  
+ro
+
+![](images/9f4bcab08b810d2af8c6fb9fce7a7c0658f0ac66510ff40f2bda0c3adf5b9ea5.jpg)  
+ia
+
+![](images/0702955a7d95d68be23e23eddbef6625e32316b21b9ebbfe96fdab5c33be72b3.jpg)  
+ar  
+Embedding position in time (s)  
+(b) Probes trained on the shuffled labels.  
+Figure 9: F-M gender probing F1 Macro performance for every contextual embedding in Whisper-large-v3 within the first 10 seconds (x axis). Logistic regression probe with L2 regularization trained on standard (a) and shuffled (b) training labels. Orange lines indicate the average length of test segments. All CV languages.
+
+![](images/32aad541155945a9081e93c2b3498e8e4ddb6720a01b6e87b2d3cdaed3b6bd7e.jpg)  
+en
+
+![](images/f0931beb6d10c24257701be04c86f6903cae0474a32d40d7f5de677e5d477ff4.jpg)  
+sr
+
+![](images/1ee23ed9de6a3bba59cdd83847bdba99798ad46faa4c130dfa024cf80c03929b.jpg)  
+it
+
+![](images/47e2f58e8d520a84019b7a85a435f766ff7ba677c9492c30fc30bd2f38c9475f.jpg)  
+ca
+
+![](images/4db75c71176e70ac00338066831a647c3d9d44627f10959977f868b5edaf2d4a.jpg)  
+sw
+
+![](images/aae0aa74894c4061c8cc078fef8c1b7ea3d974693db72493d09939f78e54297c.jpg)  
+hu  
+Embedding position in time (s)
+
+![](images/8fa78192e68f9a8ba7a6179f68dbbaf6b5d11859f72b0834695fdd42f67f4e63.jpg)  
+nl
+
+![](images/f4b76c0da8bf7a9e710e45bd51ebf4965282fbcbeb64ff979d3409582de7a29f.jpg)  
+sk
+
+![](images/bf43bd78eb171aefa17b7ffc34d87745fdf905f15222b5dbc13019baaf0bd957.jpg)  
+fr
+
+![](images/23fff0cf52edcd3bccf00e87bd7599f9e7c8c10c71a5c7ff9ec204ee67d6e5fb.jpg)  
+pt
+
+![](images/fd49016a6df57257b7094198f930ebd77c4335724c4fa03385854febe9cf7d28.jpg)  
+yo
+
+![](images/997abf1419578a0004916d18dac2472c32770c46eff048a4cc12304eaeab4109.jpg)  
+fi  
+Embedding position in time (s)

2024/Twists, Humps, and Pebbles_ Multilingual Speech Recognition Models Exhibit Gender Performance Gaps/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ee37ff3f8db37afba9e1ffc239dbadefcaacfb990a75fcab7d840dc7800f101d
+size 1146445

2024/UNICORN_ A Unified Causal Video-Oriented Language-Modeling Framework for Temporal Video-Language Tasks/1d2b554e-cf0c-46d1-9ba0-ee7589838c7b_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:edd19f6eb00c2ea9b3f1af8c7b959a9b5b7c2a0e7fea6a32635819c7105678c4
+size 5922342

2024/UNICORN_ A Unified Causal Video-Oriented Language-Modeling Framework for Temporal Video-Language Tasks/full.md

@@ -0,0 +1,367 @@
+# UNICORN: A Unified Causal Video-Oriented Language-Modeling Framework for Temporal Video-Language Tasks
+
+Yuanhao Xiong $^{1,2}$ , Yixin Nie $^{2}$ , Haotian Liu $^{3}$ , Boxin Wang $^{4}$ , Jun Chen $^{2}$ , Rong Jin $^{2}$ , Cho-Jui Hsieh $^{1}$ , Lorenzo Torresani $^{2}$ , Jie Lei $^{2}$
+
+$^{1}$ UCLA,  $^{2}$ Meta,  $^{3}$ University of Wisconsin-Madison,  $^{4}$ UIUC
+
+{yhxiong, chohsieh}@cs.ucla.edu, lht@cs.wisc.edu, boxinw2@illinois.edu
+
+{ynie, junchen20, rongjinml, torresani, jielei}@meta.com
+
+# Abstract
+
+The great success of large language models has encouraged the development of large multimodal models, with a focus on image-language interaction. Despite promising results in various image-language downstream tasks, it is still challenging and unclear how to extend the capabilities of these models to the more complex video domain, especially when dealing with explicit temporal signals. To address the problem in existing large multimodal models, in this paper we adopt visual instruction tuning to build a unified causal video-oriented language modeling framework, named UNICORN. Specifically, we collect a comprehensive dataset under the instruction-following format, and instruction-tune the model accordingly. Experimental results demonstrate that without customized training objectives and intensive pre-training, UNICORN can achieve comparable or better performance on established temporal video-language tasks including moment retrieval, video paragraph captioning and dense video captioning. Moreover, the instruction-tuned model can be used to automatically annotate internet videos with temporally-aligned captions. Compared to commonly used ASR captions, we show that training on our generated captions improves the performance of video-language models on both zero-shot and fine-tuning settings. Source code can be found at https://github.com/xyh97/UNICORN.
+
+# 1 Introduction
+
+Recent breakthroughs in large language models (LLMs) (Ouyang et al., 2022; cha, 2023; OpenAI, 2023; vic, 2023; Touvron et al., 2023a,b) have reignited the enthusiasm about the achievement of artificial general intelligence where a single foundation model can accomplish a large variety of downstream tasks based on human instructions. Towards this ultimate goal, the community has witnessed promising advances in large multimodal models (LMMs) for vision and language (Liu et al.,
+
+# Visual input example, Playing Tennis (34s in total):
+
+![](images/50be4f1a4429af3147924adcabe27931224999a613a8e98eaa2fd0921362a83b.jpg)
+
+Task 1: Moment Retrieval  
+
+<table><tr><td>Instruction</td><td>Please predict start and end time of the following moment: He hits the ball over the net several times. The output format should be &lt;start&gt;&lt;end&gt;.</td></tr></table>
+
+Response <16><48>  
+Task 2: Video Paragraph Captioning  
+
+<table><tr><td>Instruction</td><td>Provide a detailed description of the video, capturing its key moments.</td></tr><tr><td>Response</td><td>A man is bouncing a tennis ball on an outdoor court. He hits the ball over the net several times. The balls roll over to the opposing fence, broken in half from the impact.</td></tr></table>
+
+Table 1: Example of instruction-following data. The response of moment retrieval is computed by time tokenization for the window [7.7s, 22.1s] with 75 bins.
+
+2023b,a; Wang et al., 2023b; Dai et al., 2023; Bai et al., 2023; Li et al., 2023a; Zhu et al., 2023), the two essential modalities to understand the world. Most of these LMMs follow the pipeline of visual instruction tuning (Liu et al., 2023b) and demonstrate strong capabilities in vision-centric tasks like image classification and object detection (Wang et al., 2023b), and vision-language tasks like image captioning and visual question answering (Dai et al., 2023; Liu et al., 2023b).
+
+Despite impressive results in the image domain, videos, another important data format in the vision modality, are under-explored. In contrast to images, videos have an extra temporal dimension and are much more difficult to process due to increased complexity. Existing approaches either directly apply LMMs trained on image-text pairs (Dai et al., 2023) to the video domain without fine-tuning or develop video-oriented LMMs (Zhang et al., 2023a; Muhammad Maaz and Khan, 2023; Li et al., 2023c) on short trimmed videos. However, such models
+
+focus more on problems which are less dependent on temporal information like action recognition and video question answering. It still remains unclear how to solve video-language tasks that requires explicitly modeling temporal information, including moment retrieval (Hendricks et al., 2017; Lei et al., 2021), video paragraph captioning (Park et al., 2019), and dense video captioning (Krishna et al., 2017) in one single LMM.
+
+In fact, the inherent disparities among these task formats pose a challenge to the development of such models: moment retrieval requires predicting the temporal location of a moment described by language, paragraph captioning entails to write a coherent story from an untrimmed video, while the goal of dense video captioning is to generate captions and temporal locations for a series of moments simultaneously. These tasks are typically solved individually by specifically-designed models (Lei et al., 2020a; Yang et al., 2023; Lei et al., 2021; Lin et al., 2023). While attempts have been made to unify these temporal video-language tasks (Wang et al., 2023a; Yan et al., 2023), separate modules and training objectives tailored for each task are involved in these methods, making them complicated in both training and inference.
+
+To address the above challenge, we propose a UNIFIED Causal videO-oRiented laNguage modeling framework (UNICORN) that unifies the tasks as a simple yet generic language modeling problem. For moment retrieval and video paragraph captioning, we convert original training datasets into corresponding instruction-following formats, as shown in Table 1. In particular, inspired by previous efforts in discretizing bounding box coordinates (Chen et al., 2022; Peng et al., 2023; Zhang et al., 2023b), our approach represents the continuous event boundaries as a sequence of discrete time tokens and processes them similarly as language tokens. On a range of datasets and tasks, we show that this unified approach achieve comparable or better performance over previous methods.
+
+On the other hand, the development of large video-language models is hindered by the lack of semantically- and temporally-aligned videotext pairs, an issue unique to the video domain. As pointed out in (Han et al., 2022), the models pre-trained on commonly-used noisy datasets such as HowTo100M (Miech et al., 2019) and YT-Temporal-1B (Zellers et al., 2022) suffer from the misalignment between videos and ASR captions severely. Thanks to the generalization ability of
+
+LMMs, our UNICORN can be leveraged to automatically generate captions for internet videos. We demonstrate that qualitatively the generated captions are better semantically- and temporally-aligned with the videos than the original ASR captions, and quantitatively incorporating our generated captions in either instruction-tuning for moment retrieval or end-to-end video representation learning leads to significant performance gains.
+
+Our contributions are threefold: (1) We propose UNICORN, a simple and generic framework that unifies various temporal video-language tasks via language modeling; (2) Our approach achieves comparable or better performance to state-of-the-art methods on multiple downstream tasks, including moment retrieval, video paragraph captioning, and dense video captioning; (3) Compared to existing captions, those automatically generated by our method have shown to be better aligned with the videos, both semantically and temporally. Empirically, the generated captions have demonstrated to improve performance of models trained on them. Our automatic annotation pipeline is useful for empowering the development of future LMMs.
+
+# 2 Related Work
+
+Large Multimodal Models. Large language models are taking the world by storm with their incredible capabilities to answer questions in a coherent and informative way aligned with human instructions (cha, 2023; Ouyang et al., 2022; vic, 2023; OpenAI, 2023; Touvron et al., 2023a,b). The universality and generalization of LLMs make it potential to unlock the door to a foundation general-purpose model. Towards this goal, a variety of large multimodal models are emerging to bridge different modalities, in particular vision and language (Liu et al., 2023b,a; Wang et al., 2023b; Dai et al., 2023; Bai et al., 2023; Li et al., 2023a; Zhu et al., 2023). Such LMMs adopt the pipeline of visual instruction tuning (Liu et al., 2023b) by converting original datasets into the instruction-following format and casting traditional vision problems as a language modeling task. For instance, LLaVa (Liu et al., 2023b) generates multimodal language-image instructional data using GPT-4 (OpenAI, 2023) and develops an LMM connecting a pre-trained image encoder and a pre-trained large language model to deal with vision-language tasks. InstructBLIP (Dai et al., 2023) enlarges the task coverage by gathering 26 publicly available datasets and proposes
+
+an instruction-aware visual feature extraction process. These models achieve the state-of-the-art performance on numerous downstream tasks, ranging from vision-centric ones such as image classification and object detection to vision-language ones such as image captioning and visual reasoning. Despite efforts in understanding images, few attempts have been made for video-language tasks due to additional complexity. Thus, in this paper we study how to model the interaction between long untrimmed videos and captions from the perspective of language modeling.
+
+Video-Language Modeling. Video-language tasks have been widely studied, especially these require specific temporal modeling, such as moment retrieval (Lei et al., 2021; Lin et al., 2023; Mun et al., 2020; Zeng et al., 2020), video paragraph captioning (Lei et al., 2020a; Park et al., 2019; Yang et al., 2023; Wang et al., 2021), and dense video captioning (Krishna et al., 2017; Yang et al., 2023; Wang et al., 2021). Some methods (Lin et al., 2023; Yan et al., 2023; Wang et al., 2023a; Li et al., 2022) pre-train a model on large-scale corpus to generate latent video and language representations, which can be then adapted to different downstream tasks. This line of work typically requires elaborate architectural designs and multiple training objectives tailored for each target task. In contrast, we propose a more elegant unified framework to integrate various temporal video-language tasks into a simple yet generic language modeling problem. Compared with existing video-oriented LMMs targeting at short video clips (Li et al., 2023c; Zhang et al., 2023a; Muhammad Maaz and Khan, 2023), UNICORN attaches more attention to long untrimmed videos. The most relevant method to UNICORN is Vid2Seq (Yang et al., 2023), which also formulates dense video captioning as language modeling. However, it should be emphasized that Vid2Seq depends heavily on video-language pre-training and is unable to handle tasks other than captioning. On the contrary, by visual instruction tuning on high quality datasets, UNICORN demonstrates superior performance on a series of video-language tasks without intensive pre-training. Moreover, our method can be applied towards noisy video datasets to generate better-aligned captions.
+
+# 3 Method
+
+In this section, we introduce our unified framework UNICORN in detail. We start by discussing how to
+
+transform the original datasets for different downstream tasks into the general instruction-following format in Section 3.1. Then in Section 3.2, we describe the model architecture designed for videolanguage interaction. In Section 3.3, we present the training pipeline of UNICORN including datasets and training objective. Finally in Section 3.3, we demonstrate how to conduct inference with the obtained model on downstream tasks together with the process to generate captions for noisy datasets.
+
+# 3.1 Instruction-Following Data Generation
+
+As the ultimate goal is to unify various temporal video-language tasks, we cast moment retrieval and video paragraph captioning datasets into a common instruction following format. For dense video captioning, it can be regarded as a two-stage procedure of paragraphing captioning and moment retrieval and thus no specific training data are required. We provide details in following sections.
+
+Moment Retrieval In moment retrieval (MR) (Hendricks et al., 2017; Gao et al., 2017; Krishna et al., 2017; Lei et al., 2020b, 2021), a continuous time window is predicted given an untrimmed video and a language moment query. With the task definition, an example instruction can be: "Please predict start and end time of the following moment: {target}", where {target} is replaced by the specific query. We curate a template instruction list in Appendix B, to explicitly teach the underlying model the concepts of the task and the objective.
+
+A key challenge here is how to generate output sequences to represent moment locations. To reduce the exploration space for more controllable predictions, we follow previous sequence generation strategies for such continuous values (Chen et al., 2022; Peng et al., 2023; Yang et al., 2023; Wang et al., 2023b; Chen et al., 2023), and discretize the timestamp  $t$  in a  $d$ -s long video into an integer in  $\{0,1,\dots,N_{\mathrm{bin}} - 1\}$  with  $N_{\mathrm{bin}}$  equally-spaced bins by  $\lfloor t\times N_{\mathrm{bin}}\rfloor /d$ . Moreover, since recent LLMs exhibit surprising performance in mathematical reasoning, we use the original vocabulary without extra time tokens, which in turn reduces the number of trainable parameters and avoids pretraining to re-acquire the ability to reason about numbers. Meanwhile, to distinguish our discrete relative timestamps from other numerical expressions such as "5 apples", we enclose the timestamp values by "<start><end>" where start and end are replaced by corresponding converted timestamps. For instance, the moment in Table 1 starting at
+
+![](images/d6d455ae645e737033818f88932f4a028df4ff061159dd1c95dd802d6138036b.jpg)  
+Figure 1: UNICORN framework using video paragraph captioning as an example. We encode each video frame separately and concatenate their resulting visual tokens to represent the video. We highlight the encoding process of one frame in red. All modules are instruction-tuned with the language modeling loss except the image encoder.
+
+7.7s and ending at 22.1s within a 34s-long video is transformed into the desired output sequence  $"<16><48>"$  after our proposed time tokenization with 75 bins. To make output predictions consistent in format, we append a language constraint to our instruction: "The output format should be <start><end)." For a moment query associated with multiple time windows, we regard each query-location pair as an individual data sample.
+
+Video Paragraph Captioning The task of video paragraph captioning (VPC) (Park et al., 2019; Lei et al., 2020a) aims at generating a set of coherent sentences to describe an untrimmed video that contains several events. While previous pipelines (Park et al., 2019; Lei et al., 2020a) segment the video into multiple clips from ground-truth event boundary proposals, our method takes as input frames sampled from the whole video together with the instruction "Provide a detailed description of the given video, capturing its key moments". We generate a diverse template set in Appendix B to reduce overfitting and strengthen the understanding of the task. We leverage the paragraph caption of the target video as the prediction.
+
+Dense Video Captioning The goal of dense video captioning (DVC) (Krishna et al., 2017) is to generate multiple corresponding captions for a series of events together with their temporal locations from the untrimmed video. It is much harder than moment retrieval and paragraph captioning since it requires predicting events and their timestamps simultaneously. The most straightforward way to convert the task into the instruction-following format is to
+
+construct a sequence with both events and locations given a specific input prompt. However, design choices such as event serialization (e.g., chronological or random) and where to insert time windows might affect the performance significantly (Chen et al., 2022; Yang et al., 2023). Furthermore, the training of such models is challenged by the longer input sequence with both timestamps and event descriptions. It also takes extra computational costs to learn redundant information from moment retrieval and video paragraph captioning again. Considering the inherent property of DVC, we find that it can be naturally decomposed into a two-stage procedure of video paragraph captioning followed by moment retrieval. Thus, no additional training data are required and this task can be addressed at inference-time by the model instruction-tuned on two tasks above, with more details in Section 3.3.
+
+# 3.2 Model Architecture
+
+To bridge together video frames and natural language instructions as the ultimate input sequence, we propose a large multimodal language model, demonstrated in Figure 1. Specifically, a sequence of visual tokens are obtained by feeding frames and the corresponding instruction from an untrimmed video into our per-frame encoding module. Visual tokens are then processed by a projection layer to the same latent space as the large language model (LLM). The LLM takes the concatenation of visual tokens and instruction tokens as input and generates the desired output given different task instructions.
+
+Compared with image-language interaction, few attempts have been made in the video domain due
+
+to increased complexity. However, using image encoders to conduct per-frame encoding for videos by brute force will lead to an extremely long sequence of visual tokens proportional to the number of frames. On the other hand, a completely new encoder might require a considerable amount of training to align modalities of vision and language again. To strike a balance between two aforementioned issues, we resort to the recently proposed InstructBLIP (Dai et al., 2023) with a Q-Former. Specifically, Q-Former is a stack of self-attention and cross-attention layers with a number of learnable query tokens and aims at compressing the information of image embeddings and reducing the number of vision tokens. On top of Q-Former, we make some adoptions to handle the video input. In detail, our method first extracts  $n_q$  visual tokens from each frame using the frame-based encoding of the original Q-Former. For efficiency, we then apply average pooling in a frame-wise manner, which results in one token for each frame. Given a video of  $N$  frames, these  $N$  tokens are further processed by a module with two self-attention layers to integrate temporal information. Our design maintains a reasonable length of visual tokens for instruction tuning and takes advantage of pre-trained LLMs for feature alignment between the two modalities.
+
+# 3.3 Training and Inference
+
+With the data in instruction-following format, we now present a unified framework of instruction tuning on various downstream tasks.
+
+Training. The instruction-following format makes it feasible to train the model to predict next tokens with an auto-regressive language modeling loss. Given input video frames  $X = \{x_{i}\}_{i=1}^{N}$  and task instruction  $Y = \{y_{j}\}_{j=1}^{M}$ , we maximize the log likelihood of the output sequence  $Z = \{z_{k}\}_{k=1}^{L}$ :  $\max \sum_{k=1}^{L} \log p_{\theta}(z_{k}|X,Y,z_{1:k-1})$ , where  $L$  is the output sequence length,  $p_{\theta}$  is the output probability distribution over the LLM vocabulary given model parameters  $\theta$ . We finetune the whole model except the image encoder using LoRA (Hu et al., 2022).
+
+Inference. For moment retrieval and paragraph captioning, we prompt the instruction tuned model using corresponding task instructions to generate responses via beam search. For dense video captioning, we divide it into two stages, where the model first generates a paragraph caption, and then temporally locate each sentence in the paragraph with moment retrieval task instruction.
+
+# 4 Experiments
+
+In this section, we evaluate UNICORN comprehensively against state-of-the-art methods to show its effectiveness. We first introduce experimental setup in Section 4.1. Then we present results on downstream tasks including moment retrieval, video paragraph captioning and dense video captioning in Section 4.2. Ablation studies are conducted in Section 4.3 for better understanding of our designs. Finally, in Section 4.4 we investigate the quality of the automatic annotation generated by UNICORN on HowTo100M.
+
+# 4.1 Experimental Setups
+
+Architecture. The backbone of our video encoding module is adapted from InstructBLIP (Dai et al., 2023). Specifically, we implement the video encoder with the same image encoder (ViTG/14) (Fang et al., 2023), Q-Former with 32 learnable query embeddings and a fully-connected projection layer as the original InstructBLIP structure, plus a temporal modeling module with 2 self-attention layers. For the language side, we select Vicuna-7B (vic, 2023), a publicly available LLM fine-tuned from LLaMa (Touvron et al., 2023a).
+
+Datasets. Rather than intensive pre-training on a large scale noisy dataset without annotations, we directly fine-tune our model on a comprehensive set of publicly available video-language datasets, including QVHighlights (Lei et al., 2021), CharadesSTA (Gao et al., 2017), ActivityNet Captions (Krishna et al., 2017), and YouCook2 (Zhou et al., 2018a). The collection covers various domains with different length distributions. More details about datasets are included in Appendix C.
