Add Batch fff8d255-5ded-44d2-9161-739bf43781e1 data

.gitattributes

@@ -6980,3 +6980,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2024/Boosting[[:space:]]Alignment[[:space:]]for[[:space:]]Post-Unlearning[[:space:]]Text-to-Image[[:space:]]Generative[[:space:]]Models/8543d31b-4c45-4414-88f2-a92516616507_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2024/Boosting[[:space:]]Generalization[[:space:]]in[[:space:]]Parametric[[:space:]]PDE[[:space:]]Neural[[:space:]]Solvers[[:space:]]through[[:space:]]Adaptive[[:space:]]Conditioning/b1e80266-eea9-48b1-a751-62c403ef5eb8_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2024/Boosting[[:space:]]Graph[[:space:]]Pooling[[:space:]]with[[:space:]]Persistent[[:space:]]Homology/0df1c958-56f0-496b-880b-b83d4a049b97_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Boosting[[:space:]]Sample[[:space:]]Efficiency[[:space:]]and[[:space:]]Generalization[[:space:]]in[[:space:]]Multi-agent[[:space:]]Reinforcement[[:space:]]Learning[[:space:]]via[[:space:]]Equivariance/32c6ec7d-bfc7-4045-9513-b9da6e0f43e4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Boosting[[:space:]]Semi-Supervised[[:space:]]Scene[[:space:]]Text[[:space:]]Recognition[[:space:]]via[[:space:]]Viewing[[:space:]]and[[:space:]]Summarizing/95d67c07-a36c-457b-aab4-2733e2025fc0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Boosting[[:space:]]Text-to-Video[[:space:]]Generative[[:space:]]Model[[:space:]]with[[:space:]]MLLMs[[:space:]]Feedback/6a3a1350-8e96-4b88-b833-9075cc51309c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Boosting[[:space:]]Transferability[[:space:]]and[[:space:]]Discriminability[[:space:]]for[[:space:]]Time[[:space:]]Series[[:space:]]Domain[[:space:]]Adaptation/679c5ddf-dcc2-453f-813c-b5bec1720d30_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Boosting[[:space:]]Vision-Language[[:space:]]Models[[:space:]]with[[:space:]]Transduction/409473c7-326d-4c59-8e3b-87e2a4f1bbbf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Boosting[[:space:]]Weakly[[:space:]]Supervised[[:space:]]Referring[[:space:]]Image[[:space:]]Segmentation[[:space:]]via[[:space:]]Progressive[[:space:]]Comprehension/3f96f166-4232-493b-9f5f-4136794df000_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Boosting[[:space:]]the[[:space:]]Potential[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]with[[:space:]]an[[:space:]]Intelligent[[:space:]]Information[[:space:]]Assistant/b820ed92-a302-48e1-98ad-294d4073262d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Boosting[[:space:]]the[[:space:]]Transferability[[:space:]]of[[:space:]]Adversarial[[:space:]]Attack[[:space:]]on[[:space:]]Vision[[:space:]]Transformer[[:space:]]with[[:space:]]Adaptive[[:space:]]Token[[:space:]]Tuning/ba47e4ae-9806-497d-a191-2bcd53dfb393_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Bootstrapping[[:space:]]Top-down[[:space:]]Information[[:space:]]for[[:space:]]Self-modulating[[:space:]]Slot[[:space:]]Attention/ff05c636-c0b2-42c2-a962-03c695996e75_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Boundary[[:space:]]Decomposition[[:space:]]for[[:space:]]Nadir[[:space:]]Objective[[:space:]]Vector[[:space:]]Estimation/f72c3ef7-023d-48e2-b71b-3831f445ead5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Boundary[[:space:]]Matters_[[:space:]]A[[:space:]]Bi-Level[[:space:]]Active[[:space:]]Finetuning[[:space:]]Method/ec4cc988-2de2-40d9-8c85-9c3a551ceb26_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Bounds[[:space:]]for[[:space:]]the[[:space:]]smallest[[:space:]]eigenvalue[[:space:]]of[[:space:]]the[[:space:]]NTK[[:space:]]for[[:space:]]arbitrary[[:space:]]spherical[[:space:]]data[[:space:]]of[[:space:]]arbitrary[[:space:]]dimension/e1eeba5f-436f-4dc9-806d-1907bb133d1f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Brain[[:space:]]Treebank_[[:space:]]Large-scale[[:space:]]intracranial[[:space:]]recordings[[:space:]]from[[:space:]]naturalistic[[:space:]]language[[:space:]]stimuli/901c3878-1640-43ab-9da7-bc0aeff00af4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Brain-JEPA_[[:space:]]Brain[[:space:]]Dynamics[[:space:]]Foundation[[:space:]]Model[[:space:]]with[[:space:]]Gradient[[:space:]]Positioning[[:space:]]and[[:space:]]Spatiotemporal[[:space:]]Masking/6915f457-de8a-460d-ad57-4d868609a7a2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/BrainBits_[[:space:]]How[[:space:]]Much[[:space:]]of[[:space:]]the[[:space:]]Brain[[:space:]]are[[:space:]]Generative[[:space:]]Reconstruction[[:space:]]Methods[[:space:]]Using_/38268f9f-a9b9-423a-99f8-28c45a2ad39d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Breaking[[:space:]]Determinism_[[:space:]]Fuzzy[[:space:]]Modeling[[:space:]]of[[:space:]]Sequential[[:space:]]Recommendation[[:space:]]Using[[:space:]]Discrete[[:space:]]State[[:space:]]Space[[:space:]]Diffusion[[:space:]]Model/7838f4d5-5d61-4ab2-b084-8e8abcab038b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Breaking[[:space:]]Long-Tailed[[:space:]]Learning[[:space:]]Bottlenecks_[[:space:]]A[[:space:]]Controllable[[:space:]]Paradigm[[:space:]]with[[:space:]]Hypernetwork-Generated[[:space:]]Diverse[[:space:]]Experts/238044de-1255-45fc-b5f0-4a1c207b7724_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Breaking[[:space:]]Semantic[[:space:]]Artifacts[[:space:]]for[[:space:]]Generalized[[:space:]]AI-generated[[:space:]]Image[[:space:]]Detection/d3b2038e-4446-4711-9d30-e39d61afd0c0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Breaking[[:space:]]the[[:space:]]False[[:space:]]Sense[[:space:]]of[[:space:]]Security[[:space:]]in[[:space:]]Backdoor[[:space:]]Defense[[:space:]]through[[:space:]]Re-Activation[[:space:]]Attack/ddce6f7c-8bf5-4027-83ee-aab15d728066_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Breaking[[:space:]]the[[:space:]]curse[[:space:]]of[[:space:]]dimensionality[[:space:]]in[[:space:]]structured[[:space:]]density[[:space:]]estimation/eb1f749c-9f71-41f5-aa2d-86c3c08ec17b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/BricksRL_[[:space:]]A[[:space:]]Platform[[:space:]]for[[:space:]]Democratizing[[:space:]]Robotics[[:space:]]and[[:space:]]Reinforcement[[:space:]]Learning[[:space:]]Research[[:space:]]and[[:space:]]Education[[:space:]]with[[:space:]]LEGO/db877d64-7fa4-44fa-8bf2-05dfa8077892_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Bridge[[:space:]]the[[:space:]]Modality[[:space:]]and[[:space:]]Capability[[:space:]]Gaps[[:space:]]in[[:space:]]Vision-Language[[:space:]]Model[[:space:]]Selection/4c807b7c-51c8-418c-ba9b-e9bf3a993495_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Bridge[[:space:]]the[[:space:]]Points_[[:space:]]Graph-based[[:space:]]Few-shot[[:space:]]Segment[[:space:]]Anything[[:space:]]Semantically/3a7c1db9-52a6-4474-805e-b21f5fcfd18d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Bridge-IF_[[:space:]]Learning[[:space:]]Inverse[[:space:]]Protein[[:space:]]Folding[[:space:]]with[[:space:]]Markov[[:space:]]Bridges/a41b6cda-12a5-4b87-9478-f7d22fcd5511_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Bridging[[:space:]]Gaps_[[:space:]]Federated[[:space:]]Multi-View[[:space:]]Clustering[[:space:]]in[[:space:]]Heterogeneous[[:space:]]Hybrid[[:space:]]Views/0c85c49f-147d-401f-aff6-4ab7bfb795e9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Bridging[[:space:]]Geometric[[:space:]]States[[:space:]]via[[:space:]]Geometric[[:space:]]Diffusion[[:space:]]Bridge/bc70ff0d-e14e-4118-895a-aaae188a5940_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Bridging[[:space:]]Model-Based[[:space:]]Optimization[[:space:]]and[[:space:]]Generative[[:space:]]Modeling[[:space:]]via[[:space:]]Conservative[[:space:]]Fine-Tuning[[:space:]]of[[:space:]]Diffusion[[:space:]]Models/b5690155-aaf5-4cdd-8ada-8014b1ae3606_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Bridging[[:space:]]Multicalibration[[:space:]]and[[:space:]]Out-of-distribution[[:space:]]Generalization[[:space:]]Beyond[[:space:]]Covariate[[:space:]]Shift/1a8f072a-3696-4e14-84a6-f27f2355b91d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Bridging[[:space:]]OOD[[:space:]]Detection[[:space:]]and[[:space:]]Generalization_[[:space:]]A[[:space:]]Graph-Theoretic[[:space:]]View/ce8beca4-7933-4afa-b721-8d7d12bac5e4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Bridging[[:space:]]The[[:space:]]Gap[[:space:]]between[[:space:]]Low-rank[[:space:]]and[[:space:]]Orthogonal[[:space:]]Adaptation[[:space:]]via[[:space:]]Householder[[:space:]]Reflection[[:space:]]Adaptation/be02d223-6b56-4842-a137-e79c494bf659_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Bridging[[:space:]]semantics[[:space:]]and[[:space:]]pragmatics[[:space:]]in[[:space:]]information-theoretic[[:space:]]emergent[[:space:]]communication/1c132f64-c183-4855-a859-81eef2a66b62_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Bridging[[:space:]]the[[:space:]]Divide_[[:space:]]Reconsidering[[:space:]]Softmax[[:space:]]and[[:space:]]Linear[[:space:]]Attention/11d62b7b-8a0b-4ad3-86b2-f009b43ab25b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/BuckTales_[[:space:]]A[[:space:]]multi-UAV[[:space:]]dataset[[:space:]]for[[:space:]]multi-object[[:space:]]tracking[[:space:]]and[[:space:]]re-identification[[:space:]]of[[:space:]]wild[[:space:]]antelopes/07cf2a1c-f551-4166-b364-a3cc01971006_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Buffer[[:space:]]of[[:space:]]Thoughts_[[:space:]]Thought-Augmented[[:space:]]Reasoning[[:space:]]with[[:space:]]Large[[:space:]]Language[[:space:]]Models/c1e8216d-4f61-4bea-b7ed-0886d6f742d5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Building[[:space:]]Timeseries[[:space:]]Dataset_[[:space:]]Empowering[[:space:]]Large-Scale[[:space:]]Building[[:space:]]Analytics/d5563ce9-628d-41eb-9650-0cae10fae54b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Building[[:space:]]a[[:space:]]stable[[:space:]]classifier[[:space:]]with[[:space:]]the[[:space:]]inflated[[:space:]]argmax/e16c50fd-dc36-4aa0-88ec-8f532dd676a7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Building[[:space:]]on[[:space:]]Efficient[[:space:]]Foundations_[[:space:]]Effective[[:space:]]Training[[:space:]]of[[:space:]]LLMs[[:space:]]with[[:space:]]Structured[[:space:]]Feedforward[[:space:]]Layers/56af9ba9-876c-4fbf-92ea-7e0cd3a8222a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Byzantine[[:space:]]Robustness[[:space:]]and[[:space:]]Partial[[:space:]]Participation[[:space:]]Can[[:space:]]Be[[:space:]]Achieved[[:space:]]at[[:space:]]Once_[[:space:]]Just[[:space:]]Clip[[:space:]]Gradient[[:space:]]Differences/318fc1e4-9ea5-4203-9d93-74a19b78e71d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/C-GAIL_[[:space:]]Stabilizing[[:space:]]Generative[[:space:]]Adversarial[[:space:]]Imitation[[:space:]]Learning[[:space:]]with[[:space:]]Control[[:space:]]Theory/c89639f6-ee98-4175-9e8d-c7e8e4af975c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CA-SSLR_[[:space:]]Condition-Aware[[:space:]]Self-Supervised[[:space:]]Learning[[:space:]]Representation[[:space:]]for[[:space:]]Generalized[[:space:]]Speech[[:space:]]Processing/85de2e35-9799-49e3-b36e-a62bb8235b44_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CALANet_[[:space:]]Cheap[[:space:]]All-Layer[[:space:]]Aggregation[[:space:]]for[[:space:]]Human[[:space:]]Activity[[:space:]]Recognition/3a7ab39b-797b-4003-8dca-68f07580d353_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CALE_[[:space:]]Continuous[[:space:]]Arcade[[:space:]]Learning[[:space:]]Environment/70afe2dd-3bce-445e-9411-97d49f1d0b5c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CALVIN_[[:space:]]Improved[[:space:]]Contextual[[:space:]]Video[[:space:]]Captioning[[:space:]]via[[:space:]]Instruction[[:space:]]Tuning/f6af5a1a-9db5-42cd-8abe-e68f11a6a37c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CARES_[[:space:]]A[[:space:]]Comprehensive[[:space:]]Benchmark[[:space:]]of[[:space:]]Trustworthiness[[:space:]]in[[:space:]]Medical[[:space:]]Vision[[:space:]]Language[[:space:]]Models/edff5571-19dd-4d64-a3d1-d618c14537bf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CARE_[[:space:]]a[[:space:]]Benchmark[[:space:]]Suite[[:space:]]for[[:space:]]the[[:space:]]Classification[[:space:]]and[[:space:]]Retrieval[[:space:]]of[[:space:]]Enzymes/35b2ecfb-7218-45a9-a897-13bef43902f1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CAT3D_[[:space:]]Create[[:space:]]Anything[[:space:]]in[[:space:]]3D[[:space:]]with[[:space:]]Multi-View[[:space:]]Diffusion[[:space:]]Models/98175654-2af4-4106-ba53-396db5eb3dea_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CAT_[[:space:]]Coordinating[[:space:]]Anatomical-Textual[[:space:]]Prompts[[:space:]]for[[:space:]]Multi-Organ[[:space:]]and[[:space:]]Tumor[[:space:]]Segmentation/92eea05e-12a7-4b23-9af4-3e5fa7ee8762_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CE-NAS_[[:space:]]An[[:space:]]End-to-End[[:space:]]Carbon-Efficient[[:space:]]Neural[[:space:]]Architecture[[:space:]]Search[[:space:]]Framework/c9d2bc6c-ae73-45bc-8bdb-5d16b499b056_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CHASE_[[:space:]]Learning[[:space:]]Convex[[:space:]]Hull[[:space:]]Adaptive[[:space:]]Shift[[:space:]]for[[:space:]]Skeleton-based[[:space:]]Multi-Entity[[:space:]]Action[[:space:]]Recognition/fbc05f85-2c29-435a-b898-1492efd5896d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CIFD_[[:space:]]Controlled[[:space:]]Information[[:space:]]Flow[[:space:]]to[[:space:]]Enhance[[:space:]]Knowledge[[:space:]]Distillation/e6df53e0-e46c-4a1a-bd6e-45410c46d12b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CLAP4CLIP_[[:space:]]Continual[[:space:]]Learning[[:space:]]with[[:space:]]Probabilistic[[:space:]]Finetuning[[:space:]]for[[:space:]]Vision-Language[[:space:]]Models/1b2b6f59-2f3f-416f-a454-daa4af8387af_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CLAVE_[[:space:]]An[[:space:]]Adaptive[[:space:]]Framework[[:space:]]for[[:space:]]Evaluating[[:space:]]Values[[:space:]]of[[:space:]]LLM[[:space:]]Generated[[:space:]]Responses/5fd9fc51-51dd-4bad-aa6d-258984288566_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CLIP[[:space:]]in[[:space:]]Mirror_[[:space:]]Disentangling[[:space:]]text[[:space:]]from[[:space:]]visual[[:space:]]images[[:space:]]through[[:space:]]reflection/e66ef9e2-0c25-4cb9-b04b-396ecdf5509a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CLIPAway_[[:space:]]Harmonizing[[:space:]]focused[[:space:]]embeddings[[:space:]]for[[:space:]]removing[[:space:]]objects[[:space:]]via[[:space:]]diffusion[[:space:]]models/5a67a3f8-fb36-4114-a552-0e3258b558b6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CLIPCEIL_[[:space:]]Domain[[:space:]]Generalization[[:space:]]through[[:space:]]CLIP[[:space:]]via[[:space:]]Channel[[:space:]]rEfinement[[:space:]]and[[:space:]]Image-text[[:space:]]aLignment/f4f1ddb6-9e25-4c1e-b62f-50b0ce117ef7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CLIPLoss[[:space:]]and[[:space:]]Norm-Based[[:space:]]Data[[:space:]]Selection[[:space:]]Methods[[:space:]]for[[:space:]]Multimodal[[:space:]]Contrastive[[:space:]]Learning/613d2d65-5497-460b-b1c1-114d0e511b62_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CLUES_[[:space:]]Collaborative[[:space:]]Private-domain[[:space:]]High-quality[[:space:]]Data[[:space:]]Selection[[:space:]]for[[:space:]]LLMs[[:space:]]via[[:space:]]Training[[:space:]]Dynamics/e733efc6-dfad-4b1c-bfe2-14a5a15a90f1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CNCA_[[:space:]]Toward[[:space:]]Customizable[[:space:]]and[[:space:]]Natural[[:space:]]Generation[[:space:]]of[[:space:]]Adversarial[[:space:]]Camouflage[[:space:]]for[[:space:]]Vehicle[[:space:]]Detectors/7b0d0782-b437-4792-b024-79a53d55765f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CODA_[[:space:]]A[[:space:]]Correlation-Oriented[[:space:]]Disentanglement[[:space:]]and[[:space:]]Augmentation[[:space:]]Modeling[[:space:]]Scheme[[:space:]]for[[:space:]]Better[[:space:]]Resisting[[:space:]]Subpopulation[[:space:]]Shifts/813550b8-4574-44e0-b998-906c2ee68db4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CODE_[[:space:]]Contrasting[[:space:]]Self-generated[[:space:]]Description[[:space:]]to[[:space:]]Combat[[:space:]]Hallucination[[:space:]]in[[:space:]]Large[[:space:]]Multi-modal[[:space:]]Models/42a09f0a-ed8e-45e1-be6d-621dbc37ff34_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/COLD_[[:space:]]Causal[[:space:]]reasOning[[:space:]]in[[:space:]]cLosed[[:space:]]Daily[[:space:]]activities/8f029aaa-8980-4d5a-929d-5d7de6c6373c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/CONTRAST_[[:space:]]Continual[[:space:]]Multi-source[[:space:]]Adaptation[[:space:]]to[[:space:]]Dynamic[[:space:]]Distributions/9881952e-8e99-4b76-b528-da2d470dec96_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/COSMIC_[[:space:]]Compress[[:space:]]Satellite[[:space:]]Image[[:space:]]Efficiently[[:space:]]via[[:space:]]Diffusion[[:space:]]Compensation/15610b6d-89f3-45d6-a00a-4fe711e57baf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/COVE_[[:space:]]Unleashing[[:space:]]the[[:space:]]Diffusion[[:space:]]Feature[[:space:]]Correspondence[[:space:]]for[[:space:]]Consistent[[:space:]]Video[[:space:]]Editing/59785097-c29c-4004-afd5-0aea27fb56bc_origin.pdf filter=lfs diff=lfs merge=lfs -text

2024/Boosting Sample Efficiency and Generalization in Multi-agent Reinforcement Learning via Equivariance/32c6ec7d-bfc7-4045-9513-b9da6e0f43e4_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c920313e85e564ee26a925cc6fc7bc3ffd239ab77fe1dfe9901192cd685693b7
+size 1615311

2024/Boosting Sample Efficiency and Generalization in Multi-agent Reinforcement Learning via Equivariance/full.md