+
+Implementation details. We adopt LAVIS (Li et al., 2023b) under BSD 3-Clause License to run all the experiments and our usage is compatible with its license. The model is instruction tuned for 5 epochs with a batch size of 32. We randomly sample a task at a time based on data size. We use AdamW (Loshchilov and Hutter, 2019) with  $\beta_{1} = 0.9$ ,  $\beta_{2} = 0.999$ , and weight decay 0.05 for optimization. The learning rate is warmuped from  $10^{-6}$  to  $10^{-4}$  in the first epoch, followed by a cosine decay with a minimum of  $10^{-5}$ . We freeze the image encoder and fine-tune the rest of the model, with LoRA applied on the LLM. There are around 243M trainable parameters. UNICORN is trained with 8 NVIDIA A100 (80G) GPUs in 12 hours.
+
+Evaluation. For moment retrieval, we evaluate
+
+<table><tr><td rowspan="2">Method</td><td colspan="3">QVHighlights</td><td colspan="3">Charades-STA</td><td colspan="3">ActivityNet Captions</td></tr><tr><td>R@0.5</td><td>R@0.7</td><td>mAP avg</td><td>R@0.5</td><td>R@0.7</td><td>mIoU</td><td>R@0.5</td><td>R@0.7</td><td>mIoU</td></tr><tr><td>LGI (Mun et al., 2020)</td><td>—</td><td>—</td><td>—</td><td>59.5</td><td>35.5</td><td>51.4</td><td>41.5</td><td>23.1</td><td>41.1</td></tr><tr><td>2D TAN (Zhang et al., 2020b)</td><td>—</td><td>—</td><td>—</td><td>46.0</td><td>27.5</td><td>41.2</td><td>44.5</td><td>26.5</td><td>—</td></tr><tr><td>VSLNet (Zhang et al., 2020a)</td><td>—</td><td>—</td><td>—</td><td>42.7</td><td>24.1</td><td>41.6</td><td>43.2</td><td>26.2</td><td>43.2</td></tr><tr><td>MDETR (Lei et al., 2021)</td><td>59.8</td><td>40.3</td><td>36.1</td><td>52.1</td><td>30.6</td><td>45.5</td><td>—</td><td>—</td><td>—</td></tr><tr><td>GVL (Wang et al., 2023a)</td><td>—</td><td>—</td><td>—</td><td>—</td><td>—</td><td>—</td><td>48.9</td><td>27.2</td><td>46.4</td></tr><tr><td>UnLoc (Yan et al., 2023)</td><td>64.5</td><td>48.8</td><td>—</td><td>58.1</td><td>35.4</td><td>—</td><td>48.0</td><td>29.7</td><td>—</td></tr><tr><td>UniVTG (Lin et al., 2023)</td><td>58.9</td><td>40.9</td><td>35.5</td><td>58.0</td><td>35.6</td><td>50.1</td><td>—</td><td>—</td><td>—</td></tr><tr><td>UniVTG, PT (Lin et al., 2023)</td><td>65.4</td><td>50.1</td><td>43.6</td><td>60.2</td><td>38.5</td><td>52.2</td><td>—</td><td>—</td><td>—</td></tr><tr><td>UNICORN</td><td>68.4</td><td>51.9</td><td>45.0</td><td>69.0</td><td>45.6</td><td>58.9</td><td>48.4</td><td>29.8</td><td>47.1</td></tr></table>
+
+on QVHighlights, Charades-STA, and ActivityNet Captions. We report the standard metrics Recall at 1 under temporal Intersection over Union (IoU) thresholds of 0.5 and 0.7, abbreviated as R@0.5 and R@0.7. Besides, we use the average mAP over IoU thresholds [0.5:0.05:0.95] on QVHighlights with multiple ground-truth segments for one moment, and mean IoU (mIoU) for the other two datasets. For video paragraph captioning, we use commonly-adopted metrics CIDEr (Vedantam et al., 2015) (C) and METEOR (Banerjee and Lavie, 2005) (M) and report results on YouCook2 and ActivityNet Captions. As to dense video captioning, we follow the existing protocol (Krishna et al., 2017) to compute captioning metrics over the matched pairs between generated sentences and the ground truth. SODA_c (Fujita et al., 2020) (S) is also used to measure the temporal coherence for a set of captions. This task is evaluated on YouCook2 and ActivityNet Captions as well.
+
+# 4.2 Results
+
+We evaluate our instruction-tuned model on three video-language tasks: moment retrieval, video paragraph captioning, and dense video captioning. Note that all results are obtained from one shared model and different tasks are addressed by changing the prompting instructions at inference time only.
+
+Moment retrieval. In Table 2, our method is compared with state-of-the-art algorithms for this task on three representative datasets, QVHighlights (Lei et al., 2021), Charades-STA (Gao et al., 2017), and ActivityNet Captions (Krishna et al., 2017). It can be observed that our method achieves comparable (mostly better) performance on all three datasets. In particular, on QVHighlights we achieve 68.4, 51.9, and 45.0 for R@0.5, R@0.7 and average mAP respectively, improving the best-performing baseline UniVTG with pre-training substantially by +3.0,
+
+Table 2: Moment retrieval on QVHighlights (test), Charades-STA (test), and ActivityNet Captions (val_2). We bold the best, underline the second-best.  
+
+<table><tr><td rowspan="2">Method</td><td rowspan="2">Backbone</td><td colspan="2">YouCook2</td><td colspan="2">ActivityNet</td></tr><tr><td>C</td><td>M</td><td>C</td><td>M</td></tr><tr><td colspan="6">With GT Proposals</td></tr><tr><td>VTransformer (Zhou et al., 2018b)</td><td>V (ResNet-200) + F</td><td>32.3</td><td>15.7</td><td>22.2</td><td>15.6</td></tr><tr><td>Transformer-XL (Dai et al., 2019)</td><td>V (ResNet-200) + F</td><td>26.4</td><td>14.8</td><td>21.7</td><td>15.1</td></tr><tr><td>MART (Lei et al., 2020a)</td><td>V (ResNet-200) + F</td><td>35.7</td><td>15.9</td><td>23.4</td><td>15.7</td></tr><tr><td>GVDSup (Zhou et al., 2019)</td><td>V (ResNet-101) + F + O</td><td>—</td><td>—</td><td>22.9</td><td>16.4</td></tr><tr><td>AdvInf (Park et al., 2019)</td><td>V (ResNet-101) + F + O</td><td>—</td><td>—</td><td>21.0</td><td>16.6</td></tr><tr><td>PDVC (Wang et al., 2021)</td><td>V + F (TSN)</td><td>—</td><td>—</td><td>27.3</td><td>15.9</td></tr><tr><td colspan="6">With Learned Proposals</td></tr><tr><td>MFT (Xiong et al., 2018)</td><td>V + F (TSN)</td><td>—</td><td>—</td><td>19.1</td><td>14.7</td></tr><tr><td>PDVC (Wang et al., 2021)</td><td>V + F (TSN)</td><td>—</td><td>—</td><td>20.5</td><td>15.8</td></tr><tr><td>PDVC (Wang et al., 2021)</td><td>V (CLIP)</td><td>—</td><td>—</td><td>23.6</td><td>15.9</td></tr><tr><td>TDPC (Song et al., 2021)</td><td>V (ResNet-200) + F</td><td>—</td><td>—</td><td>26.5</td><td>15.6</td></tr><tr><td>Vid2Seq (Yang et al., 2023)</td><td>V (CLIP)</td><td>—</td><td>—</td><td>28.0</td><td>17.0</td></tr><tr><td>GVL (Wang et al., 2023a)</td><td>V (TSN)</td><td>—</td><td>—</td><td>26.0</td><td>16.3</td></tr><tr><td>Video-LLaMA (Zhang et al., 2023a)</td><td>V (CLIP)</td><td>36.2</td><td>16.5</td><td>32.8</td><td>15.9</td></tr><tr><td>MovieChat (Song et al., 2023)</td><td>V (CLIP)</td><td>38.5</td><td>18.8</td><td>35.6</td><td>18.2</td></tr><tr><td>UNICORN</td><td>V (CLIP)</td><td>37.8</td><td>18.3</td><td>34.8</td><td>17.3</td></tr></table>
+
+Table 3: Video paragraph captioning results on YouCook2 (val) and ActivityNet Captions (ae-test). V/F/O refers to visual/flow/object features.
+
++1.8 and +1.4. In contrast to complicated designs such as a localization loss in previous approaches, we remove most of the specification and only use a generic language modeling loss: UNICORN is mainly based on the intuition that if a model knows about where the moment is, we just need to teach it how to read the location out. In summary, UNICORN makes minimal assumptions on the task yet accomplishes it with superior performance.
+
+Video paragraph captioning. Table 3 shows the video paragraph captioning results. In UNICORN, we consider this task as a general captioning problem. Without any customized training objectives or prior knowledge on the input such as ground-truth event proposals as in previous methods (Park et al., 2019; Lei et al., 2020a), our method demonstrates outstanding performance over other baselines under both settings of ground truth or learned proposals. It further showcases the strong adaptation of LMMs to downstream tasks through instruction tuning with high-quality instruction-following data.
+
+Dense video captioning. We generate dense video captions following the procedure in Section 3.1 and
+
+(a) Comparison of training strategies.  
+
+<table><tr><td rowspan="2">Training Setup</td><td colspan="3">QVHighlights</td><td colspan="2">ActivityNet</td></tr><tr><td>R@0.5</td><td>R@0.7</td><td>mAP</td><td>C</td><td>M</td></tr><tr><td>Single-task, single-dataset</td><td>66.3</td><td>51.5</td><td>42.8</td><td>33.6</td><td>16.4</td></tr><tr><td>Single-task, multi-dataset</td><td>68.2</td><td>52.3</td><td>44.8</td><td>34.6</td><td>16.9</td></tr><tr><td>Multi-task, multi-dataset</td><td>69.5</td><td>54.4</td><td>45.3</td><td>34.8</td><td>17.3</td></tr></table>
+
+(b) The number of frames.  
+
+<table><tr><td rowspan="2">frames</td><td colspan="3">QVHighlights</td><td colspan="2">ActivityNet</td></tr><tr><td>R@0.5</td><td>R@0.7</td><td>mAP</td><td>C</td><td>M</td></tr><tr><td>25</td><td>61.5</td><td>37.4</td><td>35.0</td><td>33.4</td><td>16.9</td></tr><tr><td>50</td><td>65.4</td><td>47.9</td><td>41.4</td><td>34.5</td><td>17.0</td></tr><tr><td>75</td><td>69.5</td><td>54.4</td><td>45.3</td><td>34.8</td><td>17.3</td></tr><tr><td>100</td><td>67.8</td><td>52.8</td><td>44.7</td><td>34.6</td><td>17.3</td></tr></table>
+
+(c) LoRA & temporal modeling.  
+
+<table><tr><td rowspan="2">LoRA</td><td rowspan="2">Temporal modeling</td><td colspan="3">QVHighlights</td><td colspan="2">ActivityNet</td></tr><tr><td>R@0.5</td><td colspan="2">R@0.7 mAP</td><td>C</td><td>M</td></tr><tr><td>X</td><td>X</td><td>60.6</td><td>36.4</td><td>33.2</td><td>23.0</td><td>16.0</td></tr><tr><td>✓</td><td>X</td><td>66.7</td><td>49.2</td><td>39.8</td><td>34.4</td><td>17.2</td></tr><tr><td>X</td><td>✓</td><td>65.5</td><td>47.0</td><td>40.4</td><td>27.6</td><td>16.8</td></tr><tr><td>✓</td><td>✓</td><td>69.5</td><td>54.4</td><td>45.3</td><td>34.8</td><td>17.3</td></tr></table>
+
+Table 4: Ablation studies on training strategies and model designs of LoRA and temporal modeling.  
+
+<table><tr><td rowspan="2">Method</td><td rowspan="2">Backbone</td><td colspan="3">YouCook2</td><td colspan="3">ActivityNet</td></tr><tr><td>S</td><td>C</td><td>M</td><td>S</td><td>C</td><td>M</td></tr><tr><td>MT (Zhou et al., 2018b)</td><td>TSN</td><td>—</td><td>6.1</td><td>3.2</td><td>—</td><td>9.3</td><td>5.0</td></tr><tr><td>ECHR (Wang et al., 2020)</td><td>C3D</td><td>—</td><td>—</td><td>3.8</td><td>3.2</td><td>14.7</td><td>7.2</td></tr><tr><td>PDVC (Wang et al., 2021)</td><td>TSN</td><td>4.4</td><td>22.7</td><td>4.7</td><td>5.4</td><td>29.0</td><td>8.0</td></tr><tr><td>PDVC (Wang et al., 2021)</td><td>CLIP</td><td>4.9</td><td>28.9</td><td>5.7</td><td>6.0</td><td>29.3</td><td>7.6</td></tr><tr><td>UEDVC (Zhang et al., 2022)</td><td>TSN</td><td>—</td><td>—</td><td>—</td><td>5.5</td><td>26.9</td><td>7.3</td></tr><tr><td>E2ESG (Zhu et al., 2022)</td><td>C3D</td><td>—</td><td>25.0</td><td>3.5</td><td>—</td><td>—</td><td>—</td></tr><tr><td>Vid2Seq (Yang et al., 2023)</td><td>CLIP</td><td>5.7</td><td>25.3</td><td>—</td><td>5.9</td><td>30.2</td><td>8.5</td></tr><tr><td>GVL (Wang et al., 2023a)</td><td>TSN</td><td>4.9</td><td>26.5</td><td>5.0</td><td>6.2</td><td>32.8</td><td>8.5</td></tr><tr><td>UNICORN</td><td>CLIP</td><td>5.7</td><td>37.0</td><td>7.7</td><td>6.3</td><td>35.4</td><td>9.2</td></tr></table>
+
+Table 5: Results of DVC on YouCook2 (val) and ActivityNet Captions (val_1 and val_2).
+
+evaluate the performance in Table 5. It can be observed that our method takes the lead among the compared approaches, including Vid2Seq which leverages language models to predict captions and timestamps simultaneously. These promising results also validate our divide-and-conquer strategy for dense video captioning. Such an inference design makes the training more efficient without learning on redundant and lengthy DVC data again while still achieving competitive results.
+
+# 4.3 Ablation Studies
+
+We conduct ablation studies to analyze effects of the key components in UNICORN, including training strategies, the choice of time tokens, and various model designs. We evaluate on QVHighlights (val) for moment retrieval and ActivityNet Captions (ae-test) for video paragraph captioning. Additional analysis including base model selection can be found at Appendix D.
+
+Training strategies. We study the effects of training strategies for UNICORN. Specifically, three strategies are considered: single-task & single dataset, single task & multi-dataset, and multitask & multi-dataset. For the single-task version, we fine-tune two separate models with corresponding instructions tailored for moment retrieval and video paragraph captioning respectively, and select one representative dataset for each task for evaluation. For the single-dataset version, we train only on the training split of the evaluation dataset (i.e., QVHighlights for moment retrieval and Activ
+
+ityNet Captions for video paragraph captioning)
+
+We report detailed results in Table 4a. By introducing datasets from different domains for the same task, we can improve the model's capability on the single dataset. Besides, in contrast to traditional multi-task training strategies, instruction tuning on various descriptions works as a unified approach to integrate different tasks and can even boost the performance from understanding a video from multiple perspectives. Meanwhile, it is more convenient to store only one model to accomplish distinct tasks, which narrows the gap from constructing a general-purpose foundation model.
+
+Number of frames. By default, we evenly sample 75 frames from a video as model inputs. In Table 4b, we study the impact with #frames of 25, 50, 75, and 100. The performance generally improves when we adopt more frames while it saturates or even gets worse around 100 frames. Since the videos in the datasets we studied are usually not very long (e.g., videos in QVHighlights are on average 150 seconds long), we hypothesize that 75 frames are enough to cover the semantic information needed for the tasks. We report more results about #frames in Appendix D.
+
+LoRA. We use a parameter-efficient fine-tuning method LoRA to fine-tune the LLM of UNICORN. In Table 4c, LoRA has been proven effective in boosting performance for downstream tasks (row 1 vs. 2 and row 3 vs. 4). It is expected that frozen LLM would not work properly as we have assigned new meanings to original digit tokens to represent discrete time bins, and LoRA training mitigates the issue without tuning the whole LLM intensively.
+
+Temporal modeling. Since our model is adapted from image-based InstructBLIP, we include an additional module with self-attention layers to incorporate temporal information for videos in Figure 3. As shown in Table 4c, when temporal modeling is enabled from average pooling to self-attention interaction (row 1 vs. 3 and row 2 vs. 4), there is substantial improvement in moment retrieval and paragraph captioning, indicating the necessity of
+
+![](images/3adb8584ea4600d17402be7996e13130b9c3d8bbec1830010d374baef6343e32.jpg)  
+Figure 2: Comparison among captions from HTM-ASR, HTM-AA, and UNICORN respectively. For HTM-UNICORN, we show three sets of generated captions via beam search, coded with different colors.
+
+this module for temporal video-language tasks.
+
+Time tokens. We can either introduce new dedicated time tokens or directly
+
+<table><tr><td>Time tokens</td><td>R@0.5</td><td>R@0.7</td><td>mAP</td></tr><tr><td>Dedicated</td><td>64.1</td><td>48.2</td><td>40.7</td></tr><tr><td>Original vocab</td><td>66.3</td><td>51.5</td><td>42.8</td></tr></table>
+
+Table 6: Different time token.
+
+use the digits in the original vocabulary to represent time. We investigate the impact of the two strategies in the single-task, single-dataset setup on QVHighlights in Table 6. We observed that the original vocabulary performs better than new dedicated tokens, which indicates the knowledge of digits in LLM can be readily transferred to our tasks. Meanwhile, new tokens would introduce extra training overheads and increase the number of trainable parameters by 262M, more than double of the original value, see details in Appendix D.
+
+# 4.4 Auto Annotation of HowTo100M
+
+Thanks to the generalization of LMMs, the model to handle temporal video-language tasks can be deployed on unseen public internet videos such as HowTo100M (Miech et al., 2019). These videos are paired with auto speech recognition (ASR) transcripts, a majority of which are not visually and temporally aligned (Miech et al., 2020; Tang et al., 2021; Han et al., 2022). Since our model is capable of generating dense captions, it is promising to leverage UNICORN for annotating the dataset automatically. We use our trained model to densely caption a subset of 240K videos from HowTo100M (Han et al., 2022) and denote the dataset as HTM-UNICORN. We anonymize names with their pronouns and prompt the model not to generate offensive responses. We compare it with
+
+two variants with the same set of videos, HTM-ASR (Miech et al., 2019) with original ASR transcripts, and HTM-AA (Han et al., 2022) which has been aligned temporally via an automated process. Note that UNICORN can output diverse captions using beam search (Vijayakumar et al., 2016), which can increase the training data size and as a result improve model performance with more data.
+
+In Figure 2, we present an qualitative comparison of three variants. Our HTM-UNICORN is the best aligned with the input video both visually and temporally, compared with HTM-ASR and HTM-AA. In addition, captions from different sets can complement each other, leading to more comprehensive descriptions of the video. Quantitatively, we use three HowTo100M variants to pre-train the model for moment retrieval, and evaluate on QVHighlights under zero-shot and fine-tuning settings. We convert these datasets into the instruction-following format described in Section 3.1, and train the model from the same initialization. In Table 7, we observe that our automatically annotated HTM achieves superior zero-shot performance, which shows the better alignment of moments and timestamps. For fine-tuning, we notice that performance even degrades when pre-trained on HTM-ASR and HTM-AA, potentially due to data noise, while the model pre-trained on HTM-UNICORN outperforms other variants, reflecting the high quality of the generated dataset.
+
+Besides, we follow (Han et al., 2022) to conduct end-to-end representation learning with an InfoNCE loss (Miech et al., 2020). After contrastive pre-training, we evaluate video representations by
+
+<table><tr><td rowspan="2">Dataset</td><td rowspan="2">#queries</td><td colspan="3">Zero-shot</td><td colspan="3">Fine-tuning</td></tr><tr><td>R@0.5</td><td>R@0.7</td><td>mAP</td><td>R@0.5</td><td>R@0.7</td><td>mAP</td></tr><tr><td>InstructBLIP</td><td>—</td><td>—</td><td>—</td><td>—</td><td>66.3</td><td>51.5</td><td>42.8</td></tr><tr><td>HTM-ASR (Miech et al., 2019)</td><td>5.0M</td><td>7.7</td><td>2.8</td><td>1.9</td><td>63.5</td><td>48.3</td><td>40.3</td></tr><tr><td>HTM-AA (Han et al., 2022)</td><td>3.3M</td><td>13.0</td><td>4.8</td><td>3.5</td><td>65.8</td><td>50.6</td><td>42.1</td></tr><tr><td>HTM-UNICORN</td><td>690K</td><td>44.2</td><td>26.4</td><td>26.0</td><td>68.9</td><td>53.6</td><td>45.0</td></tr><tr><td>HTM-UNICORN × 2</td><td>1.4M</td><td>47.5</td><td>30.8</td><td>29.9</td><td>69.5</td><td>54.0</td><td>45.2</td></tr><tr><td>HTM-UNICORN × 3</td><td>2.1M</td><td>50.0</td><td>32.7</td><td>30.4</td><td>70.2</td><td>54.6</td><td>45.5</td></tr></table>
+
+Table 7: Zero-shot and fine-tuning moment retrieval evaluation on QVHighlights (val). HTM-UNICORN  $\times n$  indicates we generated  $n$  sets of captions for a video.  
+
+<table><tr><td>PT Dataset</td><td>Backbone</td><td>UCF101</td><td>HMDB51</td><td>K400</td></tr><tr><td>HTM-ASR (Miech et al., 2020)</td><td>S3D</td><td>82.1</td><td>55.2</td><td>55.7</td></tr><tr><td>HTM-AA (Han et al., 2022)</td><td>S3D</td><td>83.2</td><td>56.7</td><td>56.2</td></tr><tr><td>HTM-UNICORN</td><td>S3D</td><td>84.1</td><td>57.7</td><td>56.6</td></tr></table>
+
+Table 8: Linear probing accuracy for action recognition.
+
+linear probing on three action recognition datasets, UCF101 (Soomro et al., 2012), HMDB51 (Kuehne et al., 2011), and Kinetics-400 (K400) (Kay et al., 2017) in Table 8. UNICORN achieves the highest accuracy on all three datasets, which again demonstrates the best quality of our generated captions.
+
+Captions generated from our automated annotation pipeline has shown to be better than noisy web data both qualitatively and quantitatively. As data quality and quantity are crucial for the performance of large models (Zhou et al., 2024; Ji et al., 2023; Liu et al., 2023a), we hope such a pipeline could be useful for empowering the development of future large multimodal models.