@@ -0,0 +1,740 @@
+# Boosting Sample Efficiency and Generalization in Multi-agent Reinforcement Learning via Equivariance
+
+Joshua McClellan
+
+JHU APL*
+
+University of Maryland joshmccl@umd.edu
+
+Naveed Haghani
+
+JHU APL
+
+naveed.haghani@jhuapl.edu
+
+John Winder
+
+JHU APL
+
+john.winder@jhuapl.edu
+
+Furong Huang
+
+University of Maryland
+
+furongh@umd.edu
+
+Pratap Tokekar
+
+University of Maryland
+
+tokekar@umd.edu
+
+# Abstract
+
+Multi-Agent Reinforcement Learning (MARL) struggles with sample inefficiency and poor generalization [1]. These challenges are partially due to a lack of structure or inductive bias in the neural networks typically used in learning the policy. One such form of structure that is commonly observed in multi-agent scenarios is symmetry. The field of Geometric Deep Learning has developed Equivariant Graph Neural Networks (EGNN) that are equivariant (or symmetric) to rotations, translations, and reflections of nodes. Incorporating equivariance has been shown to improve learning efficiency and decrease error [2]. In this paper, we demonstrate that EGNNs improve the sample efficiency and generalization in MARL. However, we also show that a naive application of EGNNs to MARL results in poor early exploration due to a bias in the EGNN structure. To mitigate this bias, we present Exploration-enhanced Equivariant Graph Neural Networks or E2GN2. We compare E2GN2 to other common function approximators using common MARL benchmarks MPE and SMACv2. E2GN2 demonstrates a significant improvement in sample efficiency, greater final reward convergence, and a 2x-5x gain in over standard GNNs in our generalization tests. These results pave the way for more reliable and effective solutions in complex multi-agent systems.
+
+# 1 Introduction
+
+Multi-Agent Reinforcement Learning (MARL) has found success in various applications such as robotics [3, 4, 5], complex strategy games [6, 7, 8] and power grid management [9, 10]. However, MARL algorithms can be slow to train, difficult to tune, and have poor generalization guarantees [1, 11]. This is partially because typical implementations of MARL techniques use neural networks such as Multi-Layer Perceptrons (MLP) that do not take the underlying
+
+structure into account. The models learn simple input/output relationships with no constraints or priors on the policies learned. These generic architectures lack a strong inductive bias making them inefficient in terms of the training samples required.
+
+![](images/50fddd78980264d7d91986a86e80e93cebba5981cf4d8abf02f9e9d8e913b4fa.jpg)  
+Figure 1: An example of how using an equivariant function approximator shrinks the total search space.
+
+Symmetries are commonplace in the world. As humans, we exploit symmetries in our daily lives to improve reasoning and learning. It is a basic concept learned by children. Humans do not need to relearn how to eat an apple simply because it has been moved from the right to the left. In soccer, if you have learned to pass to someone on the right, it is easier to learn to pass to someone on the left. Our objective is to develop agents that are guaranteed to adhere to these basic principles, without needing to learn every single scenario from scratch.
+
+![](images/73284adaa8c9088989c01ffc0ae5305bf5b7ca734cae5ea708085067de83f605.jpg)  
+Figure 2: An example of rotational equivariance/symmetry in MPE simple spread environment. Note as the agent (in red) positions are rotated, the optimal actions (arrows) are also rotated.
+
+Symmetries are particularly common in MARL. These occur in the form of equivariance and invariance. Given a transformation matrix  $T$ , if  $f(Tx) = f(x)$  that function is said to be invariant. Similarly,  $f(.)$  is equivariant if  $f(Tx) = Tf(x)$  [12]. Rotational and reflection symmetries (the  $O(n)$  symmetry group) are particularly common in Reinforcement Learning (RL) scenarios. For example, consider the Multi-agent Particle Environment (MPE) [13] benchmark for MARL which has agents with simple dynamics and tasks. We note that this scenario adheres to rotational equivariance. As shown in Figure 2 rotating the agent's positions results in the op
+
+timal actions also being rotated. A policy that is rotationally equivariant effectively shrinks the state-action space for the problem, potentially making the problem easier to learn (Figure 1).
+
+One way to guarantee equivariance to rotations and reflections in MARL is to use an Equivariant Graph Neural Network [2]. Due to the more complex and equivariant structure of the EGNN, directly adapting EGNNs to MARL is not straightforward. Specifically, consider action spaces typically represented as discrete choices (e.g., up, down, left, shoot). Typically the policy over such an action space is represented using logits (to specify the probability of choosing each action). EGNN outputs a continuous equivariant output. Mapping this output to logits is non-trivial. Therefore, we ask the question What is the correct representation of a complex action space when using Equivariant GNNs?
+
+Another specific issue we observed in EGNNs is an exploration bias in the early stages of learning due to the specific nature of the equivariant computation. This bias can lead to poor initial exploration, decreasing the probability of the agent finding the optimal policy, and potentially resulting in suboptimal convergence. Therefore we ask the question How can we design a GNN equivariant to rotations, but without early exploration bias?
+
+The main contribution of this paper is Exploration-enhanced Equivariant Graph Neural Networks (E2GN2). This addresses the two research questions. Specifically, we show how to apply Equivariant GNNs to complex MARL tasks, to mixed discrete/continuous action spaces, and to mitigate the early exploration bias. Our contributions can be summarized as:
+
+(1) Our approach is the first to successfully demonstrate Equivariant GNNs for MARL on standard MARL benchmarks with complex action spaces.  
+(2) We propose E2GN2 which has no bias in early exploration for MARL and is equivariant to rotations and reflections.  
+(3) We evaluate E2GN2 on common MARL benchmarks: MPE [13] and Starcraft Multi-agent Challenge v2 [14] using PPO. It learns quicker, outperforming standard GNNs and MLPs by up to  $2\mathrm{x} - 5\mathrm{x}^2$  in sample efficiency on terran and protoss challenges. It is worth noting that E2GN2 is an improvement on the function approximation for MARL and thus is compatible with most MARL actor-critic methods.  
+(4) We showcase E2GN2's ability to generalize to scenarios it wasn't trained on, due to the equivariance guarantees. This results in 5x performance over GNNs
+
+# 2 Related Works
+
+A key theoretical foundation for our paper is [15], which formulated the structure for equivariant MDPs. One important takeaway is that if a reward is equivariant to a transformation/symmetry, we want the policy and dynamics to be equivariant to that symmetry. However, this work was limited to very simple dynamics with discrete actions such as cart-pole. [16] followed up with equivariance in multi-agent, but was again limited to simple problems. Their method is specifically formulated and tested on small discrete grid world problems with simple dynamics and discrete up, down,left actions. For example, the traffic control problem has an input of a 7x7 grid. Extending their work to continuous environments with large state spaces, large mixed discrete/continuous action spaces is not straightforward without significant modifications.
+
+Contemporary to our research, [17] demonstrated E(3) equivariant networks on simple cooperative navigation problems. However, their results on more complex tasks, such as Starcraft, did not excel over the baseline. Additionally, they used SEGNN [18] which can result in very slow training times, making tuning difficult and cumbersome. Others took the approach [19] of attempting to learn symmetries via an added loss term. However, since this approach needed to learn the symmetries in parallel with learning it did not have the same guarantees as Equivariant GNNs, and did not result in significant performance gains. Another work [20] demonstrated rotation equivariance for complex robotic manipulation tasks. This work, while promising, was for single-agent RL and used image observations and many problems don't have access to image observations.
+
+AI research in chemistry has taken a particular interest in adding symmetry constraints to GNNs. Works such as EGNN, [2] SEGNN, [18] E3NN, [21] and Equivariant transformers have demonstrated various approaches to encoding symmetries in GNNs. In this paper, we chose to focus on EGNN due to its simplicity, high performance, and quick inference time. Other works took a different approach, such as [22], which proposes Equivariant MLPs, solving a constrained optimization problem to encode a variety of symmetries. Unfortunately, in our experience, the inference time was slow, and we preferred a network with a graph structure such as EGNN.
+
+# 3 Background
+
+# 3.1 MARL
+
+Multi-agent reinforcement learning (MARL) considers problems where multiple learning agents interact in a shared environment. The goal is for the agents to learn policies that maximize long-term reward through these interactions. Typically, MARL problems are formalized as Markov games [23]. A Markov game for  $N$  agents is defined by a set of states  $\mathcal{S}$ , a set of actions  $\mathcal{A}_1, \dots, \mathcal{A}_N$  for each agent, transition probabilities  $P: \mathcal{S} \times \mathcal{A}_1 \times \dots \times \mathcal{A}_N \times \mathcal{S} \rightarrow [0,1]$ , and reward functions  $R_1, \dots, R_N$  mapping each state and joint action to a scalar reward.
+
+$$
+\begin{array}{l} \text {T h e g o a l o f e a c h a g e n t i i s t o l a r n a p o l y} \pi_ {i} (a _ {i} | s) \text {t h a t m a x i m i z e s i t s e x p e c t e d r e t u r n :} J (\pi_ {i}) = \\ \mathbb {E} \pi_ {1}, \dots , \pi_ {N} \left[ \sum_ {t = 0} ^ {T} \gamma^ {t} R _ {i} (s _ {t}, a _ {t} ^ {1}, \dots , a _ {t} ^ {N}) \right] \end{array}
+$$
+
+Where  $T$  is the time horizon,  $\gamma \in (0,1]$  is a discount factor, and  $a_{t}^{j} \sim \pi_{j}(\cdot |s_{t})$ . The presence of multiple learning agents makes this more complex than single-agent RL due to issues such as non-stationarity and multi-agent credit assignment.
+
+# 3.2 Equivariance
+
+Crucial to equivariance is group theory. A group is an abstract algebraic object describing a symmetry. For example,  $O(3)$  describes the set of continuous rotation symmetries. A group action is an element of that particular group. To describe how groups o perate on data we use representations of group actions. A representation can be described as a mapping from a group element to a matrix,  $\rho : G \to GL(m)$  where  $\rho(g) \in \mathbb{R}^{m \times m}$ . Or instead we can more simply use:  $L_g : X \to X$  where  $g \in G$  where  $L_g$  is the matrix representation of the group element  $g \in G$  [24].
+
+A function is equivariant to a particular group or symmetry if transforming the input is equivalent to transforming the function output. More formally,  $T_{g}f(x) = f(L_{g}x)$  for  $g\in G$ ,  $L_{g}:X\to X$  and  $T_{g}:Y\to Y$ . Related to equivariance is the key concept of invariance, that is a function does not change with a transformation to the input:  $f(x) = f(L_{g}x)$ . [24]
+
+Previous work [16] has shown that if a Markov game has symmetries in the dynamics and the reward function then the resulting optimal policy will be equivariant and the value function will be invariant. That is,  $V(L_{g}\pmb {s}) = V(s)$  and  $\pi (L_g\pmb {s}) = K_g\pi (\pmb {s})$  where  $L_{g}$  and  $K_{g}$ , with  $g\in G$  are transformations to the state and action respectively.
+
+# 3.3 Equivariant Graph Neural Network
+
+Equivariant Graph Neural Networks [2] (EGNN) are an extension of a standard Graph Neural Network. EGNNs are equivariant to the  $E(n)$  group, that is, rotations, translations, and reflections in Euclidean space. EGNNs have two vectors of embeddings for each graph node  $i$ : the feature embeddings denoted by  $\pmb{h}_i^l$ , and coordinate embeddings denoted by  $\pmb{u}_i^l$ , where  $l$  denotes the neural network layer. The equations describing the forward pass for a single layer are below:
+
+$$
+\boldsymbol {m} _ {i j} = \phi_ {e} \left(\boldsymbol {h} _ {i} ^ {l}, \boldsymbol {h} _ {j} ^ {l}, \| \left(\boldsymbol {u} _ {i} ^ {l} - \boldsymbol {u} _ {j} ^ {l}\right) \| ^ {2}\right) \tag {1}
+$$
+
+$$
+\boldsymbol {u} _ {i} ^ {l + 1} = \boldsymbol {u} _ {i} ^ {l} + C \sum_ {j \neq i} \left(\boldsymbol {u} _ {i} ^ {l} - \boldsymbol {u} _ {j} ^ {l}\right) \phi_ {u} (\boldsymbol {m} _ {i j}) \tag {2}
+$$
+
+$$
+\boldsymbol {m} _ {i} = \sum_ {j \neq i} \boldsymbol {m} _ {i j}, \quad \boldsymbol {h} _ {i} ^ {l + 1} = \phi_ {h} \left(\boldsymbol {h} _ {i} ^ {l}, \boldsymbol {m} _ {i}\right) \tag {3}
+$$
+
+Here  $\phi$  represents a multi-layer perceptron, where  $\phi_e: \mathbb{R}^n \mapsto \mathbb{R}^m$ ,  $\phi_u: \mathbb{R}^m \mapsto \mathbb{R}$ , and  $\phi_h: \mathbb{R}^{m+p} \mapsto \mathbb{R}^p$ . The key difference between EGNN and GNN is the addition of coordinate embeddings  $\boldsymbol{u}_i$  and equation 2, which serves to update the coordinate embeddings in a manner that is equivariant to transformations from  $E(n)$ . Note that  $\boldsymbol{u}_i$  will be equivariant and  $\boldsymbol{h}_i$  will be invariant to these transformations [2].
+
+As noted in Section 1, application of EGNN to MARL is not straightforward. In the following section, we discuss these issues in more depth and present our solution towards addressing them.
+
+# 4 Methods
+
+In this section, we address both theoretically and empirically how the output of EGNN is biased by the input, leading to suboptimal exploration in RL. To mitigate this issue, we introduce Exploration-enhanced Equivariant Graph Neural Networks (E2GN2) as a method that ameliorates this bias, leading to improved exploration.
+
+# 4.1 Biased Exploration in EGNN Policies
+
+An important component of reinforcement learning is exploration. Practitioners often use a policy parameterized by a Gaussian distribution, where the mean is determined by the policy network output, and the standard deviation is a separately learned parameter. Best practices are that the actions initially have a zero mean distribution to get a good sampling of potential state-action trajectories, i.e.,  $\pi (\pmb {a}_i|\pmb {s})\sim N(\pmb {0},\sigma)$ . Below we show that an EGNN will initially have a non-zero mean distribution, which can cause problems in early training.
+
+Theorem 1 Given a layer  $l$  of an EGNN with randomly initialized weights, with the equivariant component input vector  $\pmb{u}_i^l \in \mathbb{R}^n$ , equivariant output vector  $\pmb{u}_i^{l+1} \in \mathbb{R}^n$  and the multilayer perceptron  $\phi_u : \mathbb{R}^m \mapsto \mathbb{R}$ , where the equivariant component is updated as:  $\pmb{u}_i^{l+1} = \pmb{u}_i^l + C\sum_{j \neq i}\left(\pmb{u}_i^l - \pmb{u}_j^l\right)\phi_u(\pmb{m}_{ij})$ . Then the expected value of the output vector is approximately the expected value of the input vector:
+
+$$
+\mathbb {E} \left[ \boldsymbol {u} _ {i} ^ {l + 1} \right] \approx \mathbb {E} \left[ \boldsymbol {u} _ {i} ^ {l} \right]
+$$
+
+Furthermore, given a full EGNN with  $L$  layers then the expected value of the network output is approximately the expected value of the network input
+
+$$
+\mathbb {E} \left[ \boldsymbol {u} _ {i} ^ {L} \right] \approx \mathbb {E} \left[ \boldsymbol {u} _ {i} ^ {0} \right]
+$$
+
+(See appendix A for proof.)
+
+Corollary 1.1 Given a policy for agent  $i$  represented by an EGNN and parameterized with a Gaussian distribution, and the equivariant component of the state  $\pmb{s}_i^{eq}$  the policy will have the following distribution:
+
+$$
+\pi_ {i} (\pmb {a} _ {i} | \pmb {s}) \sim N (\pmb {s} _ {i} ^ {e q}, \sigma)
+$$
+
+(See Appendix A for proof.)
+
+In many cases  $s_i^{eq}$  is the agent's position. Corollary 1.1 indicates that an agent's action distribution will be skewed towards its own position. If the magnitude of the state representation  $s_i^{eq}$  is significantly larger than  $\sigma$ , the agent may output actions approximately equal to its position. We refer to this phenomenon, where agent actions are biased towards replicating the input state as the action, as the early exploration bias. Such bias is not conducive to effective exploration, potentially limiting the agent's ability to discover and converge to optimal solutions.
+
+Examples of early exploration bias Suppose there is an agent placed at the position  $\boldsymbol{u}_i = [2,0]$ , and its objective is to reach the origin (i.e. 0). The agent needs an action of less than 0 to move to the left. Since its action using the EGNN is close to Gaussian distributed with a mean of 2, then  $P(X < 0) \approx 0.02$ . If the agent moves 0.25 with each action, then the probability of the agent reaching the center objective in eight steps is 2.9e-8. On the other hand, the agent has a high likelihood of moving toward infinity as each move away increases the likelihood of moving towards infinity. This example shows that the bias leads to a low probability of finding the solution, which is especially harmful in sparse reward environments.
+
+![](images/fdccb35b4038d09cb1a6b7ff018c01ee599d2e9b01d640954869018e4b0095a3.jpg)  
+Figure 3: An example of biased learning in MPE simple spread environment. Left: We observed the behavior of the EGNN agents in this early training phase. Each agent moved away from the origin due to the EGNN bias. Right: Note the very low reward in early training steps due to the biased policies moving away from the goals.
+
+To confirm our hypothesis on EGNN early exploration bias, we conducted a simple experiment. We trained a Proximal Policy Optimization (PPO) [3] agent on the MPE simple-spread problem. This problem consists of three agents that seek to navigate to three goals cooperatively; the agents are rewarded based on the sum of the minimum distance from each target to an agent. We give further details of this set up and training procedure in the experiments section. Figure 3 shows the results. The EGNN policy biased the agents to move in the direction of their current position, causing them to move away from the origin and receive low rewards. Due to the dense rewards and simplicity of the problem, the EGNN agent was able to overcome this initial
+
+bias and still solve the problem. However, in more complex problems this early bias could cause more problems. Even if an agent finds an optimal trajectory, it will be a relatively small proportion of the sampled trajectories and may result in sub-optimal convergence.
+
+# 4.2 Exploration-Enhanced EGNNs
+
+As discussed previously, EGNN's severe early exploration bias can decrease learning performance. In this section, we propose our solution to this problem in the form of Exploration-enhanced Equivariant Graph Neural Networks (E2GN2). To create E2GN2 we make the following modification to Equation 2 of the equivariant component of EGNN:
+
+$$
+\boldsymbol {u} _ {i} ^ {l + 1} = \boldsymbol {u} _ {i} ^ {l} \phi_ {u _ {2}} (\boldsymbol {m} _ {i}) + C \sum_ {j \neq i} \left(\boldsymbol {u} _ {i} ^ {l} - \boldsymbol {u} _ {j} ^ {l}\right) \phi_ {u} (\boldsymbol {m} _ {i j})) \tag {4}
+$$
+
+where  $\phi_{u_2}(\pmb{m}_i):\mathbb{R}^m\to \mathbb{R}$  is an MLP parameterized by  $u_{2}$ . The remaining layer update equations of the EGNN remain the same.
+
+Theorem 2 Given a  $L$  layer E2GN2 with randomly initialized weights, where the equivariant component is updated as:  $\pmb{u}_i^{l + 1} = \pmb{u}_i^l\phi_{u2}(\pmb {m}_i) + C\sum_{j\neq i}\left(\pmb {u}_i^l -\pmb {u}_j^l\right)\phi_u(\pmb {m}_{ij}))$ . Then the expected value of the output vector is approximately the expected value of the input vector:
+
+$$
+\mathbb {E} \left[ \boldsymbol {u} _ {i} ^ {l + 1} \right] \approx \boldsymbol {0}
+$$
+
+Corollary 2.1 Given a policy for agent  $i$  represented by an E2GN2 and parameterized with a Gaussian distribution the policy will have the following distribution:
+
+$$
+\pi_ {i} (\boldsymbol {a} _ {i} | \boldsymbol {s}) \sim N (\boldsymbol {0}, \sigma)
+$$
+
+Thus we see that E2GN2 should have unbiased early exploration for the equivariant component of the actions. The primary difference between EGNN and E2GN2 is the addition of  $\phi_{u_2}$ , which serves to offset the bias from the previous layer (or input)  $u_i^l$ . To validate that this did indeed solve the exploration bias we tested E2GN2 on the same MPE simple spread environment. We observe in Figure 3 that E2GN2 did indeed have unbiased behavior when it came to early exploration, as the agents had smooth random actions, and the reward did not drastically decrease.
+
+Analysis of E2GN2 Equivariance Here we verify that E2GN2 still retains EGNN's guarantee of equivariance to rotations and reflections.
+
+Theorem 3 E2GN2 is equivariant to transformations from the  $O(n)$  group. In other words, it is equivariant to rotations and reflections.
+
+(See appendix A for proof.)
+
+Retaining our symmetries to rotations and reflections is important, since it should increase our sample efficiency and generalization. Note that E2GN2 is no longer equivariant to translations. However, in MARL translation equivariance is not necessarily a benefit. For example, consider the MPE problem with a translational equivariant policy. If we shift an agent to be 10 units right, this will add 10 to the action as well, causing it to move to the right! Essentially, this is adding an undesirable bias to the policy output:  $\pi (s + b) = \pi (s) + b$  However, we can expect  $O(n)$  policy equivariance to improve our sample efficiency in MARL, and it is key that E2GN2 retains this guarantee.
+
+# 4.3 Adapting Architectures for Complex Action Spaces
+
+Applying EGNN/E2GN2 to sophisticated MARL problems requires careful consideration. Many MARL environments have discrete action spaces or mixed continuous-discrete action spaces. Some components of these action spaces may be invariant and others may be equivariant. Further complicating the problem, MARL requires the neural networks to output parameters for a probability distribution (which is sampled for use in exploration). For continuous actions, agents typically use a gaussian and for discrete actions agents generally use logits. Unfortunately, it is not straightforward to map the distinct invariant and equivariant coordinate embeddings onto a single distribution in a manner that preserves equivariance.
+
+A core benefit of the GNN structure in MARL is the scalability afforded by permutation equivariance: the ability to handle a variable number of agents without retraining. For example, [25] demonstrates using a GNN to train  $N$  agents, then to control  $N + 1$  agents without retraining. However, this example did not operate with discrete or mixed action spaces or equivariant structures as we do here. Improperly mapping GNN outputs to logits risks losing this scalability.
+
+To address these issues, we propose leveraging the GNN's graph structure to output different components of the action space from different nodes in an equivariant manner:
+
+- Discrete Actions & Invariant Components: the invariant feature embeddings  $h_i$  of each agent's node are used to output discrete logits/actions and other invariant components. For action spaces with both equivariant and invariant components, this can be used as the 'action type selector' to select which type of action to apply at that type step (ie move action or targeting action).  
+- Continuous Spatial Actions: the equivariant coordinate embeddings  $u_{i}$  of each agent's node are used for continuous spatial actions such as movement.  
+- Targeting Specific Entities: for multi-agent environments with entities beyond the learning agents (e.g., enemies), logits for discrete actions pertaining to each entity (e.g., which enemy to target) are output from that entity's corresponding node. This enables the discrete action space to scale with the number of agents.
+
+![](images/3b959af40a5388de47a2b077aa1a8e8243c92a5de8bcd0f99d61d39899599d23.jpg)  
+Figure 4: An example of using an Equivariant Graph Neural Network in MARL. Note that the state must be structured as a graph with each node having an equivariant  $u_{i}$  and invariant  $h_i$  component. As discussed in 4.3, the output of the policy uses  $u_{i}$  for equivariant (typically spatial) actions, and the  $h_i$  for invariant components of the actions
+
+Each of these components is parameterized as a distribution (ie logits or gaussian), resulting in one to three separate distributions (depending on the environment and which components are used). After the exploration phase samples from each of these distributions, the final action can be constructed from each of these components via concatenation or the action selection component.
+
+This formulation enables us to retain the permutation equivariance and scalability of a GNN. An alternative to our approach is to add an MLP at the end of the GNN to convert the GNN outputs to the mixed action spaces. This will lose the scalability/locality of the neural network. For example, if you add two more agents how do you modify the final MLP to expand the action space? In our formulation, whenever an new entity  $i$  is added to the environment, we can simply add the node  $N + 1$  to the graph. The action space of the GNN will now be supplemented with  $\boldsymbol{u}_{N + 1}$  and  $h_{N + 1}$ , allowing us to expand the action space without retraining.
+
+By structuring the GNN output in this manner, we can handle discrete or mixed discrete-continuous action spaces while retaining the equivariance of EGNN/E2GN2 and the flexibility of GNNs to a variable number of agents and entities. This approach allows MARL agents based on equivariant GNNs to be applied to challenging environments with minimal restrictions on the action space complexity. Indeed this approach was key to enabling the success of EGNN/E2GN2 on SMACv2.
+
+# 5 Experiments
+
+We seek to answer the following questions with our experiments: (1) Do rotationally equivariant policy networks and rotationally invariant value networks improve training sample efficiency? (2) Does E2GN2 improve learning and performance over EGNN? (3) Does Equivariance indeed demonstrate improved generalization performance?
+
+To address these questions, we use common MARL benchmarks: the multi-agent particle-world environment (MPE) [13] and Starcraft Multi-agent Challenge version 2 (SMACv2) [14]. Our experiments show that equivariance does indeed lead to improved sample efficiency, with E2GN2 in particular performing especially well. We also demonstrate that generalization is guaranteed across rotational transformations.
+
+We want to focus our experiments on the neural networks' effects on MARL performance. To isolate the impact of the network architecture, we avoid using specialized tips and tricks sometimes employed in MARL [26]. This allows us to demonstrate that our proposed networks can improve performance
+
+without relying on these additional techniques. Thus we use a common standardized open source MARL training library RLlib [27]. We use the default multi-agent PPO algorithm, which does not use a centralized critic. We followed the basic hyperparameter tuning guidelines set forth in [26]. That is, we use a large training batch and mini-batch size, low numbers of SGD iterations, and a small clip rate. Further hyperparameter details are found in appendix B.
+
+We compare our results with common neural networks used for MARL benchmarks: multi-layer perceptrons (MLP), and GNNs (we use a GNN structure similar to equations 1, 3, see appendix B for details). We also compare with the approach from [17] which we refer to as E3AC in our plots. Recall that E3AC also uses neural networks that guarantee equivariance: SEGNN. We integrated E3AC into RLLIB to ensure the RL training procedure remained consistent across our comparisons. Note that the majority of MARL papers on SMAC/SMACv2 use MLPs as the base network. We also compare with GNNs to show the improvement is not primarily due to the graph structure. For the main paper results, we assume full observability since we have not explicitly tackled the partial observability problem yet. In future work, we will seek to remedy this. However, to be thorough we performed experiments with partial observability, resulting in surprising success using E2GN2 (see appendix C).
+
+# 5.1 Training Environments
+
+From the MPE environment ([13]), we use two environments to benchmark our performance: cooperative navigation also known as spread, and predator-prey, also known as tag. The MPE tag environment consists of three RL-controlled agents seeking to catch a third evader agent. For easier comparison, we use a simple heuristic algorithm to control the evader agent. The evader simply moves in the opposite direction of the pursuers. The other MPE environment is the cooperative navigation or simple spread. In this environment, each agent seeks to minimize the distance between all obstacles. To better test equivariance, for the MPE environments we use continuous actions instead of the default discrete actions.
+
+Next we test on SMACv2, a much more difficult environment than MPE. In SMACv2, the units are heterogeneous with different capabilities (with different attack ranges, total health, and sometimes action spaces). The unit types are randomized at the beginning of the scenario. The actions include more components than simply movement (such as in MPE), agents can move and attack. The goals are more complex as well. Instead of simply navigating cooperatively as in MPE, the agents must learn attack formations and strategies. Sometimes it may be optimal to sacrifice health or allies in the purpose of the greater strategic objective. SMACv2 has three basic scenarios defined by the unit types: terran, protoss, and zerg. We use 5 agents for each team, and the initial position configuration termed "surrounded" (see fig 7)
+
+The SMACv2 action space poses an interesting problem for GNN structures. A key advantage of GNNs is the permutation equivariance, which leads to scalability without retraining. The default SMACv2 agents will output simply a discrete action determining movement or target for shooting. For our purposes, we modify the action space to be a mixed action space. This consists of a continuous vector for movement actions, a discrete action determining the attack target, and a boolean determining if the agent should shoot, move or no-op (to be complete, we include plots with a discrete action space in appendix C). As discussed in section 4.3 this will both prove a greater test for our algorithms and allow for the GNN to scale to larger numbers of agents.
+
+![](images/f2af18b7adffaa99f72f12c114eee2474d0156ec9a93806866def7dbae5e1ec1.jpg)
+
+![](images/189cf80ce962c98328f4d6e11780a98e6c2b204a901b482ff7150a080560f8a7.jpg)
+
+![](images/911e7957cf25e418a7a65a5ca5c35cff089575f88c5dd1f4c22e610a596945b5.jpg)  
+Figure 5: Comparing PPO learning performance on MPE with various Neural Networks. (TOP) reward as a function of environment steps. We show the standard errors computed across 10 seeds. (BOTTOM) reward as a function of wall clock time
+
+![](images/47d65237761f8847bfd9a02fe8a7062fc986efbe9eb5b1818ddebae2f909c5d6.jpg)
+
+![](images/65bd5f379268eea3272b7cbb89b82c54c637ebc81237307fc7346ceee2eb7b19.jpg)  
+Figure 6: Comparing performance of PPO on SMACv2 with various Neural Networks representing the policy and value function. Each chart represents a different race from the SMACv2 environment. We show the standard errors computed across 10 seeds.
+
+![](images/b1dff64f745503d55388532e21d9149ee216413ae1b5ddd8bbcdfe75a3142bd1.jpg)
+
+![](images/1a7522e84a9d9d7ed5634f3f3c3afd83ff2e8cd45c92cace8063d3c9d3fca7dd.jpg)
+
+![](images/c6910acbe5cbd9bdae86da6d641ce2ed2d034658593af71867ea6f723d753f30.jpg)
+
+![](images/8113298ebf5d9bf6a20ff7ae3a3fafd44ef726885769a5d620bd9f9f6e940af2.jpg)
+
+# 5.2 Training Results
+
+We present the results from our training in this section. The results for MPE in Figure 5 are averaged across 10 seeds. As discussed previously, EGNN has poor early performance due to the early exploration bias. Despite this poor exploration, EGNN outperforms GNNs and MLPs, demonstrating the power of equivariance in MARL. Similarly, we note that in MPE tag, E2GN2 results in a strong final agent, while EGNN suffers in the initial training phases. We see that E2GN2 is able to greatly outperform EGNN's final reward, partly due to superior early exploration.
+
+In figure 5 we also compare the wall clock time required to train each algorithm. Note that these were all trained on the same hardware and using the same training algorithm (RLLIBs PPO). E3AC remains competitive in sample efficiency on MPE spread, but it requires nearly four hours to gather one million time steps, compared to less than one hour for E2GN2. This is likely due to the SEGNN [18] underlying structure used by E3AC, which is slower due to the be slower for inference time. We also note that E2GN2 does manage to outperform E3AC on MPE Tag.
+
+Next, we review the results from SMACv2 in Figure 6; these results are averaged across 10 seeds. The equivariant networks have a clear improvement in sample efficiency over the MLP and GNN. On the terran environment, EGNN once again learns slower in the initial phases, but due to equivariance, it outperforms MLP/GNN. For protoss, we note that EGNN performs well but has a high variance for its performance. E3AC struggles to perform well on SMACv2, likely since it was unable to resolve the hurdle we addressed in section 4.3. Instead for SMACv2 E3AC used a simple MLP for the policy and SEGNN for the value function [17]. The training speed and performance for E2GN2 is especially impressive since this environment is much more complex than the MPE.
+
+# 5.3 Generalization
+
+Thus far we have demonstrated that equivariance does indeed lead to improved sample efficiency. The next question to answer is whether equivariance leads to improved generalization. We test this by using the 'surrounded' initial configuration from SMACv2 shown in figure 7. By default, the agents are randomly generated in all directions (along specific axes). To test generalization, we only initialize agents on the left side
+
+of the map (i.e., surrounded left). We then test to see if the agents are able to generalize when they are tested with initial positions starting on the right (called surrounded right), and with the default surrounded configuration (termed Surrounded All or Surrounded). Theoretically, the equivariant network should see equivalent performance between the training and testing initialization, due to the guarantee of rotational equivariance for movement, and invariance for shooting actions.
+
+The results of our tests are shown in Table 1. As expected E2GN2 has stellar generalization performance. The win rate remains the same from the training configuration (Surrounded Left) to Surrounded Right. We see an increase in win rate when testing in Surrounded All. This is likely because this Surrounded All is an easier scenario; the enemy is divided into more groups, so the agents can defeat the smaller groups one at a time. As we mentioned in the abstract, this table
+
+![](images/9667d625678f01432cdf3031a65234ece9be8e311d06ea668e0de83cc33564a8.jpg)  
+Figure 7: SMACv2 initialization schemes used for testing generalization
+
+![](images/ddaf54ea1447a50f70ce645fceac067989f8eb0aa0009ab51cda7f8961cf32e8.jpg)
+
+![](images/a383cdb4c14213d3a22a3147c58660343655fe7a131ba9154cb788d0e293b7f7.jpg)
+
+Table 1: Generalization Win Rate in SMACv2. Note that E2GN2 retains high performance, while GNN and MLP lose performance when generalizing  
+
+<table><tr><td>Environment</td><td>Network</td><td>Training Initialization Surrounded Left</td><td>Testing Initialization Surrounded Right</td><td>Surrounded All</td></tr><tr><td rowspan="3">Terran</td><td>E2GN2</td><td>0.57 ±.01</td><td>0.55 ±.01</td><td>0.63 ±.01</td></tr><tr><td>GNN</td><td>0.51 ±.02</td><td>0.11 ±.02</td><td>0.32 ±.02</td></tr><tr><td>MLP</td><td>0.33 ±.02</td><td>0.12 ±.02</td><td>0.24 ±.02</td></tr><tr><td rowspan="3">Protoss</td><td>E2GN2</td><td>0.59 ±.01</td><td>0.56 ±.02</td><td>0.57 ±.02</td></tr><tr><td>GNN</td><td>0.5 ±.02</td><td>0.14 ±.01</td><td>0.32 ±.01</td></tr><tr><td>MLP</td><td>0.42 ±.02</td><td>0.17 ±.02</td><td>0.27 ±.02</td></tr><tr><td rowspan="3">Zerg</td><td>E2GN2</td><td>0.35 ±.02</td><td>0.32 ±.02</td><td>0.29 ±.02</td></tr><tr><td>GNN</td><td>0.4 ±.02</td><td>0.07 ±.01</td><td>0.21 ±.01</td></tr><tr><td>MLP</td><td>0.21 ±.02</td><td>0.05 ±.01</td><td>0.12 ±.01</td></tr></table>
+
+indicates that for the Surrounded Right generalization test E2GN2 has  $\sim 5\mathrm{x}$  the performance of GNN and MLP.
+
+Next, in Table 2 we test E2GN2's ability to scale to various numbers of agents without retraining. We use the agents trained in SMACv2 from Figure 6 for these tests. We do not include the MLP agent since it cannot scale up without retraining. Note that the method described in section 4.3 was essential to enable scaling between agents. For example, if we passed the GNN output to a final MLP to map to the discrete action space this wouldn't be scalable to more agents.