+
+# 5 Conclusion
+
+In this paper, we propose a unified causal video-oriented language modeling framework UNICORN to address temporal video-language tasks. By fine-tuning on instruction-following data constructed from existing datasets, our model achieves outstanding performance on various downstream tasks including moment retrieval, video paragraph captioning and dense video captioning. We further show that UNICORN can be leveraged in automatic annotation on internet videos such as HowTo100M for semantically- and temporally-aligned captions. These captions can be used to improve video-language model performance against ASR ones. In conclusion, UNICORN paves the way towards a general-purpose foundation model that explicitly considers temporal information.
+
+# 6 Limitations
+
+Currently, UNICORN is good at localizing an event which only appears once in the video, but would be confused when an event happens more than once.
+
+This is due to the training data mostly have events appearing once. Future work can be collecting data with events that appear more than once to improve models' ability on these scenarios. Besides, it is a promising direction to include more sophisticated and efficient temporal designs, such as leveraging a video perceiver to encode video information more comprehensively using more frames. We will also consider human evaluation to better understand the quality of generated captions.
+
+# References
+
+2023. Chatgpt. https://openai.com/blog/chatgpt.  
+2023. Vicuna. https://github.com/lm-sys/ FastChat.  
+Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. 2023. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966.  
+Satanjeev Banerjee and Alon Lavie. 2005. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In ACL workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization.  
+Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. 2023. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195.  
+Ting Chen, Saurabh Saxena, Lala Li, David J Fleet, and Geoffrey Hinton. 2022. Pix2seq: A language modeling framework for object detection. In ICLR.  
+Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. 2023. Instructblip: Towards general-purpose vision-language models with instruction tuning. In NeurIPS.  
+Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov. 2019. Transformer-xl: Attentive language models beyond a fixed-length context. In ACL.  
+Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. 2023. Eva: Exploring the limits of masked visual representation learning at scale. In CVPR.  
+Soichiro Fujita, Tsutomu Hirao, Hidetaka Kamigaito, Manabu Okumura, and Masaaki Nagata. 2020. Soda: Story oriented dense video captioning evaluation framework. In ECCV.
+
+Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevata. 2017. Tall: Temporal activity localization via language query. In ICCV.  
+Tengda Han, Weidi Xie, and Andrew Zisserman. 2022. Temporal alignment networks for long-term video. In CVPR.  
+Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. 2017. Localizing moments in video with natural language. In ICCV.  
+Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-rank adaptation of large language models. In ICLR.  
+Yunjie Ji, Yong Deng, Yan Gong, Yiping Peng, Qiang Niu, Lei Zhang, Baochang Ma, and Xiangang Li. 2023. Exploring the impact of instruction data scaling on large language models: An empirical study on real-world use cases. arXiv preprint arXiv:2303.14742.  
+Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. 2017. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950.  
+Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. 2017. Dense-captioning events in videos. In CVPR.  
+Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. 2011. Hmdb: a large video database for human motion recognition. In ICCV.  
+Jie Lei, Tamara L Berg, and Mohit Bansal. 2021. Detecting moments and highlights in videos via natural language queries. In NeurIPS.  
+Jie Lei, Liwei Wang, Yelong Shen, Dong Yu, Tamara Berg, and Mohit Bansal. 2020a. Mart: Memory-augmented recurrent transformer for coherent video paragraph captioning. In ACL.  
+Jie Lei, Licheng Yu, Tamara L Berg, and Mohit Bansal. 2020b. Tvr: A large-scale dataset for video-subtitle moment retrieval. In ECCV.  
+Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Fanyi Pu, Jingkang Yang, Chunyuan Li, and Ziwei Liu. 2023a. Mimic-it: Multi-modal in-context instruction tuning. arXiv preprint arXiv:2306.05425.  
+Dongxu Li, Junnan Li, Hung Le, Guangsen Wang, Silvio Savarese, and Steven C.H. Hoi. 2023b. LAVIS: A one-stop library for language-vision intelligence. In ACL.  
+Dongxu Li, Junnan Li, Hongdong Li, Juan Carlos Niebles, and Steven CH Hoi. 2022. Align and prompt: Video-and-language pre-training with entity prompts. In CVPR.
+
+KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. 2023c. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355.  
+Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. 2023. Univtg: Towards unified video-language temporal grounding. In ICCV.  
+Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. 2023a. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744.  
+Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023b. Visual instruction tuning. In NeurIPS.  
+Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In ICLR.  
+Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. 2020. End-to-end learning of visual representations from uncurated instructional videos. In CVPR.  
+Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. 2019. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV.  
+Salman Khan Muhammad Maaz, Hanoona Rasheed and Fahad Khan. 2023. Video-chatgpt: Towards detailed video understanding via large vision and language models. ArXiv 2306.05424.  
+Jonghwan Mun, Minsu Cho, and Bohyung Han. 2020. Local-global video-text interactions for temporal grounding. In CVPR.  
+OpenAI. 2023. Gpt-4 technical report. ArXiv, abs/2303.08774.  
+Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F Christiano, Jan Leike, and Ryan Lowe. 2022. In NeurIPS.  
+Jae Sung Park, Marcus Rohrbach, Trevor Darrell, and Anna Rohrbach. 2019. Adversarial inference for multi-sentence video description. In CVPR.  
+Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. 2023. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824.  
+Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Xun Guo, Tian Ye, Yan Lu, Jenq-Neng Hwang, et al. 2023. Moviechat: From dense token to sparse memory for long video understanding. arXiv preprint arXiv:2307.16449.
+
+Yuqing Song, Shizhe Chen, and Qin Jin. 2021. Towards diverse paragraph captioning for untrimmed videos. In CVPR.  
+Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. 2012. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402.  
+Zineng Tang, Jie Lei, and Mohit Bansal. 2021. DeCEMBERT: Learning from noisy instructional videos via dense captions and entropy minimization. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies.  
+Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023a. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971.  
+Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023b. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288.  
+Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. 2015. Cider: Consensus-based image description evaluation. In CVPR.  
+Ashwin K Vijayakumar, Michael Cogswell, Ramprasath R Selvaraju, Qing Sun, Stefan Lee, David Crandall, and Dhruv Batra. 2016. Diverse beam search: Decoding diverse solutions from neural sequence models. arXiv preprint arXiv:1610.02424.  
+Teng Wang, Jinrui Zhang, Feng Zheng, Wenhao Jiang, Ran Cheng, and Ping Luo. 2023a. Learning grounded vision-language representation for versatile understanding in untrimmed videos. arXiv preprint arXiv:2303.06378.  
+Teng Wang, Ruimao Zhang, Zhichao Lu, Feng Zheng, Ran Cheng, and Ping Luo. 2021. End-to-end dense video captioning with parallel decoding. In ICCV.  
+Teng Wang, Huicheng Zheng, Mingjing Yu, Qian Tian, and Haifeng Hu. 2020. Event-centric hierarchical representation for dense video captioning. IEEE Transactions on Circuits and Systems for Video Technology.  
+Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. 2023b. Vision-llm: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175.  
+Yilei Xiong, Bo Dai, and Dahua Lin. 2018. Move forward and tell: A progressive generator of video descriptions. In ECCV.
+
+Shen Yan, Xuehan Xiong, Arsha Nagrani, Anurag Arnab, Zhonghao Wang, Weina Ge, David Ross, and Cordelia Schmid. 2023. Unloc: A unified framework for video localization tasks. In ICCV.  
+Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. 2023. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In CVPR.  
+Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. 2022. Merlot reserve: Neural script knowledge through vision and language and sound. In CVPR.  
+Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. 2020. Dense regression network for video grounding. In CVPR.  
+Hang Zhang, Xin Li, and Lidong Bing. 2023a. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858.  
+Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. 2020a. Span-based localizing network for natural language video localization. In ACL.  
+Qi Zhang, Yuqing Song, and Qin Jin. 2022. Unifying event detection and captioning as sequence generation via pre-training. In ECCV.  
+Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Kai Chen, and Ping Luo. 2023b. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv preprint arXiv:2307.03601.  
+Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. 2020b. Learning 2d temporal adjacent networks for moment localization with natural language. In AAAI.  
+Chunting Zhou, Pengfei Liu, Puxin Xu, Srinivasan Iyer, Jiao Sun, Yuning Mao, Xuezhe Ma, Avia Efrat, Ping Yu, Lili Yu, et al. 2024. Lima: Less is more for alignment. NeurIPS, 36.  
+Luowei Zhou, Yannis Kalantidis, Xinlei Chen, Jason J Corso, and Marcus Rohrbach. 2019. Grounded video description. In CVPR.  
+Luowei Zhou, Chenliang Xu, and Jason J Corso. 2018a. Towards automatic learning of procedures from web instructional videos. In AAAI.  
+Luowei Zhou, Yingbo Zhou, Jason J Corso, Richard Socher, and Caiming Xiong. 2018b. End-to-end dense video captioning with masked transformer. In CVPR.  
+Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. 2023. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592.
+
+Wanrong Zhu, Bo Pang, Ashish V Thapliyal, William Yang Wang, and Radu Soricut. 2022. End-to-end dense video captioning as sequence generation. In COLING.
+
+# A Societal Impact
+
+Similar to many data-driven methods, the predictions from our model might be inaccurate and biased towards the distribution of data on which it is trained on. Therefore, users should not completely rely on the model in real-world scenarios.
+
+# B Instruction Templates
+
+We provide the list of instruction templates for moment retrieval and video paragraph captioning respectively in Table 9 and Table 10.
+
+- "Please predict start and end time of the following moment."  
+- "Can you tell me the time window of this event?"  
+What is the location of the moment?
+
+Table 9: The list of instructions for moment retrieval.
+
+- "Provide a detailed description of the given video, capturing its key moments."  
+Describe the following video in detail, including the actions and scenes.  
+- "Clarify the contents of the displayed video with great detail, focusing on its progression."  
+- "Offer a thorough analysis of the video, discussing its various elements and storyline."
+
+# C Datasets
+
+In this section, we present more details about datasets used for both instruction-tuning and evaluation. An overview of statistics of training data is presented in Table 11. We mix all samples of the same task across datasets and obtain two large training sets: one for moment retrieval  $S_{\mathrm{MR}}$  with  $|S_{\mathrm{MR}}| = 71829$  video-query pairs, and the other for paragraph captioning  $S_{\mathrm{VPC}}$  with  $|S_{\mathrm{VPC}}| = 16533$  videos. These datasets are introduced comprehensively below.
+
+QVHighlights (Lei et al., 2021). This dataset includes 10,148 trimmed videos with an average length of 150 sec that covers daily vlogs, travel vlogs, and news events scenarios. There are in total 10,310 queries associated with 18, 367 moments. Following (Lei et al., 2021), we use train split for
+
+Table 10: The list of instructions for video paragraph captioning.  
+
+<table><tr><td>Dataset</td><td>Domain</td><td>#Videos</td><td>#Queries</td><td>MR</td><td>VPC</td></tr><tr><td>QVHighlights</td><td>Vlog</td><td>7100</td><td>12803</td><td>✓</td><td>✗</td></tr><tr><td>Charades-STA</td><td>Activity</td><td>5336</td><td>12404</td><td>✓</td><td>✓</td></tr><tr><td>ActivityNet Captions</td><td>Activity</td><td>10009</td><td>37421</td><td>✓</td><td>✓</td></tr><tr><td>Youcook2</td><td>Instruction</td><td>1188</td><td>9201</td><td>✓</td><td>✓</td></tr></table>
+
+Table 11: Statistics of training data.
+
+instruction tuning of moment retrieval, test split for evaluation, and val split for ablation studies. The license is Attribution-NonCommercial-ShareAlike 4.0 International and our usage is consistent with its license.
+
+Charades-STA (Gao et al., 2017). The dataset contains 6,672 videos with an average duration of 30.6 sec and 16,128 moment/caption pairs. Each video is annotated with 2.4 segments on average. We use train split for instruction tuning and test for evaluation. The license is License Non-Commercial Use and our usage is consistent with its license.
+
+ActivityNet Captions (Krishna et al., 2017). The dataset contains 14,934 untrimmed videos of various human activities from YouTube. On average, each video lasts 120s and is annotated with 3.7 temporally-localized sentences. The dataset is split into 10,009 and 4,925 videos for training and validation, respectively. train split is included in instruction tuning for both moment retrieval and video paragraph captioning. The validation set has two independent dense video captioning annotations (val_1 and val_2). For moment retrieval, we evaluate on val_2 according to prior work (Yan et al., 2023). For video paragraph captioning, we report results on the ae-test split following (Lei et al., 2020a; Zhou et al., 2019). For dense video captioning, we use both val_1 and val_2 for evaluation, by computing the average of the scores over each set for SODA_c and by using the standard evaluation tool (Krishna et al., 2017) for all other dense event captioning metrics. The license is not specified by the original authors.
+
+YouCook2 (Zhou et al., 2018a). It has 1,790 untrimmed videos of cooking procedures. On average, each video lasts 320s and is annotated with 7.7 temporally-localized sentences. The dataset is split into 1,333 videos for training and 457 videos for validation. We use train split for instruction tuning and evaluate on val split. The license is MIT License and our usage is consistent with its license.
+
+Besides, we adopt a subset of HowTo100M (Han et al., 2022) with 240K videos for automatic an
+
+<table><tr><td rowspan="2">#frames</td><td colspan="3">QVHighlights</td><td colspan="3">Charades-STA</td><td colspan="3">ActivityNet Captions</td></tr><tr><td>R@0.5</td><td>R@0.7</td><td>mAP avg</td><td>R@0.5</td><td>R@0.7</td><td>mIoU</td><td>R@0.5</td><td>R@0.7</td><td>mIoU</td></tr><tr><td>25</td><td>61.5</td><td>37.4</td><td>35.0</td><td>63.4</td><td>38.0</td><td>55.0</td><td>43.6</td><td>25.9</td><td>43.9</td></tr><tr><td>50</td><td>65.4</td><td>47.9</td><td>41.4</td><td>67.3</td><td>46.0</td><td>58.1</td><td>46.2</td><td>28.3</td><td>46.1</td></tr><tr><td>75</td><td>69.5</td><td>54.4</td><td>45.3</td><td>69.0</td><td>45.6</td><td>58.9</td><td>48.4</td><td>29.8</td><td>47.1</td></tr><tr><td>100</td><td>67.8</td><td>52.8</td><td>44.7</td><td>68.4</td><td>46.0</td><td>58.4</td><td>48.5</td><td>29.3</td><td>46.5</td></tr></table>
+
+Table 12: Effects of the number of frames on moment retrieval.  
+
+<table><tr><td rowspan="2">LoRA</td><td rowspan="2">Temporal modeling</td><td colspan="3">QVHighlights</td><td colspan="3">Charades-STA</td><td colspan="3">ActivityNet Captions</td></tr><tr><td>R@0.5</td><td>R@0.7</td><td>mAP avg</td><td>R@0.5</td><td>R@0.7</td><td>mIoU</td><td>R@0.5</td><td>R@0.7</td><td>mIoU</td></tr><tr><td>X</td><td>X</td><td>60.6</td><td>36.4</td><td>33.2</td><td>62.3</td><td>36.8</td><td>55.0</td><td>43.1</td><td>25.9</td><td>43.8</td></tr><tr><td>✓</td><td>X</td><td>66.7</td><td>49.2</td><td>39.8</td><td>67.6</td><td>44.5</td><td>58.2</td><td>44.3</td><td>27.5</td><td>45.1</td></tr><tr><td>X</td><td>✓</td><td>65.5</td><td>47.0</td><td>40.4</td><td>66.8</td><td>43.4</td><td>57.3</td><td>44.9</td><td>27.7</td><td>45.3</td></tr><tr><td>✓</td><td>✓</td><td>69.5</td><td>54.5</td><td>45.3</td><td>69.0</td><td>45.6</td><td>58.9</td><td>48.4</td><td>29.8</td><td>47.1</td></tr></table>
+
+notation. It is a large-scale dataset of narrated videos with an emphasis on instructional videos where content creators teach complex tasks with an explicit intention of explaining the visual content on screen (Miech et al., 2019). The license is not specified by the original authors. For evaluation, we leverage three action recognition tasks: UCF101 (license not specified) (Soomro et al., 2012), HMDB51 (CC BY 4.0) (Kuehne et al., 2011) and Kinetics-400 (CC BY 4.0) (Kay et al., 2017). Our usage is consistent with their licenses.
+
+# D Additional Results
+
+Additional experimental results are reported in this section, including the analysis of dedicated time tokens, effects of the number of frames, effects of LoRA and temporal modeling, and the influence of different pre-training data ratios. We only run instruction tuning once and all results in Section 4 and this section are from this single model.
+
+Extra overheads of dedicated time tokens. As mentioned in Section 4.3, new dedicated time tokens would introduce a considerably larger number of trainable parameters. In particular, given the current implementation of LLMs, it is challenging to train new tokens only without affecting the rest of parameters in the embedding layer and the final output layer. Thus, we take an alternative to tune all parameters in these two layers: given the original vocabulary size of 32000, the number of new time tokens of 75, and the hidden dimension of 4096, the total number of trainable parameters is computed as:  $(32000 + 75) \times 4096 \times 2 = 262\mathrm{M}$ .
+
+Number of frames. In addition to results presented in Table 4b, we show more complete experiments on moment retrieval and video paragraph captioning in Table 12 and Table 14. The trends are consistent with what we observed in Table 4b, where 75 frames are enough to cover all the semantic information needed for these two tasks.
+
+Table 13: Effects of LoRA and temporal modeling on moment retrieval.  
+
+<table><tr><td rowspan="2">#frames</td><td colspan="2">YouCook2</td><td colspan="2">ActivityNet</td></tr><tr><td>C</td><td>M</td><td>C</td><td>M</td></tr><tr><td>25</td><td>29.2</td><td>16.9</td><td>33.4</td><td>16.9</td></tr><tr><td>50</td><td>34.3</td><td>17.8</td><td>34.5</td><td>17.0</td></tr><tr><td>75</td><td>37.8</td><td>18.3</td><td>34.8</td><td>17.3</td></tr><tr><td>100</td><td>37.4</td><td>18.5</td><td>34.6</td><td>17.3</td></tr></table>
+
+LoRA and temporal modeling. We present more thorough and comprehensive experimental results to understand the effects of LoRA and temporal modeling. In Table 13 and 15, we can conclude that both LoRA training and temporal modeling contribute to performance gains in moment retrieval and video paragraph captioning.
+
+Table 14: Effects of the number of frames on video paragraph captioning.  
+
+<table><tr><td rowspan="2">LoRA</td><td rowspan="2">Temporal modeling</td><td colspan="2">YouCook2</td><td colspan="2">ActivityNet</td></tr><tr><td>C</td><td>M</td><td>C</td><td>M</td></tr><tr><td>X</td><td>X</td><td>25.7</td><td>16.9</td><td>23.0</td><td>16.0</td></tr><tr><td>✓</td><td>X</td><td>32.3</td><td>17.7</td><td>34.4</td><td>17.2</td></tr><tr><td>X</td><td>✓</td><td>26.5</td><td>17.4</td><td>27.6</td><td>16.8</td></tr><tr><td>✓</td><td>✓</td><td>37.8</td><td>18.3</td><td>34.8</td><td>17.3</td></tr></table>
+
+Table 15: Effects of LoRA and temporal modeling on video paragraph captioning.
+
+Base model selection. It should be emphasized that UNICORN is a generic framework to which we can flexibly utilize various LMMs as the base model with a simple re-design to take video inputs. We analyze the effects of adopting different base models like LLaVA (Liu et al., 2023b) here to justify our framework design. Specifically, we instruction-tuned LLaVA for moment retrieval and video paragraph captioning. Results are shown in Table 16 and it is expected that the performance of LLaVA variant drops compared with our InstructBLIP variant, due to the information loss from 256 frame-level tokens pooled to one token. Besides, InstructBLIP has a QFormer while LLaVA only uses a simple projection layer, which may be insufficient to align video and language. We also added an experiment with InstrutBLIP-13B and observed performance gains with a larger model size.
+
+<table><tr><td rowspan="2">Base model</td><td colspan="3">QVHighlights</td><td colspan="2">ActivityNet</td></tr><tr><td>R@0.5</td><td>R@0.7</td><td>mAP</td><td>C</td><td>M</td></tr><tr><td>LLaVA-7B (Liu et al., 2023b)</td><td>66.3</td><td>51.5</td><td>42.8</td><td>33.6</td><td>16.4</td></tr><tr><td>InstructBLIP-7B (Dai et al., 2023)</td><td>68.2</td><td>52.3</td><td>44.8</td><td>34.6</td><td>16.9</td></tr><tr><td>InstructBLIP-13B (Dai et al., 2023)</td><td>69.5</td><td>54.4</td><td>45.3</td><td>34.8</td><td>17.3</td></tr></table>
+
+Table 16: Comparison of different base models.
+
+Ratios of PT dataset. We also alter the ratio of pre-training dataset and record corresponding performance in Figure 3. With only  $25\%$  of the videos, the model using UNICORN captions far outperforms other counterparts trained on all videos, demonstrating the effectiveness of our captions compared to ASR captions and its de-noised version. Meanwhile, UNICORN can generate multiple captions for the same video, with more sets of captions, we see a consistent performance gain from the model. Notably, when using 3 sets of captions  $(\times 3)$ , the performance is improved from 44.2 to 50.0 for R@0.5.
+
+![](images/60eb1fa921fd52c2061335164cbd6f15c7161d8bc3049b9127d33b31047df8ca.jpg)  
+Figure 3: Zero-shot moment retrieval on QVHighlights (val) under different data ratios.

2024/UNICORN_ A Unified Causal Video-Oriented Language-Modeling Framework for Temporal Video-Language Tasks/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:005cb9ea53515076e3967f87cac6eafba983c707c117c6dd722b9d149ef9c38a
+size 648118

2024/UNO Arena for Evaluating Sequential Decision-Making Capability of Large Language Models/24977746-f195-40b6-ae76-06cf317c5f92_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d6e57042e0be2e3a6dcbf107f8b5c2e1a46b6fb468a8a403fe9d301a47425af8
+size 1970343

2024/UNO Arena for Evaluating Sequential Decision-Making Capability of Large Language Models/full.md

@@ -0,0 +1,538 @@
+# UNO Arena for Evaluating Sequential Decision-Making Capability of Large Language Models
+
+Zhanyue Qin $^{1}$ , Haochuan Wang $^{1}$ , Deyuan Liu $^{1}$ , Ziyang Song $^{1}$ , Cunhang Fan $^{2}$ , Zhao Lv $^{2}$ , Jinlin Wu $^{3,4}$ , Zhen Lei $^{3,4,5}$ , Zhiying Tu $^{1}$ , Dianhui Chu $^{1}$ , Xiaoyan Yu $^{6}$ , Dianbo Sui $^{1}$  *
+
+<sup>1</sup> Harbin Institute of Technology, <sup>2</sup> Anhui University,
+
+$^{3}$  CAIR, HKISI-CAS,  $^{4}$ CASIA,  $^{5}$ UCAS,
+
+$^{6}$  Beijing Institute of Technology
+
+johnneyqin@gmail.com, suidianbo@hit.edu.cn
+
+# Abstract
+
+Sequential decision-making refers to algorithms that take into account the dynamics of the environment, where early decisions affect subsequent decisions. With large language models (LLMs) demonstrating powerful capabilities among various tasks, we cannot help but ask: Can Current LLMs Make Sequential Decisions Effectively? In order to answer this question, we propose the UNO Arena based on the card game UNO for evaluating the sequential decision-making capability of LLMs and explain in detail why we choose the UNO game. In the UNO Arena, we also involve some novel metrics based on Monte Carlo methods for evaluating the sequential decision-making capability of LLMs dynamically. Besides, we set up random players, DQN-based reinforcement learning players, and LLM players (e.g. GPT-4, Gemini-pro) for comparison testing. Furthermore, in order to improve the sequential decision-making capability of LLMs, we propose the TUTRI player, which can involve enabling LLMs to reflect on their actions with the summary of game history and the game strategy. Various experimental results demonstrate that the TUTRI player can achieve a notable breakthrough in the performance of sequential decision-making compared to the vanilla LLM player.<sup>1</sup>
+
+# 1 Introduction
+
+In artificial intelligence, sequential decision-making refers to algorithms that take the dynamics of the world into consideration (Frankish and Ramsey, 2014), and it can be described as a procedural approach to decision-making, or as a step by step decision theory. As a consequence, sequential decision-making has the intertemporal choice problem, where earlier decisions influences the later available choices (Amir, 2014).
+
+In recent years, Large language models (LLMs) are gaining increasing popularity in both academia and industry, owing to their unprecedented performances in various applications (Chang et al., 2023), ranging from chatbots to medical diagnoses (Wang et al., 2023a) to robotics (He et al., 2022). From robots handling complex tasks (Amiri et al., 2020) to entrepreneurial action (McMullen, 2015), sequential decision-making permeates diverse domains. Hence, an interesting question arises: Can Current LLMs Make Sequential Decisions Effectively?
+
+To answer this question, we need to design a benchmark to evaluate the sequential decision-making ability of LLMs. However, evaluating LLMs' abilities is not trivial. Many studies have been proposed to test LLMs' performances on either a large-scale static benchmark such as MMLU (Hendrycks et al., 2021), or with A/B tests judged by humans (Ganguli et al., 2023). One common and evident limitation of these methods, however, is that the environment for LLMs to be tested is static (Aiyappa et al., 2023; Zhou et al., 2023), which can not reflect the domino effect in sequential decision-making. Besides, data contamination (Sainz et al., 2023; Zeng et al., 2024; Xu et al., 2024), which means the inclusion of test data examples and labels in the pre-training data, also challenges the efficacy of these static benchmarks in differentiating model capabilities.
+
+Unlike static evaluation, dynamic evaluation by treating LLMs as game-playing agents attracted more and more attention of researchers recently, such as beauty contests and private-value second price auctions (Guo et al., 2024a), Warewolf (Xu et al., 2023), Avalon (Wang et al., 2023b; Light et al., 2023) and Leduc Hold'em (Guo et al., 2023). However, current attempts do not account for sequential decision-making, and these games are either challenging to evaluate for intermediate results (such as Werewolf) or have too few decision points
+
+![](images/27a0f21c337fb43266f9de85ffb21689704de0a8f4898a20e6c47707e0f6511f.jpg)  
+Figure 1: In this figure, (A) demonstrates the sequential decision-making process in the UNO Arena, (B) presents the execution process of the vanilla LLM player, and (C) shows the execution process of the TUTRI player. Note that, the Module (D) and the Module (E) are completely identical.
+
+per round (such as Leduc Hold'em). Meanwhile, we should also note that studies of dynamically evaluating sequential decision-making capability in the reinforcement learning, such as games like Go (Silver et al., 2017), Dou Di Zhu (You et al., 2019), and Mahjong (Li et al., 2020). However, these games present an excessively large action space. For instance, in Dou Di Zhu, players can use any combination of their cards in each round, posing significant challenges for current LLMs (Zhai et al., 2024).
+
+Considering the above aspects, we make the following efforts in this paper:
+
+First, we build the UNO Arena to dynamically evaluate the sequential decision-making capability of current LLMs. In the UNO Arena, we allow LLMs to participate as players in the UNO game $^2$ , aiming to play all the cards in their hand as quickly as possible. Compared to games like Leduc Hold'em, which have fewer moves per game, UNO features an average of dozens of moves per game, making it an ideal testbed for sequential decision-making (Pfann, 2021). Additionally, unlike common games in the reinforcement learning, legal actions in the UNO Arena are limited, only including drawing cards, playing cards, selecting colors to convert, and choosing whether to challenge the wild draw four card. Furthermore, to monitor the behaviours of LLMs in the UNO Arena, we propose some real-time quantitative evaluation metrics by leveraging the Monte Carlo method (Kroese et al., 2014) as the reference, which provide a window to observe the intermediate results and various phenomena (like domino effect) in LLMs' se
+
+quental decision-making.
+
+Second, based on the proposed UNO Arena, we set up a family of strong and representative players. In detail, we first build the random player, which makes decisions based on chance rather than a specific, consistent plan, without considering the game's current state or potential outcomes. Then, we built heuristic players based on a conservative heuristic UNO play strategy and implement the reinforcement learning based player, which leverages DQN (Mnih et al., 2013) to develop sophisticated strategies for playing UNO. Finally, to probing the capability of LLMs in sequential decision-making, we provides the task description and then prompt LLMs, like GPT-4 (Achiam et al., 2023) and Gemini-pro (Team et al., 2023), to generate their reasoning steps that lead to the final action.
+
+Third, to unleash the fully potential capability of LLMs in sequential decision-making, we propose the TUTRI player with reflection mechanism (Shinn et al., 2024), which can enable LLMs to analyze their own actions based on the game history and the game strategy. In detail, the proposed agent framework consists of two key reflection modules: the game history reflection module and the game strategy reflection module. In the game history reflection module, we provide the statistical data of game history and then prompt the LLMs to rethink their decision, which simulates the process of card memorization by humans when playing UNO. In the game strategy reflection module, LLMs further take into account the game strategy, like saving wild draw four, and proceed to make the final decision, which simulates the use and adherence to strategies by humans when playing UNO.
+
+In the experiment, we comprehensively evaluate some mainstream LLMs' ability of sequential decision-making, including GPT-3.5 (OpenAI, 2022), GPT-4 (Achiam et al., 2023), Gemini-pro (Team et al., 2023), Llama 2 (Touvron et al., 2023), and ChatGLM3 (Du et al., 2021; Zeng et al., 2022). Our experimental results show that among these LLMs, GPT-4 is the most effective sequential decision-maker.
+
+In summary, our contributions are as follows:
+
+- We propose a dynamic evaluation method, named UNO Arena, for assessing the sequential decision-making capability of large language models (LLMs) based on the card game UNO. This method supports the evaluation of 2-10 LLM players, reinforcement learning players, heuristic players, or random players engaged in a single UNO game.  
+- We introduce multiple unique evaluation metrics based on the Monte Carlo method for evaluating the sequential decision-making capabilities of players in the UNO Arena.  
+- To improve the sequential decision-making capabilities of LLMs and enhance their performance in the highly dynamic and complex UNO game, we have developed the TUTRI player and compared it horizontally with the vanilla LLM player.
+
+# 2 Related Work
+
+Evaluate LLMs Dynamically with Game: LLMs have presented increasingly emerging ability on game-playing (Brookins and DeBacker, 2023; Akata et al., 2023) in recent development and iterations. Wang et al. (2023b) use the Avalon, which contains elements of deception, to evaluate the capability of LLMs to recognize and handle deceptive information. Gong et al. (2023) leverage the Cuisine World and Minecraft to assess the planning and emergency cooperation capabilities of LLMs. Guo et al. (2024a) employ beauty contests and auction games to evaluate the rationality, strategic reasoning capability, and adherence to instructions of LLMs. Xu et al. (2023) use the Werewolf game to evaluate the capability of LLMs to infer player roles. Despite evaluating with game becoming a popular trend, exploring into sequential decision-making capability is still of scarcity in current works.
+
+Development of Agent Framework: LLM agents have been perceived as a promising way to realiz
+
+ing Artificial General Intelligence(AGI) (Xi et al., 2023) and recently have shown emergent abilities to execute various tasks in the complex environment (Wei et al., 2022). SiLLM (Guo et al., 2024b) merges large language models with synchronous machine translation, using policy decision agents and translation agents. LLM-Vectorizer (Taneja et al., 2024) uses multiple agents to generate vectorized code by leveraging large language models and test-based feedback. We tailor a special framework for UNO in this paper, featuring self-refinement and iterative thinking.
+
+Sequential Decision-Making Capability: Sequential decision-making refers to the process of making a series of decisions over time, where each decision may impact future choices and outcomes (Amir, 2014). Though certain algorithms or reinforcement learning provide solutions for some sequential decision-making problems (Littman, 1996), LLM-based sequential decision-making are only employed in limited field like recommendation (Wang et al., 2023c). In our work, we utilize the UNO game, which is not an easy one even for human (Demaine et al., 2014), to explore the sequential decision-making ability of LLMs. With certain methods like integrating past experiences and expert advice or demonstrations (Chen et al., 2023), we make efforts to maximally leverage the decision making ability as possible in a sequential manner.
+
+# 3 The UNO Arena
+
+In this section, we first provide a brief overview of the version of UNO we adopt in the subsection §3.1. Then, we present the four different types of players in the UNO Arena in the subsection §3.2. Next, we detail how to use Monte Carlo methods to determine whether a player has made an optimal decision in subsection §3.3. In the end, we introduce our evaluation metrics in subsection §3.4.
+
+# 3.1 The UNO Game
+
+We select the UNO as the foundation within our arena due to its widespread popularity, simplicity and mathematical value. There are various versions of the UNO game. In this section, we briefly introduce the rules of the version we adopt in this work.
+
+UNO Cards: A deck of UNO cards comprises a total of 108 cards. UNO cards are divided into three types: number cards, function cards, and wild
+
+cards. A number card is composed of a color (Red, Blue, Yellow and Green) and a number (ranging from 0 to 9). A function card is composed of a color (Red, Blue, Yellow and Green) and a function (Skip, Reverse, Draw Two). The wild cards have no color and is only composed of Wild cards and Wild Draw Four cards. The effects of the function cards and wild cards are shown in the Table 1.
+
+UNO Process: First, deal each player 7 initial cards in clockwise order, then continue drawing cards until a number card is drawn and set as the top card of the initial discard pile. All players take rounds playing cards in clockwise order (it will be reversed by a reverse card) until a player runs out of his cards or the draw pile is exhausted, signaling the end of the game.
+
+UNO Action: From the beginning to the end of the game, players continuously take actions in UNO. In our work, UNO includes the following types of actions:
+
+- Select Card: When a player comes his playing round, he need to play a card that matches the color, number, or function of the top card in the discard pile, or play a Wild card. If he does not have a card to play, he must draw one card.  
+- Select Color: After a player plays a Wild card or a Wild Draw Four card, the player need to change the color of the current top card to one of Red, Yellow, Blue or Green.  
+- Select ChallengeFlag: After a player's previous opponent plays a Wild Draw Four card, the player need to decide whether to challenge the legality of the previous opponent's Wild Draw Four card.
+
+For more details about the UNO games, please refer to Appendix A. The Figure 1 (A) shows the workflow diagrams of UNO Arena.
+
+# 3.2 Players in the UNO Arena
+
+In the UNO Arena, we initially involve three types of players: random player, reinforcement learning based player, vanilla LLM player. To further unleash the potential capability of LLMs in sequential decision-making, we propose TUTRI player, which involves reflection mechanism.
+
+Random Player: As like its name suggests, the random player performs all actions randomly, such as randomly selecting a regulative card to play when it is his turn. The random player can be considered the baseline of the UNO Arena, mainly
+
+<table><tr><td>Card</td><td>Sample</td><td>Effect</td></tr><tr><td>Skip</td><td></td><td>The next player in sequence misses a round.</td></tr><tr><td>Reverse</td><td></td><td>Order of play switches directions (clockwise to counterclockwise, or vice versa).</td></tr><tr><td>Draw Two</td><td>+2</td><td>The next player in sequence draws two cards and misses a round.</td></tr><tr><td>Wild</td><td></td><td>Player declares the next color to be matched (it can be used on any round even if the player has any card of matching color).</td></tr><tr><td>Wild Draw Four</td><td>+4</td><td>Player declares the next color to be matched. The next player in sequence draws four cards and misses a round. May be legally played if the player has cards of the current color.</td></tr></table>
+
+Table 1: The effects of function and wild cards.
+
+serving to maintain the flow of the UNO game. If some players outperform the random player, we can infer that these players are consciously playing UNO with an understanding of the game rules.
+
+Heuristic Player: The heuristic player here is a player who follows the following strategy: (1) The number cards are matched first, and since the numbers 0 and 9 are half the number of other number cards, 0 number cards and 9 number cards have higher priority than other number cards. (2) Keep the function cards as much as possible, first match the type of function cards, and the priority is: Reverse cards, Skip cards and Draw Two cards. (3) Keep the wild cards as much as possible, and choose the most color in your hand when using the wild card to change the color of the top card in the current discard pile.
+
+Reinforcement Learning Based Player: Previous research has sought breakthroughs in the UNO game by using reinforcement learning models (Pfann, 2021). We built our reinforcement learning player with DQN (Mnih et al., 2013) model based on the open-source project RLcard (Zha et al., 2019).
+
+Vanilla LLM Player: During the vanilla LLM player's turn, the game host transmits all publicly available information through a prompt to the LLM. The LLM then returns a JSON containing the decision and reasoning as required by the prompt. The Figure 1 (B) shows the workflow diagrams of vanilla LLM player.
+
+TUTRI Player: While LLMs do not always generate the best output on their first try just as hu
+
+man (Madaan et al., 2023), iterative feedback and refinement could be a necessity for a better agent framework. Moreover, human-like thinking patterns, such as introspective reflections foster divergent thinking processes (Zhang et al., 2023), inspire us to propose the TUTRI player. This advanced framework is designed to navigate the intricacies of UNO game play, offering a more structured approach to strategic sequential decision-making. The original decision for TUTRI player is exactly the same as the vanilla LLM player's decision, after that are two additional reflection modules.
+
+- The Game History Reflection Module: In this module, we provide statistical information about game history to the TUTRI player, and the player is told to reflect the action you just selected with these auxiliary information. Just like human thinking when playing UNO, if there is a large number of green cards that have already been played in the game's history, it is very advantageous for the player to play a green card. After reflection, the player need to output both reflection thoughts and the updated action.  
+- The Game Strategy Reflection Module: In this module, we provide additional useful game strategies to the TUTRI player, and the player is again told to reflect the action you just selected based on game strategies. For example, since wild cards can be played at any situations and disrupt other players, saving the wild cards in your hand as long as possible is a very useful game strategy. After reflection, the player should output both reflection thoughts and updated action (the final action).
+
+It must be emphasized that the TUTRI player should work in a conversational manner, with exactly 3 times Q&A per round. Moreover, the TUTRI player may keep their original decision, in other words, literally updating the action is not a necessity, nevertheless, the reflection process, instead of simple I-O prompting of interaction, providing more opportunities for mistake correcting and divergent thinking. The Figure 1(C) shows the workflow diagrams of TUTRI player.
+
+# 3.3 Monte Carlo Simulation Method for Monitoring Players' Behavior
+
+In the game play, the change in each player's winning rate after making a decision is the key for tracking. In the classical combinatorial games, like
+
+Nim (Bouton, 1901) or Wythoff's Game (Wythoff, 1907), positions space are limited and thus computationally affordable, while UNO is more intricate, where positional space exponentially increases as cards number increases and the calculation gets tougher (Demaine et al., 2014).
+
+To make a plausible ranking mechanism of the candidate decisions, we define the concept of optimal decision, meaning the state transferred by the decision from last state, has a highest winning rate concerning all subsequent outcomes, and thus adopt Monte Carlo Simulation (Mooney, 1997) to calculate the estimated winning rate.
+
+Detailedly, with  $S_{i}$  representing the state of the game after the  $i$ -th step taken,  $D_{i,j}$  representing the  $j$ -th legal decision candidates at state  $S_{i}$ ,  $\mathcal{T}$  representing the state transfer function,  $\mathcal{E}$  representing the estimate function of state, thereby we have the definition of the optimal decision  $D_{i,opt}$  at the  $i$ -th step where
+
+$$
+o p t = \underset {j} {\arg \max } \mathcal {E} \left(\mathcal {T} \left(S _ {i - 1}, D _ {i, j}\right)\right) \tag {1}
+$$
+
+In calculation of  $\mathcal{E}(S_i)$ , we massively randomly generate the subsequent decision sequence  $\{D_{i+1}, D_{i+2}, \ldots\}$  and thus obtain the subsequent state sequence  $\{S_{i+1}, S_{i+2}, \ldots\}$ . Then  $\mathcal{E}(S_i)$  is assigned to the ratio of number of sequences where the player plays the state  $S_{i-1}$  comes as the winner, to the total number of sequences simulated.
+
+As the times we simulate the subsequent sequence increases, the approximate value  $\mathcal{E}(S_i)$  gets more precise, though we could not enumerate all the possible situations. To balance the time expenditure and the precision of the metrics, we control the simulation times in a certain range. Additionally, a threshold parameter  $p$  is set to identify critical decisions. We say a decision  $D \coloneqq D_i$  is critical if among its all candidate choices  $D_j$
+
+$$
+\max  \mathcal {E} (\mathcal {T} (S, D _ {j})) - \min  \mathcal {E} (\mathcal {T} (S, D _ {j})) \geq p \tag {2}
+$$
+
+Actual decisions made on critical positions may have a huge effect on the winning rate, which is consistent with the game nature.
+
+# 3.4 Evaluation Metrics in the UNO Arena
+
+In our work, we design three evaluation metrics, including WR, ODHR@K and ADR@K, in conjunction with the UNO game to comprehensively evaluate the sequential decision-making capability of LLMs. Among these metrics, ODHR@K and
+
+ADR@K can off a glimpse into the intermediate results in the sequential decision-making of LLMs.
+
+Winning Rate (WR). WR denotes the proportion of player wins to total game innings, and can be represented as:
+
+$$
+\mathrm {W R} = \frac {N _ {\text {W i n}}}{N _ {\text {G a m e}}} \tag {3}
+$$
+
+where  $N_{\mathrm{Win}}$  represents the total number of times the given player has won, and  $N_{\mathrm{Game}}$  denotes the total game innings.
+
+Optimal Decision Hit Rate at K Decision Points (ODHR@K): This metric measure the proportion of times players make the best decision to all decision times, when facing K decision points:
+
+$$
+\mathrm {O D H R} @ \mathrm {K} = \frac {N _ {\text {H i t} @ \mathrm {K}}}{N _ {\text {D e c i s i o n} @ \mathrm {K}}} \tag {4}
+$$
+
+where  $N_{\mathrm{Hit@K}}$  is the number of times the player makes the optimal decision when facing K optional decision points, and  $N_{\mathrm{Decision@K}}$  represents the total number of times the agent player makes decision when it faces K optional decision points.
+
+Average Decision Rank at K Decision Points (ADR@K). This metric looks at the rank of output decision made by the player, and can be denoted as:
+
+$$
+\mathrm {A D R} @ \mathrm {K} = \frac {\sum_ {i = 1} ^ {N _ {\text {D e c i s i o n}} @ \mathrm {K}} \operatorname {R a n k} \left(D _ {i}\right)}{N _ {\text {D e c i s i o n}} @ \mathrm {K}} \tag {5}
+$$
+
+where  $\operatorname{Rank}(D_i)$  represents the rank from best to worst among all legal decisions in its decision-making process, and  $N_{\text{Decision}@\mathbf{K}}$  represents the total number of times the agent player makes decision when it faces  $\mathbf{K}$  optional decision points.
+
+For metrics ODHR@K and ADR@K, according to the characteristics of UNO, we only focus on the situations where  $K$  is equal to 2, 3 or 4, because the vast majority of decisions in UNO do not exceed 4 (Pfann, 2021).