+
+The results for scalability confirm that both E2GN2 and GNN are able to maintain performance with larger numbers of agents. As the number of agents gradually increases the win rate does slowly decrease. E2GN2 does seem to have a slightly steeper loss in performance in the Zerg domain, which was also the most difficult domain for the initial training phase. We believe this is due to the higher complexity of the Zerg scenario, with certain unit types (banelings) that can explode to damage many other agents.
+
+# 6 Conclusion and Future Work
+
+In this paper, we have demonstrated that E2GN2 merits strong consideration for many MARL applications. We addressed several important problems including the early exploration bias and how to apply it to complex action spaces. The sample-efficiency has dramatically improved, and there are now guarantees of generalization built into the network. There is still further work to do, such as handling the partial or incomplete symmetries. Furthermore this approach would need to be improved further to be applied to environments with angular momentum and accelerations, which are not included in these benchmarks. Currently, we expect that the improvement gained from E2GN2 in MARL will depend on the amount of rotational symmetry applicable in the observations. We believe that building on this work can be helpful to deploying MARL agents with a greater degree of trust (due to the guarantees). This could be helpful in many fields such as robotics, medicine, and power systems. In summation, the results of this paper provide a solid foundation upon which to build.
+
+Table 2: Generalization Win Rate: Testing RL agents ability to scale to different numbers of agents (originally trained with 5 agents)  
+
+<table><tr><td rowspan="2">Environment</td><td rowspan="2">Network</td><td rowspan="2">Training Setup 5 Agents</td><td rowspan="2">4 Agents</td><td colspan="3">Testing Setup</td></tr><tr><td>6 Agents</td><td>7 Agents</td><td>8 Agents</td></tr><tr><td rowspan="2">Terran</td><td>E2GN2</td><td>0.69 ± .02</td><td>0.65 ± .02</td><td>0.63 ± .02</td><td>0.62 ± .02</td><td>0.54 ± .04</td></tr><tr><td>GNN</td><td>0.45 ± .01</td><td>0.44 ± .01</td><td>0.42 ± .02</td><td>0.39 ± .01</td><td>0.32 ± .02</td></tr><tr><td rowspan="2">Protoss</td><td>E2GN2</td><td>0.62 ± .03</td><td>0.61 ± .02</td><td>0.59 ± .03</td><td>0.47 ± .04</td><td>0.37 ± .03</td></tr><tr><td>GNN</td><td>0.39 ± .02</td><td>0.37 ± .01</td><td>0.36 ± .03</td><td>0.30 ± .02</td><td>0.21 ± .02</td></tr><tr><td rowspan="2">Zerg</td><td>E2GN2</td><td>0.36 ± .03</td><td>0.32 ± .03</td><td>0.31 ± .01</td><td>0.23 ± .01</td><td>0.18 ± .03</td></tr><tr><td>GNN</td><td>0.28 ± .04</td><td>0.28 ± .02</td><td>0.25 ± .03</td><td>0.2 ± .02</td><td>0.15 ± .02</td></tr></table>
+
+# References
+
+[1] Kaiqing Zhang, Zhuoran Yang, and Tamer Başar. Multi-Agent Reinforcement Learning: A Selective Overview of Theories and Algorithms, pages 321-384. Springer International Publishing, Cham, 2021.  
+[2] Victor Garcia Satorras, Emiel Hoogeboom, and Max Welling. E(n) equivariant graph neural networks, 2 2021.  
+[3] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. CoRR, abs/1707.06347, 2017.  
+[4] Shai Shalev-Shwartz, Shaked Shammah, and Amnon Shashua. Safe, multi-agent, reinforcement learning for autonomous driving. ArXiv, abs/1610.03295, 2016.  
+[5] Pinxin Long, Tingxiang Fan, Xinyi Liao, Wenxi Liu, Hao Zhang, and Jia Pan. Towards optimally decentralized multi-robot collision avoidance via deep reinforcement learning. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 6252–6259, 2018.  
+[6] O. Vinyals, I. Babuschkin, and W.M. et al Czarnecki. Grandmaster level in starcraft ii using multi-agent reinforcement learning. Nature, 575, 2019.  
+[7] David Silver, Thomas Hubert, Julian Schrittwieser, Ioannis Antonoglou, Matthew Lai, Arthur Guez, Marc Lanctot, Laurent Sifre, Dharshan Kumaran, Thore Graepel, Timothy Lillicrap, Karen Simonyan, and Demis Hassabis. A general reinforcement learning algorithm that masters chess, shogi, and go through self-play. Science, 362(6419):1140-1144, 2018.  
+[8] Julian Schrittwieser, Ioannis Antonoglou, Thomas Hubert, Karen Simonyan, L. Sifre, Simon Schmitt, Arthur Guez, Edward Lockhart, Demis Hassabis, Thore Graepel, Timothy P. Lillicrap, and David Silver. Mastering atari, go, chess and shogi by planning with a learned model. Nature, 588:604 - 609, 2019.  
+[9] Esmat Samadi, Ali Badri, and Reza Ebrahimpour. Decentralized multi-agent based energy management of microgrid using reinforcement learning. International Journal of Electrical Power & Energy Systems, 122:106211, 2020.  
+[10] Ziming Yan and Yan Xu. A multi-agent deep reinforcement learning method for cooperative load frequency control of a multi-area power system. IEEE Transactions on Power Systems, 35(6):4599-4608, 2020.  
+[11] Karl Cobbe, Oleg Klimov, Chris Hesse, Taehoon Kim, and John Schulman. Quantifying generalization in reinforcement learning. In Kamalika Chaudhuri and Ruslan Salakhutdinov, editors, Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 1282–1289. PMLR, 09–15 Jun 2019.  
+[12] Andrew M. Pitts. Nominal Sets: Names and Symmetry in Computer Science. Cambridge Tracts in Theoretical Computer Science. Cambridge University Press, 2013.  
+[13] Ryan Lowe, YI WU, Aviv Tamar, Jean Harb, OpenAI Pieter Abbeel, and Igor Mordatch. Multi-agent actor-critic for mixed cooperative-competitive environments. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017.  
+[14] Benjamin Ellis, Jonathan Cook, Skander Moalla, Mikayel Samvelyan, Mingfei Sun, Anuj Mahajan, Jakob N. Foerster, and Shimon Whiteson. Smacv2: An improved benchmark for cooperative multi-agent reinforcement learning, 2023.  
+[15] Elise van der Pol, Daniel E. Worrall, Herke van Hoof, Frans A. Oliehoek, and Max Welling. Mdp homomorphic networks: Group symmetries in reinforcement learning. Advances in Neural Information Processing Systems, 2020-December, 6 2020.  
+[16] Elise Van Der Pol, Herke Van Hoof, Uva-Bosch Deltalab, Frans A Oliehoek, and Max Welling. Multi-agent mdp homomorphic networks, 10 2021.
+
+[17] Dingyang Chen and Qi Zhang. e(3)-equivariant actor-critic methods for cooperative multi-agent reinforcement learning, 2024.  
+[18] Johannes Brandstetter, Rob Hesselink, Elise van der Pol, Erik J. Bekkers, and Max Welling. Geometric and physical quantities improve E(3) equivariant message passing. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022.  
+[19] Xin Yu, Rongye Shi, Pu Feng, Yongkai Tian, Simin Li, Shuhao Liao, and Wenjun Wu. Leveraging partial symmetry for multi-agent reinforcement learning. Proceedings of the AAAI Conference on Artificial Intelligence, 38(16):17583-17590, Mar. 2024.  
+[20] Dian Wang, Robin Walters, and Robert Platt.  $\{SO\}(2)$ -equivariant reinforcement learning, 3 2022.  
+[21] Mario Geiger and Tess Smidt. e3nn: Euclidean neural networks, 7 2022.  
+[22] Marc Finzi, Max Welling, and Andrew Gordon Wilson. A practical method for constructing equivariant multilayer perceptrons for arbitrary matrix groups, 2021.  
+[23] Michael L. Littman. Markov games as a framework for multi-agent reinforcement learning. Morgan Kaufmann, San Francisco (CA), 1994.  
+[24] Michael M. Bronstein, Joan Bruna, Taco Cohen, and Petar Velicković. Geometric deep learning: Grids, groups, graphs, geodesics, and gauges, 2021.  
+[25] Akshat Agarwal, Sumit Kumar, and Katia P. Sycara. Learning transferable cooperative behavior in multi-agent teams. CoRR, abs/1906.01202, 2019.  
+[26] Chao Yu, Akash Velu, Eugene Vinitsky, Jiaxuan Gao, Yu Wang, Alexandre Bayen, and Yi Wu. The surprising effectiveness of ppo in cooperative, multi-agent games, 2022.  
+[27] Eric Liang, Richard Liaw, Robert Nishihara, Philipp Moritz, Roy Fox, Ken Goldberg, Joseph E. Gonzalez, Michael I. Jordan, and Ion Stoica. RLlib: Abstractions for distributed reinforcement learning. In International Conference on Machine Learning (ICML), 2018.  
+[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In 2015 IEEE International Conference on Computer Vision (ICCV), pages 1026-1034, 2015.  
+[29] Mikayel Samvelyan, Tabish Rashid, Christian Schroeder de Witt, Gregory Farquhar, Nantas Nardelli, Tim G. J. Rudner, Chia-Man Hung, Philip H. S. Torr, Jakob Foerster, and Shimon Whiteson. The starcraft multi-agent challenge, 2019.
+
+# Appendix and Supplemental Material
+
+# A Proofs
+
+# A.1 EGNN Bias Proof
+
+Here we prove our assertion that the EGNN layers have biased outputs, and thus poor exploration.
+
+We start by showing that  $\mathbb{E}\left[\phi_u(m_{ij})\right] \approx 0$ . We do this by computing the expected value of a single neural network layer.
+
+An individual neural network layer is defined as  $\pmb{y} = W\pmb{x} + \pmb{b}$  where  $\pmb{b}$  is a N-by-1 vector,  $\pmb{W}$  is a N-by-M matrix and  $\pmb{x}$  is a M-by-1 vector. Note that similar to [28] we initialize the elements in  $\pmb{W}$  to be mutually independent and share the same distribution.  $\pmb{x}$  is initialized similarly. and  $\pmb{x}$  and  $\pmb{W}$  are independent. Thus for each element (subscript  $l$ ) of the equation, we have:
+
+$$
+\mathbb {E} \left[ y _ {l} \right] = \sum \mathbb {E} \left[ w _ {l} x _ {l} \right] + \mathbb {E} \left[ b _ {l} \right] = n _ {l} \mathbb {E} \left[ w _ {l} x _ {l} \right] + \mathbb {E} \left[ b _ {l} \right] = n _ {l} \mathbb {E} \left[ w _ {l} \right] \mathbb {E} \left[ x _ {l} \right] + \mathbb {E} \left[ b _ {l} \right]
+$$
+
+We can expect  $w_{l}$  and  $b_{l}$  to be zero mean, since many commonly used initializers use either uniform, or normal distribution, with zero mean [28]. This results in:
+
+$$
+= n _ {l} * 0 * \mathbb {E} \left[ x _ {l} \right] + 0 = 0
+$$
+
+Thus, we can assume that for an individual layer the output is 0, since the final layer of any mlp is 0, we can see that  $\mathbb{E}\left[\phi_u(m_{ij})\right] \approx 0$
+
+Next we show that  $\mathbb{E}\left[\boldsymbol{u}_i^L\right] \approx \mathbb{E}\left[\boldsymbol{u}_i^0\right]$ . Recall that  $\boldsymbol{u}_i^L$  is the  $L$ th layer of  $i$ th node of the equivariant component of the network. We take the expectation over a sampling of the inputs  $u_i$  and  $h_i$ , treating each as a random variable. Note that  $u_i$  and  $u_j$  will have identical distributions, since each node will be sampled in the same manner. We start our proof by taking the expected value of the output:
+
+$$
+\begin{array}{l} \mathbb {E} \left[ \boldsymbol {u} _ {i} ^ {l + 1} \right] = \mathbb {E} \left[ \boldsymbol {u} _ {i} ^ {l} + C \sum_ {j \neq i} \left(\boldsymbol {u} _ {i} ^ {l} - \boldsymbol {u} _ {j} ^ {l}\right) \phi_ {x} (\boldsymbol {m} _ {i j}) \right] = \mathbb {E} \left[ \boldsymbol {u} _ {i} ^ {l} \right] + C \sum_ {j \neq i} \mathbb {E} \left[ \left(\boldsymbol {u} _ {i} ^ {l} - \boldsymbol {u} _ {j} ^ {l}\right) \phi_ {x} (\boldsymbol {m} _ {i j}) \right] \\ = \mathbb {E} \left[ \boldsymbol {u} _ {i} ^ {l} \right] + C \sum_ {j \neq i} \mathbb {E} \left[ \boldsymbol {u} _ {i} ^ {l} \phi_ {u} \left(\boldsymbol {m} _ {i j}\right) \right] - \mathbb {E} \left[ \boldsymbol {u} _ {j} ^ {l} \phi_ {u} \left(\boldsymbol {m} _ {i j}\right) \right] \\ \approx \mathbb {E} \left[ \boldsymbol {u} _ {i} ^ {l} \right] + C \sum_ {j \neq i} \mathbb {E} \left[ \boldsymbol {u} _ {i} ^ {0} \right] \mathbb {E} \left[ \phi_ {u} \left(\boldsymbol {m} _ {i j}\right) \right] - \mathbb {E} \left[ \boldsymbol {u} _ {j} ^ {0} \right] \mathbb {E} \left[ \phi_ {u} \left(\boldsymbol {m} _ {i j}\right) \right] \\ \approx \mathbb {E} \left[ \boldsymbol {u} _ {i} ^ {l} \right] + C \sum_ {j \neq i} \boldsymbol {u} _ {i} ^ {l} 0 - \boldsymbol {u} _ {j} ^ {l} 0 = \boldsymbol {u} _ {i} ^ {l} \\ \mathbb {E} \left[ \boldsymbol {u} _ {i} ^ {l + 1} \right] \approx \mathbb {E} \left[ \boldsymbol {u} _ {i} ^ {l} \right] \\ \end{array}
+$$
+
+Thus,
+
+$$
+\mathbb {E} \left[ \boldsymbol {u} _ {i} ^ {L} \right] \dots \approx \mathbb {E} \left[ \boldsymbol {u} _ {i} ^ {1} \right] \approx \mathbb {E} \left[ \boldsymbol {u} _ {i} ^ {0} \right]
+$$
+
+Where we used the assumption that  $\mathbb{E}\left[\pmb{u}_j^l\phi_u(\pmb{m}_{ij})\right] \approx \mathbb{E}\left[\pmb{u}_j^l\right]\mathbb{E}\left[\phi_u(\pmb{m}_{ij})\right]$
+
+The corollary:  $\pi_i(\pmb{a}_i|\pmb{s})\sim N(\pmb{s}_i^{eq},\sigma)$  is a simple result from the fact that the equivariant action for an agent is the output of the coordinate embedding of the EGNN:  $u_{i}^{L}$ . Here we assume the standard deviation is a separately trained parameter and not a function of the EGNN. If this is the case, then the policy mean  $E[\pi (\pmb {a}_i|\pmb {s})] = \mathbb{E}\left[\pmb {u}_i^L\right]\approx \mathbb{E}\left[\pmb {u}_i^0\right] = \mathbb{E}\left[\pmb {s}_i^{eq}\right]$  (by definition)
+
+# A.2 Proof E2GN2 is Unbiased
+
+Here we show that E2GN2 leads to unbiased early exploration. We begin with the equation for the E2GN2 layer:
+
+$$
+\boldsymbol {u} _ {i} ^ {l + 1} = \boldsymbol {u} _ {i} ^ {l} \phi_ {u 2} (\boldsymbol {m} _ {i j}) + C \sum_ {j \neq i} \left(\boldsymbol {u} _ {i} ^ {l} - \boldsymbol {u} _ {j} ^ {l}\right) \phi_ {u} (\boldsymbol {m} _ {i j}))
+$$
+
+Next we take the expected value of the output:
+
+$$
+\begin{array}{l} E \left[ \boldsymbol {u} _ {i} ^ {l + 1} \right] = \mathbb {E} \left[ \boldsymbol {u} _ {i} ^ {l + 1} \phi_ {u 2} (\boldsymbol {m} _ {i j}) \right] + C \sum_ {j \neq i} \mathbb {E} \left[ \boldsymbol {u} _ {i} ^ {l} \phi_ {u} (\boldsymbol {m} _ {i j}) \right] - \mathbb {E} \left[ \boldsymbol {u} _ {j} ^ {l} \phi_ {u} (\boldsymbol {m} _ {i j}) \right] \\ \approx \mathbb {E} \left[ \boldsymbol {u} _ {i} ^ {l + 1} \right] \mathbb {E} \left[ \phi_ {u 2} (\boldsymbol {m} _ {i j}) \right] + C \sum_ {j \neq i} \mathbb {E} \left[ \boldsymbol {u} _ {i} ^ {l} \right] \mathbb {E} \left[ \phi_ {u} (\boldsymbol {m} _ {i j}) \right] - \mathbb {E} \left[ \boldsymbol {u} _ {j} ^ {l} \right] \mathbb {E} \left[ \phi_ {u} (\boldsymbol {m} _ {i j}) \right] \\ \approx \mathbb {E} \left[ \boldsymbol {u} _ {i} ^ {l + 1} \right] \boldsymbol {0} + C \sum_ {j \neq i} \mathbb {E} \left[ \boldsymbol {u} _ {i} ^ {l} \right] \boldsymbol {0} - \mathbb {E} \left[ \boldsymbol {u} _ {j} ^ {l} \right] \boldsymbol {0} = \boldsymbol {0} \\ \end{array}
+$$
+
+Thus we see that for each layer  $\mathbb{E}\left[\boldsymbol{u}_i^{l + 1}\right] \approx \mathbf{0}$  and therefore  $\mathbb{E}\left[\boldsymbol{u}_i^L\right] \approx \mathbf{0}$ . Similar to before, we used the assumption that Where we used the assumption that  $\mathbb{E}\left[\boldsymbol{u}_j^l\phi_u(\boldsymbol{m}_{ij})\right] \approx \mathbb{E}\left[\boldsymbol{u}_j^l\right]\mathbb{E}\left[\phi_u(\boldsymbol{m}_{ij})\right]$  and  $\mathbb{E}\left[\boldsymbol{u}_j^l\phi_{u2}(\boldsymbol{m}_{ij})\right] \approx \mathbb{E}\left[\boldsymbol{u}_j^l\right]\mathbb{E}\left[\phi_{u2}(\boldsymbol{m}_{ij})\right]$
+
+Similar to the first corollary the result:  $\pi_i(\pmb{a}_i|\pmb{s})\sim N(\pmb {0},\sigma)$  follows mostly from the definition and the above proof. The corollary  $\pi_i(\pmb {a}_i|\pmb {s})\sim N(\pmb {0},\sigma)$  is a simple result from the fact that the equivariant action for an agent is the output of the coordinate embedding of the EGNN:  $u_{i}^{L}$ . Here we assume the standard deviation is a separately trained parameter and not a function of the EGNN. If this is the case, then the policy mean  $E[\pi (\pmb {a}_i|\pmb {s})] = \mathbb{E}\left[\pmb {u}_i^L\right]\approx \mathbf{0}$  (by definition)
+
+# A.3 Equivariance and Invariance of E2GN2
+
+We follow the structure of [2] for how to show equivariance. Note that [2] showed that  $\phi_u(m_{ij})$  will be invariant to  $E(n)$  transformations. This should still hold in our case as we make no modifications to  $m_{ij}$ , similarly  $\phi_{u2}(m_{ij})$  will be invariant to translations. To show equivariance to rotations and reflections we show that applying a transformation  $T$  to the input, results in a transformation to the output. That is:  $f(T\pmb{u}_i^l) = T\pmb{u}_i^l$  where  $f(.)$  is the update equation of E2GN2:
+
+$$
+\begin{array}{l} \boldsymbol {T} \boldsymbol {u} _ {i} ^ {l} \phi_ {u 2} (\boldsymbol {m} _ {i}) + C \sum_ {j \neq i} \left(\boldsymbol {T} \boldsymbol {u} _ {i} ^ {l} - \boldsymbol {T} \boldsymbol {u} _ {j} ^ {l}\right) \phi_ {u} (\boldsymbol {m} _ {i j}) \\ = \boldsymbol {T} \boldsymbol {u} _ {i} ^ {l} \phi_ {u 2} (\boldsymbol {m} _ {i}) + T C \sum_ {j \neq i} \left(\boldsymbol {u} _ {i} ^ {l} - \boldsymbol {u} _ {j} ^ {l}\right) \phi_ {u} \left(\boldsymbol {m} _ {i j}\right) \\ = \boldsymbol {T} \boldsymbol {u} _ {i} ^ {l} \phi_ {u 2} (\boldsymbol {m} _ {i}) + T C \sum_ {j \neq i} \left(\boldsymbol {u} _ {i} ^ {l} - \boldsymbol {u} _ {j} ^ {l}\right) \phi_ {u} \left(\boldsymbol {m} _ {i j}\right) \\ = \boldsymbol {T} \left(\boldsymbol {u} _ {i} ^ {l} \phi_ {u 2} (\boldsymbol {m} _ {i}) + C \sum_ {j \neq i} \left(\boldsymbol {u} _ {i} ^ {l} - \boldsymbol {u} _ {j} ^ {l}\right) \phi_ {u} (\boldsymbol {m} _ {i j})\right) = \boldsymbol {T} \boldsymbol {u} _ {i} ^ {l + 1} \\ \end{array}
+$$
+
+Since there are no modifications to E2GN2 update equation for  $h_i$ , then E2GN2 should retain invariance to transformations from  $E(n)$ . More precisely (using the fact that the initial  $m_{ij}$  is invariant to translations from [2]), if we translate the input  $u_i$  by  $b$ :
+
+$$
+\begin{array}{l} \left(\boldsymbol {u} _ {i} ^ {l} + \boldsymbol {b}\right) \phi_ {u 2} (\boldsymbol {m} _ {i}) + C \sum_ {j \neq i} \left(\left(\boldsymbol {u} _ {i} ^ {l} + \boldsymbol {b}\right) - \left(\boldsymbol {u} _ {j} ^ {l} + \boldsymbol {b}\right)\right) \phi_ {u} \left(\boldsymbol {m} _ {i j}\right) \\ = \boldsymbol {b} \phi_ {u 2} (\boldsymbol {m} _ {i}) + \boldsymbol {u} _ {i} ^ {l} \phi_ {u 2} (\boldsymbol {m} _ {i}) + C \sum_ {j \neq i} \left(\boldsymbol {u} _ {i} ^ {l} - \boldsymbol {u} _ {j} ^ {l}\right) \phi_ {u} (\boldsymbol {m} _ {i j}) = \boldsymbol {b} \phi_ {u 2} (\boldsymbol {m} _ {i}) + \boldsymbol {u} _ {i} ^ {l + 1} \neq \boldsymbol {b} + \boldsymbol {u} _ {i} ^ {l + 1} \\ \end{array}
+$$
+
+This shows that E2GN2 is not translation equivariant.
+
+# B Additional Training Details
+
+<table><tr><td>Hyperparameters</td><td>value</td></tr><tr><td>train batch size</td><td>8000</td></tr><tr><td>mini-batch size</td><td>2000</td></tr><tr><td>PPO clip</td><td>0.1</td></tr><tr><td>learning rate</td><td>25e-5</td></tr><tr><td>num SGD iterations</td><td>15</td></tr><tr><td>gamma</td><td>0.99</td></tr><tr><td>lambda</td><td>0.95</td></tr></table>
+
+Table 3: PPO Hyperparameters for SMACv2  
+
+<table><tr><td>Hyperparameters</td><td>value</td></tr><tr><td>train batch size</td><td>2000</td></tr><tr><td>mini-batch size</td><td>1000</td></tr><tr><td>PPO clip</td><td>0.2</td></tr><tr><td>learning rate</td><td>30e-5</td></tr><tr><td>num SGD iterations</td><td>10</td></tr><tr><td>gamma</td><td>0.99</td></tr><tr><td>lambda</td><td>0.95</td></tr></table>
+
+Table 4: hyperparameters for MPE
+
+MLP uses two hidden layers of 64 width each for MPE and 128 for SMACv2. All MLPs in the GNNs use 2 layers with a width of 32. For both MPE and GNN structures we use separate networks for the policy and value functions.
+
+Graph Structure The graph structure for MPE environments is set as a complete graph. The graph structure for SMACv2 is complete among the agent controlled units (or all of the friendly nodes), and each friendly node also has an edge to each of the enemy controlled units. This was an attempt to model what we imagine a real world scenario to look like (agents see the enemies, communicate with allies, but do not see the enemies communication signals)
+
+Graph Inputs For MPE environments the input invariant feature for each node  $h_i^0$  is the id (pursuer, evader, or landmark). For SMACv2,  $h_i^0$  is made up of the features: health, shield unit type, and team. For each  $u_i^0$  is set as the position of node  $i$ . For MPE there is also a velocity feature, which we incorporate following the procedure described in [2]
+
+Graph Outputs for Value function Similar to the visual in section 4.3, the value function output comes from the invariant component of the agent's node of final layer of the EGNN/E2GN2. In other words the value function is:  $h_i^L$  where  $L$  is the final layer, and  $i$  is the agent in making the decision.
+
+Graph Outputs for Policy For MPE environments the actions are forces in a particular direction, thus the action output for agent  $i$  is  $\boldsymbol{u}_i^L$  where  $L$  is the final layer.
+
+Section 4.3 discussed applying EGNN to more sophisticated action spaces. Specifically for SMACv2 the discrete action space has an equivariant component (movements) and an invariant component (shoot). The outputs for the equivariant component of EGNN/E2GN2 are continuous values, mapping these values to logits is not straightforward. Furthermore, they would need to be mapped onto the same distribution of logits being represented by the shoot commands. We solve this problem by having three distributions output by the GNN structure: the continuous/equivariant movement gaussian distribution, the discrete/invariant distribution, and a third distribution that determines whether we should move or shoot. After the RL sampling from each distribution is performed for the actions, then the three components of the action can be converted to the final action for either a mixed discrete-continuous action space or a discrete action space.
+
+Graph Neural Network details We used the same GNN as the EGNN paper: Specifically, we use equations 1 and 3, but update equation 1 to be  $m_{ij} = \phi (h_i,h_j,x_i,x_j)$ . For convenience we rewrite the GNN equations describing a single layer here:
+
+$$
+\pmb {m} _ {i j} = \phi_ {e} (\pmb {h} _ {i} ^ {l}, \pmb {h} _ {j} ^ {l}), \quad \pmb {m} _ {i} = \sum_ {j \neq i} \pmb {m} _ {i j}, \quad \pmb {h} _ {i} ^ {l + 1} = \phi_ {h} (\pmb {h} _ {i} ^ {l}, \pmb {m} _ {i})
+$$
+
+Recall that  $\phi_h$  and  $\phi_e$  are two layer MLPs. Note that the GNN only has one variable  $h_i$  per node (compared to the two variables of  $h_i$  and  $u_i$  of EGNN/E2GN2). The features inputs are similar to an EGNN but with the two invariant/equivariant components concatenated together. The output for GNN will all come from the variable  $h$  as there is no need to distinguish between equivariant and invariant outputs (the action may be composed across several nodes, see above and figure 4).
+
+Hardware For training hardware, we trained the graph-structured networks using various GPUs. Many were trained on an A100, but that is certainly not necessary, they didn't need that much space. MLPs were trained on CPUs. We typically used 4 rollout workers across 4 CPU threads, so each training run used 5 CPU threads.
+
+Further Notes/details on SMACv2 SMACv2 ([14]) is a modification of the original SMAC ([29]). Since SMAC had deterministic initialization and static scenarios it turns out policies could memorize the solution. In each scenario, there is a pre-specified number of agents on each team (we use 5 agents for each team). In each scenario, the agents are randomly initialized in a specific formation. For this work, we use the initial position configuration termed "surrounded" (see fig 7). Finally, it is important to note that SMACv2 will randomly sample unit types from 3 different units (for each scenario). We use the same policy network for each agent but have the unit type as an input in the observation space, so each policy must learn to condition its behavior on the unit type.
+
+# C Supplementary Results
+
+Below we include supplementary results we believe interesting to understanding further E2GN2's performance.
+
+![](images/32891b4dadd66b932869b8830b422f81442eebf9d11174e7188f1680057a4457.jpg)  
+Figure 8: These results are using SMACv2 with the 'surrounded' initialization and 5 agents. Specifically, here we map the actions from the mixed discrete-continuous actions back to the default SMACv2 discrete action space. This is done by simply rounding the continuous actions to the nearest axis. These results are using the standard SMACv2 action input.
+
+![](images/b2a9c5e15f41d3c510b49ba37cbba1679a170b07727d477779b8c3c642bee4b9.jpg)  
+Figure 9: These results are using SMACv2 with the 'surrounded' initialization and 5 agents. Specifically, the observations here are partially observable. Note that E2GN2 still performs well with partial observability.
+
+![](images/41137268ebb170cf21dc17469c0ed2b7f66d0dca1f579e76a064c7ff12f1db15.jpg)
+
+![](images/0a58487781c95478f1a0b5007c556e03ca52e7757fccb7ac4c22853c17e553a5.jpg)
+
+![](images/fdde1ed00307fa601639324b63278e1025c28958d4edb448f7359a60726fcb6f.jpg)  
+Figure 10: These results are using SMACv2 with the 'surrounded' initialization and 5 agents. Specifically, the observations here are partially observable. In all other experiments we recenter the states at 0 (standard in MARL). Here we did perform the recentering. Note how important this step is for EGNN. Without the centering of observations the EGNN is assuming rotational equivariance around the wrong center. Interestingly, this seems to have little impact on E2GN2's performance.
+
+![](images/a480aa76aabefa3b83c0cfb1ab2b010a81c669b27ba0b2d1a8ac4409c1687dcb.jpg)
+
+![](images/4fffc9f672a71a2a501b9b7ce2ebbdbbdfe71da707d1699ceb02f004c6eb83b9.jpg)
+
+![](images/aeb25e1ac8dc7477083dbbc114ca5d2390b5ec626bae2e2eb757e0680d429860.jpg)  
+Figure 11: These were trained on SMACv2 using the 'surrounded' initialization with 5 agents. We were curious how an MLP would perform using the mixed continuous/discrete action space vs all discrete. Note that our E2GN2 is still able to outperform both.
+
+![](images/73c3122f86d0e4f8886775831231b6dc0f18007d0f000017b970a9a89e2e3b24.jpg)
+
+![](images/3f049f47ddd389aa7ab75420761223714a980d97629565d9bb8d98cf55d12165.jpg)
+
+![](images/81b130329c1974a738a5da5c340c3b1cb6c54152e1aa7e4de5f8764b0ea21405.jpg)
+
+![](images/3f58af82422e830d9d73774dc793f0a8d1d6e07313431a1d62a92e5bdccf76af.jpg)
+
+![](images/61df61ec40ad5b15ad980f2beac3aa34ff55a8c98841b30e1d4157d803114570.jpg)
+
+![](images/e06540a8128cc09f8a1ad5eaffc228d497af4868ff25bf2fba48deddd45fe15d.jpg)  
+Figure 12: SMACv2 has various initial position configurations. Here we show the results using the 'Surrounded and Reflect' initial configuration.
+
+![](images/3eafbf91bb599562679fd9ea8d93c6af2e30be22295163513e61eb64053aa845.jpg)
+
+![](images/cc3e933eea0c61ec2db56e78822503ac605dcb846d03ca12a22d23eaaf9cff91.jpg)
+
+![](images/5116e75ca36bb04aaee171d8adb66d07331275cd083381d288b4290665a4f3a2.jpg)
+
+![](images/5d0b7c679c1b20ca029f4d7aab3d93e1355c54a3741b177e6bd5518c6969609f.jpg)
+
+![](images/ad112c9f6816ca043a8840d409de3daa9fc8e846622f1f75b0ea76d85596bb06.jpg)
+
+![](images/28522f0d94e93e79c77e2b14346652f9ce5c829535a29d9f068f31ad5db097e6.jpg)  
+Figure 13: SMACv2, Surrounded and Reflect, 5 agents. Here we dive deeper into the performance when trained on individual unit types (instead of a mix of unit types). Top Row: Only Marine, stalker or Zergling units Bottom Row: Only Marauder, Zealot or hydralisk unit
+
+![](images/f4ae9f2090e7eafff5c9687e8c3ad7e4304981272eb9cd04dbc41a6dfbc36052.jpg)
+
+![](images/8b64535ae997c3b7ddea69d05a00dbe932c79f826b74916d49a57741d038f3f2.jpg)
+
+Table 5: Generalization Win Rate: Training on 5 agents in Surrounded-Left configuration, Testing on 7 agents in all configurations  
+
+<table><tr><td rowspan="2">Environment</td><td rowspan="2">Network</td><td rowspan="2">Training Initialization (but with 7 agents) 
+Surrounded Left</td><td colspan="2">Testing Initialization</td></tr><tr><td>Surrounded Right</td><td>Surrounded All</td></tr><tr><td rowspan="2">Terran</td><td>E2GN2</td><td>0.362</td><td>0.354</td><td>0.432</td></tr><tr><td>GNN</td><td>0.383</td><td>0.095</td><td>0.246</td></tr><tr><td rowspan="2">Protoss</td><td>E2GN2</td><td>0.431</td><td>0.418</td><td>0.424</td></tr><tr><td>GNN</td><td>0.415</td><td>0.206</td><td>0.083</td></tr><tr><td rowspan="2">Zerg</td><td>E2GN2</td><td>0.207</td><td>0.156</td><td>0.145</td></tr><tr><td>GNN</td><td>0.320</td><td>0.047</td><td>0.153</td></tr></table>
+
+# NeurIPS Paper Checklist
+
+# 1. Claims
+
+Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope?
+
+Answer: [Yes]
+
+Justification: We believe that our charts and tables reflect the claims made in the abstract and introduction.
+
+Guidelines:
+
+- The answer NA means that the abstract and introduction do not include the claims made in the paper.  
+- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers.  
+- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings.  
+- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper.
+
+# 2. Limitations
+
+Question: Does the paper discuss the limitations of the work performed by the authors?
+
+Answer: [Yes]
+
+Justification: Yes, we discuss these briefly in the conclusion and results.
+
+Guidelines:
+
+- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper.  
+- The authors are encouraged to create a separate "Limitations" section in their paper.  
+- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be.  
+- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated.  
+- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon.  
+- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size.  
+- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness.  
+- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations.
+
+# 3. Theory Assumptions and Proofs
+
+Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof?
+
+Answer: [Yes]
+
+Justification: Yes we provide our proofs and rationale in the appendix
+
+# Guidelines:
+
+- The answer NA means that the paper does not include theoretical results.  
+- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced.  
+- All assumptions should be clearly stated or referenced in the statement of any theorems.  
+- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition.  
+- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material.  
+- Theorems and Lemmas that the proof relies upon should be properly referenced.
+
+# 4. Experimental Result Reproducibility
+
+Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)?
+
+# Answer: [Yes]
+
+Justification: We specifically used an open-source MARL library (RLlib) for our RL training. The code for EGNN is available online. We described any modifications we made to the environments, as well as hyperparameters and training details.
+
+# Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not.  
+- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable.  
+- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed.  
+- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example  
+(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm.  
+(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully.  
+(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset).  
+(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results.
+
+# 5. Open access to data and code
+
+Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material?
+
+Answer: [No]
+
+Justification: Much of the necessary code is available online. Part of this work was developed for a private company, and permission for the full code release was not given.
+
+Guidelines:
+
+- The answer NA means that paper does not include experiments requiring code.  
+- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details.  
+- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark).  
+- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details.  
+- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc.  
+- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why.  
+- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable).  
+- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted.
+
+# 6. Experimental Setting/Details
+
+Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results?
+
+Answer: [Yes]
+
+Justification: Yes, we mention that we selected hyperparameters following guidelines set by [26], and included the hyperparameters used, the unspecified ones were simply the RLlib default parameters.
+
+Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them.  
+- The full details can be provided either with the code, in appendix, or as supplemental material.
+
+# 7. Experiment Statistical Significance
+
+Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments?
+
+Answer: [Yes]
+
+Justification: The charts all had confidence intervals (95 percent), computed using the standard seaborn package.
+
+Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper.
+
+- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions).  
+- The method for calculating the error bars should be explained (closed form formula call to a library function, bootstrap, etc.)  
+- The assumptions made should be given (e.g., Normally distributed errors).  
+- It should be clear whether the error bar is the standard deviation or the standard error of the mean.  
+- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a  $96\%$  CI, if the hypothesis of Normality of errors is not verified.  
+- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates).  
+- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text.
+
+# 8. Experiments Compute Resources
+
+Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments?
+
+Answer: [Yes]
+
+Justification: We describe this in the appendix.
+
+Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- The paper should indicate the type of compute workers CPU or GPU, internal cluster or cloud provider, including relevant memory and storage.  
+- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute.  
+- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper).
+
+# 9. Code Of Ethics
+
+Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines?
+
+Answer: [Yes]
+
+Justification: Yes, we believe this work adheres to the code of ethics.
+
+Guidelines:
+
+The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics  
+- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics.  
+- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction).
+
+# 10. Broader Impacts
+
+Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed?
+
+Answer: [Yes]
+
+Justification: We discuss this in the conclusion.
+
+Guidelines:
+
+- The answer NA means that there is no societal impact of the work performed.  
+- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact.
+
+- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations.  
+- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster.  
+- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology.  
+- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML).
+
+# 11. Safeguards
+
+Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)?
+
+Answer: [NA]
+
+Justification: We did not release data or code.
+
+# Guidelines:
+
+- The answer NA means that the paper poses no such risks.  
+- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters.  
+- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images.  
+- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort.
+
+# 12. Licenses for existing assets
+
+Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected?
+
+Answer: [Yes]
+
+Justification: We cited the code library used. RLlib uses an Apache license 2.0
+
+# Guidelines:
+
+- The answer NA means that the paper does not use existing assets.  
+- The authors should cite the original paper that produced the code package or dataset.  
+- The authors should state which version of the asset is used and, if possible, include a URI.  
+- The name of the license (e.g., CC-BY 4.0) should be included for each asset.  
+- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided.  
+- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset.
+
+- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided.  
+- If this information is not available online, the authors are encouraged to reach out to the asset's creators.
+
+# 13. New Assets
+
+Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets?
+
+Answer: [NA]
+
+Justification: There were no new assets introduced.
+
+Guidelines:
+
+- The answer NA means that the paper does not release new assets.  
+- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc.  
+- The paper should discuss whether and how consent was obtained from people whose asset is used.  
+- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file.
+
+# 14. Crowdsourcing and Research with Human Subjects
+
+Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)?
+
+Answer: [NA]
+
+Justification: There were no human subjects.
+
+Guidelines:
+
+- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects.  
+- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper.  
+- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector.
+
+# 15. Institutional Review Board (IRB) Approvals or Equivalent for Research with Human Subjects
+
+Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained?
+
+Answer: [NA]
+
+Justification: No human subjects in this work.
+
+Guidelines:
+
+- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects.  
+- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper.  
+- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution.  
+- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review.

2024/Boosting Sample Efficiency and Generalization in Multi-agent Reinforcement Learning via Equivariance/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ed12ef94b6086ef55d7d20880a7af7ba6996ddd132573fd838cc1765b6f69907
+size 867258

2024/Boosting Semi-Supervised Scene Text Recognition via Viewing and Summarizing/95d67c07-a36c-457b-aab4-2733e2025fc0_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:24dd1bef374919013147583fb3ed011aa4dae6dd5c1dc6350303cd9834cfe3a9
+size 866473

2024/Boosting Semi-Supervised Scene Text Recognition via Viewing and Summarizing/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:53c7079458eb13a35b33ae6b7fb228c8e9c74fee3fccb3e507a5b2e2a1b1204f
+size 1135163

2024/Boosting Text-to-Video Generative Model with MLLMs Feedback/6a3a1350-8e96-4b88-b833-9075cc51309c_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f261d3740f70e94422eedfecac5baa4b114f53bdaead94a87e643c6e7acdc2c9
+size 4552022

2024/Boosting Text-to-Video Generative Model with MLLMs Feedback/full.md

@@ -0,0 +1,775 @@
+# Boosting Text-to-Video Generative Model with MLLMs Feedback
+
+Xun Wu<sup>1</sup>, Shaohan Huang<sup>1*</sup>, Guolong Wang<sup>2</sup>, Jing Xiong<sup>3</sup>, Furu Wei<sup>1</sup>  
+<sup>1</sup> Microsoft Research Asia, <sup>2</sup> University of International Business and Economics  
+<sup>3</sup> The University of Hong Kong  
+xunwu@microsoft.com, shaohanh@microsoft.com, fuwei@microsoft.com
+
+# Abstract
+
+Recent advancements in text-to-video generative models, such as Sora [3], have showcased impressive capabilities. These models have attracted significant interest for their potential applications. However, they often rely on extensive datasets of variable quality, which can result in generated videos that lack aesthetic appeal and do not accurately reflect the input text prompts. A promising approach to mitigate these issues is to leverage Reinforcement Learning from Human Feedback (RLHF), which aims to align the outputs of text-to-video models with human preferences. However, the considerable costs associated with manual annotation have led to a scarcity of comprehensive preference datasets. In response to this challenge, our study begins by investigating the efficacy of Multimodal Large Language Models (MLLMs) generated annotations in capturing video preferences, discovering a high degree of concordance with human judgments. Building upon this finding, we utilize MLLMs to perform fine-grained video preference annotations across two dimensions, resulting in the creation of VIDEOPREFER, which includes 135,000 preference annotations. Utilizing this dataset, we introduce VIDEORM, the first general-purpose reward model tailored for video preference in the text-to-video domain. Our comprehensive experiments confirm the effectiveness of both VIDEOPREFER and VIDEORM, representing a significant step forward in the field.
+
+# 1 Introduction
+
+Diffusion models have significantly enhanced the quality of generation across various media formats, including images [30, 31, 55] and videos [13, 22, 42]. In the context of text-to-video generation, recent advancements in text-to-video diffusion models, exemplified by Sora [4], have achieved remarkable success in generating high-quality videos from textual prompts. However, despite recent progress, the visual fidelity of generated videos still presents opportunities for refinement [38]. Video generation poses greater challenges than image generation, as it entails modeling a higher-dimensional spatio-temporal output space while remaining conditioned solely on a textual prompt. Consequently, existing text-to-video generative models often produce results that are visually unappealing and inadequately aligned with the provided textual prompts.
+
+To address these limitations, a pivotal solution is Reinforcement Learning from Human Feedback (RLHF), which has proven its efficacy in text-to-image diffusion models [51, 2, 26]. These methods aim to fine-tune a diffusion model to maximize a reward function corresponding to specific aspects of image quality or alignment with text prompts. Despite this, when a Reward Model trained in the image domain is transferred to the video domain, it can exhibit significant disparities with the objectives of video optimization (e.g., InstructVideo [53]). Training a reward model directly in the video domain is becoming increasingly urgent, but the scale of preference datasets in the video field is small (see in Table 1).
+
+Table 1: Statistics of existing preference datasets for text-to-video generative models. * denotes annotated by human while † denotes annotated by GPT-4 V.  
+
+<table><tr><td>Dataset</td><td>Prompts</td><td>Videos</td><td>Preference Choices</td></tr><tr><td>VBench [15]*</td><td>1K</td><td>4K</td><td>44K</td></tr><tr><td>TVGE [56]*</td><td>0.5K</td><td>2.5K</td><td>2.5K</td></tr><tr><td>T2VQA-DB [18]*</td><td>1K</td><td>45K</td><td>45K</td></tr><tr><td>VIDEOPREFER†</td><td>14K</td><td>54K</td><td>135K</td></tr></table>
+
+Table 2: Correlations between MLLMs and human judgment on text-video alignment on TVGE datasets [56]. * denotes image-based MLLMs while † denotes Video-based MLLMs.  
+
+<table><tr><td>Model</td><td>Spearman ρ</td><td>Kendall τ</td><td>Acc (%)</td></tr><tr><td>Video-LLaMA [54]†</td><td>0.288</td><td>0.206</td><td>-</td></tr><tr><td>mPLUG-OWL2-V [52]†</td><td>0.394</td><td>0.285</td><td>61.87</td></tr><tr><td>InstructBLIP [10]*</td><td>0.342</td><td>0.246</td><td>54.33</td></tr><tr><td>mPLUG-OWL2-I [52]*</td><td>0.358</td><td>0.257</td><td>53.36</td></tr><tr><td>Gemini pro Vision*</td><td>0.3921</td><td>0.2993</td><td>64.71</td></tr><tr><td>LLaVA 1.6-34B*</td><td>0.3139</td><td>0.2278</td><td>53.20</td></tr><tr><td>GPT-4 V*</td><td>0.486</td><td>0.360</td><td>69.65</td></tr></table>
+
+Recognizing the importance of addressing these challenges in text-to-video generative models, we first construct a large-scale fine-grained preference benchmark by utilizing MLLMs as annotators, namely VIDEOPREFER. VIDEOPREFER contains following strength: (1) VIDEOPREFER is the largest open-source video preference dataset, containing 135,000 preference choices (see in Table 1). (2) Different from existing video preference datasets, VIDEOPREFER contains true video captured by human, making VIDEOPREFER more generalizable. (3) VIDEOPREFER is annotated using Multimodal Large Language Models, which is cost-effective and easily scalable.
+
+Based on VIDEOPREFER, we release the first general-purpose text-to-video preference reward model, VIDEORM. Unlike the preference rewards in the image domain used by previous methods for video generation alignment [53], our VIDEORM automatically captures the temporal information in videos, enhancing the modeling of quality assessment and alignment. Furthermore, we investigate the alignment of text-to-video generative models using VIDEORM and conduct extensive experiments that demonstrate its efficacy as a reward model and metric for video alignment. This significantly improves the generation quality of video generation models across multiple aspects. Our main contributions are:
+
+- We systematically identify the challenges for text-to-video human preference annotation. Consequently, we employ MLLMs for preference annotation and construct the largest and most comprehensive video preference dataset, VIDEOPREFER.  
+- We systematically identify the lack of general-purpose text-to-video preference reward model and propose VIDEORM, outperforming existing reward models for text-to-video alignment.  
+- Extensive experimental results validate the effectiveness of both the VIDEOPREFER and VIDEORM. Furthermore, through detailed discussions, we demonstrate that MLLMs are effective and cost-effective as annotators for video preferences, revealing the promising future prospects of RLAIF in the domain of video preference alignment.
+
+# 2VIDEOPREFER
+
+We introduce VIDEOPREFER, a large-scale, fine-grained video preference dataset constructed by collecting feedback from multimodal large language model annotators. In total, VIDEOPREFER contains 135K pairs of binary preference choices for 54K videos. In § 2.1, we provide evidence demonstrating that GPT-4 V is a human-aligned preference annotator in the video domain. The construction pipeline of VIDEOPREFER is introduced in § 2.2. Detailed analysis of the statistics of VIDEOPREFER can be found in Appendix § B.1.
+
+# 2.1 Why GPT-4 V(ision) can act as a human-aligned preference annotator?
+
+GPT-4 V has already demonstrated annotation performance aligned with human consistency in text-to-image generation [7, 46, 44]. To further validate GPT-4 V's ability to provide reliable preference annotations in the video domain, we used different MLLMs to annotate a subset of the TVGE dataset and calculated their correlation with the ground truth annotations, e.g., accuracy and kendall. The results are shown in Table 2. We find that (1) GPT-4 V's annotation accuracy and correlation are the best among all MLLMs. (2) The accuracy of GPT-4 V is  $69.65\%$ , which is very close to the previously reported agreement rate between qualified human annotators (approximately  $70\%$  [9, 46, 25, 11]). This demonstrates that GPT-4 V is a reliable annotator.
+
+# 2.2 Construction Pipeline
+
+The construction pipeline of VIDEOPREFER can be split into following three steps:
+
+Step-1: Prompts Collection. We collect prompts from VidProM [40] which contains 1.67 million unique text-to-video prompts, as well as from two video-captioning benchmarks: ActivityNet[5] and MSR-VTT[49]. For VidProM, we just randomly sample 12.9k prompts from its corpus. For video-captioning benchmarks, we directly utilize the provided video segment captions from the original dataset as prompts. For instance, in the ActivityNet [5], we utilize the text provided in caption corpus, comprising 0.9K thousand text instances corresponding to distinct video segments. Consequently, we obtain a total of 14K prompts, constituting the prompt set for our VIDEOPREFER.
+
+Step-2: Video Collection & Generation. We generate videos with text-to-video models and collect real-world videos based on prompts collected in Step-1 for preference annotation. Through these two ways, we obtained 54K video candidates:
+
+- Model-generated videos. We selected the top-ranked text-to-video generation models on Hugging-Face<sup>1</sup> (e.g., ModelScopeT2V [38]), as well as the open-source state-of-the-art text-to-video models (e.g., Open-Sora [57]), to constitute our video generation model pool. Details regarding these models, as well as the proportions and resolutions of the generated videos, can be found in Table 5. For each prompt, we randomly sample models from the model pools, along with class-free guidance scales, to generate four different corresponding videos. This approach results in a high degree of diversity, facilitating the training of a more generalizable and comprehensive preference reward model.
+
+Real-world videos. Given the substantial quality disparity between videos generated by existing text-to-video generation algorithms and real videos, we also incorporate real-world videos to enhance the generalizability and diversity of our VIDEOPREFER. Specifically, we incorporate video segments from the two datasets introduced in Step-1 (ActivityNet [5] and MSR-VTT [49]), as one of the four candidate videos, with the remaining three being generated as above.
+
+Step-3: Preference Generation. Due to the prohibitively high cost of manual annotation, we utilize the state-of-the-art multimodal large language models (MLLMs), i.e., GPT-4 V to annotate video preferences, enabling extensive and fine-grained annotation at scale. The reliability of GPT-4 V as annotators is thoroughly discussed in § 2.1. Specifically, for each prompt and its corresponding four video candidates, we employ GPT-4 V to assign preference scores on a 1-to-5 Likert scale to each video candidate for two aspects: Prompt-Following and Video Quality. Lower scores indicate lower preference and vice versa. Finally, we obtain 135K preference choices. Detailed input instructions used for GPT-4 V's annotation can be found in the Appendix E.
+
+By performing the above three steps, we finally have 14K data items in VIDEOPREFER, while each data item contains a textual prompt and four corresponding generated videos. For each video, there are preference annotations for two aspects: Prompt-Following and Video-Quality provided. More details of VIDEOPREFER, e.g., visualization of example data item (see in Figure 10), can be found in Appendix B.
+
+# 3 VIDEORM: A General-Purpose Video Preference Reward Model
+
+Based on VIDEOPREFER, we implement the reward model training and derive the first general-purpose video preference reward model, VIDEORM. We detail the architecture of VIDEORM in § 3.1, its optimization objectives in § 3.2, and the methodology for fine-tuning video generative models with VIDEORM in § 3.3.
+
+# 3.1 Architecture
+
+Existing text-to-video alignment work [53] adopts HPS v2 [45] as the reward model, which is finetuned from CLIP [27] and optimized upon a large image preference benchmark and demonstrates state-of-the-art preference evaluation capabilities for text-to-image generation.
+
+However, we contend that the direct application of HPS v2 may lead to a one-sided assessment of video preferences, as it lacks the capability to evaluate the overall attributes of a video, such
+
+as temporal coherence and dynamics. Nevertheless, given the strong correlation between video preferences and the preferences of individual frames, the proficiency of HPS v2 [45] in evaluating single-frame image preferences can still aid in enhancing the assessment of video preferences. Thus, to achieve a better evaluation for video preferences, we modify the structure of HPS v2 by integrating several temporal modeling modules, to develop a specialized reward model VIDEORM for the video domain.
+
+The full architecture of VIDEORM is shown in Figure 1. Specifically, inspired by recent advancements [39, 21] in enhancing the temporal modeling capabilities of the CLIP [27], we add two kinds of temporal modeling modules into HPS v2 to modeling videos:
+
+Temporal Shift [20], a parameter-free module which shifts part of the feature channels along the temporal dimension and facilitates information exchange between neighboring input frames. Following [39], we insert the module between every two ViT Layers of the image encoder of HPS v2.
+
+Temporal Transformer. The sequence of frameswise features extracted by the image encoder are then fed into a temporal transformer to modeling the temporal features of the video.
+
+![](images/9f377f0c960fe86a3ee2c044ccc50f09a680c76303e5584ed67b7c0671b016f8.jpg)  
+Figure 1: Overview of VIDEORM. By incorporating temporal modeling modules, VIDEORM is capable of not only capturing the preference scores of individual frames but also modeling the temporal features of the video, thereby better evaluating the overall preference score of the video.
+
+# 3.2 Optimization
+
+During the optimization process of VIDEORM, we freeze the text encoder and ViT layers contained in the image encoder to retain the single-frame preference modeling capability of HPS v2, while only optimizing the temporal transformer module. Thus, VIDEORM achieves the capability to model video preferences from both the perspectives of individual frames and temporal dynamics, allowing for a more comprehensive understanding of video content.
+
+Similar to previous works [36, 25], to train VIDEORM on VIDEOPREFER, we first average the scores of all 16 aspects for each video candidate, obtaining the final preference score for that video candidate. Then we formulate the preference score for each pair-wise video candidates as rankings. Thus, for each data item (consisting of one prompt  $\mathbf{T}$  and its corresponding four video candidates  $\mathbf{v}_1, \mathbf{v}_2, \mathbf{v}_3$  and  $\mathbf{v}_4$ ), we get at most  $C_4^2$  comparison pairs if there are no ties between two video candidates. For each comparison, if  $\mathbf{v}^+$  is better and  $\mathbf{v}^-$  is worse ( $\mathbf{v}^+ > \mathbf{v}^-$ ), the loss function can be formulated as:
+
+$$
+\left. \left. \log (\theta) = - \mathbb {E} _ {\left(\mathbf {T}, \mathbf {v} ^ {+}, \mathbf {v} ^ {-}\right) \sim \mathcal {D}} \left[ \log \left(\sigma \left(R _ {\theta} \left(\mathbf {T}, \mathbf {v} ^ {+}\right) - R _ {\theta} \left(\mathbf {T}, \mathbf {v} ^ {-}\right)\right)\right) \right] \right. \right. \tag {1}
+$$
+
+where  $R_{\theta}(\mathbf{T},\mathbf{v})$  is a scalar value of preference model for prompt  $\mathbf{T}$  and generated video  $\mathbf{v}$ .
+
+# 3.3 Fine-tuning Text-to-Video Models with VIDEORM
+
+Current exploration of reward reinforcement learning algorithms for fine-tuning text-to-video generative models is quite limited. The only related work is InstructVideo [53], which utilizes an image-domain reward model (HPS v2 [45]) to fine-tune text-to-video models.
+
+Consequently, InstructVideo may have the following shortcomings: (1) Due to the inherent gap between images and videos, directly using image-domain reward models cannot accurately calculate the reward for generated videos, leading to visual artifacts such as structural twitching and color jittering [53]. (2) Additionally, InstructVideo [53] calculates the reward value for all frames selected from generated videos throughout the full DDIM [34] sampling procedure, making the fine-tuning process highly sample inefficient.
+
+VIDEORM is specifically designed for the video domain. It is initialized with the weights of the best image-domain reward model and trained on video preference datasets, utilizing temporal modeling mechanisms (e.g., temporal transformers) to capture temporal features of videos. This enables it to evaluate the reward score of generated videos more effectively from both individual frame and
+
+temporal perspectives. Additionally, unlike the image-domain reward model used in InstructVideo, which takes single video frames as input, VIDEORM processes the entire video as input. Therefore, our approach is more computational efficient and can directly leverage various effective image-domain reward reinforcement learning algorithms (such as DRaFT [8]) without the need to balance rewards between frames generated at each step, as required by InstructVideo.
+
+Based on the above, we integrated VIDEORM into the image-domain DRaFT [8] algorithm and design a reward reinforcement learning algorithm suitable for text-to-video generative models, named DRaFT-V. The specific algorithm details are shown in Algorithm 1. DRaFT-V truncates the backward pass, differentiating and computing reward score from VIDEORM through only the last  $K$  steps, making the full fine-tuning process more efficient.
+
+In implementation, rather than fine-tuning the full set of model parameters, we follow [53] to adopt LoRA [14] to further accelerate fine-tuning and circumvent the issue of computational intensity as well as the risk of catastrophic forgetting associated with the reward loss in diffusion models.
+
+# 4 Experiments
+
+We conduct extensive experiments to validate the effectiveness of VIDEOPREFER and VIDEORM. We first train VIDEORM and evaluate it on existing human-preference benchmark (§ 4.1). Next, we fine-tune existing text-to-video diffusion models for aligning human preference by utilizing DRaFT-V with VIDEORM (§ 4.2). Finally, we present ablation studies (§ 4.3).
+
+# 4.1 Task 1: Reward Modeling
+
+Setup. Based on VIDEOPREFER, we develop VIDEORM, an advanced open-source general-purpose reward model that provides preferences for generated videos. Specifically, we train three versions of VIDEORM. (1) Firstly, to validate the effectiveness of VIDEOPREFER, we average the preference scores in each aspect of VIDEOPREFER to get a final preference score, and train VIDEORM-V with the merging
+
+version of VIDEOPREFER. (2) Then, we train VIDEORM-H on several open-source human-crafted preference datasets listed in Table 1. (3) Finally, to build a stronger RM for text-to-video generation, we mix these open-source human-crafted preference datasets with VIDEOPREFER to train VIDEORM. The details for dataset processing can be found in Appendix A. All VIDEORM series models are trained in half-precision on  $8 \times 32$ GB NVIDIA V100 GPUs, with a learning rate of 1e-5 and batch size of 64 in total. We set the input frames  $N = 8$ .
+
+Compared Baselines. Due to the lack of reward models in the video domain, we compare VIDEORM with state-of-the-art reward models from the image domain, i.e., CLIP ViT-H/14 [27], ImageReward [48], PickScore [17] and HPS v2 [45]. For all these reward models from the image domain, we calculate the scores for all video frames and then take the average as the final reward score for the video.
+
+Preference Accuracy. The preference prediction accuracy across test benchmarks from three human-crafted preference datasets are reported in Table 3. As we can see, the VIDEORM series outperform baseline reward models by a large margin, indicating that VIDEORM series are the best open-source reward models for text-to-video domain. We also find that VIDEORM-V which does not train on any open-source video preference datasets also surpasses all other baselines. This result validates the high quality of VIDEOPREFER enables strong out-of-distribution generalization and validate
+
+# Algorithm 1 DRaFT-V: Reward Reinforcement Learning for Fine-tuning Text-to-Video Models with VIDEORM
+
+1: Dataset: Prompt set  $\mathcal{Y} = \{y_1, y_2, \dots, y_n\}$  
+2: Pre-training Dataset: Text-Video pairs dataset  $\mathcal{D} = \{(t_1,v_1),\dots (t_n,v_n)\}$  
+3: Input: Text-to-video models  $\Upsilon$  with pretrained parameters  $\theta_0$ , VIDEORM  $R$ , reward-to-loss map function  $\theta$ , Text-to-video models pre-training loss function  $\psi$ , reward re-weight scale  $\lambda$  
+4: Initialization: The number of noise scheduler time steps  $T$ , and the truncating step  $K$  
+5: for  $y_{i}\in \mathcal{V}$  and  $(\mathrm{t}_i,\mathrm{v}_i)\in \mathcal{D}$  do  
+6:  $\mathcal{L}_{pre}\gets \psi_{\theta_i}(\mathfrak{t}_i,\mathbf{v}_i)$  
+7:  $w_{i}\gets w_{i} / / \mathrm{Update}\Upsilon_{\theta_{i}}$  using  $\mathcal{L}_{pre}$  
+8:  $x_{T}\sim \mathcal{N}(0,I) / /$  Sample noise as latent  
+9: for  $j = T, \dots, 1$  do  
+10: if  $j > K$  then  
+11: no grad:  $x_{j - 1}\gets \Upsilon_{\theta_i}\{x_j\}$  
+12: else  
+13: with  $\mathbf{grad}\colon x_{j - 1}\leftarrow \Upsilon_{\theta_i}\{x_j\}$  
+14:  $x_0\gets x_j$  //Predict the original latent by noise scheduler  
+15:  $z_{i}^{j}\gets x_{0} / /$  From latent to image  
+16:  $\mathcal{L}_{reward} \gets \lambda \phi(r(y_i, z_i^j))$  // Reward loss  
+17: end if  
+18: end for  
+19:  $\theta_{i + 1}\gets \theta_i / / \mathrm{Update}\Upsilon_{\theta_i}$  using  $\mathcal{L}_{reward}$  
+20: end for
+
+Table 3: Pair-wise preference prediction accuracy across human-crafted preference datasets. The Aesthetic Classifier (simplified as Aesthetic) makes prediction without seeing the text prompt. The best results are in blod and the second are underlined.  
+
+<table><tr><td>Model</td><td>TVGE [56]</td><td>VBench [15]</td><td>T2VQA-DB [18]</td><td>Avg</td></tr><tr><td>CLIP ViT-H/14 [27]</td><td>57.3</td><td>52.7</td><td>52.1</td><td>54.0</td></tr><tr><td>Aesthetic [33]</td><td>56.1</td><td>51.0</td><td>52.7</td><td>53.3</td></tr><tr><td>ImageReward [48]</td><td>66.8</td><td>54.3</td><td>53.9</td><td>58.3</td></tr><tr><td>PickScore [17]</td><td>64.7</td><td>53.9</td><td>61.2</td><td>59.9</td></tr><tr><td>HPS v2 [45]</td><td>69.5</td><td>55.7</td><td>52.8</td><td>59.3</td></tr><tr><td>VIDEORM-H</td><td>72.8</td><td>60.2</td><td>64.3</td><td>65.8</td></tr><tr><td>VIDEORM-V</td><td>73.0</td><td>61.1</td><td>65.1</td><td>66.4</td></tr><tr><td>VIDEORM</td><td>73.7</td><td>63.5</td><td>65.4</td><td>67.5</td></tr></table>
+
+![](images/ec35e0d0e3f001daee025e20863bec7cde2595f972743e6969bfaef723129d32.jpg)  
+Figure 2: Best-of-  $n$  experiments on the T2VQA-DB [18] test benchmark. We sample  $n$  generated videos and choose the one with the highest reward.
+
+![](images/a6440a16edeb30461a5c2635fc5fdc0ac6cea6ce605bcb56ed8fc6ef95287f69.jpg)  
+Figure 3: Win rates of VIDEORM compared to other reward models across three test benchmarks. On average,  $72\%$  to CLIP,  $67\%$  to ImageReward,  $62\%$  to PickScore and  $65\%$  to HPS v2.
+
+![](images/057516caa2c2800965888e8a7d0ddfc399f88d585c8389c9748c4f711a28fc04.jpg)
+
+![](images/5bcc41f7ed364dd07ab329e3b6ee15ed79d4a8f91ec7710972fd2cf013f72f6e.jpg)
+
+the effectiveness of treating state-of-the-art MLLMs, i.e., GPT-4 V, as preference annotator for text-to-video generation domain.
+
+Best-of- $n$  Experiments. To verify that our VIDEORM could serve as a good indicator of video generation quality, we conduct best-of- $n$  experiments on the T2VQA-DB [18] test benchmark. For each data item, which includes a prompt and ten corresponding videos, we calculate the reward score for each of the ten videos with VIDEORM. Thereafter, we select the best-of-{1, 2, 4, 8} responses and calculate their scores. The final results are presented at Figure 2, we find the win rate at the test benchmark increases proportionally with rewards, which validates that our VIDEORM gives rigorous rewards that reflect the overall preference score.
+
+Human Evaluation. To evaluate the ability of VIDEORM to select the more preferred videos among large amounts of generated videos, we produce human study. We randomly select 200 textual prompts from these three test benchmarks and generate 32 different videos for each prompt by utilizing ModelScopeT2V [38]. Then, we perform different reward to select from those videos to get top3 results. After that, five annotators are asked to identify which video was superior or if both were of equal quality (denoted as 'Tie') and we show the corresponding win rates against other reward models [45] at Figure 3. Qualitative results can be seen at Figure 11 in Appendix. All of these results show that VIDEORM can select videos that are more aligned to text and with higher fidelity and avoid toxic contents.
+
+# 4.2 Task 2: Fine-tuning Text-to-Video Generative Models
+
+Setup. Based on VIDEORM, we validate the effectiveness of DRaFT-V. We adopt the publicly available text-to-video diffusion model ModelScopeT2V [38] as base model, which is trained on WebVid10M with  $\mathrm{T} = 1000$  and is able to generate videos of  $16 \times 256 \times 256$  resolution. We random sample 2K prompt-video pair from the mixture of existing video preference datasets (TVGE [56], VBench [15] and T2VQA-DB [18]) and VIDEOPREFER as the training data. Each group of experiment is trained in half-precision on  $4 \times 32$  GB NVIDIA V100 GPUs, with a learning rate of 1e-5, batch size of 8, fine-tuning step of 400 and  $K$  adopted in DRaFT-V as 10 in total.
+
+Compared Baselines. Since the only existing reward-based fine-tuning algorithm for text-to-video models is InstructVideo [53]. We compare our DRaFT-V with the baseline text-to-video model (i.e., ModelScopeT2V [38]) and InstructVideo implemented by ourselves. Note that in the aforementioned algorithms, InstructVideo uses HPS v2 [45] as the reward model. Additionally, to verify the effectiveness of using VIDEORM as a reward model for fine-tuning video generation models, we
+
+![](images/e57122412cdd770c80e837007553e2d74499b6db118d1abd9373f9efc3ebe527.jpg)  
+Figure 5: Win rates of text-to-video models fine-tuned with DRaFT-V compared to other baselines. Here baseline denotes the base text-to-video model without any fine-tuning.
+
+![](images/7443271d34d67a59f2375262f16fda36e49e28f1d659abd070ef89e0b1ca4de0.jpg)
+
+![](images/3da38225e2d2f75405aa40d8fec6198c9727e1b2667eae79551d904bcab98ad2.jpg)
+
+![](images/ec65e86173a531bcec31f64f932452ca19607966c4bad0d8d303ceb5fdbb0ee0.jpg)  
+(a)
+
+![](images/ee054f5852dd03620819e4c81c63eebab1741eb3852799f74524ecfbda602a6d.jpg)  
+(b)
+
+![](images/9930ee911d9c0602a256a91eb4463b445ce98807fec36cf88429a5053fefb2d3.jpg)  
+(c)  
+Figure 6: (a) Visualization of reward model values change with the training steps for both DRaFT-V and DRaFT-H. (b) Evaluation result across three test benchmarks for the size of training data used in optimizing VIDEORM. (c) Ablation study for the number of input frames  $N$  in VIDEORM.
+
+introduced two comparison groups: (1) DRaFT-H, which uses the same algorithm as our DRaFT-V (i.e., DRaFT [8]), but with HPS v2 as the reward model. (2) InstructVideo-V, which uses the same algorithm as InstructVideo [53], but with VIDEORM as the reward model.
+
+Generation Quality Evaluated by Multiple Reward Models. At first, we visualize the evolution of model's generation quality across multiple reward models as the model training steps increase when using our DRaFT-V at Figure 4. We observe that with the progress of training, all reward models show an increasing trend, indicating that VIDEORM can serve as a reliable reward model to enable the generated videos to align more closely with human preferences.
+
+Human Evaluation. To evaluate the generative ability of text-to-video models fine-tuned with DRaFT-V and other compared baselines, we produce human study. We randomly select 100 prompts from these three test benchmarks (TVGE [56], VBench [15] and T2VQA-DB [18]), and for each prompt, we generate videos using different fine-tuned models. Five annotators are asked to identify
+
+which video was superior or if both were of equal quality (denoted as 'Tie') and we show the corresponding win rates at Figure 5. We find that: (1) All of these results show that DRaFT-V can generate videos that are more preferred by human. (2) By comparing DRaFT-V with DRaFT-H, as well as InstructVideo-V with InstructVideo, we demonstrate that VIDEORM is a more effective reward model for fine-tuning text-to-video models.
+
+Besides, qualitative results are presented at Figure 12 in Appendix D. We find that: (1) The base generation model often fails to align videos with prompt descriptions, but fine-tuning with a reward model (image or video domain) improves quality and prompt consistency, demonstrating the effectiveness of reward-based fine-tuning. (2) Fine-tuning with VIDEORM (DRaFT-V and InstructVideo-V) significantly outperforms fine-tuning with an image domain reward model HPS v2, indicating VIDEORM's superiority for text-to-video model fine-tuning.
+
+Efficiency Evaluation. We compare the efficiency of our algorithm on two aspects: (1) Convergence speed: we visualize the changes in reward values with training steps for DRaFT-H and DRaFT-V in Figure 6 (a). We find that under the same reward-based fine-tuning algorithm, using VIDEORM leads to earlier convergence and better performance (as demonstrated by the results above) compared to HPS v2, further validating the effectiveness of VIDEORM in fine-tuning text-to-video models. (2) Inference speed. For a single video with 8 frames, HPS v2 requires approximately an average of 5 seconds for inference, while VIDEORM only needs an average of 1.3 seconds. Besides, InstructVideo requires 20 DDIM steps for a single fine-tuning step, taking approximately an average of 8.8 seconds, while DRaFT-V only needs an average of 2.3 seconds.
+
+![](images/ee28ac2961abcb11a8e2c040964c7ec76372c655277fa9640c275d7ebb0d7ce2.jpg)  
+Figure 4: Evaluation results of the text-to-video model's generation quality across multiple reward models when maximizing scores from VIDE-ORM during the DRaFT-V fine-tuning process.
+
+# 4.3 Ablation Study
+
+Scalability. To investigate the effect of training dataset sizes on the performance of VIDEORM, and to verify the scalability of VIDEOPREFER, comparative experiments are conducted. Note that VIDEORM in this experiment are only trained on VIDEOPREFER. Figure 6 (b) shows that adding up the scale of the dataset significantly improves the preference accuracy of VIDEORM. It's promising that if we employ GPT-4 V to collect more annotation data in the future, VIDEORM will get better performance.
+
+Number of Video Frames  $N$  Input to VIDEORM. We investigate the impact of different numbers of input video frames on the performance of the RM model. Specifically, we set the number of input frames to  $N = 1, 4, 8$ , and 12 and test corresponding model at TVGE [56] test benchmark. The corresponding results are presented at Figure 6 (c), we find that the VIDEORM's performance significantly decreases when  $N$  is small. This is promising because too few input frames prevent the model from accurately capturing all the information and temporal features of the video, e.g., the degree of motion and temporal consistency.
+
+Temporal Feature Modeling method in VIDEORM. We conduct experiments to explore the impact of different video frame temporal feature modeling methods on the performance of the reward model. In VIDEORM, we use a temporal transformer trained from scratch to model the temporal features of video frames. Additionally, we conducted ablation experiments by replacing this transformer with 1D Convolutional Layer (denoted as  $\text{VIDEOM}^\dagger$ ) and LSTM (denoted as  $\text{VIDEOM}^\ddagger$ ). Note that we adjusted the number of layers in the temporal feature modeling module to ensure that the number of trainable parameters remains equal. Besides, Temporal Shift are
+
+![](images/6d09693b0049b6f037c14cbcfaae84e0e4682ab96e2ddc0b85ab95e62dfed46f.jpg)  
+Figure 7: Ablation over  $K$  adopted in DRaFT-V during finetuning text-to-video model.
+
+employed for all compared methods. The results are presented at Table 4. We find that: (1) All models outperform the image-domain reward model. (2) Replacing the temporal transformer with LSTM achieved comparable performance, indicating that VIDEORM is robust to the choice of model architecture for temporal feature modeling. (3) Replacing the temporal transformer with Conv1D resulted in a significant performance drop. We hypothesize that this is because Conv1D cannot effectively model temporal features, leading to VIDEORM†'s inability to accurately evaluate video generation quality from a temporal perspective. This highlights the effectiveness of using a temporal transformer for modeling temporal features in VIDEORM.
+
+Backbone of VIDEORM. We explore the impact of different backbone on the final performance of VIDEORM. In our design, VIDEORM is initialized with the weights of HPS v2 [45], which is demonstrated as the best reward model for image domain. Here we replace HPS v2 with other reward models in image domain, e.g., PickScore [17] (denoted as  $\text{VIDEORM}^\sigma$ ) and ImageReward [48] (denoted as  $\text{VIDEORM}^\beta$ ) while keep the count of trainable parameter equal. The corresponding results are presented at Table 4. We find that the final performance of the corresponding VIDEORM is proportional to the performance of the initialized backbone weights. For example, ImageReward, which has the lowest performance among the three (see Table 3), corresponds to the worst  $\text{VIDEORM}^\beta$ , while PickScore, which has the best performance among the three (see Table 3), corresponds to the best  $\text{VIDEORM}^\sigma$ .
+
+Step  $K$  adopted in DRaFT-V. We investigate the impact of different  $K$  adopted in DRaFT-V on the final generative performance. The results are presented at Figure 7, we find that: (1) Performance degrades as K increases for  $K > 10$ . (2) Even with smaller values of  $K$ , e.g.,  $K = 1$ , DRaFT-V can achieve good results. Both findings are aligned with the results in DRaFT [8].
+
+# 5 Analysis
+
+In this section, we analyze the impact of different parameter selections—specifically, the number of frames  $(N)$  sampled from a single video and the temperature setting  $\tau$  of the GPT-4 V, on the accuracy of GPT-4 V-generated annotations. Due to GPT-4 V's limitation of processing a maximum of 10 frames, we analyze annotation accuracy for  $N = 4, 6, 8,$  and 10 frames. The temperature  $\tau$  controls the diversity and randomness of GPT-4 V's outputs. We analyze annotation accuracy for  $\tau = 0.3, 0.5, 0.7,$  and 0.9.
+
+Table 4: Ablation study for VIDEORM. The Aesthetic Classifier (simplified as Aesthetic) makes prediction without seeing the text prompt.  
+