+
+# 4 Experiments
+
+In this section, we first conduct preliminary experiments with vanilla LLM players, RL players, and random players in subsection §4.1. Then, we have multiple different LLM-based vanilla players compete in UNO Arena to identify the best LLM in subsection §4.2. Next, we test the superiority of the TUTRI players compared to the vanilla LLM players in subsection §4.3. Finally, we perform
+
+<table><tr><td rowspan="2">Metrics</td><td colspan="6">Vanilla LLM Players &amp; RL Player with DNQ</td></tr><tr><td>GPT-3.5</td><td>GPT-4</td><td>Gemini-Pro</td><td>Llama 2</td><td>ChatGLM3</td><td>DNQ</td></tr><tr><td>WR (↑)</td><td>55.80</td><td>63.20</td><td>53.80</td><td>53.60</td><td>48.80</td><td>57.40</td></tr><tr><td>ODHR@2 (↑)</td><td>57.34</td><td>61.47</td><td>53.94</td><td>53.69</td><td>49.75</td><td>54.96</td></tr><tr><td>ADR@2 (↓)</td><td>1.427</td><td>1.385</td><td>1.461</td><td>1.463</td><td>1.503</td><td>1.450</td></tr><tr><td>ODHR@3 (↑)</td><td>32.15</td><td>39.30</td><td>34.42</td><td>33.84</td><td>34.45</td><td>35.98</td></tr><tr><td>ADR@3 (↓)</td><td>2.010</td><td>1.904</td><td>2.017</td><td>1.994</td><td>2.034</td><td>1.947</td></tr><tr><td>ODHR@4 (↑)</td><td>27.20</td><td>36.99</td><td>31.05</td><td>27.39</td><td>25.36</td><td>37.74</td></tr><tr><td>ADR@4 (↓)</td><td>2.399</td><td>2.142</td><td>2.331</td><td>2.436</td><td>2.460</td><td>2.247</td></tr></table>
+
+ablation experiments on the TUTRI player in subsection  $\S 4.4$
+
+To ensure the generalization of the experiments, we take the mainstream LLMs mentioned in the introduction: (1) gpt-3.5-turbo-16k-0613 (OpenAI, 2022); (2) gpt-4-1106-preview (Achiam et al., 2023); (3) Gemini-pro (Team et al., 2023); (4) Llama-2-7b-chat (Touvron et al., 2023); (5) ChatGLM3-6b (Du et al., 2021; Zeng et al., 2022).
+
+# 4.1 1v1 UNO Arena between vanilla LLM players, RL players and random players
+
+In order to verify the rationality of using UNO Arena to evaluate the sequential decision-making ability of LLMs, we first conduct experiments on vanilla LLM players, RL players and random players in 1V1 UNO Arena. We randomly generate 500 sets of UNO initial decks. Each vanilla LLM player or RL player have to play with the random player in these 500 initial decks. In addition, the random players are the first to play cards in all games. The results are shown in the Table 2.
+
+Table 2: Statistical results of random player VS vanilla LLM player or RL player with DNQ. The decision threshold  $p$  for critical decision in ODHR@K and ADR@K is 0.15. Bold indicates the best result, underline the second best result, and the Table 3 below follows this pattern.  
+
+<table><tr><td rowspan="2">Metrics</td><td colspan="5">Vanilla LLM Players</td></tr><tr><td>GPT-3.5</td><td>GPT-4</td><td>Gemini-Pro</td><td>Llama 2</td><td>ChatGLM3</td></tr><tr><td>WR (↑)</td><td>22.80</td><td>24.20</td><td>20.40</td><td>20.00</td><td>15.60</td></tr><tr><td>ODHR@2 (↑)</td><td>52.57</td><td>54.77</td><td>49.88</td><td>54.08</td><td>50.52</td></tr><tr><td>ADR@2 (↓)</td><td>1.474</td><td>1.452</td><td>1.501</td><td>1.459</td><td>1.495</td></tr><tr><td>ODHR@3 (↑)</td><td>39.56</td><td>41.41</td><td>33.14</td><td>34.78</td><td>33.13</td></tr><tr><td>ADR@3 (↓)</td><td>1.889</td><td>1.885</td><td>2.034</td><td>1.978</td><td>2.043</td></tr><tr><td>ODHR@4 (↑)</td><td>26.75</td><td>29.03</td><td>25.74</td><td>24.90</td><td>25.04</td></tr><tr><td>ADR@4 (↓)</td><td>2.407</td><td>2.366</td><td>2.516</td><td>2.471</td><td>2.477</td></tr></table>
+
+Table 3: Statistical results of competition among 5 vanilla LLM players in UNO Arena. The decision threshold  $p$  for critical decision in ODHR@K and ADR@K is 0.00.
+
+From the Table 2, we can find that (1) Except for
+
+<table><tr><td>LLM</td><td>WR (↑)</td><td>ODHR@2 (↑)</td><td>ADR@2 (↓)</td><td>ODHR@3 (↑)</td><td>ADR@3 (↓)</td><td>ODHR@4 (↑)</td><td>ADR@4 (↓)</td></tr><tr><td>GPT-3.5 (vanilla)</td><td>48.00</td><td>53.05</td><td>1.4695</td><td>34.97</td><td>1.9508</td><td>34.47</td><td>2.2340</td></tr><tr><td>GPT-3.5 (TUTRI)</td><td>52.50 (+4.50%)</td><td>54.01 (+0.06%)</td><td>1.4599 (-0.06%)</td><td>43.13 (+8.16%)</td><td>1.8563 (-4.73%)</td><td>32.92 (-1.55%)</td><td>2.2667 (+1.09%)</td></tr><tr><td>GPT-4 (vanilla)</td><td>49.00</td><td>56.27</td><td>1.4373</td><td>39.38</td><td>1.9375</td><td>36.24</td><td>2.2140</td></tr><tr><td>GPT-4 (TUTRI)</td><td>51.00 (+2.00%)</td><td>56.60 (+0.33%)</td><td>1.4340 (-0.33%)</td><td>40.14 (+0.76%)</td><td>1.8592 (-3.92%)</td><td>36.33 (+0.09%)</td><td>2.1510 (-2.10%)</td></tr><tr><td>Gemini-pro (vanilla)</td><td>44.00</td><td>50.62</td><td>1.4938</td><td>37.04</td><td>2.0159</td><td>25.44</td><td>2.4737</td></tr><tr><td>Gemini-pro (TUTRI)</td><td>56.50 (+12.50%)</td><td>53.64 (+3.02%)</td><td>1.4636 (-3.02%)</td><td>34.13 (-2.91%)</td><td>1.9461 (-3.49%)</td><td>30.36 (+4.92%)</td><td>2.3482 (-4.18%)</td></tr><tr><td>Llama 2 (vanilla)</td><td>47.00</td><td>49.54</td><td>1.5046</td><td>33.11</td><td>1.9595</td><td>29.11</td><td>2.3944</td></tr><tr><td>Llama 2 (TUTRI)</td><td>54.00 (+7.00%)</td><td>55.07 (+5.53%)</td><td>1.4493 (-5.53%)</td><td>37.31 (+4.20%)</td><td>1.8507 (-5.44%)</td><td>26.75 (-2.36%)</td><td>2.4650 (+2.35%)</td></tr><tr><td>ChatGLM3 (vanilla)</td><td>47.00</td><td>55.82</td><td>1.4418</td><td>29.05</td><td>2.0541</td><td>31.84</td><td>2.2935</td></tr><tr><td>ChatGLM3 (TUTRI)</td><td>54.00 (+7.00%)</td><td>57.24 (+1.42%)</td><td>1.4276 (-1.42%)</td><td>39.51 (+10.46%)</td><td>1.8642 (-9.50%)</td><td>30.62 (-1.22%)</td><td>2.4689 (+5.85%)</td></tr></table>
+
+Table 4: Statistical results of vanilla LLM players VS TUTRI players. The decision threshold  $p$  for critical decision in ODHR@K and ADR@K is 0.00. Red annotations indicate favorable experimental results, while blue annotations indicate unfavorable experimental results.  
+
+<table><tr><td>LLM</td><td>WR (↑)</td><td>ODHR@2 (↑)</td><td>ADR@2 (↓)</td><td>ODHR@3 (↑)</td><td>ADR@3 (↓)</td><td>ODHR@4 (↑)</td><td>ADR@4(↓)</td></tr><tr><td>Gemini-pro (TUTRI)</td><td>56.50</td><td>53.64</td><td>1.4636</td><td>34.13</td><td>1.9461</td><td>30.36</td><td>2.3482</td></tr><tr><td>Gemini-pro + TUTRI&#x27;</td><td>52.50 (-4.00%)</td><td>54.33 (+0.69%)</td><td>1.4567 (-0.69%)</td><td>29.88 (-4.25%)</td><td>2.0610 (+5.75%)</td><td>31.25 (+0.89%)</td><td>2.4219 (+2.46%)</td></tr><tr><td>Gemini-pro + TUTRI&#x27;&#x27;</td><td>53.50 (-3.00%)</td><td>54.59 (+0.95%)</td><td>1.4541 (-0.95%)</td><td>31.95 (-2.18%)</td><td>2.0384 (+4.62%)</td><td>27.59 (-2.77%)</td><td>2.3824 (-1.14%)</td></tr></table>
+
+Table 5: Statistical results of the ablation study. Where TUTRI' represents the TUTRI player which remove the game history reflection module, and TUTRI'' represents the TUTRI player which remove the game strategy reflection module. The decision threshold  $p$  for critical decision in ODHR@K and ADR@K is 0.15. Red annotations indicate favorable experimental results, while blue annotations indicate unfavorable experimental results.
+
+ChatGLM3, the WR of other vanilla LLM players and RL players are all above  $50.00\%$ ; (2) The performance of GPT-4 is the best, and GPT-4 performs excellently on the 7 evaluation metrics. Especially, the WR of GPT-4 is  $63.20\%$ ,  $13.20\%$  higher than  $50.00\%$ .
+
+# 4.2 5-players UNO Arena with 5 LLMs
+
+To find the best LLM, we place 5 LLMs in a 5- players UNO Arena to compete against each other. We fix the initial playing order of UNO Arena in the sequence of GPT-3.5, GPT-4, Gemini-Pro, Llama 2, and ChatGLM3. We conduct experiment on 200 decks generated randomly. All players are the vanilla LLM players. The results are shown in the Table 3.
+
+From the Table 3, we can find that (1) GPT-4 has the best performance, with a WR of  $24.20\%$ ,  $4.2\%$  higher than the average  $(20.00\%)$  and  $1.4\%$  higher than the second highest ranked GPT-3.5. Not only that, GPT-4 also performs the best in other 6 evaluation metrics; (2) ChatGLM3 has the worst performance, with a WR of  $15.60\%$ , which is  $4.4\%$  lower than the average  $(20.00\%)$  and  $8.6\%$  lower than the highest ranked GPT-4. Not only that, ChatGLM3 also performs the worst in ODHR@2, ADR@2, ADR@3, ODHR@4, and ADR@4.
+
+![](images/fdddf7b2c83ded2db980b327b29648619a9e827ad070dd414b8e8cc3e642ea53.jpg)  
+Figure 2: The Pearson Correlation Heatmap among WR, ODHR@K (K=2,3,4), and ADR@K (K=2,3,4).
+
+# 4.3 Validation of the superiority of the TUTRI player compared to the vanilla LLM player
+
+To verify that our TUTRI player can improve the sequential decision-making ability of LLMs, we compare the vanilla LLM players (baseline) with TUTRI players. We let 5 LLMs serve as the backend LLMs for both the vanilla LLM players and TUTRI players, and play two-players UNO Arena on 200 decks generated randomly. The results are shown in the Table 4.
+
+From the Table 4, we can find that: (1) All LLMs (the TUTRI player) are better than LLM (the vanilla LLM player) on WR, ODHR@2, and ADR@3.
+
+![](images/fed3a04d45febe539a5c1d570d287fe00c41f46e539d9162a21a9edabe865f68.jpg)
+
+![](images/f1f161d85fbf4241170883963aecc64709f337e9d8d08e0c48d9459f82fcf82f.jpg)
+
+![](images/f83268fc7bac53192b26b3e6dcbf3cb85ecb06ac690a13768e3f59f4303f417c.jpg)  
+Figure 3: GPT-4 (the vanilla LLM player) real-time winning rate variations on 4 decks.
+
+![](images/86bcc60a1bbd1202a4425a880fbcd7efab245422f8163f923198148d544227be.jpg)
+
+Gemini-Pro (the TUTRI player) has a  $12.50\%$  higher than Gemini-Pro (the vanilla LLM player) on WR; (2) For ODHR@3, except for Gemini-Pro which performed slightly worse  $(-2.91\%)$ , the other 4 LLMs achieved good results. For ODHR@4 and ADR@4, GPT-4 and Gemini-Pro both performed well. It can be seen that the TUTRI player based on reflection can significantly improve its abilities of sequential decision-making after two rounds of reflection on the summary of game history and the game strategies. The experimental results strongly support the superiority of our TUTRI player based on the reflection mechanism over the vanilla LLM player.
+
+# 4.4 Ablation studies on TUTRI player
+
+To illustrate the necessity of the two reflection modules in the TUTRI player, we further conduct some ablation studies. We remove the game history reflection module and the game strategy reflection module from the TUTRI players, and conduct two-players UNO Arena with vanilla LLM player respectively. The results are shown in the Table 5.
+
+From the Table 5, we can find that: (1) After removing the game history reflection module, the WR of decreased by  $4\%$ , the ODHR@3 of decreased  $4.25\%$ , and the ADR@3 of increase by  $5.75\%$ . (2) After removing the game strategy reflection module, the WR of decreased by  $3\%$ , the ODHR@3 of decreased  $2.18\%$ , and the ADR@3 of increase by  $4.62\%$ . The game history holds significant potential information for incomplete information games. Therefore, removing the game history reflection module has a greater adverse impact on the TUTRI player.
+
+# 5 Discussion
+
+# 5.1 Further Exploration of ODHR@K and ADR@K
+
+To better analyze the relationship between our unique evaluation metrics (ODHR@K and ADR@K), and the evaluation metric WR, we conduct a Pearson correlation analysis of the experimental results from the Table 2. The results are shown in the Figure 2. From the Figure 2, we can find that (1) WR shows a positive correlation with ODHR@K  $(\mathrm{K} = 2,3,4)$ , and simultaneously, WR shows a negative correlation with ADR@K  $(\mathrm{K} = 2,3,4)$ ; (2) The strongest positive correlation, reaching 0.85, exists between WR and ODHR@3, while the strongest negative correlation, reaching -0.86, exists between WR and ADR@3. Overall, our unique ODHR@K and ADR@K have a good correlation with WR, so they can serve as reference evaluation metrics for evaluating LLMs in the UNO Arena.
+
+# 5.2 Case Study
+
+In order to more intuitively see the advantages of LLM versus random player, we conduct a case study. We utilized GPT-4 as the backend LLM for the vanilla LLM player to engage in the game across 4 decks generated randomly, with the random player plays first. We recorded all decision points (for both the vanilla LLM player and the random player) and employed the Monte Carlo method to calculate the real-time percentage change in winning rate for both sides following each decision point. The results are shown in the Figure 3.
+
+From the Figure 3, we can find that: (1) In the
+
+<table><tr><td>Player</td><td>WR (↑)</td><td>ODHR@2 (↑)</td><td>ADR@2 (↓)</td><td>ODHR@3 (↑)</td><td>ADR@3 (↓)</td><td>ODHR@4 (↑)</td><td>ADR@4(↓)</td></tr><tr><td>Heuristic Players</td><td>63.60</td><td>62.68</td><td>1.373</td><td>37.55</td><td>1.931</td><td>37.03</td><td>2.227</td></tr><tr><td>Random Players</td><td>36.40</td><td>37.12</td><td>1.630</td><td>28.92</td><td>2.153</td><td>16.64</td><td>2.615</td></tr><tr><td>Heuristic Players</td><td>60.80</td><td>59.74</td><td>1.402</td><td>36.37</td><td>1.932</td><td>34.81</td><td>2.252</td></tr><tr><td>Vanilla LLM Players</td><td>39.60</td><td>38.31</td><td>1.618</td><td>29.12</td><td>2.148</td><td>17.23</td><td>2.651</td></tr></table>
+
+Table 6: The experimental results of heuristic players compared to random players and vanilla LLM players.
+
+UNO Arena, winning rates fluctuate significantly. For example, in deck 1, from round 28 to 29, the random player's winning rate dropped by  $34.5\%$ , while the vanilla LLM player by  $34.67\%$ ; (2) Turning points, like rounds 44 and 67 in deck 3, show shifts in dominance. Initially, the vanilla LLM player leads until round 44, then loses advantage until round 67, before regaining control; (3) Brief game durations occur, notably in deck 4, where the agent player consistently makes exceptional decisions, steadily increasing its winning rate until achieving victory. These findings underscore LLM's adeptness at identifying crucial decision junctures and exploiting its capabilities, highlighting its potential in sequential decision-making scenarios.
+
+# 5.3 Exploring the Capabilities of Heuristic Players
+
+We conduct two experiments: one where heuristic players face random players in 500 different UNO decks, and another where heuristic players compete against vanilla LLM players using Llama-2 in the same 500 decks. The experimental results are shown in the Table 6.
+
+From the Table 6, we can find that: (1) Heuristic players significantly outperform random players across all evaluation metrics. (2) Heuristic players also demonstrate strong capabilities compared to vanilla LLM players (Llama 2), although the performance of vanilla LLM players is still superior to that of random players.
+
+# 5.4 Concerns about Data Contamination
+
+As previous studies suggest (Topsakal et al., 2024; Karvonen, 2024), we usually assume LLMs are not trained explicitly for games, like UNO and Tic-Tac-Toe. The LLMs know the description/rules of the games but the experimental results show that even the strongest LLM, like GPT-4, struggles. Besides, As we mention in the introduction section, there are static methods and dynamic methods (such as UNO Arena and Warewolf) in evaluating LLMs'
+
+ability. Compared to static methods, UNO Arena is less prone to data contamination due to vast game sampling space and vast game state (for a given game) space. Specifically, we sample 500 UNO games and find that among these 500 games, the maximum game state is  $1.610^{34}$ , and the average game state is  $5.310^{31}$ . Considering the enormous game state space and almost infinite game space, LLMs are extremely difficult to learn or overfit genuinely useful patterns or strategies. Therefore, we don't need to worry about data contamination potentially skewing the results.
+
+# 6 Conclusion
+
+In conclusion, LLMs possess the capability for sequential decision-making, as evidenced by the experimental results of LLMs playing the UNO game. Our proposed UNO Arena and unique evaluation metrics enable LLMs to compete with each other in the same UNO Arena game, thereby providing a better dynamic assessment of LLMs' sequential decision-making abilities. Furthermore, we propose that the TUTRI player effectively addresses how to enhance LLMs' sequential decision-making abilities for better performance in playing UNO Arena.
+
+# Acknowledgements
+
+This work is supported by the National Key R&D Program of China (Grant No. 2023YFB3307500). This work is also supported by the National Natural Science Foundation of China (Grant No. 62306087, No. 62472121 and No.62306313), the Natural Science Foundation of Shandong Province (Grant No. ZR2023QF154), Special Funding Program of Shandong Taishan Scholars Project, the Open Project of Anhui Provincial Key Laboratory of Multimodal Cognitive Computation, Anhui University (Grant No. MMC202420) and the InnoHK program. Besides, we sincerely thank the anonymous reviewers for their valuable feedback.
+
+# Limitations
+
+The method of dynamically evaluating the sequential decision-making ability of LLMs using the UNO Arena, as well as the TUTRI player, is only applicable to LLMs that support chat. The unique evaluation metrics, ODHR@K and ADR@K, introduced in this paper are only applicable to games or tasks with a limited action space.
+
+# References
+
+Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774.  
+Rachith Aiyappa, Jisun An, Haewoon Kwak, and Yong-Yeol Ahn. 2023. Can we trust the evaluation on chatgpt?  
+Elif Akata, Lion Schulz, Julian Coda-Forno, Seong Joon Oh, Matthias Bethge, and Eric Schulz. 2023. Playing repeated games with large language models. arXiv preprint arXiv:2305.16867.  
+Eyal Amir. 2014. Reasoning and decision making. The Cambridge handbook of artificial intelligence, pages 191-212.  
+Saeid Amiri, Mohammad Shokrolah Shirazi, and Shiqi Zhang. 2020. Learning and reasoning for robot sequential decision making under uncertainty. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 2726-2733.  
+Charles L Bouton. 1901. Nim, a game with a complete mathematical theory. The Annals of Mathematics, 3(1/4):35-39.  
+Philip Brookins and Jason Matthew DeBacker. 2023. Playing games with gpt: What can we learn about a large language model from canonical strategic games? Available at SSRN 4493398.  
+Yupeng Chang, Xu Wang, Jindong Wang, Yuan Wu, Linyi Yang, Kaijie Zhu, Hao Chen, Xiaoyuan Yi, Cunxiang Wang, Yidong Wang, et al. 2023. A survey on evaluation of large language models. ACM Transactions on Intelligent Systems and Technology.  
+Liting Chen, Lu Wang, Hang Dong, Yali Du, Jie Yan, Fangkai Yang, Shuang Li, Pu Zhao, Si Qin, Saravan Rajmohan, et al. 2023. Introspective tips: Large language model for in-context decision making. arXiv preprint arXiv:2305.11598.  
+Erik D Demaine, Martin L Demaine, Nicholas JA Harvey, Ryuhei Uehara, Takeaki Uno, and Yushi Uno. 2014. Uno is hard, even for a single player. Theoretical Computer Science, 521:51-61.
+
+Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. 2021. Glm: General language model pretraining with autoregressive blank infilling. arXiv preprint arXiv:2103.10360.  
+Keith Frankish and William M Ramsey. 2014. The Cambridge handbook of artificial intelligence. Cambridge University Press.  
+Deep Ganguli, Nicholas Schiefer, Marina Favaro, and Jack Clark. 2023. Challenges in evaluating AI systems.  
+Ran Gong, Qiuyuan Huang, Xiaojian Ma, Hoi Vo, Zane Durante, Yusuke Noda, Zilong Zheng, Song-Chun Zhu, Demetri Terzopoulos, Li Fei-Fei, et al. 2023. Mindagent: Emergent gaming interaction. arXiv preprint arXiv:2309.09971.  
+Jiaxian Guo, Bo Yang, Paul Yoo, Bill Yuchen Lin, Yusuke Iwasawa, and Yutaka Matsuo. 2023. Suspicion-agent: Playing imperfect information games with theory of mind aware gpt-4.  
+Shangmin Guo, Haoran Bu, Haochuan Wang, Yi Ren, Dianbo Sui, Yuming Shang, and Siting Lu. 2024a. Economics arena for large language models. arXiv preprint arXiv:2401.01735.  
+Shoutao Guo, Shaolei Zhang, Zhengrui Ma, Min Zhang, and Yang Feng. 2024b. Sillm: Large language models for simultaneous machine translation. arXiv preprint arXiv:2402.13036.  
+Zexue He, Yu Wang, Julian McAuley, and Bodhisattwa Prasad Majumder. 2022. Controlling bias exposure for fair interpretable predictions. arXiv preprint arXiv:2210.07455.  
+Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring massive multitask language understanding.  