+<table><tr><td>Model</td><td>TVGE [56]</td><td>VBranch [15]</td><td>T2VQA-DB [18]</td><td>Avg</td></tr><tr><td>VIDEORM†</td><td>70.1</td><td>58.4</td><td>61.5</td><td>63.3</td></tr><tr><td>VIDEORM‡</td><td>73.9</td><td>62.7</td><td>65.2</td><td>67.3</td></tr><tr><td>VIDEORMσ</td><td>72.2</td><td>64.2</td><td>66.9</td><td>67.7</td></tr><tr><td>VIDEORMβ</td><td>72.4</td><td>63.6</td><td>64.2</td><td>66.7</td></tr><tr><td>VIDEORM</td><td>73.7</td><td>63.5</td><td>65.4</td><td>67.5</td></tr></table>
+
+![](images/6bf5da3447be41336c5bf022f28fcc9741f4571dd94ba18bc079016ea157d6f3.jpg)  
+Figure 8: The impact of different hyperparameter choices on the annotation accuracy of GPT-4 V.
+
+The summarized results are shown in Figure 8. For  $N$ , accuracy initially increases with more frames but decreases beyond  $N = 8$ , achieving the highest accuracy at  $N = 8$ . This is because more frames allow for a comprehensive evaluation, but too many frames can lead to overly long contexts, reducing accuracy due to GPT-4 V's limited capacity. For  $\tau$ , lower values yield higher accuracy, likely because reduced randomness leads to more comprehensive predictions, while higher randomness might cause GPT-4 V to overly focus on specific frames or aspects, resulting in less comprehensive outcomes.
+
+# 6 Related Work
+
+Generative models, particularly those based on diffusion techniques, have demonstrated high-quality generation capabilities [32, 28, 31, 30, 35] by training on extensive internet-scale datasets, but the mixed quality of these datasets often leads to visually unappealing and misaligned outputs.
+
+Aligning text-to-image generative models. Aligning text-to-image generative models [37, 17, 8, 26, 37] has garnered increasing attention in recent years and has shown promising results in producing outputs that are more aligned with human preference. Imagereward [48], HPS v2 [47, 45], and PickScore [17] are the three most commonly used reward models for align text-to-image models. They are respectively trained on three major image domain preference datasets, namely Imagereward, ImageRewardDB [48], HPD [47, 45] and Pick-a-Pic [17]. The effectiveness of these reward models has been validated across alignment algorithms like AlignProp [26], DRAFT [8] and ReFL [48].
+
+Aligning text-to-video generative models. Compared to text-to-image, exploration related to aligning text-to-video generative models is relatively sparse. InstructVideo [53] instruct text-to-video diffusion models with fine-tuning with existing image-wise preference reward model HPS v2. We hypothesize that this may restrict effective preference evaluations for generated videos, as image-based reward models cannot adequately capture temporal features, impairing assessments of video coherence and dynamics. Additionally, preference datasets for text-to-video generation are also scarce. The lack of large-scale and effective open-source preference datasets severely restricts the development of research related to align text-to-video generative models.
+
+Reinforcement Learning from AI Feedback. LLMs have been extensively used for data generation [43, 24], augmentation [12] and in self-training setups [41, 23]. Some works [1] introduced the idea of reinforcement learning from AI feedback (RLAIF), which used LLMs labeled preferences in conjunction with human labeled preferences to jointly optimize for the two objectives of helpfulness and harmlessness. Recent works have also explored related techniques for generating rewards from LLMs [29, 19, 50]. These works demonstrate that LLMs can generate useful signals for reinforcement learning fine-tuning. However, leveraging the feedback from MLLMs for aligning text-to-image generative model or for text-to-video generative model is less explored.
+
+# 7 Conclusion
+
+In this paper, we identify the current obstacles faced in aligning text-to-video generative models, i.e., lacking of large-scale preference datasets and reward models specifically tailored for videos. Thus, we introduce VIDEOPREFER and VIDEORM to address the aforementioned issues. Experimental validation confirms that GPT-4 V can act as a human-aligned preference annotator. We utilized it to label 135K video preference annotation, forming VIDEOPREFER. Based on this, we trained a videospecific reward model, VIDEORM. Extensive analytical usage has demonstrated the effectiveness of both VIDEOPREFER and VIDEORM.
+
+# References
+
+[1] Y. Bai, S. Kadavath, S. Kundu, A. Askell, J. Kernion, A. Jones, A. Chen, A. Goldie, A. Mirhoseini, C. McKinnon, C. Chen, C. Olsson, C. Olah, D. Hernandez, D. Drain, D. Ganguli, D. Li, E. Tran-Johnson, E. Perez, J. Kerr, J. Mueller, J. Ladish, J. Landau, K. Ndousse, K. Lukosuite, L. Lovitt, M. Sellitto, N. Elhage, N. Schiefer, N. Mercado, N. DasSarma, R. Lasenby, R. Larson, S. Ringer, S. Johnston, S. Kravec, S. E. Showk, S. Fort, T. Lanham, T. Telleen-Lawton, T. Conerly, T. Henighan, T. Hume, S. R. Bowman, Z. Hatfield-Dodds, B. Mann, D. Amodei, N. Joseph, S. McCandlish, T. Brown, and J. Kaplan. Constitutional ai: Harmlessness from ai feedback, 2022.  
+[2] K. Black, M. Janner, Y. Du, I. Kostrikov, and S. Levine. Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301, 2023.  
+[3] T. Brooks, B. Peebles, C. Holmes, W. DePue, Y. Guo, L. Jing, D. Schnurr, J. Taylor, T. Luhman, E. Luhman, C. Ng, R. Wang, and A. Ramesh. Video generation models as world simulators. 2024.  
+[4] T. Brooks, B. Peebles, C. Holmes, W. DePue, Y. Guo, L. Jing, D. Schnurr, J. Taylor, T. Luhman, E. Luhman, C. Ng, R. Wang, and A. Ramesh. Video generation models as world simulators. 2024.  
+[5] F. Caba Heilbron, V. Escorcia, B. Ghanem, and J. Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015.  
+[6] H. Chen, Y. Zhang, X. Cun, M. Xia, X. Wang, C. Weng, and Y. Shan. Videocrafter2: Overcoming data limitations for high-quality video diffusion models, 2024.  
+[7] J. Chen, C. Ge, E. Xie, Y. Wu, L. Yao, X. Ren, Z. Wang, P. Luo, H. Lu, and Z. Li. Pixart-  $\backslash$  sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024.  
+[8] K. Clark, P. Vicol, K. Swersky, and D. J. Fleet. Directly fine-tuning diffusion models on differentiable rewards. arXiv preprint arXiv:2309.17400, 2023.  
+[9] G. Cui, L. Yuan, N. Ding, G. Yao, W. Zhu, Y. Ni, G. Xie, Z. Liu, and M. Sun. Ultrafeedback: Boosting language models with high-quality feedback. arXiv preprint arXiv:2310.01377, 2023.  
+[10] W. Dai, J. Li, D. Li, A. M. H. Tiong, J. Zhao, W. Wang, B. Li, P. N. Fung, and S. Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. Advances in Neural Information Processing Systems, 36, 2024.  
+[11] Y. Dubois, C. X. Li, R. Taori, T. Zhang, I. Gulrajani, J. Ba, C. Guestrin, P. S. Liang, and T. B. Hashimoto. Alpacafarm: A simulation framework for methods that learn from human feedback. Advances in Neural Information Processing Systems, 36, 2024.  
+[12] S. Y. Feng, V. Gangal, J. Wei, S. Chandar, S. Vosoughi, T. Mitamura, and E. Hovy. A survey of data augmentation approaches for NLP. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 968-988, Online, Aug. 2021. Association for Computational Linguistics.  
+[13] J. Ho, T. Salimans, A. Gritsenko, W. Chan, M. Norouzi, and D. J. Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022.  
+[14] E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, and W. Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021.  
+[15] Z. Huang, Y. He, J. Yu, F. Zhang, C. Si, Y. Jiang, Y. Zhang, T. Wu, Q. Jin, N. Chanpaisit, Y. Wang, X. Chen, L. Wang, D. Lin, Y. Qiao, and Z. Liu. VBench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024.
+
+[16] L. Khachatryan, A. Movsisyan, V. Tadevosyan, R. Henschel, Z. Wang, S. Navasardyan, and H. Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators, 2023.  
+[17] Y. Kirstain, A. Polyak, U. Singer, S. Matiana, J. Penna, and O. Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. arXiv preprint arXiv:2305.01569, 2023.  
+[18] T. Kou, X. Liu, Z. Zhang, C. Li, H. Wu, X. Min, G. Zhai, and N. Liu. Subjective-aligned dataset and metric for text-to-video quality assessment. arXiv preprint arXiv:2403.11956, 2024.  
+[19] M. Kwon, S. M. Xie, K. Bullard, and D. Sadigh. Reward design with language models. In The Eleventh International Conference on Learning Representations, 2022.  
+[20] J. Lin, C. Gan, and S. Han. Tsm: Temporal shift module for efficient video understanding, 2019.  
+[21] H. Luo, L. Ji, M. Zhong, Y. Chen, W. Lei, N. Duan, and T. Li. Clip4clip: An empirical study of clip for end to end video clip retrieval, 2021.  
+[22] Z. Luo, D. Chen, Y. Zhang, Y. Huang, L. Wang, Y. Shen, D. Zhao, J. Zhou, and T. Tan. Videofusion: Decomposed diffusion models for high-quality video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10209-10218, 2023.  
+[23] A. Madaan, N. Tandon, P. Gupta, S. Hallinan, L. Gao, S. Wiegreffe, U. Alon, N. Dziri, S. Prabhumoye, Y. Yang, et al. Self-refine: Iterative refinement with self-feedback. arXiv preprint arXiv:2303.17651, 2023.  
+[24] Y. Meng, M. Michalski, J. Huang, Y. Zhang, T. Abdelzaher, and J. Han. Tuning language models as training data generators for augmentation-enhanced few-shot learning. In International Conference on Machine Learning, pages 24457-24477. PMLR, 2023.  
+[25] L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730-27744, 2022.  
+[26] M. Prabhudesai, A. Goyal, D. Pathak, and K. Fragkiadaki. Aligning text-to-image diffusion models with reward backpropagation. arXiv preprint arXiv:2310.03739, 2023.  
+[27] A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In ICML, 2021.  
+[28] A. Ramesh, P. Dhariwal, A. Nichol, C. Chu, and M. Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2):3, 2022.  
+[29] P. Roit, J. Ferret, L. Shani, R. Aharoni, G. Cideron, R. Dadashi, M. Geist, S. Girgin, L. Hussenot, O. Keller, et al. Factually consistent summarization via reinforcement learning with textual entailment feedback. arXiv preprint arXiv:2306.00186, 2023.  
+[30] R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022.  
+[31] C. Saharia, W. Chan, S. Saxena, L. Li, J. Whang, E. L. Denton, K. Ghasemipour, R. Gontijo Lopes, B. Karagol Ayan, T. Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022.  
+[32] C. Saharia, W. Chan, S. Saxena, L. Li, J. Whang, E. L. Denton, K. Ghasemipour, R. Gontijo Lopes, B. Karagol Ayan, T. Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022.
+
+[33] C. Schuhmann, R. Beaumont, R. Vencu, C. Gordon, R. Wightman, M. Cherti, T. Coombes, A. Katta, C. Mullis, M. Wortzman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402, 2022.  
+[34] J. Song, C. Meng, and S. Ermon. Denoising diffusion implicit models. arXiv:2010.02502, October 2020.  
+[35] J. Song, C. Meng, and S. Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.  
+[36] N. Stiennon, L. Ouyang, J. Wu, D. Ziegler, R. Lowe, C. Voss, A. Radford, D. Amodei, and P. F. Christiano. Learning to summarize with human feedback. Advances in Neural Information Processing Systems, 33:3008-3021, 2020.  
+[37] Z. Tang, D. Rybin, and T.-H. Chang. Zeroth-order optimization meets human feedback: Provable learning via ranking oracles. arXiv preprint arXiv:2303.03751, 2023.  
+[38] J. Wang, H. Yuan, D. Chen, Y. Zhang, X. Wang, and S. Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023.  
+[39] M. Wang, J. Xing, and Y. Liu. Actionclip: A new paradigm for video action recognition, 2021.  
+[40] W. Wang and Y. Yang. Vidprom: A million-scale real prompt-gALLERY dataset for text-to-video diffusion models. arXiv preprint arXiv:2403.06098, 2024.  
+[41] X. Wang, J. Wei, D. Schuurmans, Q. V. Le, E. H. Chi, S. Narang, A. Chowdhery, and D. Zhou. Self-consistency improves chain of thought reasoning in language models. In The Eleventh International Conference on Learning Representations, 2022.  
+[42] Y. Wang, X. Chen, X. Ma, S. Zhou, Z. Huang, Y. Wang, C. Yang, Y. He, J. Yu, P. Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023.  
+[43] Z. Wang, A. W. Yu, O. First, and Y. Cao. Towards zero-label language learning. arXiv preprint arXiv:2109.09193, 2021.  
+[44] S. Wen, G. Fang, R. Zhang, P. Gao, H. Dong, and D. Metaxas. Improving compositional text-to-image generation with large vision-language models. arXiv preprint arXiv:2310.06311, 2023.  
+[45] X. Wu, Y. Hao, K. Sun, Y. Chen, F. Zhu, R. Zhao, and H. Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023.  
+[46] X. Wu, S. Huang, and F. Wei. Multimodal large language model is a human-aligned annotator for text-to-image generation. arXiv preprint arXiv:2404.15100, 2024.  
+[47] X. Wu, K. Sun, F. Zhu, R. Zhao, and H. Li. Human preference score: Better aligning text-to-image models with human preference. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2096-2105, 2023.  
+[48] J. Xu, X. Liu, Y. Wu, Y. Tong, Q. Li, M. Ding, J. Tang, and Y. Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. arXiv preprint arXiv:2304.05977, 2023.  
+[49] J. Xu, T. Mei, T. Yao, and Y. Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5288-5296, 2016.  
+[50] K. Yang, D. Klein, A. Celikyilmaz, N. Peng, and Y. Tian. Rlcd: Reinforcement learning from contrast distillation for language model alignment, 2023.  
+[51] K. Yang, J. Tao, J. Lyu, C. Ge, J. Chen, Q. Li, W. Shen, X. Zhu, and X. Li. Using human feedback to fine-tune diffusion models without any reward model. arXiv preprint arXiv:2311.13231, 2023.
+
+[52] Q. Ye, H. Xu, J. Ye, M. Yan, H. Liu, Q. Qian, J. Zhang, F. Huang, and J. Zhou. mplug-owl2: Revolutionizing multi-modal large language model with modality collaboration. arXiv preprint arXiv:2311.04257, 2023.  
+[53] H. Yuan, S. Zhang, X. Wang, Y. Wei, T. Feng, Y. Pan, Y. Zhang, Z. Liu, S. Albanie, and D. Ni. Instructvideo: Instructing video diffusion models with human feedback. arXiv preprint arXiv:2312.12490, 2023.  
+[54] H. Zhang, X. Li, and L. Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023.  
+[55] L. Zhang, A. Rao, and M. Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023.  
+[56] J. Zhangjie Wu, G. Fang, H. Wu, X. Wang, Y. Ge, X. Cun, D. Junhao Zhang, J.-W. Liu, Y. Gu, R. Zhao, et al. Towards a better metric for text-to-video generation. arXiv e-prints, pages arXiv-2401, 2024.  
+[57] Z. Zheng, X. Peng, and Y. You. Open-sora: Democratizing efficient video production for all, March 2024.
+
+# A Datasets Processing
+
+T2VQA-DB [18]. The T2VQA-DB dataset only comprises a single overall preference evaluation score for each data item. We randomly selected 2.3K samples to serve as the final test benchmark.
+
+TVGE [56]. The TVGE dataset comprises preference evaluations from 2 aspects, namely: text alignment and video quality. We just simply average the scores across these two aspects to obtain a final overall evaluation score as with the aforementioned datasets. We randomly selected 15K samples to serve as the final test benchmark.
+
+VBench [15]. The VBench dataset comprises preference evaluations from 16 aspects, namely: subject consistency, background consistency, temporal flickering, motion smoothness, dynamic degree, aesthetic quality, imaging quality, object class, multiple objects, human action, color, spatial relationship, scene, temporal style, appearance style, overall consistency.
+
+Due to the fact that the videos and prompts evaluated for preference in each aspect are largely non-overlapping, we cannot simply average the scores across all aspects to obtain a final overall evaluation score as with the aforementioned datasets. Instead, we mixed all samples from the 16 aspects and randomly selected 10K samples to serve as the final test benchmark.
+
+# B VIDEOPREFER
+
+# B.1 Statistic ofVIDEOPREFER
+
+In Table 5, we present the sources and distribution of the videos in our VIDEOPREFER. We find that VIDEOPREFER comprises a diverse range of video sources, including videos generated by state-of-the-art text-to-video models as well as real videos. This extensive variety of video sources enhances the robustness and generalization capabilities of VIDEOPREFER.
+
+We visualize the distribution of preference annotations across two evaluation aspects (prompt-following an video-quality) in VIDEOPREFER in Figure 9. We find that the overall score distribution of the dataset is close to the normal distribution for the Prompt-Following evaluation perspective. From the perspective of Video-quality evaluation, the overall distribution of data sets tends to be low, which indicates that the generation effect of the existing text-to-video model is not satisfactory.
+
+Besides, we also visualize some data examples in VIDEOPREFER at Figure 10.
+
+Table 5: Video sources of VIDEOPREFER.  ${}^{ \dagger  }$  denotes the realistic videos from existing benchmarks.  
+
+<table><tr><td>Source</td><td>Proportion</td><td>Video Length</td><td>Type</td><td>Resolution</td></tr><tr><td>LaVie [42]</td><td>16.4%</td><td>2.0s</td><td>Diffusion</td><td>512×512</td></tr><tr><td>ModelScope [22]</td><td>21.0%</td><td>2.0s</td><td>Diffusion</td><td>256×256</td></tr><tr><td>VideoCrafter2 [6]</td><td>17.2%</td><td>1.6s</td><td>Diffusion</td><td>320×320</td></tr><tr><td>Open-Sora [57]</td><td>2.4%</td><td>-</td><td>DiT</td><td>512×512</td></tr><tr><td>Pika2</td><td>7.2%</td><td>3.0s</td><td>DiT</td><td>1088×640</td></tr><tr><td>Text2Video-Zero [16]</td><td>15.3%</td><td>2.0s</td><td>Diffusion</td><td>512×512</td></tr><tr><td>ZeroScope</td><td>13.8%</td><td>-</td><td>Diffusion</td><td>1024×576</td></tr><tr><td>Gen-2</td><td>0.7%</td><td>-</td><td>-</td><td>1792×1024</td></tr><tr><td>ActivityNet [5]†</td><td>1.7%</td><td>-</td><td>Human-captured</td><td>-</td></tr><tr><td>MSR-VTT [49]†</td><td>0.03%</td><td>-</td><td>Human-captured</td><td>-</td></tr></table>
+
+![](images/c5ee00ef83c00617560fd7b551039e5d62b7dac71477397e8f7690864a1f58ae.jpg)  
+Figure 9: Score distribution across two annotation aspects in VIDEOPREFER.
+
+# Prompt: A man working in his room, typing a portable computer, and a cat watching him
+
+![](images/4f2522fda1fd0fa5d3a28d230c301f68b59c191af3175766bf021ff3b322fd6e.jpg)
+
+![](images/ce1a248481db3097575648d1c0a139a424ba6bdfef2cceeccf99ace4fc5e9dca.jpg)
+
+![](images/e152c86b5e5bfc38d7ca5b7076cbda2530a1e2f82714bd628f8d2d9cb64c9b08.jpg)
+
+![](images/2a9b98338173e76c34a3d14d83157b9e67b399e36acea5ff9779a31fda064a59.jpg)
+
+![](images/5e6998e49700d4f82bfda7bf7cfde7f361b2c99c9bfea0563b68d3d5679ebd4e.jpg)
+
+![](images/ebb77edeb0e35849bb6b743c0f617d48a0e4dbe1687eb4110fccd930adfff405.jpg)
+
+# Prompt-Following
+
+![](images/e2602988247cbb6b2055b4567a9b1f47bc88f318f79cfabf22f1518f6d177112.jpg)
+
+# Rating: 5
+
+Rationale: The sampled frames from the video show a man working on a portable computer with a cat watching him, which aligns perfectly with the original text provided. The setting appears to be a room, as indicated by the presence of a bookshelf in the background, which further supports the text description. There are no deviations from the text in the visuals provided.
+
+# Video-Quality
+
+![](images/66c1c30fe2b814f643b180fe6bb04148616fb5fddf12c515bfd5a6421693e54f.jpg)
+
+# Rating: 3
+
+Rationale: The sampled frames showed that there was no obvious artifacts in this video as a whole, but from the perspective of time sequence, both humans and cats were almost immobile, lacking the dynamic continuity of the video, so the final score is 3.
+
+![](images/d6950790e1cd93f2a6fceee5ba30bd2d4dc31b34d2a767603c141977791e9f4a.jpg)
+
+![](images/34ef44c1a6130c2ee046c00436285f5531f5c135a30d6fedbdfb18f74f911e24.jpg)
+
+![](images/dcbc880a55c1167fc7625868352771bc203daa205aaba36383ac079435d7123d.jpg)
+
+![](images/dac15166ea1fcf88d3b6ec28c99a30010e4dc0ed7d37479f55be0b14d6f4b946.jpg)
+
+![](images/8796c1835aaf8853a78ffb162b13e2d9123e9210dfcb011e2cb9c7e27998d6bd.jpg)
+
+![](images/ef6d19e3c1c03b47b8eedf19c628af635610a1fb80ba9163dd5dcc601a57cb6c.jpg)
+
+# Prompt-Following
+
+![](images/f752cb0d045bed6f3a1d774c09cfd75398fb27a3fc4d41702758a69f06666620.jpg)
+
+# Rating: 4
+
+Rationale: The video frames show a cat watching and interacting with a laptop, which aligns with the text describing a cat watching a man working on a portable computer. However, the man is not visible in the frames, only his hand is shown, which leads to a minor deviation from the text. The setting appears to be a room, as mentioned in the text. Overall, the video is nearly aligned with the original text, but the absence of a clear depiction of the man working results in a rating of 4.
+
+# Video-Quality
+
+![](images/dbe146ecc9a2415c8478a1bc56c531af4e2a2f0e10073f4e923130b1a8791548.jpg)
+
+# Rating: 4
+
+Rationale: On the whole, the video is rich in details, beautiful and strong in color, which is in line with human aesthetics. In terms of temporal dynamics, the hands in the video have tapping movements, so it can meet the dynamic requirements of the video.
+
+![](images/1f1b905c30b67a1a7306ac52789f6f44db4c1053fde34c75f36c839eca194974.jpg)
+
+![](images/ecf67e60dd177d6c703798dbae8b28fbf787d76677fc9ca46035be478125e9b2.jpg)
+
+![](images/3f097471845734805401c2c1906aa942bca2442c8085d34d304823b42dcae333.jpg)
+
+![](images/0b9ebd8ecd33ea33e1e1cc347c5299c8dd5ffa093cebbd5320b1f79385576ff2.jpg)
+
+![](images/4e7e428e929152aff523ab099308dab386b0f8cf83b732327999dafdad969458.jpg)
+
+![](images/2e3f79a0925d0eb2502214d647be8aeb920180f1ba87c0ee22d12e5b6e28697e.jpg)
+
+# Prompt-Following
+
+![](images/86b4f823b30889b497b5062125797188e48215c13189a93b8b943bb30357fec9.jpg)
+
+# Rating: 4
+
+Rationale: The video frames show a man working in a room on a portable computer, which aligns well with the text description. However, the cat is not actively watching the man but is instead lying on the couch, partially visible in some frames. This constitutes a minor deviation from the text, which specifies the cat should be watching the man. Hence, the rating is  $\backslash$  "Almost There" due to this slight discrepancy.
+
+# Video-Quality
+
+![](images/401bd81a095f7c35bb8b7827424b8127de496c1d8bfe6ab2857ddc43fa3010c0.jpg)
+
+# Rating: 2
+
+Rationale: In the sampled frames of this video, there are obvious object deformations such as laptop computers. In addition, in terms of timing features, the task has the action of tapping the keyboard, which conforms to the timing features, but objects disappear, which is not good in the continuity of time series, so the final score is 2
+
+![](images/3fe0a52db017a2df72c9b4784b718a2aaf85b0066e596b1f90f7dba4da772484.jpg)  
+Figure 10: Visualization of example data item in VIDEOPREFER. Here we show one data item which contains a prompt and four corresponding generated videos. For each video, there are two annotations from different annotation aspects (Prompt-Following and Video-Quality) are provided by GPT-4 V
+
+![](images/d1cf5bd9df508596988a9efe00215b2b78ac23318a20b3532391667471ee18ea.jpg)
+
+![](images/78912d9ef387f46b69e9b09a1da6b62383136bc7910ce41ec7293b194acc1a02.jpg)
+
+![](images/200fd84e59338f78cde4ae53e23b591fd8eaebb0a8d331e907a94f28ea10b77e.jpg)
+
+![](images/5bc100b70088d9779a409a8c39fde19126e7a7b273a6c7cf9ac26c87f0639b26.jpg)
+
+# Prompt-Following
+
+![](images/00b699476cbe720360e0b6570c9544384f1d4dc1950ec37a3fcda4442c706b01.jpg)
+
+# Rating: 5
+
+Rationale: The frames from the video show a person typing on a portable computer with a cat watching, which is in full alignment with the provided text. There are no deviations from the described scenario, thus the video fully aligns with the original text requirements.
+
+# Video-Quality
+
+![](images/1f98986b67be5f6a73547dfe68feeb4793fc724b42f260d0067b8b244439ccae.jpg)
+
+# Rating: 3
+
+Rationale: The overall contrast of light and dark light in the video is strong, and the object details are rich, which is in line with human aesthetics, but the sequence dynamics is not good, so the final score is 3
+
+# C Visualization of Selection Results
+
+![](images/9ea967b85204c1e368212f515ce73e91d6b4409fb11f8259b70d7a751417008e.jpg)  
+Figure 11: Top-1 videos from 32 generated videos select by CLIP, ImageReward, PickScore, HPS v2 and VIDEORM. VIDEORM is capable of selecting higher-quality generated videos, e.g., those that better match the prompt descriptions and exhibit more dynamic content.
+
+# D Visualization Results for Fine-tuning Text-to-Video Models
+
+![](images/97cd9a9210591d726f1806b4910bfa6679444d7d69d9566fdc49ae1a707173ad.jpg)  
+Figure 12: Visualization Results for different fine-tuning methods. We find that compared to fine-tuning with an image domain reward model, fine-tuning with VIDEORM significantly enhances the performance of text-to-video models (DRAFT-V and InstructVideo-V).
+
+# E Instruction Template
+
+# Preference Instruction for Prompt-Following
+
+# Prompt-Following:
+
+You are an AI assistant programmed to assess videos with impartial and balanced standards. A video has been created based on a piece of text. Your task is to evaluate how well The content of the video aligns with the original text provided as ("Input") The video evaluation is based on sampled frames shown in sequence.
+
+Scoring: Rating outputs 1 to 5:
+
+1. Irrelevant: No alignment.  
+2. Partial Focus: Addresses one aspect poorly.  
+3. Partial Compliance:
+
+- (1) Meets goal or restrictions, neglecting other.  
+- (2) Acknowledges both but slight deviations.
+
+4. Almost There: Near alignment, minor deviations.  
+5. Comprehensive Compliance: Fully aligns, meets all requirements.
+
+Please present your assessment as follows:
+
+# Output
+
+Rating: [Provide the rating]
+
+Rationale: [Explain the reason for your rating in concise sentences]
+
+Now, review the following video and its corresponding text.
+
+# Input:
+
+##### Text: [INSERT PROMPT HERE]
+
+Frames sampled from video: [INSERT THE FRAMES OF VIDEO HERE]
+
+# Preference Instruction for Video-Quality:
+
+# Video-Quality:
+
+You are an AI assistant trained to impartially assess temporal consistency quality, dynamic quality and aesthetic quality in videos. A video has been created based on a piece of text. Your task is to analyze the quality of this video based on the following guidelines and provide a comprehensive evaluation.
+
+Scoring: Rating outputs 1 to 5:
+
+1. Bad: blurry, underexposed with significant noise, indiscernible subjects, exhibits significant inconsistencies and noticeable discrepancies in appearance of subjects.  
+2. Poor: Noticeable blur, poor lighting, washed-out colors, and awkward composition with cut-off subjects, suffers from noticeable issues in maintaining uniformity of subjects and backgrounds.  
+3. Fair: In focus with adequate lighting, dull colors, decent composition but lacks creativity. Subjects and backgrounds maintain a reasonable degree of uniformity throughout most of the video, with only minor discrepancies.  
+4. Good: Sharp, good exposure, vibrant colors, thoughtful composition with a clear focal point. Good video dynamics and temporal consistency.  
+5. Excellent: Exceptional clarity, perfect exposure, rich colors, masterful composition with emotional impact. perfect temporal consistency and excellent dynamics.
+
+Please present your assessment as follows:
+
+# Output
+
+Rating: [Provide the rating]
+
+Rationale: [Explain the reason for your rating in concise sentences]
+
+Now, review the following video and its corresponding text.
+
+# Input:
+
+##### Text: [INSERT PROMPT HERE]
+
+Frames sampled from video: [INSERT THE FRAMES OF VIDEO HERE]
+
+# NeurIPS Paper Checklist
+
+# 1. Claims
+
+Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope?
+
+Answer: [Yes]
+
+Justification: The paper includes our motivation, preference data construction details, experimental settings, quantitative experimental results, and qualitative visual examples that reflect and justify the claims in our abstract and introduction.
+
+Guidelines:
+
+- The answer NA means that the abstract and introduction do not include the claims made in the paper.  
+- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers.  
+- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings.  
+- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper.
+
+# 2. Limitations
+
+Question: Does the paper discuss the limitations of the work performed by the authors?
+
+Answer: [Yes]
+
+Justification: The analysis section contains a discussion of limitations in our work.
+
+Guidelines:
+
+- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper.  
+- The authors are encouraged to create a separate "Limitations" section in their paper.  
+- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be.  
+- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated.  
+- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon.  
+- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size.  
+- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness.  
+- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations.
+
+# 3. Theory Assumptions and Proofs
+
+Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof?
+
+# Answer: [Yes]
+
+Justification: We state the reference before illustrating formulations.
+
+# Guidelines:
+
+- The answer NA means that the paper does not include theoretical results.  
+- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced.  
+- All assumptions should be clearly stated or referenced in the statement of any theorems.  
+- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition.  
+- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material.  
+- Theorems and Lemmas that the proof relies upon should be properly referenced.
+
+# 4. Experimental Result Reproducibility
+
+Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)?
+
+# Answer: [Yes]
+
+Justification: We provide a detailed description of the dataset construction process, implementation for training the reward model, and offer a multitude of visualization examples.
+
+# Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not.  
+- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable.  
+- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed.  
+- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example  
+(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm.  
+(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully.  
+(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset).  
+(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results.
+
+# 5. Open access to data and code
+
+Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material?
+
+Answer: [No]
+
+Justification: We will public our data and code upon paper acceptance, due to the management regulations of our institution.
+
+Guidelines:
+
+- The answer NA means that paper does not include experiments requiring code.  
+- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details.  
+- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark).  
+- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://nips.cc/public/guides/CodeSubmissionPolicy) for more details.  
+- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc.  
+- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why.  
+- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable).  
+- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted.
+
+# 6. Experimental Setting/Details
+
+Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results?
+
+Answer: [Yes]
+
+Justification: Full training and testing details are in  $\S 4$  and Appendix E.
+
+Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them.  
+- The full details can be provided either with the code, in appendix, or as supplemental material.
+
+# 7. Experiment Statistical Significance
+
+Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments?
+
+Answer: [No]
+
+Justification: It would take too long with our available computational resources to repeat all experiments multiple times.
+
+Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper.  
+- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions).
+
+- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.)  
+- The assumptions made should be given (e.g., Normally distributed errors).  
+- It should be clear whether the error bar is the standard deviation or the standard error of the mean.  
+- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a  $96\%$  CI, if the hypothesis of Normality of errors is not verified.  
+- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates).  
+- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text.
+
+# 8. Experiments Compute Resources
+
+Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments?
+
+Answer: [Yes]
+
+Justification: We include the name of GPU we used for experiments in Setup.
+
+Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage.  
+- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute.  
+- The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper).
+
+# 9. Code Of Ethics
+
+Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines?
+
+Answer: [Yes]
+
+Justification: All relevant human questions and relevant datasets have been checked for privacy compliance prior to experiments and submission.
+
+Guidelines:
+
+- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics.  
+- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics.  
+- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction).
+
+# 10. Broader Impacts
+
+Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed?
+
+Answer: [Yes]
+
+Justification: Our work supports the alignment of generative models with human values, and its societal impact is discussed in the introduction.
+
+Guidelines:
+
+- The answer NA means that there is no societal impact of the work performed.  
+- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact.
+
+- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations.  
+- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster.  
+- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology.  
+- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML).
+
+# 11. Safeguards
+
+Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)?
+
+Answer: [NA]
+
+Justification: Our contribution does not include new datasets or pre-trained models that pose a risk of misuse.
+
+Guidelines:
+
+- The answer NA means that the paper poses no such risks.  
+- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters.  
+- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images.  
+- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort.
+
+# 12. Licenses for existing assets
+
+Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected?
+
+Answer: [Yes]
+
+Justification: Code that we derive from earlier work is properly licensed and referenced.
+
+Guidelines:
+
+- The answer NA means that the paper does not use existing assets.  
+- The authors should cite the original paper that produced the code package or dataset.  
+- The authors should state which version of the asset is used and, if possible, include a URL.  
+- The name of the license (e.g., CC-BY 4.0) should be included for each asset.  
+- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided.
+
+- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset.  
+- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided.  
+- If this information is not available online, the authors are encouraged to reach out to the asset's creators.
+
+# 13. New Assets
+
+Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets?
+
+Answer: [Yes]
+
+Justification: We provide detailed illustration of new data benchmark and visualized examples.
+
+Guidelines:
+
+- The answer NA means that the paper does not release new assets.  
+- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc.  
+- The paper should discuss whether and how consent was obtained from people whose asset is used.  
+- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file.
+
+# 14. Crowdsourcing and Research with Human Subjects
+
+Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)?
+
+Answer: [No]
+
+Justification: We only recruited participants for user experiments to validate the effectiveness of our model, where they were asked to choose from generated images. No human participants were involved in the dataset construction or model training process.
+
+Guidelines:
+
+- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects.  
+- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper.  
+- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector.
+
+# 15. Institutional Review Board (IRB) Approvals or Equivalent for Research with Human Subjects
+
+Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained?
+
+Answer: [NA]
+
+Justification: Our experiment solely involves measurement and does not entail behavioral manipulation; therefore, we did not apply for IRB approval.
+
+Guidelines:
+
+- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects.
+
+- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper.  
+- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution.  
+- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review.

2024/Boosting Text-to-Video Generative Model with MLLMs Feedback/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c37f048edfff2b9fd0975aacfa1832f573f1d1587a3135d70df9cca003f92d92
+size 1040027

2024/Boosting Transferability and Discriminability for Time Series Domain Adaptation/679c5ddf-dcc2-453f-813c-b5bec1720d30_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:431a738966f17a39dc01c47bb36aaadb0988b5654523ade07fa0f3adc52f84b3
+size 1249791

2024/Boosting Transferability and Discriminability for Time Series Domain Adaptation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0ccfe5a125f4993eee262895cbe22a8bcb346c8ff011eb977dc7a982895fc789
+size 1811126

2024/Boosting Vision-Language Models with Transduction/409473c7-326d-4c59-8e3b-87e2a4f1bbbf_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:09b87f2ebdfdf140ca907b9a63406b693bbe2265e67e33a02ab4e24ebb7dea8c
+size 627041

2024/Boosting Vision-Language Models with Transduction/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1bf3e882e88068787a92d0b18be9260de88b9e6bfd8f5deb9e4ec152e0c6d713
+size 2760804

2024/Boosting Weakly Supervised Referring Image Segmentation via Progressive Comprehension/3f96f166-4232-493b-9f5f-4136794df000_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:768f0c2231bd64574d8f8b028069c296219564577ee33b3a2e8eae703181b6b0
+size 4146346

2024/Boosting Weakly Supervised Referring Image Segmentation via Progressive Comprehension/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2bb15c5b7f21219c1cd1ab9f3ca9b800384cffa13663cc07098ffc8e5fb9fd33
+size 1373687

2024/Boosting the Potential of Large Language Models with an Intelligent Information Assistant/b820ed92-a302-48e1-98ad-294d4073262d_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bdd760ddc215d9fc25dee137e18104aba4056220f928994bcde0aa54dc886ec8
+size 4950993

2024/Boosting the Potential of Large Language Models with an Intelligent Information Assistant/full.md