+Adam Karvonen. 2024. Emergent world models and latent variable estimation in chess-playing language models.  
+Dirk P Kroese, Tim Brereton, Thomas Taimre, and Zdravko I Botev. 2014. Why the monte carlo method is so important today. Wiley Interdisciplinary Reviews: Computational Statistics, 6(6):386-392.  
+Junjie Li, Sotetsu Koyamada, Qiwei Ye, Guoqing Liu, Chao Wang, Ruihan Yang, Li Zhao, Tao Qin, Tie-Yan Liu, and Hsiao-Wuen Hon. 2020. Suphx: Mastering mahjong with deep reinforcement learning. arXiv preprint arXiv:2003.13590.  
+Jonathan Light, Min Cai, Sheng Shen, and Ziniu Hu. 2023. Avalonbench: Evaluating llms playing the game of avalon.  
+Michael Lederman Littman. 1996. Algorithms for sequential decision-making. Brown University.
+
+Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al. 2023. Self-refine: Iterative refinement with self-feedback. arXiv preprint arXiv:2303.17651.  
+Jeffery S McMullen. 2015. Entrepreneurial judgment as empathic accuracy: A sequential decision-making approach to entrepreneurial action. Journal of Institutional Economics, 11(3):651-681.  
+Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, and Martin Riedmiller. 2013. Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602.  
+Christopher Z Mooney. 1997. Monte carlo simulation. 116. Sage.  
+OpenAI. 2022. Introducing chatgpt. https://openai.com/blog/chatgpt. Accessed: 2023-09-30.  
+Bernhard Pfann. 2021. Tackling the uno card game with reinforcement learning. Towards Data Science: tackling-uno-card-game-with-reinforcement-learning.  
+Oscar Sainz, Jon Campos, Iker García-Ferrero, Julien Etxaniz, Oier Lopez de Lacalle, and Eneko Agirre. 2023. NLP evaluation in trouble: On the need to measure LLM data contamination for each benchmark. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 10776-10787, Singapore. Association for Computational Linguistics.  
+Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2024. Reflexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems, 36.  
+David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez, Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, et al. 2017. Mastering the game of go without human knowledge. nature, 550(7676):354-359.  
+Jubi Taneja, Avery Laird, Cong Yan, Madan Musuvathi, and Shuvendu K Lahiri. 2024. Llm-vectorizer: Llm-based verified loop vectorizer. arXiv preprint arXiv:2406.04693.  
+Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. 2023. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805.  
+Oguzhan Topsakal, Colby Jacob Edell, and Jackson Bailey Harper. 2024. Evaluating large language models with grid-based game competitions: An extensible lmm benchmark and leaderboard.
+
+Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971.  
+Sheng Wang, Zihao Zhao, Xi Ouyang, Qian Wang, and Dinggang Shen. 2023a. Chatcad: Interactive computer-aided diagnosis on medical image using large language models. arXiv preprint arXiv:2302.07257.  
+Shenzhi Wang, Chang Liu, Zilong Zheng, Siyuan Qi, Shuo Chen, Qisen Yang, Andrew Zhao, Chaofei Wang, Shiji Song, and Gao Huang. 2023b. Avalon's game of thoughts: Battle against deception through recursive contemplation.  
+Yu Wang, Zhiwei Liu, Jianguo Zhang, Weiran Yao, Shelby Heinecke, and Philip S Yu. 2023c. Drdt: Dynamic reflection with divergent thinking for lmbased sequential recommendation. arXiv preprint arXiv:2312.11336.  
+Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. 2022. Emergent abilities of large language models. arXiv preprint arXiv:2206.07682.  
+Willem A Wythoff. 1907. A modification of the game of nim. Nieuw Arch. Wisk, 7(2):199-202.  
+Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, Boyang Hong, Ming Zhang, Junzhe Wang, Senjie Jin, Enyu Zhou, et al. 2023. The rise and potential of large language model based agents: A survey. arXiv preprint arXiv:2309.07864.  
+Cheng Xu, Shuhao Guan, Derek Greene, and M-Tahar Kechadi. 2024. Benchmark data contamination of large language models: A survey.  
+Yuzhuang Xu, Shuo Wang, Peng Li, Fuwen Luo, Xiaolong Wang, Weidong Liu, and Yang Liu. 2023. Exploring large language models for communication games: An empirical study on werewolf.  
+Yang You, Liangwei Li, Baisong Guo, Weiming Wang, and Cewu Lu. 2019. Combinational q-learning for dou di zhu. arXiv preprint arXiv:1901.08925.  
+Aohan Zeng, Xiao Liu, Zhengxiao Du, Zihan Wang, Hanyu Lai, Ming Ding, Zhuoyi Yang, Yifan Xu, Wendi Zheng, Xiao Xia, et al. 2022. Glm-130b: An open bilingual pre-trained model. arXiv preprint arXiv:2210.02414.  
+Zhongshen Zeng, Pengguang Chen, Shu Liu, Haiyun Jiang, and Jiaya Jia. 2024. Mr-gsm8k: A meta-reasoning revolution in large language model evaluation.
+
+Daochen Zha, Kwei-Herng Lai, Yuanpu Cao, Songyi Huang, Ruzhe Wei, Junyu Guo, and Xia Hu. 2019. Rlcard: A toolkit for reinforcement learning in card games. arXiv preprint arXiv:1910.04376.
+
+Yuexiang Zhai, Hao Bai, Zipeng Lin, Jiayi Pan, Shengbang Tong, Yifei Zhou, Alane Suhr, Saining Xie, Yann LeCun, Yi Ma, and Sergey Levine. 2024. Finetuning large vision-language models as decision-making agents via reinforcement learning.
+
+Jintian Zhang, Xin Xu, and Shumin Deng. 2023. Exploring collaboration mechanisms for llm agents: A social psychology view. arXiv preprint arXiv:2310.02124.
+
+Kun Zhou, Yutao Zhu, Zhipeng Chen, Wentong Chen, Wayne Xin Zhao, Xu Chen, Yankai Lin, Ji-Rong Wen, and Jiawei Han. 2023. Don't make your llm an evaluation benchmark theater.
+
+# Appendix
+
+# A UNO Game
+
+In this section of appendix, you will learn what UNO Game is, and in order to facilitate the evaluation of LLMs with the UNO Game, we have made some slight modifications to it.
+
+# A.1 Game Objective
+
+In the modified UNO game, We simply set the game objective as to be the first player to clear out the hand. Players play alternatively(2 players) or in circle manner(3 or more players) and strive to achieve the unique goal. It should be noted that if the cards in the deck are exhausted by players, the player with the fewest number of cards in hand wins, so there may be multiple winners in the same game.
+
+# A.2 UNO Cards
+
+UNO comprises 3 categories of cards: number cards, function cards, and wild cards. In total, UNO features 108 cards.
+
+- The Number Cards: the number cards can be expressed in the form of COLOR + NUMBER, where COLOR is belong to the set  $\{Red, Blue, Yellow, Green\}$ , and NUMBER is an integer from 0 to 9. It is important to note that there is only one 0-number card per color, while there are two 1-9 number cards per color. There are a total of 76 number cards.  
+- The Function Cards: the function cards can be expressed in the form of COLOR
+
++ FUNCTION, where COLOR is belong to the set  $\{Red,Blue,Yellow,Green\}$  and FUNCTION is belong to the set  $\{Skip,Reverse,DrawTwo\}$ . There are two cards of the same COLOR for each FUNCTION. There are a total of 24 function cards.
+
+- **Skip:** the player's next player skips this round of play.  
+- Reverse: the player reverses the order of play (from clockwise to counterclockwise, or from counterclockwise to clockwise).  
+- Draw Two: The player's next player draws two cards and skips this round of play.
+
+- The Wild Cards: the wild card includes 4 Black Wild cards and 4 Black Wild Draw Four cards, totaling 8 cards.
+
+- Wild: the player selects one color from the COLOR set  $\{Red,Blue,Yellow,Green\}$  as the new color for the top card in the discard pile.  
+- Wild Draw Four: the player selects one color from the COLOR set  $\{\text{Red}, \text{Blue}, \text{Yellow}, \text{Green}\}$  as the new color for the top card in the discard pile, and the player's next player draws 4 cards.
+
+# A.3 Game Progress
+
+First, deal each player 7 initial cards in clockwise order, then continue drawing cards until a number card is drawn and set as the top card of the initial discard pile. All players take rounds playing cards in clockwise order (it will be reversed by a reverse card) until a player runs out of his cards or the draw pile is exhausted, signaling the end of the game.
+
+# A.4 Legal Decision(Action)
+
+In every round of the game, player in charge can using rules to match the top card of the discard pile otherwise pick up a new card into hand. The rules, or say, the legal decisions consists of several sorts:
+
+- Draw Card: If a player does not have any cards to play during their playing round, they must draw a card, or their previous
+
+player used Draw Two or Wild Draw Four cards to make the player draw multiple cards.
+
+- Select Card: In a player's playing round, they need to play a card that matches either the COLOR, NUMBER, or FUNCTION of the top card in the discard pile, or play a Wild card(include Wild Draw Four card) to match. The card played by the player then becomes the new top card.  
+- Select Color: After selecting either Wild Card or Wild Draw Four Card, the player needs to convert the color of the current top card to one of  $\{Red, Blue, Yellow, Green\}$ .  
+- Select ChallengeFlag: The use of the Wild Draw Four card may be illegal. After a player plays a Wild Draw Four card, their next player can choose to challenge its use. If the player who played the Wild Draw Four card still holds non-Wild cards matching the color of the current top card, the use of the Wild Draw Four card is illegal. Possible scenarios are as follows: (1) If the player's play is illegal and their next player challenges it, the player must draw 4 cards, and their next player faces no penalty; (2) If the player's play is legal and their next player challenges it, the player's next player must draw 6 cards. (3) If the player's next player does not challenge, regardless of the legality of the player's play, the player's next player must draw 4 cards. Note that Challenge is not a stand-alone action to complete turns, it should be accompanied by a card draw or card match action.
+
+# B Prompt
+
+Here is the prompt design for the entire experiment.
+
+# B.1 Select Card
+
+The input1 prompt of the select card shared by the vanilla LLM player and the TUTRI player is shown in the Figure 4. The game history Reflection module prompt of the select card for the TUTRI player is shown in the Figure 5. The game strategy reflection module prompt of the select card for the TUTRI player is shown
+
+in the Figure 6.
+
+# B.2 Select Color
+
+The input1 prompt of the select color shared by the vanilla LLM player and the TUTRI player is shown in the Figure 7. The game history reflection module prompt of the select color for the TUTRI player is shown in the Figure 8. The game strategy reflection module prompt of the select color for the TUTRI player is shown in the Figure 9.
+
+# B.3 Select ChallengeFlag
+
+The input1 prompt of the select ChallengeFlag shared by the vanilla LLM player and the TUTRI player is shown in the Figure 10. The game history reflection module prompt of the select ChallengeFlag for the TUTRI player is shown in the Figure 11. The game strategy reflection module prompt of the select ChallengeFlag for the TUTRI player is shown in the Figure 12.
+
+# Select Card Input Prompt
+
+You are playing a two-player UNO game.
+
+- You are the player{player_id}, and your opponent is the player{opponent_id}.  
+- Currently, there are {len_deck} cards in the deck, and the discard pile has {len_discard_pile} cards.  
+- The number of cards in the hand of your opponent is  $\{\text{len\_opponent\_hand}\}$ .  
+- The game history of the last {len_history} rounds is {history}.  
+- Your entire hand consists of  $\{\text{hand}\}$ .  
+- The cards you can play are: {playable_card}.
+
+Please note that you are not playing a normal UNO game, if the cards in the deck are depleted, the person who has the minimum cards will win directly. The goal of the game is to minimize the number of cards in your possession. In order to win the UNO game, you must consider all the provided information and select the best card from the cards you can play.
+
+The output should strictly be a JSON object with two keys: 'thoughts' and 'action'.
+
+- In this context, the value corresponding to the 'thoughts' key represents your thoughts and considerations, with its data type being a string.  
+- Simultaneously, the value corresponding to the 'action' key is a int which represents the card index of the card you have selected.
+
+Figure 4: The input1 prompt of the select card shared by the vanilla LLM player and the TUTRI player.
+
+# Select Card Reflection1 Prompt
+
+Here is the statistical data of the game history: {history_summary}
+
+Now, in order to win the game, you must consider the statistical data carefully and reflect the action you just selected. You should strictly output a JSON object with two keys: 'reflection' and 'action'.
+
+- The value corresponding to the 'reflection' key is your reflection.  
+- The value corresponding to the 'action' key is the card index you currently select.
+
+Figure 5: The game history reflection module prompt of the select card for TUTRI player.
+
+# Select Card Reflection2 Prompt
+
+Here is an useful tip that you can follow:
+
+- The card values range from low to high, starting with number cards 0, followed by number cards (1-9), reverse cards, skip cards and wild cards.  
+- It is better to start with low-value cards before playing high-value cards.  
+Unless your opponent is on the verge of victory, it is time to play some high-value cards to disrupt your opponent's strategy.
+
+Now, in order to win the game, you should reflect the action you just selected based on the tip. You should strictly output a JSON object with two keys: 'reflection' and 'action'.
+
+- The value corresponding to the 'reflection' key is your reflection.  
+- The value corresponding to the 'action' key is the final card index you currently select.
+
+Figure 6: The game strategy reflection module prompt of the select card for TUTRI player.
+
+# Select Color Input Prompt
+
+You are playing a two-player UNO game.
+
+- You are the player{player_id}, and your opponent is the player{opponent_id}.  
+- Currently, there are {len_deck} cards in the deck, and the discard pile has {len_discard_pile} cards.  
+- The number of cards in the hand of your opponent is  $\{\text{len\_opponent\_hand}\}$ .  
+- The game history of the last  $\{\text{len\_history}\}$  rounds is  $\{\text{history}\}$ .  
+- Your entire hand consists of  $\{\text{hand}\}$ .
+
+Please note that you are not playing a normal UNO game, if the cards in the deck are depleted, the person who has the minimum cards will win directly. The goal of the game is to minimize the number of cards in your possession. In order to win the UNO game, you just played a {wild_type} card, and you must consider all the provided information and select the best color from Red, Yellow, Blue and Green to switch.
+
+The output should strictly be a JSON object with two keys: 'thoughts' and 'action'.
+
+- In this context, the value corresponding to the 'thoughts' key represents your thoughts and considerations, with its data type being a string.  
+- Simultaneously, the value corresponding to the 'action' key is one of Red, Yellow, Blue or Green, indicating the color you have selected.
+
+Figure 7: The input1 prompt of the select color shared by the vanilla LLM player and the TUTRI player.
+
+# Select Color Reflection1 Prompt
+
+Here is the statistical data of the game history:
+
+{history_summary}
+
+Now, in order to win the game, you must consider the statistical data carefully and reflect the action you just selected.
+
+You should strictly output a JSON object with two keys: 'reflection' and 'action'.
+
+- The value corresponding to the 'reflection' key is your reflection.  
+- The value corresponding to the 'action' key is the color you currently select.
+
+Figure 8: The game history reflection module prompt of the select color for TUTRI player.
+
+# Select Color Reflection2 Prompt
+
+Here are some useful tips that you can follow:
+
+It is better to select the color with the highest frequency of occurrence in your hand.  
+- It is better to avoid selecting the color with the lowest frequency of occurrence in your hand.  
+- Consider carefully which color of cards is relatively more frequent in your opponent's hand and try to avoid selecting that color.
+
+Now, in order to win the game, you should reflect the action you just selected based on these tips.
+
+You should strictly output a JSON object with two keys: 'reflection' and 'action'.
+
+- The value corresponding to the 'reflection' key is your reflection.  
+- The value corresponding to the 'action' key is the final color you currently select."
+
+Figure 9: The game strategy reflection module prompt of the select color for TUTRI player.
+
+# Select ChallengeFlag Input Prompt
+
+You are playing a two-player UNO game.
+
+- You are the player{player_id}, and your opponent is the player{opponent_id}.  
+- Currently, there are {len_deck} cards in the deck, and the discard pile has {len_discard_pile} cards.  
+- The number of cards in the hand of your opponent is  $\{\text{len\_opponent\_hand}\}$ .  
+- The game history of the last  $\{\text{len\_history}\}$  rounds is  $\{\text{history}\}$ .  
+- Your entire hand consists of  $\{\text{hand}\}$ .
+
+Your opponent played a Wild Draw Four card, and changed the color of the current discard pile's top card to {new_color}. But the use of the Wild Draw Four card may be illegal, when your opponent still has cards in {old_color}. Please note that you are not playing a normal UNO game, if the cards in the deck are depleted, the person who has the minimum cards will win directly. The goal of the game is to minimize the number of cards in your possession. In order to win the UNO game, you must consider all the provided information and select whether to challenge the use of the Wild Draw Four card which played by your opponent.
+
+The output should strictly be a JSON object with two keys: 'thoughts' and 'action'.
+
+- In this context, the value corresponding to the 'thoughts' key represents your thoughts and considerations, with its data type being a string.  
+- Simultaneously, the value corresponding to the 'action' key is 'Yes' or 'No', indicating that you select to challenge or not to challenge, respectively."
+
+Figure 10: The input1 prompt of the select ChallengeFlag shared by the vanilla LLM player and the TUTRI player.
+
+# Select ChallengeFlag Reflection1 Prompt
+
+Here is the statistical data of the game history:
+
+{history_summary}
+
+Now, in order to win the game, you must consider the statistical data carefully and reflect the action you just selected. You should strictly output a JSON object with two keys: 'reflection' and 'action'.
+
+- The value corresponding to the 'reflection' key is your reflection.  
+- The value corresponding to the 'action' key is the choice you currently select.
+
+Figure 11: The game history reflection module prompt of the select ChallengeFlag for TUTRI player.
+
+# Select ChallengeFlag Reflection2 Prompt
+
+Here are some useful tips that you can follow:
+
+- Please remember the penalty for a failed challenge: you must draw 6 cards.  
+- Please remember the benefits of a successful challenge: your opponent must draw 4 cards.  
+- Wild Draw Four is only illegal if your opponent has cards of {old_color} color in his hand.  
+Please carefully consider whether your opponent's Wild Draw Four card is genuinely illegal. Now, in order to win the game, you should reflect the action you just selected based on these tips.
+
+You should strictly output a JSON object with two keys: 'reflection' and 'action'.
+
+- The value corresponding to the 'reflection' key is your reflection.  
+- The value corresponding to the 'action' key is the final choice you currently select.
+
+Figure 12: The game strategy reflection module prompt of the select ChallengeFlag for TUTRI player.