@@ -0,0 +1,633 @@
+# Boosting the Potential of Large Language Models with an Intelligent Information Assistant
+
+Yujia Zhou*
+
+Tsinghua University
+
+zhouuyujia@mail.tsinghua.edu.cn
+
+Zheng Liu*
+
+The Hong Kong Polytechnic University
+
+zhengliu1026@gmail.com
+
+Zhicheng Dou
+
+Renmin University of China
+
+dou@ruc.edu.cn
+
+# Abstract
+
+The emergence of Large Language Models (LLMs) has significantly advanced natural language processing, but these models often generate factually incorrect information, known as "hallucination". Initial retrieval-augmented generation (RAG) methods like the "Retrieve-Read" framework was inadequate for complex reasoning tasks. Subsequent prompt-based RAG strategies and Supervised Fine-Tuning (SFT) methods improved performance but required frequent retraining and risked altering foundational LLM capabilities. To cope with these challenges, we propose Assistant-based Retrieval-Augmented Generation (AssISTRAG), integrating an intelligent information assistant within LLMs. This assistant manages memory and knowledge through tool usage, action execution, memory building, and plan specification. Using a two-phase training approach—Curriculum Assistant Learning and Reinforced Preference Optimization—ASSISTRAG enhances information retrieval and decision-making. Experiments show AssISTRAG significantly outperforms benchmarks, especially benefiting less advanced LLMs, by providing superior reasoning capabilities and accurate responses.
+
+# 1 Introduction
+
+The emergence of Large Language Models (LLMs) has significantly advanced the field of natural language processing, demonstrating an impressive ability to mimic human-like language patterns [1]. However, despite their extensive knowledge acquired during training, LLMs can occasionally generate factually incorrect information, a phenomenon referred to as "hallucination" [2, 3]. To address this, the integration of retrieval systems with LLMs has been suggested, allowing these models to tap into external databases to generate more reliable responses [4].
+
+Initially, retrieval-augmented generation (RAG) relied on a simple "Retrieve-Read" framework [5], which was adequate for basic question-answering but insufficient for complex, multi-step reasoning tasks. As language models advanced, various prompt-based RAG strategies emerged [6, 7], incorporating pre-retrieval and post-retrieval prompts to refine the process. However, these strategies heavily relied on the foundational capabilities of the language models. Consequently, the focus shifted to Supervised Fine-Tuning (SFT)-based RAG methods [8], which involve fine-tuning language models specifically for RAG tasks to enhance their performance.
+
+While SFT-based methods have improved the quality of generated responses, they face two limitations that hinder their practical application. Firstly, these fine-tuned models are not easily adaptable to
+
+![](images/10eaa98aeb097a1a270121515ec7749879735e1d3922d00b3ad16c00658e609e.jpg)  
+Figure 1: Comparisons of Naive, Prompt-based, SFT-based and our Assistant-based RAG frameworks.
+
+emerging LLMs, requiring retraining for each new foundational LLM. Secondly, directly fine-tuning a foundational LLM in the RAG scenario may change its innate abilities, potentially leading to negative impacts on the model's performance on other tasks. To address these challenges, we propose Assistant-based Retrieval-Augmented Generation (ASSISTRAG), which integrates an intelligent information assistant as a plugin within LLMs. This approach comprises a trainable assistant for information management and a static main LLM dedicated to task execution, as depicted in Figure 1.
+
+As an intelligent information assistant, ASSISTRAG operates in two primary categories to handle complex tasks: memory management and knowledge management. Memory management involves integrating and analyzing content from internal memory, while knowledge management focuses on leveraging external knowledge. These two main functions are supported by four core capabilities of ASSISTRAG: (1) Tool usage, which involves recalling relevant information from both internal memory and external knowledge bases through a retriever; (2) Action execution, which involves processing, analyzing, and extracting information; (3) Memory building, which involves recording essential knowledge and reasoning patterns from historical interactions; (4) Plan specification, which involves determining the necessity of each step in the process. These four capabilities work together to ensure that ASSISTRAG can provide accurate and comprehensive support to the main LLM.
+
+To implement ASSISTRAG, we adopt a two-phase training approach. The first phase, Curriculum Assistant Learning, enhances the assistant's capabilities in note-taking, question decomposition, and knowledge extraction through progressively complex tasks. The second phase, Reinforced Preference Optimization, uses reinforcement learning to tailor the assistant's feedback to the main LLM's specific needs, optimizing knowledge extraction based on feedback from the main LLM.
+
+During the inference stage, ASSISTRAG operates through a three-step process: (1) Information Retrieval and Integration: The assistant understands the main LLM's needs, retrieves relevant knowledge from internal and external sources, and extracts valuable information. (2) Decision Making: The assistant evaluates and decides whether to provide the retrieved memories and knowledge to the main LLM based on their relevance. (3) Answer Generation and Memory Updating: The main LLM generates an answer using its internal knowledge and the assistant's information, while the assistant updates its memory with crucial reasoning steps.
+
+Results from experiments across three complex question-answering datasets reveal that ASSISTRAG exhibits superior reasoning capabilities and markedly outperforms existing benchmarks. Notably, when applied to different foundational LLMs, ASSISTRAG appears to confer more pronounced benefits on less advanced LLMs.
+
+# 2 Related Work
+
+# 2.1 Retrieval-Augmented Generation
+
+RAG represents a significant advancement in the domain of LLMs, particularly for tasks demanding extensive knowledge. This paradigm begins with a retrieval step, where the LLM accesses an external database to gather relevant information before addressing queries. Traditionally, RAG follows a
+
+"Retrieve-Read" framework [9, 5, 10, 11], with efforts focused on refining either the retriever or the generator through pre-training approaches to augment RAG's accuracy. Building on this foundation, new RAG strategies have emerged, including the use of prompt-based methods like Chain-of-Thought (CoT) reasoning [7, 12], iterative retrieval processes [13, 14, 15, 16], and leveraging LLM-generated content for dynamic retrieval [6, 17, 18]. These strategies underscore the LLMs' ability to select relevant information adaptively in response to specific contexts. Concurrently, research on fine-tuning LLMs for RAG applications is rapidly expanding [19, 20], focusing on enhancing skills such as query reformulation [21] and knowledge integration [22, 23, 24], as well as developing critical functions like determining the necessity of retrieval and appraising the value of retrieved data [8, 25]. Departing from these approaches, our paper introduces Assistant-based RAG, integrating an intelligent information assistant with the main LLM to boost its potential.
+
+# 2.2 LLM-based Autonomous Agents
+
+Recent advancements in LLMs have facilitated the development of LLM-based autonomous agents such as AutoGPT [26], Toolformer [27], and MetaGPT [28], which utilize LLMs for effective decision-making. Notably, ReAct [29] combines LLMs with external tools to manage knowledge-intensive tasks, allowing for dynamic responses to environmental changes. Additionally, models like WebGPT [30] integrate reinforcement learning with GPT-3, enabling the autonomous operation of search engines during text generation. Innovative methods used by Flare [17] and Self-Ask [6] determine optimal times for information retrieval, while Reflexion [31] endows LLMs with introspective mechanisms that continually refine their outputs. Our proposed Assistant-based RAG model further enhances LLM capabilities by combining memory management and knowledge management, thus providing robust support to the main LLM in tackling complex tasks.
+
+# 3 Methodology
+
+In this section, we first define the task of RAG and then introduce our proposed framework, ASSISTRAG. ASSISTRAG enhances the capabilities of LLMs through the support of an intelligent information assistant. With abilities to use tools, execute actions, build memory, and plan, ASSISTRAG can provide precise memory and knowledge management services for LLMs.
+
+# 3.1 Task Definition
+
+Given a question  $q$  and a collection of documents  $D = \{d_i\}_{i=1}^{|D|}$ , the main LLM aims to generate an answer  $y$  based on both the question and the relevant documents. This can be formalized as  $y = \mathrm{LLM}_{\mathrm{main}}([D_q, q])$ , where  $D_q$  represents the set of documents retrieved for the query  $q$ , and  $[\cdot, \cdot]$  denotes the concatenation of the retrieved documents with the query. Expanding this concept, ASSISTRAG employs an intelligent information assistant,  $\mathrm{LLM}_{\mathrm{Assist}}$ , to enhance the main LLM's responses by providing relevant information, formalized as  $y = \mathrm{LLM}_{\mathrm{main}}([\mathrm{LLM}_{\mathrm{Assist}}(q), q])$ . In the following sections, we will detail the capabilities of the ASSISTRAG framework, along with its training and inference procedures.
+
+# 3.2 ASSISTRAG Overview
+
+By incorporating an intelligent information assistant, ASSISTRAG aims to boost the potential of LLMs in handling complex reasoning tasks. As illustrated in Figure 2, this framework consists of two main components: a frozen main LLM tasked with generating answers based on the information provided, and a trainable assistant LLM responsible for information management. This assistant LLM is designed with two tasks: Memory Management involves storing interactions with the main LLM and retrieving relevant past memories to assist in addressing similar questions. Knowledge Management encompasses retrieving relevant information from external databases and processing it to support the main LLM in formulating responses to new questions.
+
+To effectively accomplish these tasks, we have endowed the assistant with four key capabilities:
+
+- Tool Usage: Retrieving relevant information from internal memory and external knowledge bases.  
+- Action Execution: Reasoning, analyzing information need, and extracting knowledge.  
+- Memory Building: Recording essential knowledge and reasoning patterns from past interactions.
+
+![](images/817c230cf9645055a70c2795fc9c8c2964cffcc9e024c3938525cfc8e2a3e4fe.jpg)  
+Figure 2: Overview of ASSISTRAG. ASSISTRAG enhances LLMs by providing an intelligent information assistant. Endowed with the ability of tool usage, action execution, memory building and plan specification, it can achieve effective memory and knowledge management.
+
+- Plan Specification: Determining the necessity of assistance during answer generation.
+
+These four capabilities synergize to ensure that ASSISTRAG offers precise and comprehensive support to the main LLM. In the following sections, we will provide a detailed examination of the role and implementation of each capability.
+
+# 3.2.1 Memory Management
+
+Effective memory management is crucial for enhancing the main LLM's performance by storing and retrieving historical interactions. This functionality comprises two key processes: capturing new insights and retrieving previously stored information. This stage activates the following three capabilities of AssistRAG:
+
+- Action I: Note Taking. This action  $\mathcal{F}_{\mathrm{NT}}$  records critical information and the reasoning patterns behind each historical interaction. Given the historical interactions of the main LLM, which include question  $q$ , reference  $r$ , and answer  $y$ , the assistant is tasked with memorizing the key reasoning process behind the answer into the memory slot  $m_q$ :  $m_q \longleftarrow \mathcal{F}_{\mathrm{NT}}(q,r,y)$ . The accumulation of memory slots for all prior questions forms the assistant's memory  $\mathcal{M}$ , which is utilized for subsequent memory retrieval.  
+- Tool I: Memory Retriever. Given the question  $q$  and the assistant's memory  $\mathcal{M}$ , the memory retriever retrieves historically relevant memories, represented as:  $\mathcal{M}_q \longleftarrow \mathcal{R}_{\mathrm{memory}}(q, \mathcal{M})$ .  
+- Plan I: Assessing the Usefulness of Retrieved Memory. If the question is entirely new, the retrieved memories may not only be unhelpful but also negatively impact the main LLM's response. Therefore, we implement this plan to determine whether the retrieved memory slots should be provided to the main LLM. Using a prompt, the assistant evaluates whether the retrieved memories are beneficial for answering the current question. Only if the answer is affirmative will the retrieved memories be supplied to the main LLM.
+
+![](images/7b71ff88b00ac50bd78031a490affadadba1197bb00699529616b5b4958155c4.jpg)  
+Figure 3: Training framework of ASSISTRAG. It undergoes a two-stage training pipeline through curriculum assistant learning and reinforced preference optimization.
+
+# 3.2.2 Knowledge Management
+
+Effective knowledge management is essential for an intelligent information assistant, involving the efficient gathering of necessary knowledge to support the main LLM. This process includes analyzing the information needs of the main LLM, retrieving relevant knowledge, and integrating it. This process involves the following four capabilities of AssistRAG:
+
+- Action II: Question Decomposition. This action  $\mathcal{F}_{\mathrm{QD}}$  aims to break down the current question into multiple sub-queries to facilitate the retrieval of knowledge across various aspects:  $Q^{\prime}\longleftarrow \mathcal{F}_{\mathrm{QD}}(q)$  where  $Q^{\prime}$  represents a series of sub-queries derived from the question  $q$ .  
+- Tool II: Knowledge Retriever. Utilizing a batch of sub-queries  $Q'$ , the knowledge retriever sources relevant documents from external knowledge bases  $D$ , denoted as:  $D_{Q'} \leftarrow \mathcal{R}_{\text{knowledge}}(Q', D)$ .  
+- Action III: Knowledge Extraction. This action  $\mathcal{F}_{\mathrm{KE}}$  involves extracting essential knowledge from a large number of retrieved documents. Given the question  $q$  and the retrieved documents  $D_{Q'}$ , the assistant is responsible for extracting the relevant knowledge  $\mathcal{K}_q$  from the search results:  $\mathcal{K}_q \longleftarrow \mathcal{F}_{\mathrm{KE}}(q, D_{Q'})$ .  
+- Plan II: Evaluating the Relevance of Extracted Knowledge. To ensure the accuracy and relevance of the information provided to the main LLM, this plan determines whether the extracted knowledge should be included in the response generation process. Similarly, we prompt the assistant to assess whether the extracted knowledge is relevant to the current question.
+
+To summarize, we have endowed the assistant with memory capabilities and designed three actions, two retrieval tools, and two planning strategies integral to ASSISTRAG. Next, we will introduce the training strategies developed for ASSISTRAG, focusing on enhancing the accuracy of these actions and ensuring their compatibility with the main LLM.
+
+# 3.3 AssistRAG Training
+
+The training objectives of ASSISTRAG focus on two main goals: (1) enhancing the effectiveness of each action within the RAG process, and (2) ensuring that its outputs align with the main LLM's requirements. To achieve these two goals, as depicted in Figure 3, we implement curriculum-based assistant learning and reinforced preference optimization to optimize the training of ASSISTRAG.
+
+Several studies have demonstrated that GPT-4 can achieve human-like annotation accuracy [32]. Based on this consideration, we leverage it to collect training data for the three actions. The supervised training samples for each specific action are cataloged as  $\mathcal{C}_{\mathrm{QD}}$ ,  $\mathcal{C}_{\mathrm{KE}}$ , and  $\mathcal{C}_{\mathrm{NT}}$ , preparing these for the assistant's subsequent training phase.
+
+# 3.3.1 Curriculum Assistant Learning
+
+Motivation. The tasks of question decomposition, knowledge extraction, and note-taking are interconnected, each contributing towards navigating the reasoning path from a question to its answer.
+
+To equip the assistant with a comprehensive understanding of the RAG process, we devise a step-wise curriculum assistant learning strategy designed to evolve from simpler to more complex tasks to foster a deepened mastery over time.
+
+Training Objective. The curriculum learning strategy integrates training samples across three sequential phases  $\mathcal{C}_{\mathrm{QD}} \to \mathcal{C}_{\mathrm{KE}} \to \mathcal{C}_{\mathrm{NT}}$ . Each phase dedicates  $60\%$  of its focus to the task at hand, with the remaining  $40\%$  evenly divided between the other two tasks. The assistant's training employs the standard next token prediction target based on the training set  $D_{gen}$ :
+
+$$
+\mathbb {E} _ {(x, y) \sim D _ {\text {g e n}}} \log p _ {\phi} (y | x), \tag {1}
+$$
+
+where  $\phi$  symbolizes the generator's adjustable parameters, and  $(x,y)$  is the pair of input and expected output. This methodical training strategy is designed to progressively refine the assistant's proficiency in each component of the RAG process, thereby boosting its effectiveness.
+
+# 3.3.2 Reinforced Preference Optimization
+
+Motivation. Although ASSISTRAG effectively handles RAG tasks after assistant learning, its output may sometimes not fully meet the downstream LLM's specific needs. To enhance integration, we implement reinforced preference optimization, a technique that adjusts the assistant's output based on feedback from the main LLM, ensuring tailored assistance that better meets its requirements.
+
+Training Objective. To optimize the assistant for better alignment with the main LLM, we adopt Direct Preference Optimization (DPO) [33]. This approach involves generating two sets of references, one from externally retrieved knowledge and the other generated by the assistant itself. The main LLM evaluates these sets, with a preference determined by comparing the F1 scores of its responses against correct answers. For reinforced preference optimization, we leverage the DPO algorithm's optimization objective, utilizing paired preference data  $D_{dpo}$ :
+
+$$
+\mathbb {E} _ {(x, y _ {1}, y _ {2}) \sim D _ {d p o}} \left[ \log \sigma \left(\log r _ {\theta} (x, y _ {1}) - \log r _ {\theta} (x, y _ {2})\right) \right] \tag {2}
+$$
+
+where  $r_{\theta}(x,y_i) = \beta \frac{\pi_{\theta}(y_i|x)}{\pi_{\mathrm{ref}}(y_i|x)}$  is the reward implicitly defined by the language model  $\pi_{\theta}$  and the reference model  $\pi_{\mathrm{ref}}$ . This reinforced training stage enhances the assistant's capability to deliver assistance that aligns more closely with the main LLM's preferences, enhancing overall efficacy.
+
+# 3.4 ASSISTRAG Inference
+
+Upon completing its training phase, ASSISTRAG initiates its inference process through three steps:
+
+Information Retrieval and Integration. At this initial stage, ASSISTRAG first activates Action II to understand the main LLM's information needs. It then uses Tool I and Tool II to retrieve relevant information from internal memory and external knowledge bases, respectively. Subsequently, it invokes Action III to extract essential knowledge from the retrieved documents.
+
+Decision Making. In this stage, ASSISTRAG decides whether to provide the retrieved memories and extracted knowledge to the main LLM. It activates Plan I and Plan II to evaluate the relevance and usefulness of the retrieved memories and knowledge for the current question. If the assistant deems them helpful, they are supplied to the main LLM to aid in answer generation.
+
+Answer Generation and Memory Updating. In the final phase, we prompt the main LLM to generate an answer based on the question, its internal knowledge, and the information provided by the assistant. Following this, ASSISTRAG activates Action I to utilize its note-taking feature, capturing crucial reasoning steps from the interaction and incorporating them into its memory. This ensures the assistant's knowledge base remains up-to-date.
+
+# 4 Experimental Setup
+
+# 4.1 Datasets and Evaluation Metrics
+
+In this study, we evaluate the effectiveness of our proposed method through experiments on three intricate question-answering datasets: HotpotQA [34], 2WikiMultiHopQA [35], and Bamboogle [6]. These datasets, all derived from Wikipedia documents, provide a uniform corpus and retrieval
+
+Table 1: The evaluation results for three datasets. The "Main LLM" indicates the LLM employed for question answering. The best results are shown in bold, while the second-best results are underlined.  
+
+<table><tr><td rowspan="2">Method</td><td rowspan="2">Main LLM</td><td colspan="3">HotpotQA</td><td colspan="3">2Wiki</td><td colspan="3">Bamboogle</td></tr><tr><td>EM</td><td>F1</td><td>Prec.</td><td>EM</td><td>F1</td><td>Prec.</td><td>EM</td><td>F1</td><td>Prec.</td></tr><tr><td colspan="11">Baselines without retrieval</td></tr><tr><td>CloseBook</td><td>LLaMA2-chat7B</td><td>13.2</td><td>18.4</td><td>17.8</td><td>14.4</td><td>18.2</td><td>17.8</td><td>10.4</td><td>16.3</td><td>16.7</td></tr><tr><td>CloseBook</td><td>ChatGLM6B</td><td>15.6</td><td>20.4</td><td>19.9</td><td>15.8</td><td>19.5</td><td>20.0</td><td>12.6</td><td>17.6</td><td>16.9</td></tr><tr><td>CloseBook</td><td>ChatGPT3.5</td><td>20.0</td><td>25.8</td><td>26.4</td><td>21.6</td><td>25.7</td><td>24.5</td><td>14.4</td><td>22.0</td><td>22.3</td></tr><tr><td colspan="11">Baselines with retrieval</td></tr><tr><td>Naive RAG</td><td>LLaMA2-chat7B</td><td>18.2</td><td>23.0</td><td>22.5</td><td>17.4</td><td>23.7</td><td>22.8</td><td>15.2</td><td>20.4</td><td>20.3</td></tr><tr><td>Naive RAG</td><td>ChatGLM6B</td><td>21.8</td><td>27.2</td><td>25.8</td><td>17.8</td><td>25.0</td><td>25.2</td><td>15.8</td><td>21.1</td><td>20.8</td></tr><tr><td>Naive RAG</td><td>ChatGPT3.5</td><td>24.6</td><td>33.0</td><td>34.5</td><td>23.8</td><td>30.2</td><td>31.1</td><td>18.4</td><td>24.4</td><td>24.7</td></tr><tr><td>ReAct</td><td>ChatGPT3.5</td><td>26.8</td><td>41.7</td><td>42.6</td><td>25.0</td><td>33.0</td><td>31.6</td><td>28.8</td><td>37.7</td><td>38.2</td></tr><tr><td>IRCoT</td><td>ChatGPT3.5</td><td>31.4</td><td>40.3</td><td>41.6</td><td>30.8</td><td>42.6</td><td>42.3</td><td>30.2</td><td>38.8</td><td>37.9</td></tr><tr><td>Self-Ask</td><td>ChatGPT3.5</td><td>28.2</td><td>43.1</td><td>44.8</td><td>28.6</td><td>37.5</td><td>42.8</td><td>23.2</td><td>32.8</td><td>30.8</td></tr><tr><td>SELF-RAG</td><td>SELF-RAG7B</td><td>31.0</td><td>42.4</td><td>42.3</td><td>35.0</td><td>40.7</td><td>41.0</td><td>29.8</td><td>35.5</td><td>37.8</td></tr><tr><td>LLMLingua</td><td>ChatGPT3.5</td><td>28.2</td><td>40.2</td><td>40.0</td><td>29.4</td><td>38.6</td><td>37.8</td><td>25.2</td><td>31.3</td><td>30.8</td></tr><tr><td>ASSISTRAG</td><td>LLaMA2-chat7B</td><td>32.4</td><td>41.5</td><td>42.6</td><td>36.2</td><td>41.0</td><td>40.5</td><td>33.0</td><td>39.6</td><td>38.7</td></tr><tr><td>ASSISTRAG</td><td>ChatGLM6B</td><td>33.0</td><td>42.4</td><td>43.5</td><td>38.0</td><td>43.2</td><td>42.8</td><td>32.8</td><td>39.8</td><td>39.0</td></tr><tr><td>ASSISTRAG</td><td>ChatGPT3.5</td><td>34.4</td><td>44.8</td><td>46.5</td><td>39.6</td><td>45.6</td><td>45.7</td><td>34.6</td><td>41.4</td><td>41.1</td></tr></table>
+
+mechanisms to supply external references for LLMs. To manage costs, we follow a similar approach as previous studies [17] by selecting a maximum of 500 questions from each dataset's validation set for our experiments. To assess the performance, we employ Exact Match (EM), F1 score, and Precision (Prec.).
+
+# 4.2Baselines
+
+We benchmark our model against three foundational models: LLaMA2-chat $_{7B}$  [36], ChatGLM $_{3_{6B}}$  [37], and ChatGPT, assessing their performance in Closebook, Naive RAG, and ASSISTRAG settings. To compare our RAG framework with other RAG models, we include advanced prompt-based methods ReAct [29], IRCoT [7], Self-Ask [6], an SFT-based RAG model Self-RAG [8], and a knowledge extraction model LLMLingua [38]. We ensure a fair comparison by standardizing evaluation conditions across all models.
+
+# 4.3 Implementation Details
+
+Training Settings: In the assistant learning phase, we create a dataset comprising 50k training samples based on instruction-following input-output pairs across three distinct task types. The assistant LLM, which is based on ChatGLM3-6B [37], is fully fine-tuned across all parameters. The training is conducted over 2 epochs with a batch size of 32 and a peak learning rate of 2e-5. For the preference optimization phase, we employ a DPO trainer and uses LoRA for fine-tuning, with a learning rate set to 1e-5 and the training duration extended to 2 epochs. The training code and data can be accessed at https://github.com/smallporridge/AssistRAG.
+
+Inference Settings: For employing ChatGPT, we opt for the gpt-35-turbo-16k model, accessed through its API at a temperature setting of 0. We use a Wikipedia dump as our document corpus, breaking down articles into 100-token passages. For both memory and knowledge retrieval, we deploy the off-the-shelf LLM Embedder [39] to fetch up to 5 documents per input.
+
+# 5 Results and Analysis
+
+# 5.1 Main Results
+
+The main results are presented in Table 1. Several key findings can be observed as follows:
+
+Comparison among Different Reasoning Types. Applying our ASSISTRAG framework to ChatGPT demonstrates a significant performance advantage over other models across all datasets. Specifi
+
+cally, ASSISTRAG showcases its adaptability with different base LLMs, consistently outperforming them in both Closebook and Naive RAG settings. This result highlights the advantage of ASSISTRAG in effectively assisting a variety of downstream LLMs. Additionally, our method surpasses contemporary approaches employing prompt engineering or supervised fine-tuning, validating the efficacy of our curriculum assistant learning and reinforced preference optimization training strategies.
+
+Comparison among Different Base LLMs. By comparing the performance of ASSISTRAG across various base LLMs, it is observed that stronger base LLMs yield higher quality responses across all reasoning types. Notably, compared to Naive RAG settings, ASSISTRAG achieves performance improvements of  $78\%$ ,  $51\%$ , and  $40\%$  for LLaMA, ChatGLM, and ChatGPT, respectively. This indicates that ASSISTRAG brings more substantial benefits to weaker base LLMs. A likely reason is that weaker models inherently have less robust noise resistance. Benefiting from the assistant's knowledge extraction capability, the main LLM only receives relevant knowledge to generate answers, leading to improved responses.
+
+# 5.2 Analysis
+
+Ablation Studies. ASSISTRAG integrates memory and knowledge management to support the main LLM, encompassing three actions: note-taking, question decomposition, and knowledge extraction. To evaluate their contribution, we conduct ablation studies by removing each action or freezing the parameters of the assistant. Additionally, we assess the effects of not implementing planning (w/o. Planning), curriculum learning (w/o. Curriculum), and reinforced preference optimization (w/o. DPO) to explore there contribution to the F1 score. Table 2 illustrates that removing or freezing any of the ASSISTRAG's actions results in decreased performance, underscoring the value of the assistant learning in the RAG context. Notably, maintaining these actions in a frozen state still
+
+outperforms completely removing them, highlighting their critical role in the RAG process. Concerning training strategies, the absence of planning, curriculum learning, and preference optimization slightly diminishes performance, indicating that a structured progression from simple to complex tasks and aligning with downstream LLM preferences contribute to the assistant providing more accurate information to the downstream LLM, thereby enhancing the accuracy of LLM responses.
+
+Table 2: Ablation Studies of ASSISTRAG.  
+
+<table><tr><td>Method</td><td>Hotpot.</td><td>2Wiki</td><td>Bamb.</td></tr><tr><td colspan="4">Memory Management</td></tr><tr><td>Remove FN</td><td>40.2</td><td>42.0</td><td>39.0</td></tr><tr><td>Freeze FN</td><td>41.3</td><td>43.1</td><td>39.9</td></tr><tr><td colspan="4">Knowledge Management</td></tr><tr><td>Remove FQD</td><td>39.5</td><td>37.8</td><td>37.0</td></tr><tr><td>Freeze FQD</td><td>41.3</td><td>40.3</td><td>37.8</td></tr><tr><td>Remove FKE</td><td>39.2</td><td>38.5</td><td>38.7</td></tr><tr><td>Freeze FKE</td><td>40.9</td><td>39.7</td><td>39.4</td></tr><tr><td>ASSISTRAG</td><td>44.8</td><td>45.6</td><td>41.4</td></tr><tr><td>w/o. Planning</td><td>43.0</td><td>44.5</td><td>40.7</td></tr><tr><td>w/o. Curriculum</td><td>43.2</td><td>44.3</td><td>40.0</td></tr><tr><td>w/o. DPO</td><td>42.5</td><td>43.2</td><td>40.5</td></tr></table>
+
+Token Usage of Different Methods. A notable benefit of our ASSISTRAG framework is its efficiency in preprocessing extensive information prior to engaging the main LLM. This not only enhances the inference speed of the main LLM but also minimizes token usage, which is particularly valuable when utilizing online API services like ChatGPT. We select a representative model for each reasoning type to compare their token consumption in terms of online API and SFT model, alongside their performance metrics F1 on the 2Wiki dataset. The results, as outlined in Table 3, reveal significant differences in token