2024/UNO Arena for Evaluating Sequential Decision-Making Capability of Large Language Models/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:81e8b821130aeed0fa11798abb661a473665b0bda158f74b369fc0f839cbba86
+size 477157

2024/UOUO_ Uncontextualized Uncommon Objects for Measuring Knowledge Horizons of Vision Language Models/3fd78e6a-da2e-45dd-83ff-a49b6a122da9_content_list.json

@@ -0,0 +1,1350 @@
+[
+    {
+        "type": "text",
+        "text": "UOUO: Uncontextualized Uncommon Objects for Measuring Knowledge Horizons of Vision Language Models",
+        "text_level": 1,
+        "bbox": [
+            122,
+            89,
+            875,
+            130
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Xinyu Pi*1 Mingyuan Wu*2 Jize Jiang*2 Haozhen Zheng*2",
+        "bbox": [
+            220,
+            146,
+            779,
+            162
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Beitong Tian² Chengxiang Zhai² Klara Nahrstedt² Zhiting Hu¹",
+        "bbox": [
+            203,
+            162,
+            791,
+            180
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "<sup>1</sup>University of California San Diego <sup>2</sup>University of Illinois Urbana-Champaign",
+        "bbox": [
+            181,
+            180,
+            815,
+            197
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "xpi@ucsd.edu, {mw34, jizej2, haozhen3}@illinois.edu",
+        "bbox": [
+            238,
+            197,
+            754,
+            212
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "* indicates equal contribution",
+        "bbox": [
+            344,
+            214,
+            650,
+            229
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            260,
+            339,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Smaller-scale Vision-Language Models (VLMs) often claim to perform on par with larger models in general-domain visual grounding and question-answering benchmarks while offering advantages in computational efficiency and storage. However, their ability to handle rare objects, which fall into the long tail of data distributions, is less understood. To rigorously evaluate this aspect, we introduce the \"Uncontextualized Uncommon Objects\" (UOOU) benchmark. This benchmark focuses on systematically testing VLMs with both large and small parameter counts on rare and specialized objects. Our comprehensive analysis reveals that while smaller VLMs maintain competitive performance on common datasets, they significantly underperform on tasks involving uncommon objects. We also propose an advanced, scalable pipeline for data collection and cleaning, ensuring the UOOU benchmark provides high-quality, challenging instances. These findings highlight the need to consider long-tail distributions when assessing the true capabilities of VLMs. Code and project details for UOOU can be found at https://zoezheng126.github.io/UOOU-Website/.",
+        "bbox": [
+            144,
+            284,
+            460,
+            667
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            687,
+            258,
+            702
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The advent of Vision-Language Models (VLMs) has marked a revolutionary leap in the integration of natural language processing and computer vision, largely due to the capabilities of the self-attention mechanism and the Transformer architecture (Vaswani et al., 2023). These technologies allow VLMs to effectively process and fuse information from both text and images, leading to significant advancements in tasks that require multimodal understanding, such as visual question answering and image captioning (Radford et al., 2021; Li et al., 2023; Alayrac et al., 2022; Xu et al., 2023; Young et al., 2014).",
+        "bbox": [
+            112,
+            711,
+            489,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "VLMs, trained on large-scale datasets, typically boast high performance on general tasks involving everyday objects and common scenarios (Li et al., 2024; Du et al., 2022; Wang et al., 2023). However, models of smaller scale, defined here as having fewer than 70 billion parameters, often claim to match the capabilities of their larger counterparts on general domain tasks (Lin et al., 2015; Agrawal et al., 2016; Yu et al., 2016; Liu et al., 2024; Goyal et al., 2017; Yu et al., 2023b) while offering advantages in computational efficiency and storage. Despite these claims, the No-Free-Lunch Theorem (Wolpert and Macready, 1997) suggests that these smaller models may compromise on their ability to handle less common or more complex scenarios that lie in the long tail of data distributions.",
+        "bbox": [
+            507,
+            260,
+            884,
+            518
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "One natural and intuitive hypothesis is that they are sacrificing their fitness to the elements on the long tail of the distribution. Empirical observations of real-world data frequently align with Zipf's and Power Law (Piantadosi, 2014; Clauset et al., 2009), which indicates that while some objects and concepts are exceedingly common, a vast number of them are rare and fall into the long tail of the distribution. Understanding how well VLMs handle these rare and uncommon instances is crucial for assessing their true robustness and applicability across diverse and nuanced contexts.",
+        "bbox": [
+            507,
+            519,
+            885,
+            709
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Despite the importance of this evaluation, there is currently a lack of dedicated benchmarks that systematically test VLMs on objects and concepts that are significantly outside the everyday norm. To address this gap, we introduce the \"Uncontextualized Uncommon Objects\" (UOUO) benchmark. The object class distribution of UOUO is systematically out of common image sources such as ImageNet (Russakovsky et al., 2015), COCO (Lin et al., 2015), and Open Image Dataset (Kuznetsova et al., 2020). Our goal is to rigorously test and quantify the performance of both large-scale and small-scale VLMs on elements from the long tail of",
+        "bbox": [
+            507,
+            711,
+            885,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "6432",
+        "bbox": [
+            478,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 6432-6441 November 12-16, 2024 ©2024 Association for Computational Linguistics",
+        "bbox": [
+            159,
+            945,
+            835,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/6898fd3fb1b89d0bb395714add135a57deb41ef29b6939a1c09716bfc297982d.jpg",
+        "image_caption": [
+            "Figure 1: UOUO Data Curation Pipeline. Snowflake means frozen weights, and fire means tune-able weights."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            119,
+            58,
+            880,
+            272
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "the distribution to showcase their knowledge gap.",
+        "bbox": [
+            112,
+            316,
+            480,
+            331
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The contribution of our work is three-fold. (1) We compile a million-scale dataset specifically designed to include uncommon and uncontextualized objects, which are rarely encountered in everyday contexts but are significant in specialized domains. (2) We evaluate the performance gap between large-scale and small-scale VLMs when dealing with these rare elements, showcasing the significant knowledge and performance gap between large-and small-scale model on the long-tail distributions. (3) We propose a systematic pipeline for automatic and scalable data collection and cleaning, ensuring high-quality and representative testing instances.",
+        "bbox": [
+            112,
+            332,
+            489,
+            541
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            114,
+            556,
+            270,
+            570
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Real-world VQA Benchmarks Based on our survey, the typical real-world visual question answering datasets (excluding mathematics, celebrity, landmark, place, OCR and chart-reading) used in popular open-source VLMs such as LLaVa (Li et al., 2024), CogVLM (Wang et al., 2023) BLIP2 (Li et al., 2023), Qwen VL (Bai et al., 2023) and MiniCPM-V (Yu et al., 2023a) includes the following: COCO (Lin et al., 2015), RefCOCO (Yu et al., 2016), NoCAPs (Agrawal et al., 2019), MMBench (Liu et al., 2024), VQA-v2 (Goyal et al., 2017), OK-VQA (Marino et al., 2019), MME (Fu et al., 2024), GQA (Hudson and Manning, 2019).",
+        "bbox": [
+            112,
+            583,
+            489,
+            791
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Much to our surprise, it turns out that the image sources of GQA, RefCoCo, OK-VQA, MME Coarse-Grained Recognition, VQA-v2, and a significant proportion of MMBench are all direct random samples from COCO. Only NoCAPs features novel object classes (sourced from the 600-categories Open Image Dataset (Kuznetsova et al., 2020) outside COCO's less-than-100 common",
+        "bbox": [
+            112,
+            793,
+            489,
+            920
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "classes. This showcases the significant limitation of categorical diversity of extant VQA datasets. The knowledge and performance gap between the small- and large- scale VLMs might be concealed in such low coverage and diversity.",
+        "bbox": [
+            507,
+            316,
+            884,
+            395
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Existing Datasets with Uncommon Object Labels In extant datasets, Stanford Cars (Krause et al., 2013), CUB-bird (Wah et al., 2011), Deepfish (Saleh et al., 2020), ROCOv2 (Rückert et al., 2024), FGVC-Aircraft (Maji et al., 2013) also features rare object labels. Some non-academic mine & stone datasets, and chemical objects datasets can also be found on internet. However, the typical emphasis of these datasets is either fine-grained subtype or subspecies of common objects, or domain-specific expert knowledge. In realistic use cases such as autonomous car or embodied robotics, such knowledge might have limited generalizability.",
+        "bbox": [
+            507,
+            407,
+            884,
+            617
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3 Data Curation and Filtering",
+        "text_level": 1,
+        "bbox": [
+            507,
+            630,
+            788,
+            646
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3.1 Domain Selection and Scraping",
+        "text_level": 1,
+        "bbox": [
+            507,
+            657,
+            803,
+            673
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To construct the UOUO (Uncontextualized Uncommon Objects) benchmark, we began by selecting specific domains that are rich in specialized knowledge yet contain objects and tools that are rarely encountered by the general public. Our focus was on the industry sector, given its diversity and the presence of numerous specialized tools and equipment. These artificial tools are significantly out of the distribution of ImageNet, COCO, and Open Image Dataset.",
+        "bbox": [
+            505,
+            678,
+            884,
+            839
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We used Wikipedia as a starting point, targeting the page dedicated to manufacturing (https://en.wikipedia.org/wiki/Manufacturing). For each sub-sector identified within this domain, we employed GPT-4-Turbo (OpenAI, 2024) to gener",
+        "bbox": [
+            507,
+            841,
+            884,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "6433",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "ate a list of the top 50 objects or tools pertinent to experts in the field but obscure to the general populace. This list was generated through prompt-based querying, asking the model to identify objects that are crucial within the industry but not commonly known.",
+        "bbox": [
+            112,
+            84,
+            487,
+            179
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Once we had our list of uncommon objects, we performed a Google Image Search for each object name. For each query, we collected the top 50 image results. This approach allowed us to gather a diverse set of images representing each object under different conditions and contexts. For detailed dataset statistics of UOUO, we refer readers to Appendix C.",
+        "bbox": [
+            112,
+            181,
+            487,
+            309
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Manuall Annotation The image instances collected from Google Image Search can be noisy, with perhaps one fifth irrelevant instances for each queried uncommon category. To ensure the quality and relevance of the dataset, we implemented a rigorous annotation and cleaning process, combining manual and automated techniques. Our team manually reviewed and annotated on a subset of the collected categories of images to identify and remove outliers and noisy data. Categories with consistent visual representation across examples were retained, while those filled with ambiguous or irrelevant images were discarded. This initial curation aimed to maintain high fidelity to the object's intended representation. The instruction for manual annotation of UOUO can be found in Appendix B.",
+        "bbox": [
+            112,
+            317,
+            489,
+            589
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Automatic Data Cleaning We utilized the CLIP model to further enhance the dataset. CLIP (Contrastive Language-Image Pre-training) provides embeddings for both images and text, enabling us to compute similarities within and across categories. For each image, we extracted its CLIP image embedding  $E_{i}^{c}$  and the text embedding  $T_{c}$  of its corresponding category name (Radford et al., 2021; Sun et al., 2023). We calculated the cosine similarity between all pairs of image embeddings within each category to construct a GRAM matrix  $G$ , where  $G_{i,j} = \\mathrm{Cosine}(E_i^c,E_j^c)$ . Additionally, we computed the image-text similarity for each image as  $\\mathrm{Cosine}(E_i^c,T_c)$ , alongside statistical metrics such as the percentile, mean, and variance of the average similarity within each category.",
+        "bbox": [
+            112,
+            599,
+            489,
+            854
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The complete feature set includes image embeddings, GRAM percentiles (25th, 50th, and 75th), GRAM mean and variance, the instance's mean similarity with other images, and the percentiles of",
+        "bbox": [
+            112,
+            857,
+            489,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/766e6b66afdd9ab692d8b2f66b4707124e5b06e3f3099aea6b017377400b4985.jpg",
+        "image_caption": [
+            "Figure 2: With MMD, we can retrieve harder negative examples and construct higher-quality test instances."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            521,
+            79,
+            694,
+            250
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/963c7614d4b68d951a398098f8366cc8a1023d1d458b36f49a1c498bd886d0c5.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            702,
+            80,
+            880,
+            250
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "its pairwise similarities. Additionally, it incorporates image-text similarity metrics, corresponding percentiles, z-scores, and the instance label.",
+        "bbox": [
+            507,
+            313,
+            882,
+            362
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Using these computed features, we applied an XGBoost classifier to label each image instance. This classifier was trained on manually cleaned data from 500 categories to distinguish between high-quality and low-quality instances based on their similarity scores.",
+        "bbox": [
+            507,
+            363,
+            882,
+            458
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We optimized our XGBoost classifier (Chen and Guestrin, 2016) through 5-fold cross-validation and grid search to identify the best hyperparameters. The optimal configuration consisted of a maximum tree depth of 6, 200 estimators, a learning rate of 0.15, a subsample ratio of 1.0, gamma value of 0.1, and a colsample-bytree of 1.0. Additionally, the regularization parameters included reg-lambda of 1.5 and reg-alpha of 0.0, with a minimum child weight of 1.0.",
+        "bbox": [
+            507,
+            461,
+            882,
+            620
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The classifier achieved an accuracy of 0.8754 on cross-validation, closely aligning with human judgment, and exhibited Macro-Average Precision, Recall, and F1-Score of 0.8631, 0.8353, and 0.8460, respectively.",
+        "bbox": [
+            507,
+            621,
+            882,
+            702
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4 Test Instances Generation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            715,
+            766,
+            732
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Background Removal and Decontextualization Connectionist neural networks (including VLMs) are notoriously known for their tendency of overfitting to spurious correlations present in the training data. For instance, in our collected data, bulldozers are often seen in construction scenes laden with materials such as sand, concrete, and bricks. This high co-occurrence can lead models to rely on these contextual cues rather than truly understanding and recognizing the bulldozer itself. To mitigate this issue and ensure that models focus on the objects rather",
+        "bbox": [
+            507,
+            744,
+            882,
+            920
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "6434",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "than their typical environments, we implement a robust background removal process to decontextualize all candidate objects in our dataset. To achieve effective background removal, we utilize a state-of-the-art, off-the-shelf background removal model (BRIA-AI, 2024).",
+        "bbox": [
+            112,
+            84,
+            489,
+            180
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Testing Instances Generation To assess the performance of Vision-Language Models on our UOVO benchmark, we generated challenging test instances designed to probe the models' capabilities beyond common knowledge. Specifically, we employ the CLIP embeddings combined with the Maximum Mean Discrepancy (MMD) with a Gaussian RBF kernel (Dziugaite et al., 2015) to identify and retrieve hard negative examples.",
+        "bbox": [
+            112,
+            187,
+            487,
+            332
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Let  $\\mathbf{x}$  and  $\\mathbf{y}$  be the sets of CLIP embeddings for two different object categories, each of shape  $(n,d)$ , where  $n$  is the number of embeddings and  $d$  is the embedding dimension.",
+        "bbox": [
+            112,
+            332,
+            487,
+            398
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The Maximum Mean Discrepancy (MMD) between sets of embeddings  $\\mathbf{x}$  and  $\\mathbf{y}$  is calculated as follows:",
+        "bbox": [
+            112,
+            398,
+            487,
+            445
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {M M D} (\\mathbf {x}, \\mathbf {y}) = k (\\mathbf {x}, \\mathbf {x}) + k (\\mathbf {y}, \\mathbf {y}) - 2 \\cdot k (\\mathbf {x}, \\mathbf {y})\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            122,
+            454,
+            478,
+            472
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where the Gaussian Radial Basis Function (RBF) kernel value  $k(\\mathbf{a},\\mathbf{b})$  is defined as:",
+        "bbox": [
+            112,
+            479,
+            487,
+            512
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nk (\\mathbf {a}, \\mathbf {b}) = \\frac {1}{n ^ {2}} \\sum_ {i = 1} ^ {n} \\sum_ {j = 1} ^ {n} \\exp \\left(- \\frac {1}{2 \\sigma^ {2}} \\left\\| \\mathbf {a} _ {i} - \\mathbf {b} _ {j} \\right\\| ^ {2}\\right)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            117,
+            518,
+            478,
+            561
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "For our calculations, we set  $\\sigma = 10$",
+        "bbox": [
+            112,
+            568,
+            384,
+            581
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We use the Mosaic Image Augmentation Technique (Ge et al., 2021) to generate testing data in a scalable way. Each testing data point is created from four images, each background-removed. The four images contain objects of different categories but share some similar visual properties such as structures, colors, or textures. The selection of these images is determined by the Maximum Mean Discrepancy (MMD) distance between the categories they belong to. The closer the MMD distance, the more similar in features they might appear. We create an  $800 \\times 800$  canvas large enough to accommodate all four images. Then, each of the four images is augmented and positioned on the canvas's top-left, top-right, bottom-left, or bottom-right. The ground-truth bounding box for the object grounding is generated from the segmentation mask of background removal and normalized to be dimension-insensitive, accounting for potential differences in the VLM's rescaling process. Figure 2 showcases an exemplar test instance.",
+        "bbox": [
+            112,
+            583,
+            489,
+            921
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/a017c16dfcf550f87196c2cfe2c6c3821219ffe8026860a5a0348407f5403de0.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>mIoU-mmd</td><td>mIoU-rand</td><td>acc-mmd</td><td>acc-rand</td></tr><tr><td>llava-v1.5-7b</td><td>0.18</td><td>0.41</td><td>0.42</td><td>0.70</td></tr><tr><td>llava-v1.5-13b</td><td>0.23</td><td>0.47</td><td>0.44</td><td>0.73</td></tr><tr><td>llava-v1.6-vicuna-7b</td><td>0.28</td><td>0.48</td><td>0.49</td><td>0.75</td></tr><tr><td>llava-v1.6-vicuna-13b</td><td>0.28</td><td>0.49</td><td>0.52</td><td>0.78</td></tr><tr><td>llava-v1.6-34b</td><td>0.38</td><td>0.55</td><td>0.57</td><td>0.83</td></tr><tr><td>cogvlm-llama3-chat-19b</td><td>0.49</td><td>0.69</td><td>0.43</td><td>0.60</td></tr><tr><td>gemini-1.5-pro</td><td>0.27</td><td>0.27</td><td>0.63</td><td>0.80</td></tr><tr><td>gpt-4-turbo</td><td>0.34</td><td>0.38</td><td>0.67</td><td>0.90</td></tr><tr><td>gpt-4o</td><td>0.33</td><td>0.35</td><td>0.68</td><td>0.88</td></tr></table>",
+        "bbox": [
+            510,
+            80,
+            873,
+            166
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Mosaic Grounding Performance Metrics",
+        "bbox": [
+            527,
+            175,
+            862,
+            190
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "5 Experiment",
+        "text_level": 1,
+        "bbox": [
+            507,
+            206,
+            648,
+            223
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Procedures Following the aforementioned test instance generation, we test both open source VLMs that are trained to perform grounding, including: llava-v1.5-7b, llava-v1.5-13b (Liu et al., 2023), llava-v1.6-vicuna-7b, llava-v1.6-vicuna-13b, llava-v1.6-34b (Li et al., 2024), cogglm-v1.5-vicuna-7b (Wang et al., 2023), and proprietary VLMs including: gemini-1.5-pro (Team, 2024), gpt-4-turbo, gpt-4o (OpenAI, 2024).",
+        "bbox": [
+            507,
+            233,
+            884,
+            378
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We test VLMs' performance on both randomly generated test instances and the MMD-augmented hard instances. We employ two metrics to quantify the performance:  $mIoU$  - Mean IoU (Intersection over Union), a standard metric for object segmentation; and Accuracy, which we prompt the VLM to output one positions from \"top-left, top-right, bottom-left, bottom-right\", and directly evaluate whether the answer matches the ground truth. The prompts used in this experiment can be found in Appendix A.",
+        "bbox": [
+            507,
+            379,
+            884,
+            556
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Observations and Analysis We present all experimental results in Table 1. (a) Comparing horizontally across columns, we observe significant performance drops of smaller-scale models in both mIoU and Accuracy with the application of MMD-based hard instance generation. Notably, the performance drops of many of them are around  $30\\%$ . This provides solid support for our initial hypothesis that smaller-scale models have some, but insufficient fitness to the long-tail distribution objects. Furthermore, the drastic performance change showcases MMD's effectiveness in generating hard instances and non-robustness of existing grounding models. (b) Comparing vertically within columns, the central tendency is that larger scale models (except Genimi which might not be trained to perform grounding) perform much better than small-scale models in accuracy. This reveals the concealed gap of knowledge horizon of small- and large- scale models, which is usually unobservable in benchmarks consist of common objects. (c) The observation that GPT-4 series can still handle the task",
+        "bbox": [
+            507,
+            567,
+            884,
+            920
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "6435",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/b51991b040de317c10856fcf2bfa5048a31576fa7079136d1d5bbdbfdd39183f.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            126,
+            77,
+            868,
+            231
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/d8f8db19821976746a200574caa8288bf9ae784f4f1c101b0f16257204308e8c.jpg",
+        "image_caption": [
+            "Figure 3: A glimpse into COCO and UOUO, with demo images of COCO cited from official website."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            127,
+            236,
+            868,
+            394
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "remarkably well (near  $90\\%$  and  $70\\%$  on random and MMD settings, respectively) showcases the task's solvability, revealing the soundness of our automatically constructed test instances.",
+        "bbox": [
+            112,
+            455,
+            487,
+            519
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "6 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            112,
+            533,
+            245,
+            546
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In our work, we introduced the UOUO benchmark to assess VLMs on objects out of everyday distributions. Our findings show that while smaller VLMs perform well on tasks of common objects, they struggle significantly with uncommon objects, unlike larger models which handle these challenges much better. This highlights the need to consider long-tail distributions in evaluations. The systematic data curation, filtering, and hard test instance generation pipeline for UOUO construction has high extensibility, paving the road of future research of long-tail distribution objects. UOUO itself could also be expanded in this way, extending beyond the domain of manufacturing and to other broad category of objects.",
+        "bbox": [
+            112,
+            558,
+            489,
+            802
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "7 Limitations",
+        "text_level": 1,
+        "bbox": [
+            112,
+            814,
+            248,
+            829
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "One limitation of our work is the reliance on automated data collection and cleaning processes, though efficient, may introduce biases or fail to capture nuanced representations compared to fully manual curation. We also note that the Mosaic",
+        "bbox": [
+            112,
+            841,
+            489,
+            919
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Image Augmentation was applied with the assumption that the model takes single-image inputs. Our preliminary experiment showed most VLMs have limited to none multi-image inference support, thus multi-image inputs results are not included in UOUO benchmark. The UOUO benchmark currently emphasizes static images, potentially overlooking the dynamic and context-dependent nature of object recognition in real-world scenarios. Future extensions should explore a wider range of uncommon objects across various fields and consider the inclusion of video or sequential data to better reflect real-world applications. Addressing these limitations will enhance the comprehensiveness and applicability of the UOUO benchmark.",
+        "bbox": [
+            507,
+            455,
+            884,
+            697
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgement",
+        "text_level": 1,
+        "bbox": [
+            509,
+            714,
+            672,
+            730
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "This work was supported by DARPA ECOLE HR00112390063 and by the National Science Foundation grants NSF CNS 19-00875, NSF CNS 21-06592, NSF OAC 18-35834 KN, NSF CCF 22-17144. This research used the Delta advanced computing and data resource which is supported by the National Science Foundation (award OAC 2005572) and the State of Illinois. Any results and opinions are our own and do not represent views of National Science Foundation.",
+        "bbox": [
+            507,
+            744,
+            882,
+            904
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "6436",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            115,
+            84,
+            213,
+            98
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Aishwarya Agrawal, Jiasen Lu, Stanislaw Antol, Margaret Mitchell, C. Lawrence Zitnick, Dhruv Batra, and Devi Parikh. 2016. Vqa: Visual question answering. Preprint, arXiv:1505.00468.",
+            "Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. 2019. nocaps: novel object captioning at scale. In Proceedings of the IEEE International Conference on Computer Vision, pages 8948-8957.",
+            "Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. 2022. Flamingo: a visual language model for few-shot learning. Preprint, arXiv:2204.14198.",
+            "Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. 2023. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. Preprint, arXiv:2308.12966.",
+            "BRIA-AI. 2024. Bria background removal v1.4 model card. https://huggingface.co/briaai/RMBG-1. 4. Accessed: 2024-06-13.",
+            "Tianqi Chen and Carlos Guestrin. 2016. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD '16. ACM.",
+            "Aaron Clauset, Cosma Rohilla Shalizi, and Mark EJ Newman. 2009. Power-law distributions in empirical data. SIAM review, 51(4):661-703.",
+            "Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. 2022. Glm: General language model pretraining with autoregressive blank infilling. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 320-335.",
+            "Gintare Karolina Dziugaite, Daniel M Roy, and Zoubin Ghahramani. 2015. Training generative neural networks via maximum mean discrepancy optimization. arXiv preprint arXiv:1505.03906.",
+            "Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. 2024. Mme: A comprehensive evaluation benchmark for multimodal large language models. Preprint, arXiv:2306.13394."
+        ],
+        "bbox": [
+            115,
+            108,
+            489,
+            919
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. 2021. Yolox: Exceeding yolo series in 2021. Preprint, arXiv:2107.08430.",
+            "Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. 2017. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. Preprint, arXiv:1612.00837.",
+            "Drew A. Hudson and Christopher D. Manning. 2019. Gqa: A new dataset for real-world visual reasoning and compositional question answering. *Preprint*, arXiv:1902.09506.",
+            "Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 2013. 3d object representations for fine-grained categorization. In 2013 IEEE International Conference on Computer Vision Workshops, pages 554-561.",
+            "Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, and Vittorio Ferrari. 2020. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision, 128(7):1956-1981.",
+            "Bo Li, Kaichen Zhang, Hao Zhang, Dong Guo, Renrui Zhang, Feng Li, Yuanhan Zhang, Ziwei Liu, and Chunyuan Li. 2024. Llava-last: Stronger llms supercharge multimodal capabilities in the wild.",
+            "Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023. Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. Preprint, arXiv:2301.12597.",
+            "Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dolar. 2015. Microsoft coco: Common objects in context. Preprint, arXiv:1405.0312.",
+            "Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. 2023. Improved baselines with visual instruction tuning.",
+            "Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. 2024. Mmbench: Is your multi-modal model an all-around player? Preprint, arXiv:2307.06281.",
+            "S. Maji, J. Kannala, E. Rahtu, M. Blaschko, and A. Vedaldi. 2013. Fine-grained visual classification of aircraft. Technical report.",
+            "Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. 2019. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Conference on Computer Vision and Pattern Recognition (CVPR).",
+            "OpenAI. 2024. Gpt-4 technical report. Preprint, arXiv:2303.08774."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            919
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "6437",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Steven T. Piantadosi. 2014. Zipf's word frequency law in natural language: a critical review and future directions. Psychonomic Bulletin & Review, 21(5):1112-1130.",
+            "Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning transferable visual models from natural language supervision. Preprint, arXiv:2103.00020.",
+            "Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. 2015. Imagenet large scale visual recognition challenge. Preprint, arXiv:1409.0575.",
+            "Johannes Rückert, Louise Bloch, Raphael Brüngel, Ahmad Idrissi-Yaghir, Henning Schäfer, Cynthia S. Schmidt, Sven Koitka, Obioma Pelka, Asma Ben Abacha, Alba G. Seco de Herrera, Henning Müller, Peter A. Horn, Felix Nensa, and Christoph M. Friedrich. 2024. Rocov2: Radiology objects in context version 2, an updated multimodal image dataset. Preprint, arXiv:2405.10004.",
+            "Alzayat Saleh, Issam H Laradji, Dmitry A Konovalov, Michael Bradley, David Vazquez, and Marcus Sheaves. 2020. A realistic fish-habitat dataset to evaluate algorithms for underwater visual analysis. Scientific Reports, 10(1):14671.",
+            "Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. 2023. Eva-clip: Improved training techniques for clip at scale. Preprint, arXiv:2303.15389.",
+            "Gemini Team. 2024. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. Preprint, arXiv:2403.05530.",
+            "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2023. Attention is all you need. Preprint, arXiv:1706.03762.",
+            "Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. 2011. The caltech-ucsd birds-200-2011 dataset.",
+            "Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. 2023. Cogvlm: Visual expert for pretrained language models. Preprint, arXiv:2311.03079.",
+            "David H Wolpert and William G Macready. 1997. No free lunch theorems for optimization. IEEE transactions on evolutionary computation, 1(1):67-82.",
+            "Peng Xu, Xiatian Zhu, and David A. Clifton. 2023. Multimodal learning with transformers: A survey. Preprint, arXiv:2206.06488."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            919
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. 2014. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78.",
+            "Licheng Yu, Patrick Poirson, Shan Yang, Alexander C. Berg, and Tamara L. Berg. 2016. Modeling context in referring expressions. Preprint, arXiv:1608.00272.",
+            "Tianyu Yu, Yuan Yao, Haoye Zhang, Taiwen He, Yifeng Han, Ganqu Cui, Jinyi Hu, Zhiyuan Liu, Hai-Tao Zheng, Maosong Sun, et al. 2023a. Rlhf-v: Towards trustworthy mllms via behavior alignment from fine-grained correctional human feedback. arXiv preprint arXiv:2312.00849.",
+            "Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. 2023b. Mm-vet: Evaluating large multimodal models for integrated capabilities. Preprint, arXiv:2308.02490."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            360
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Appendix",
+        "text_level": 1,
+        "bbox": [
+            510,
+            370,
+            596,
+            387
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "A Prompt Details",
+        "text_level": 1,
+        "bbox": [
+            510,
+            395,
+            678,
+            411
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "mIoU Prompt Please provide the bounding box coordinate (x1,y1,x2,y2) of {object name} in the image with the format \\n item1:(x1,y1,x2,y2).",
+        "bbox": [
+            509,
+            420,
+            880,
+            468
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Accuracy Prompt Identify the location of the given object in this 2x2 mosaic image. The possible answers are: 'top left', 'top right', 'bottom left', 'bottom right', or 'none'. Only give a deterministic response as one of the possible answers. If the object is not present, the response should be 'none'. Please do not give more than one response.  $\\backslash n$  object name: {object name}  $\\backslash n$  Location:",
+        "bbox": [
+            509,
+            476,
+            882,
+            605
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Prefix setting All other settings follow the model's defaults. For instance, in the case of llava, the prompt prefix is: A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions. USER: <image>\\n{question} ASSISTANT:",
+        "bbox": [
+            509,
+            613,
+            882,
+            725
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "B Instructions for manual annotation",
+        "text_level": 1,
+        "bbox": [
+            510,
+            737,
+            848,
+            752
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Consider removing individual images, or removing the entire category completely. Any category or image meeting the following criterion should be removed.",
+        "bbox": [
+            509,
+            760,
+            880,
+            824
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Category Exclusion Principles",
+        "text_level": 1,
+        "bbox": [
+            532,
+            834,
+            786,
+            850
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "1. Lack of Sufficiently Uniform Images",
+        "text_level": 1,
+        "bbox": [
+            564,
+            857,
+            873,
+            872
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Categories should be excluded if the collected images do not show enough consistency in appearance.",
+        "bbox": [
+            584,
+            873,
+            884,
+            920
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "6438",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "2. Ambiguous Collections of Objects Categories representing a collection multiple objects (e.g. Wax Work Tools, Tools, Kits, etc.) should be cluded.",
+            "3. Insufficient Number of Images Collected Categories should be removed if there are not enough images available.",
+            "4. Not a Tool Items that are not standalone tools (e.g. pastes or liquids that always need to be stored in a container) should be excluded."
+        ],
+        "bbox": [
+            164,
+            84,
+            489,
+            312
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "- Image Sample Exclusion Principles",
+        "text_level": 1,
+        "bbox": [
+            136,
+            323,
+            431,
+            340
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. Distinctive Image Images that are significantly different from other images within the same category should be excluded.",
+            "2. Cluttered Composition Images with cluttered backgrounds that make it difficult or impossible to isolate the tool.",
+            "3. Partial Display Images that only show part of the object should be excluded.",
+            "4. Not a Real Object Images that depict diagrams, 3D renders, or other non-realistic representations should be excluded.",
+            "5. Excessive Text Images that contain excessive text, which obscures the main object, should be excluded."
+        ],
+        "bbox": [
+            166,
+            347,
+            489,
+            659
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "C Important statistics of UOUO",
+        "text_level": 1,
+        "bbox": [
+            114,
+            670,
+            410,
+            686
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Number of categories:",
+        "bbox": [
+            136,
+            696,
+            329,
+            711
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Number of categories originally: 27,926",
+            "- Number of categories kept: 25,864",
+            "Percentage of categories kept:  $92.6\\%$"
+        ],
+        "bbox": [
+            168,
+            719,
+            485,
+            771
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "- Total number of images:",
+        "bbox": [
+            136,
+            800,
+            347,
+            816
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Number of images originally: 956,167",
+            "- Number of images kept: 678,535",
+            "- Percentage of images kept:  $71.0\\%$"
+        ],
+        "bbox": [
+            168,
+            824,
+            473,
+            876
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Images per category stats:",
+        "bbox": [
+            136,
+            904,
+            359,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "- Original dataset:",
+        "bbox": [
+            564,
+            84,
+            729,
+            99
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "* Average: 34.382",
+            "* Minimum: 11",
+            "* Maximum: 48"
+        ],
+        "bbox": [
+            594,
+            103,
+            736,
+            151
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "-Filtered dataset:",
+        "bbox": [
+            564,
+            156,
+            715,
+            168
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "* Average: 26.235",
+            "* Minimum: 5",
+            "* Maximum: 48"
+        ],
+        "bbox": [
+            594,
+            175,
+            734,
+            223
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "- Average percentage of images kept in each category:  $76.0\\%$",
+        "bbox": [
+            532,
+            255,
+            880,
+            287
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "D Wikipedia Industry List",
+        "text_level": 1,
+        "bbox": [
+            509,
+            298,
+            756,
+            315
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "See figure 4.",
+        "bbox": [
+            509,
+            324,
+            606,
+            338
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "E Randomly sampled 100 categories",
+        "text_level": 1,
+        "bbox": [
+            509,
+            351,
+            840,
+            369
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "See table 2.",
+        "bbox": [
+            509,
+            376,
+            600,
+            391
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "6439",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/7bd38cf9d8aa350a94ad687668f6d022f6cbfea3b242152e0edf2305066190f6.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"3\">V·T·E Major industries [hide]</td></tr><tr><td colspan=\"3\">Natural sector [hide]</td></tr><tr><td rowspan=\"3\">Biotic</td><td>Agriculture</td><td>Arable farming (Cereals · Legumes · Vegetables · Fiber crops · Oilseeds · Sugar · Tobacco) · Permanent crops (Apples et al · Berries · Citrus · Stone fruits · Tropical fruit · Viticulture · Cocoa · Coffee · Tea · Nuts · Olives · Medicinal plants · Spices) · Horticulture (Flowers · Seeds) · Animal husbandry (Beef cattle · Dairy farming · Fur farming · Horses · Other livestock · Pig · Wool · Poultry · Beekeeping · Cochineal · Shellac · Silk) · Hunting (Fur trapping)</td></tr><tr><td>Forestry</td><td>Silviculture (Bamboo) · Logging (Firewood) · Rattan · Tree tapping (Frankincense · Gum arabic · Gutta-percha · Maple syrup · Mastic · Natural rubber · Palm sugar, syrup, &amp; wine · Pine resin) · Wild mushrooms (Fungiculture · Truffles)</td></tr><tr><td>Aquatic</td><td>Fishing (Anchovies · Herring · Sardines · Cod · Haddock · Pollock · Mackerel · Shark · Swordfish · Tuna · Crabs · Lobsters · Sea urchins · Squid · Whaling) · Aquaculture (Carp · Catfish · Tilapia · Abalone · Mussels · Oysters · Pearls · Microalgae · Seaweed) · Both (Clams · Sea cucumbers · Scallops · Salmon · Shrimp)</td></tr><tr><td>Geological</td><td colspan=\"2\">Fossil fuels (Coal · Peat · Natural gas · Oil shale · Petroleum · Tar sands) · Mining of ores (Aluminum · Copper · Iron · Gold · Silver · Palladium · Platinum · Lithium · Rare-earth metals · Uranium) · Other minerals (Gemstones · Phosphorus · Potash · Salt · Sulfur) · Quarrying (Gravel · Sand · Chalk · Clay · Gypsum · Limestone · Dimension stone · Granite · Marble)</td></tr><tr><td colspan=\"3\">Industrial sector [hide]</td></tr><tr><td rowspan=\"5\">Manufacturing</td><td>Light industry</td><td>Food (Animal feed · Baking · Canning · Dairy products · Flour · Meat · Prepared · Preserved · Sweets · Vegetable oils) · Beverages (Beer · Bottled water · Liquor · Soft drinks · Wine) · Textiles (Carding · Dyeing · Pins · Spinning · Weaving · Carpets · Lace · Linens · Rope) · Clothing (Accessories · Dressmaking · Furs · Hatmaking · Sewing · Shoemaking · Tailoring) · Printing (Bookbinding · Embossing · Engraving · Secure · Typesetting) · Media reproduction (Cassette tapes · Phonographs · Optical discs) · Metal fabrication (Boilermaking · Builders&#x27; &amp; household hardware · Cutlery · Gunsmithing · Locksmithing · Machining · Other smithing · Powder metallurgy · Prefabrication · Surface finishing) · Other fabrication (3D printing · Blowing molding · Drawing · Extrusion · Glassblowing · Injection moulding · Pottery · Sintering · Stonemasonry · Woodworking) · Furniture · Other goods (Baggage · Bicycles · Jewellery · Medical supplies · Musical instruments · Office supplies · Outdoors &amp; sports equipment · Personal protective equipment · Toys)</td></tr><tr><td>Electrical &amp; optical</td><td>Electronics (Components · Circuit boards · Semiconductors) · Computers (Computer systems · Parts &amp; peripherals · Blank storage media) · Communications equipment (Mobile phones · Network infrastructure) · Consumer electronics (Televisions · Video game consoles) · Instrumentation (Clocks &amp; watches · GPS devices · Scientific instruments) · Medical imaging systems · Optical instruments (Cameras · Gun &amp; spotting scopes · Laser construction · Lens grinding · Microscopes · Telescopes) · Electrical equipment (Batteries · Electrical &amp; fiber optic cables · Electric lighting · Electric motors · Home appliances · Transformers)</td></tr><tr><td>Chemicals</td><td>Coal &amp; oil refining (Bitumen · Coke · Diesel fuel · Fuel oil · Gasoline · Jet fuel · Kerosene · Mineral oil · Paraffin wax · Petrochemicals · Petroleum jelly · Propane · Synthetic oil · Tar) · Commodity chemicals (Fertilizers · Industrial gases · Pigments · Pure elements) · Speciality chemicals (Adhesives · Agrochemicals · Aroma compounds · Cleaning products · Cosmetics · Explosives · Fireworks · Paints &amp; inks · Perfumes · Soap · Toiletries) · Fine chemicals · Pharmaceuticals (Antibiotics · Blood products · Chemical &amp; hormonal contraceptives · Generic drugs · Illegal drugs · Supplements · Vaccines)</td></tr><tr><td>Materials</td><td>Leather (Liming &amp; delimining · Tanning · Currying &amp; oiling) · Wood (Drying · Sawmillings · Engineered · Lumber · Composite) · Paper (Sizing · Cardboard · Pulp · Tissue) · Rubber (Tires · Vulcanized rubber) · Plastics (Commodity · Engineered · Specialty · Pellets · Synthetic fibers · Thermoplastics &amp; thermosets) · Glass (Borosilicate · Fused quartz · Soda-lime · Float glass · Glass fiber · Glass wool &amp; fiberglass · Safety glass) · Ceramics (Brick · Earthenware · Porcelain · Refractory · Tile) · Cement (Mortar · Plaster · Ready-mix concrete) · Other mineral (Abraxives · Carbon fibers &amp; advanced materials · Mineral wool · Synthetic gems) · Metal refining (Iron · Aluminum · Copper) · Alloys (Steel) · Formed metal (Rolled · Forged) · Cast metal</td></tr><tr><td>Heavy industry</td><td>Machinery (Conveyors · Heavy · Hydraulic · Machine tools · Power &amp; wind turbines) · Automobiles · Other heavy vehicles (Aerospace &amp; space · Rail vehicles · Ships &amp; offshore platforms) · Weapons</td></tr><tr><td>Utilities</td><td colspan=\"2\">Power (Electric · Gas distribution · Renewable) · Water (Sewage) · Waste management (Collection · Dumping · Hazardous · Recycling) · Remediation · Telecom networks (Cable TV · Internet · Mobile · Satellite · Telephone)</td></tr><tr><td>Construction</td><td colspan=\"2\">Buildings (Commercial · Industrial · Residential) · Civil engineering (Bridges · Railways · Roads · Tunnels · Canals · Dams · Shedding · Harbors) · Specialty trades (Cabinetry · Demolition · Electrical wiring · Elevators · HVAC · Painting and decorating · Plumbing · Site preparation)</td></tr><tr><td colspan=\"3\">Service sector [hide]</td></tr><tr><td>Sales</td><td colspan=\"2\">Retail (Car dealership · Consumer goods · General store · Grocery store · Department store · Mail order · Online shopping · Specialty store) · Wholesale (Auction · Brokerage · Distribution)</td></tr><tr><td>Transport &amp; Storage</td><td colspan=\"2\">Cargo (Air cargo · Intermodal · Mail · Moving company · Rail · Trucking) · Passenger transport (Airlines · Car rentals · Passenger rail · Rideshaaring · Taxis) · Warehousing (Self storage)</td></tr><tr><td>Hospitality</td><td colspan=\"2\">Foodservice (Drink service · Cafés · Catering · Fast food · Food delivery · Restaurants · Teahouses) · Hotels</td></tr><tr><td>Asset management</td><td colspan=\"2\">Financial services (Banking · Credit · Financial advice · Holding company · Money transfer · Payment cards · Risk management · Securities) · Insurance (Health · Life · Pension funding · Property · Reinsurance) · Real estate (Brokerage · Property management)</td></tr><tr><td>Professional</td><td colspan=\"2\">Accounting (Assurance · Audit · Bookkeeping · Tax advice) · Architecture &amp; engineering (Inspection · Surveying · Physical, product, &amp; system testing) · Design (Fashion · Interior · Product) · Legal services · Management (Consulting · Public relations) · Marketing (Advertising)</td></tr><tr><td>Healthcare</td><td colspan=\"2\">Medicine (Dentist offices · Hospitals · Nursing) · Residential care · Veterinary medicine</td></tr><tr><td>Entertainment &amp; leisure</td><td colspan=\"2\">Gambling (Online) · Sport · Venues (Arcades · Amusement parks · Fairgrounds · Nightclubs)</td></tr><tr><td>Other</td><td colspan=\"2\">Administrative (Customer service · Leasing · Renting · Staffing · Private investigation &amp; security) · Maintenance (Janitors · Landscaping) · Repairs · Personal services (Beauty · Dry cleaning · Funeral · Maid service · Pet care · Sex) · Poverty · Travel (Business travel · Cruise lines · Tourism)</td></tr><tr><td colspan=\"3\">Information sector [hide]</td></tr><tr><td>Publishing &amp; Mass media</td><td colspan=\"2\">Written (Books · Periodicals · Software) · Audio-visual (Film · Music · Video games) · Broadcasting (News · Radio · Television) · Internet (Hosting · Social networks · Streaming · Websites)</td></tr><tr><td>Education</td><td colspan=\"2\">Primary · Secondary · Tertiary (Vocational school · University) · Testing · Tutoring</td></tr><tr><td>Other</td><td colspan=\"2\">Creative · Language · Research and development (Basic research)</td></tr></table>",
+        "bbox": [
+            213,
+            80,
+            786,
+            925
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Figure 4: Wikipedia Industry List",
+        "bbox": [
+            384,
+            938,
+            613,
+            951
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "6440",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "table",
+        "img_path": "images/b0aa3b0228c2c422f90aaa393d567f4cf7f5901d3a7a8bfff026256f53b6f669.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>2D pantograph</td><td>AC Recharge Kit</td><td>Adhesive scale</td><td>Aluminum dross pro-cessing machine</td></tr><tr><td>Artificial insemination gun</td><td>Ballistic clipboard</td><td>Ballot Box (for collect-ing anonymous feed-back)</td><td>Banjo rim lathe</td></tr><tr><td>Bingo balls</td><td>Broodstock tanks</td><td>Broom</td><td>Burnishing Stone</td></tr><tr><td>Cable Retention Sleeve</td><td>Carding Machine</td><td>Cattle Curtain</td><td>Cell Model</td></tr><tr><td>Climbing rope</td><td>Coal centrifuge</td><td>Coffee roaster</td><td>Cold Storage Back-pack</td></tr><tr><td>Compressor (hard-ware)</td><td>Cooling Incubator</td><td>Copy Stand</td><td>Culture trays</td></tr><tr><td>Dehooking tool</td><td>Deposit Slip Printer</td><td>Disc golf basket welder</td><td>Disc repair kit</td></tr><tr><td>Display Turntables</td><td>Distillation column</td><td>Electronic rate board</td><td>Evaporating Dish</td></tr><tr><td>Extrusion laminator</td><td>Fiber disc</td><td>Fishing rod holders</td><td>Flange spreader</td></tr><tr><td>Flower press</td><td>Foundation crack ruler</td><td>Fume Extraction Hood</td><td>Goniophotometer</td></tr><tr><td>Graduated cylinders</td><td>Granule Filler</td><td>Inductively Coupled Plasma (ICP) Spectrometer</td><td>Irrigation pipelayer</td></tr><tr><td>Lacquer polishing brush</td><td>Leachate Collection Pipe</td><td>Live Feed Incubator</td><td>Longlines and ropes</td></tr><tr><td>Martingale</td><td>Metal scribe</td><td>Mobile manufacturing unit (MMU)</td><td>Mushroom grow tent</td></tr><tr><td>Music on hold player</td><td>Network Firewall Hardware</td><td>Offshore aquaculture cage</td><td>Ore skip</td></tr><tr><td>Oscillating shaker</td><td>Oxygen concentrators</td><td>Packing Gauge</td><td>Pellets coating system</td></tr><tr><td>Pellicle Formation Tool</td><td>Pillory</td><td>Pin beater</td><td>Pointer stick</td></tr><tr><td>Portable battery booster</td><td>Pressure vessels</td><td>Print Quality Inspection Scope</td><td>Pulling post</td></tr><tr><td>Purging compound dispenser</td><td>Queue stanchion</td><td>Quick release hook</td><td>Roll Coating Paint Line</td></tr><tr><td>Rope pump</td><td>Rotary drum bauxite washer</td><td>Rotary impeller feeder</td><td>Sand filter</td></tr><tr><td>Scale Breaker</td><td>Schlenk flask</td><td>Security drone</td><td>Security token device</td></tr><tr><td>Shear Line</td><td>Shock Absorber</td><td>Sign language interpreter gloves</td><td>Slab Tongs</td></tr><tr><td>Slush ice machines</td><td>Soap scum remover</td><td>Spin Welder</td><td>Spoke cutting machine</td></tr><tr><td>Spot meter</td><td>Springform pan</td><td>Tabbing shears for composite test specimens</td><td>Texture sprayer</td></tr><tr><td>Tower Climbing Harness</td><td>Violin varnish brush</td><td>Vixen Plate</td><td>Wall Hooks for Art</td></tr><tr><td>Waste basket</td><td>Water jet cutter for stone</td><td>Whalebone Scaper</td><td>Wire Mesh Cable Trays</td></tr></table>",
+        "bbox": [
+            121,
+            117,
+            875,
+            856
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Table 2: List of 100 Randomly Sampled Categories",
+        "bbox": [
+            321,
+            866,
+            672,
+            881
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "6441",
+        "bbox": [
+            480,
+            928,
+            517,
+            940
+        ],
+        "page_idx": 9
+    }
+]