+
+usage among the methods. Prompt-based RAG methods tend to consume a large number of tokens due to their dependency on multiple API calls. On the other hand, SFT-based methods are more economical in terms of API calls but require retraining for new LLM adaptations. In contrast, our ASSISTRAG demonstrates a balanced approach by reducing API token costs while maintaining adaptability across different LLMs without the need for retraining. This method not only lowers the overall costs associated with API usage but also achieves superior performance.
+
+Table 3: Token usage comparison. "tok." is the average token input length preceding the answer.  
+
+<table><tr><td>Method</td><td>API tok.</td><td>SFT tok.</td><td>F1</td></tr><tr><td>CloseBook</td><td>18</td><td>0</td><td>25.7</td></tr><tr><td>Naive RAG</td><td>782</td><td>0</td><td>30.2</td></tr><tr><td>IR-CoT</td><td>1890</td><td>0</td><td>42.6</td></tr><tr><td>SELF-RAG</td><td>0</td><td>1456</td><td>40.7</td></tr><tr><td>LLMLingua</td><td>176</td><td>780</td><td>38.6</td></tr><tr><td>ASSISTRAG</td><td>90</td><td>1528</td><td>45.6</td></tr></table>
+
+![](images/30ab2ae03d8457636c0739fb7288303e419939647a3a3a3a3c04fa060c4645d0.jpg)  
+Figure 4: The relationships between inference time, cost, and F1 accuracy for different methods.
+
+![](images/7a470db09515601a696081a965c0207384e17200fe6fb8bcdae97307b1a3920b.jpg)
+
+![](images/2c16d84177b4527ba5b39b73a2e96d9464d70938731b8ea01cfee0fe17368d5e.jpg)
+
+![](images/6e4e5626710039ebeb8d3a81028dabb342d8657e48ea54578e59ccef5870564f.jpg)  
+(a) HotpotQA
+
+![](images/a74dc0089a1aaf957bc2f2bd32bb449cf60003413142eef85507fb92613a37fe.jpg)  
+(b) 2WikiMultiHopQA
+
+![](images/cba2e1a749f0d97fa071451635e4d49cf91ee9114137a674d16a18763fb89a60.jpg)  
+(c) Bamboogle  
+Figure 5: Performance with different training data sizes.
+
+Accuracy, Efficiency, and Cost Analysis. When evaluating an algorithm's value, we consider three dimensions: accuracy, efficiency, and cost. To compare different RAG methods, we calculate each method's F1 accuracy, inference speed, and cost. We then illustrate the relationships between these variables using three separate plots. From Figure 4, we observe that ASSISTRAG stands out as the most balanced method, achieving the highest F1 accuracy of 45.6, while maintaining a comparable inference time of 5.73 seconds and a low cost of 0.009 cents per question. Although methods like IR-CoT show higher costs and longer inference times, they do not surpass ASSISTRAG in accuracy. These results demonstrate that ASSISTRAG is advantageous for applications requiring high accuracy without incurring significant costs.
+
+Impact of Dataset Size and Training Strategy. We examine the effect of training dataset size on model performance by creating subsets of 5k, 10k, and 20k instances from our original 50k training samples. These subsets fine-tune three separate model versions, evaluated on three datasets, and compared to the model trained on the full 50k dataset. We also compare the impact of curriculum learning and DPO training by evaluating performance with each strategy omitted. Figure 5 shows a clear performance improvement for ASSISTRAG as the training dataset size increases from 5k to 50k across both datasets, indicating potential further gains with larger datasets. The curriculum learning strategy performs better than random mixed training, especially with smaller datasets, showing its advantage when data is limited. In contrast, DPO training benefits more from larger datasets, likely because more data enables better training for high-quality data generation.
+
+Case Study. Table 4 is a case study that highlights the capabilities of AssistRAG in processing and answering complex comparative questions. In this case study, the main question "Who is older, Danny Green or James Worthy?" is systematically broken down by ASSISTRAG into simpler sub-questions regarding the birth dates of both individuals. This decomposition enables targeted information retrieval, allowing the system to accurately locate and extract relevant birth date information from the corpus. ASSISTRAG effectively retrieves multiple pieces of information, including relevant and irrelevant entries, and filters through them to extract the necessary facts. For instance, it identifies the birth dates of both Danny Green and James Worthy and ignores unrelated entries, such as those concerning another individual named Danny Green who is a boxer. The memory retrieval capability is then utilized to access previous similar questions and their answers, which aids in reinforcing
+
+Table 4: Case Study of AssistRAG.  
+
+<table><tr><td>Question: Who is older, Danny Green or James Worthy?</td></tr><tr><td>Question Decomposition: AssistRAG initially breaks down the main question into sub-questions to facilitate targeted information retrieval:
+- When was Danny Green born?
+- When was James Worthy born?</td></tr><tr><td>Knowledge Retrieval: AssistRAG retrieves relevant information from the corpus:
+- [1] James Worthy ...
+- [2] Danny Green (basketball) ...
+- [3] Danny Green (boxer) ...
+- ...</td></tr><tr><td>Knowledge Extraction: From the retrieved information, AssistRAG extracts key facts:
+- Danny Green (born June 22, 1987) is an American professional basketball player who last played for the Philadelphia 76ers of the National Basketball Association (NBA). In his NBA career, Green has played for six teams. As of 2020, Green is one of just four players in history to have won NBA championships with three different teams.
+- James Worthy James Ager Worthy (born February 27, 1961) is an American former professional basketball player who is currently a commentator, television host, and analyst. A standout at the University of North Carolina, the 6 ft 9 in (2.06 m) small forward shared College Player of the Year honors en route to leading the Tar Heels to the 1982 NCAA Championship.</td></tr><tr><td>Memory Retrieval: AssistRAG retrieves similar questions and answers from the memory:
+1. question: Who is older, Danny Jones or David Coverdale? answer: David Coverdale was born on 22 September 1951, while Danny Jones was born on 12 March 1986. Since David Coverdale was born earlier, he is older than Danny Jones. So the answer is David Coverdale.
+2. question: Who is older, Danny Shirley or Kevin Parker? answer: Danny Shirley was born on August 12, 1956, while Kevin Parker was born on January 20, 1986. Since Danny Shirley was born earlier than Kevin Parker, he is older. So the answer is Danny Shirley.
+3. ...</td></tr><tr><td>Planning: Both retrieved memory and extracted knowledge are useful for the question.</td></tr><tr><td>Final Output: Based on the analysis, Main LLM outputs:
+- James Worthy was born on February 27, 1961, while Danny Green was born on June 27, 1987. Since James Worthy was born earlier, he is older. So the answer is James Worthy.</td></tr><tr><td>Final Answer: James Worthy</td></tr><tr><td>Ground-truth: James Worthy</td></tr></table>
+
+the current decision-making process. This step provides context and supports consistency in the reasoning pattern applied by the system. Combining the extracted knowledge and retrieved memory, ASSISTRAG plans the final response by confirming the usefulness of both sources of information. The Main LLM then generates a comprehensive answer, stating that James Worthy, born on February 27, 1961, is older than Danny Green, born on June 22, 1987. This case study showcases ASSISTRAG's ability to manage complex tasks by leveraging its multi-step reasoning process, ensuring the delivery of accurate and reliable answers. The superior performance in accurately answering comparative questions is a testament to the system's robust architecture and its effective integration of question decomposition, information retrieval, knowledge extraction, planning, and memory capabilities.
+
+# 6 Conclusion
+
+In this study, we introduce ASSISTRAG to augment LLMs with an intelligent information assistant, significantly improving their ability to tackle tasks requiring complex reasoning. By implementing a two-stage training methodology that integrates curriculum assistant learning with reinforced preference optimization, we enhance the assistant's skills in memory and knowledge management. Experiments demonstrate that ASSISTRAG surpasses existing baselines with a notable margin. Looking ahead, we plan to further expand the assistant's skills to include long-text processing [40] and personalized support [41], thereby providing more effective assistance to the main LLM.
+
+# References
+
+[1] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll L. 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. Training language models to follow instructions with human feedback. In NeurIPS, 2022.  
+[2] Chunting Zhou, Graham Neubig, Jiatao Gu, Mona T. Diab, Francisco Guzmán, Luke Zettlemoyer, and Marjan Ghazvininejad. Detecting hallucinated content in conditional neural sequence generation. In ACL/IJCNLP (Findings), volume ACL/IJCNLP 2021 of Findings of ACL, pages 1393–1404. Association for Computational Linguistics, 2021.  
+[3] Weihang Su, Changyue Wang, Qingyao Ai, Yiran Hu, Zhijing Wu, Yujia Zhou, and Yiqun Liu. Unsupervised real-time hallucination detection based on the internal states of large language models. arXiv preprint arXiv:2403.06448, 2024.  
+[4] Zheng Liu, Yujia Zhou, Yutao Zhu, Jianxun Lian, Chaozhuo Li, Zhicheng Dou, Defu Lian, and Jian-Yun Nie. Information retrieval meets large language models. In Companion Proceedings of the ACM Web Conference 2024, WWW '24, page 1586-1589, 2024.  
+[5] Gautier Izacard and Edouard Grave. Leveraging passage retrieval with generative models for open domain question answering. In EACL, pages 874-880. Association for Computational Linguistics, 2021.  
+[6] Ofir Press, Muru Zhang, Sewon Min, Ludwig Schmidt, Noah A. Smith, and Mike Lewis. Measuring and narrowing the compositionality gap in language models. CoRR, abs/2210.03350, 2022.  
+[7] Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions. arXiv preprint arXiv:2212.10509, 2022.  
+[8] Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. Self-rag: Learning to retrieve, generate, and critique through self-reflection. arXiv preprint arXiv:2310.11511, 2023.  
+[9] Gautier Izacard, Patrick S. H. Lewis, Maria Lomeli, Lucas Hosseini, Fabio Petroni, Timo Schick, Jane Dwivedi-Yu, Armand Joulin, Sebastian Riedel, and Edouard Grave. Few-shot learning with retrieval augmented language models. CoRR, abs/2208.03299, 2022.  
+[10] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems, 33:9459-9474, 2020.  
+[11] Yujia Zhou, Yan Liu, Xiaoxi Li, Jiajie Jin, Hongjin Qian, Zheng Liu, Chaozhuo Li, Zhicheng Dou, Tsung-Yi Ho, and Philip S Yu. Trustworthiness in retrieval-augmented generation systems: A survey. arXiv preprint arXiv:2409.10102, 2024.  
+[12] Omar Khattab, Keshav Santhanam, Xiang Lisa Li, David Hall, Percy Liang, Christopher Potts, and Matei Zaharia. Demonstrate-search-predict: Composing retrieval and language models for knowledge-intensive NLP. CoRR, abs/2212.14024, 2022.  
+[13] Urvashi Khandelwal, Omer Levy, Dan Jurafsky, Luke Zettlemoyer, and Mike Lewis. Generalization through memorization: Nearest neighbor language models. In ICLR. OpenReview.net, 2020.  
+[14] Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Rutherford, Katie Millican, George van den Driessche, Jean-Baptiste Lespiau, Bogdan Damoc, Aidan Clark, Diego de Las Casas, Aurelia Guy, Jacob Menick, Roman Ring, Tom Hennigan, Saffron Huang, Loren Maggiore, Chris Jones, Albin Cassirer, Andy Brock, Michela Paganini, Geoffrey Irving, Oriol Vinyals, Simon Osindero, Karen Simonyan, Jack W. Rae, Erich Elsen, and Laurent Sifre. Improving language models by retrieving from trillions of tokens. In ICML, volume 162 of Proceedings of Machine Learning Research, pages 2206-2240. PMLR, 2022.
+
+[15] Ori Ram, Yoav Levine, Itay Dalmedigos, Dor Muhlgay, Amnon Shashua, Kevin Leyton-Brown, and Yoav Shoham. In-context retrieval-augmented language models. CoRR, abs/2302.00083, 2023.  
+[16] Zhihong Shao, Yeyun Gong, Yelong Shen, Minlie Huang, Nan Duan, and Weizhu Chen. Enhancing retrieval-augmented large language models with iterative retrieval-generation synergy. CoRR, abs/2305.15294, 2023.  
+[17] Zhengbao Jiang, Frank F. Xu, Luyu Gao, Zhiqing Sun, Qian Liu, Jane Dwivedi-Yu, Yiming Yang, Jamie Callan, and Graham Neubig. Active retrieval augmented generation. CoRR, abs/2305.06983, 2023.  
+[18] Tushar Khot, Harsh Trivedi, Matthew Finlayson, Yao Fu, Kyle Richardson, Peter Clark, and Ashish Sabharwal. Decomposed prompting: A modular approach for solving complex tasks. In ICLR. OpenReview.net, 2023.  
+[19] Weijia Shi, Sewon Min, Michihiro Yasunaga, Minjoon Seo, Rich James, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. REPLUG: retrieval-augmented black-box language models. CoRR, abs/2301.12652, 2023.  
+[20] Hongyin Luo, Yung-Sung Chuang, Yuan Gong, Tianhua Zhang, Yoon Kim, Xixin Wu, Danny Fox, Helen Meng, and James Glass. Sail: Search-augmented instruction learning. arXiv preprint arXiv:2305.15225, 2023.  
+[21] Xinbei Ma, Yeyun Gong, Pengcheng He, Hai Zhao, and Nan Duan. Query rewriting for retrieval-augmented large language models. arXiv preprint arXiv:2305.14283, 2023.  
+[22] Fangyuan Xu, Weijia Shi, and Eunsol Choi. Recomp: Improving retrieval-augmented lms with compression and selective augmentation. arXiv preprint arXiv:2310.04408, 2023.  
+[23] Jiajie Jin, Yutao Zhu, Yujia Zhou, and Zhicheng Dou. Bider: Bridging knowledge inconsistency for efficient retrieval-augmented llms via key supporting evidence. arXiv preprint arXiv:2402.12174, 2024.  
+[24] Xiaoxi Li, Yujia Zhou, and Zhicheng Dou. Unigen: A unified generative framework for retrieval and question answering with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 8688-8696, 2024.  
+[25] Yujia Zhou, Zheng Liu, Jiajie Jin, Jian-Yun Nie, and Zhicheng Dou. Metacognitive retrieval-augmented large language models. In Proceedings of the ACM on Web Conference 2024, pages 1453-1463, 2024.  
+[26] Significant Gravitas. AutoGPT.  
+[27] Timo Schick, Jane Dwivedi-Yu, Roberto Dessi, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools. arXiv preprint arXiv:2302.04761, 2023.  
+[28] Sirui Hong, Xiawu Zheng, Jonathan Chen, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, et al. Metagpt: Meta programming for multi-agent collaborative framework. arXiv preprint arXiv:2308.00352, 2023.  
+[29] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629, 2022.  
+[30] Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina Kim, Christopher Hesse, Shantanu Jain, Vineet Kosaraju, William Saunders, et al. Webgpt: Browser-assisted question-answering with human feedback. arXiv preprint arXiv:2112.09332, 2021.  
+[31] Noah Shinn, Federico Cassano, Beck Labash, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflection: Language agents with verbal reinforcement learning. arXiv preprint arXiv:2303.11366, 2023.
+
+[32] Jaromir Savelka, Kevin D Ashley, Morgan A Gray, Hannes Westermann, and Huihui Xu. Can gpt-4 support analysis of textual data in tasks requiring highly specialized domain expertise? arXiv preprint arXiv:2306.13906, 2023.  
+[33] Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D Manning, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. arXiv preprint arXiv:2305.18290, 2023.  
+[34] Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W Cohen, Ruslan Salakhutdinov, and Christopher D Manning. Hotpotqa: A dataset for diverse, explainable multi-hop question answering. arXiv preprint arXiv:1809.09600, 2018.  
+[35] Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, and Akiko Aizawa. Constructing A multi-hop QA dataset for comprehensive evaluation of reasoning steps. In COLING, pages 6609-6625. International Committee on Computational Linguistics, 2020.  
+[36] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi ..., and Thomas Scialom. Llama 2: Open foundation and fine-tuned chat models. CoRR, abs/2307.09288, 2023.  
+[37] Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. 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, 2022.  
+[38] Huiqiang Jiang, Qianhui Wu, Chin-Yew Lin, Yuqing Yang, and Lili Qiu. Lmlingua: Compressing prompts for accelerated inference of large language models. arXiv preprint arXiv:2310.05736, 2023.  
+[39] Peitian Zhang, Shitao Xiao, Zheng Liu, Zhicheng Dou, and Jian-Yun Nie. Retrieve anything to augment large language models. arXiv preprint arXiv:2310.07554, 2023.  
+[40] Hongjin Qian, Zheng Liu, Kelong Mao, Yujia Zhou, and Zhicheng Dou. Grounding language model with chunking-free in-context retrieval. arXiv preprint arXiv:2402.09760, 2024.  
+[41] Yujia Zhou, Qiannan Zhu, Jiajie Jin, and Zhicheng Dou. Cognitive personalized search integrating large language models with an efficient memory mechanism. In Proceedings of the ACM on Web Conference 2024, pages 1464–1473, 2024.
+
+# A Appendix / supplemental material
+
+# A.1 Implementation Details
+
+To achieve the effective functioning of ASSISTRAG, we meticulously designed and executed the training and inference phases, ensuring optimal use of computational resources and robust fine-tuning processes.
+
+# A.1.1 Training Settings
+
+The training of the assistant LLM involved a two-phase approach: Assistant Learning and Preference Optimization.
+
+# 1. Assistant Learning Phase:
+
+- Dataset Creation: We created a dataset comprising 50,000 training samples based on instruction-following input-output pairs. These pairs were categorized across three distinct task types: question decomposition, note-taking, and knowledge extraction.  
+- Model Architecture: The assistant LLM is based on ChatGLM3-6B [37], a state-of-the-art language model known for its robust performance in various NLP tasks.  
+- Training Procedure: The assistant LLM was fully fine-tuned across all parameters over 2 epochs, with a batch size of 32 and a peak learning rate of 2e-5. This phase focused on enhancing the model's ability to decompose complex queries, take notes, and extract relevant knowledge.
+
+Table 5: Training samples of note-taking.  
+
+<table><tr><td>Input
+You are given 1) the question, 2) the answer, 3) the supporting facts where the answer can be derived. You are supposed to figure out the reasoning process towards the answer step-by-step without other content. Be concise and direct.
+Question: American politician Joe Heck ran unsuccessfully against Democrat Catherine Cortez Masto, a woman who previously served as the 32nd Attorney General of where?
+Answer: Nevada
+Supporting Facts: Joseph John Heck (born October 30, 1961) is an American politician, physician, and U.S. Army Brigadier General who had served as the U.S. Representative for Nevada&#x27;s 3rd congressional district from 2011 to 2017. He ran unsuccessfully against Democrat Catherine Cortez Masto in the general election for the open Nevada United States Senate seat in 2016. She previously served as the 32nd Attorney General of Nevada from 2007 to 2015.
+Reasoning:</td></tr><tr><td>Output
+Catherine Cortez Masto previously served as the 32nd Attorney General of Nevada. Joe Heck ran unsuccessfully against her in the general election for the open Nevada United States Senate seat in 2016. Therefore, the answer is Nevada.&quot;</td></tr></table>
+
+# 2. Preference Optimization Phase:
+
+- Optimization Technique: We employed a DPO (Distributed Preference Optimization) trainer to refine the assistant's feedback mechanisms.  
+- Fine-tuning with LoRA: Low-Rank Adaptation (LoRA) was utilized for fine-tuning, which helps in adjusting a subset of model parameters efficiently, reducing the computational load.  
+- Learning Rate and Duration: The learning rate was set to 1e-5, with the training duration extending to 2 epochs.
+
+Training Resources: The entire training process was conducted using 8 A800 GPUs, providing substantial computational power to handle the intensive training tasks efficiently.
+
+# A.1.2 Inference Settings
+
+During the inference phase, the ASSISTRAG system was fine-tuned to operate seamlessly with the main LLM, ensuring effective retrieval and processing of information.
+
+- Model Selection: For the main LLM, we opted for the gpt-35-turbo-16k model, accessed through its API. This model was chosen for its extended context window and advanced capabilities in handling complex queries.  
+- Temperature Setting: The temperature was set to 0, ensuring deterministic outputs which are crucial for consistency in inference tasks.  
+- Document Corpus: A Wikipedia dump was used as the primary document corpus. Articles were segmented into 100-token passages to facilitate efficient retrieval.  
+- Retrieval tool: Both memory and knowledge retrieval processes utilized the off-the-shelf LLM Embedder [39], capable of fetching up to 5 documents per input, ensuring comprehensive coverage of relevant information.
+
+Inference Resources: The inference tasks were efficiently handled using a single A800 GPU, highlighting the system's capability to deliver high performance with minimal resource usage.
+
+# A.1.3 Data Annotation and Samples
+
+The training data annotation process was critical in ensuring the quality and relevance of the dataset. We utilized GPT4-turbo for annotating the dataset, ensuring high accuracy and consistency across the training samples. The dataset included approximately 50k training samples, labeled to cover a wide range of instruction-following tasks.
+
+Instructions and Training Data Samples: The following tables (Table 6, 5, and 7) provide detailed examples of the instruction types and training data samples used in the training phase.
+
+<table><tr><td>Input
+Please generate a series of search queries that can be used to find information relevant to the given question.
+Question: The 1968 Baltimore Orioles season included a replacement of Hank Bauer by the manager who spent how many seasons with the Orioles?.
+Search queries:</td></tr><tr><td>Output Who replaced Hank Bauer as the manager of the 1968 Baltimore Orioles?
+How many seasons did Earl Weaver spend with the Baltimore Orioles?</td></tr></table>
+
+Table 7: Training samples of knowledge extraction.  
+
+<table><tr><td>Input
+Please extract relevant snippets from search results that would be helpful in answering this question.
+Question: Musician and satirist Allie Goertz wrote a song about the The Simpsons character Milhouse, who Matt Groening named after who??.
+Search results:
+- [1] Allie Goertz – Allie Goertz Allison Beth Allie Goertz (born March 2, 1991) is an American musician. Goertz is known for her satirical songs based on various pop culture topics. Her videos are posted on YouTube under the name of Cossbysweater. Subjects of her songs have included the film The Room ...
+- [2] Allie Goertz – and logo of the series. The album was made possible through K Islandarter. She is co-host of Everything&#x27;s Coming Up Simpsons podcast, formerly known as Everything&#x27;s Coming Up Podcast, a Simpsons-focused podcast along with Julia Prescott. Goertz also served as a social media producer for ...
+- [3] ...
+Snippets:</td></tr><tr><td>Output Allie Goertz – Allie Goertz Allison Beth Allie Goertz (born March 2, 1991) is an American musician. Goertz is known for her satirical songs based on various pop culture topics. Milhouse Van Houten Milhouse Mussolini Van Houten is a fictional character in the animated television series The Simpsons, voiced by Pamela Hayden, and created by Matt Groening who named the character after ...</td></tr></table>
+
+# A.2 Error Analysis
+
+We have conducted a comprehensive error analysis on the performance of AssistRAG. To facilitate this analysis, we selected 50 erroneous examples from the HotpotQA dataset and calculated the proportion of each error type:
+
+1. Insufficient Knowledge Retrieval: Instances where the retrieved knowledge does not contain the answer.  
+2. Knowledge Extraction Errors: Cases where the answer is present in the retrieved knowledge, but the assistant fails to extract this information.  
+3. Answer Reasoning Mistakes: Situations where the assistant extracts the correct information but the main LLM produces an incorrect answer.  
+4. Other: Including errors such as non-exact match answers.
+
+From Table 8, our findings indicate that more than half of the errors stem from insufficient knowledge retrieval, which is likely linked to the performance of the retriever and the manner in which questions are reformulated. Additionally, a significant portion of errors are due to reasoning mistakes, highlighting the importance of the main LLM's reasoning capabilities. Given that HotpotQA involves multi-hop question-answering tasks, these findings underscore the high demands placed on reasoning abilities.
+
+# A.3 Limitations
+
+Despite the advancements offered by ASSISTRAG, several limitations warrant consideration. Firstly, relying on an intelligent information assistant introduces additional computational complexity and latency. The two-phase training approach and its operation during inference require substantial computational resources, which may limit the practical application of ASSISTRAG in environments with restricted processing capabilities or where real-time responses are critical.
+
+Table 8: Proportion of each error type in the analysis.  
+
+<table><tr><td>Error Type</td><td>Proportion</td></tr><tr><td>Insufficient Knowledge Retrieval</td><td>58%</td></tr><tr><td>Knowledge Extraction Errors</td><td>12%</td></tr><tr><td>Answer Reasoning Mistakes</td><td>20%</td></tr><tr><td>Other</td><td>10%</td></tr></table>
+
+Secondly, the effectiveness of ASSISTRAG depends on the quality and comprehensiveness of the external knowledge bases and internal memory it accesses. In scenarios where the available data is sparse, outdated, or biased, the assistant's ability to retrieve and integrate relevant information may be compromised, leading to suboptimal or erroneous outputs from the main LLM. This dependence on data quality underscores the need for continuous updates and maintenance of the knowledge sources.
+
+Lastly, the decision-making process during inference, which involves the assistant evaluating the relevance of retrieved information, is inherently difficult. The assistant's ability to accurately determine the necessity and applicability of specific knowledge is crucial for effective support. However, this process is susceptible to errors, particularly in scenarios involving ambiguous or multifaceted queries. Enhancing the precision and reliability of this decision-making mechanism is a key area for further research.
+
+# NeurIPS Paper Checklist
+
+# 1. Claims
+
+Question: Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope?
+
+Answer: [Yes]
+
+Justification: [TODO]
+
+Guidelines:
+
+- The answer NA means that the abstract and introduction do not include the claims made in the paper.  
+- The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers.  
+- The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings.  
+- It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper.
+
+# 2. Limitations
+
+Question: Does the paper discuss the limitations of the work performed by the authors?
+
+Answer: [Yes]
+
+Justification: [TODO]
+
+Guidelines:
+
+- The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper.  
+- The authors are encouraged to create a separate "Limitations" section in their paper.  
+- The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be.  
+- The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated.  
+- The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon.  
+- The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size.  
+- If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness.  
+- While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren't acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations.
+
+# 3. Theory Assumptions and Proofs
+
+Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof?
+
+Answer: [Yes]
+
+# Justification: [TODO]
+
+# Guidelines:
+
+- The answer NA means that the paper does not include theoretical results.  
+- All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced.  
+- All assumptions should be clearly stated or referenced in the statement of any theorems.  
+- The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition.  
+- Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material.  
+- Theorems and Lemmas that the proof relies upon should be properly referenced.
+
+# 4. Experimental Result Reproducibility
+
+Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)?
+
+Answer: [Yes]
+
+Justification: [TODO]
+
+# Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not.  
+- If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable.  
+- Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed.  
+- While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example  
+(a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm.  
+(b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully.  
+(c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset).  
+(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results.
+
+# 5. Open access to data and code
+
+Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material?
+
+Answer: [Yes]
+
+Justification: [TODO]
+
+Guidelines:
+
+- The answer NA means that paper does not include experiments requiring code.  
+- Please see the NeurIPS code and data submission guidelines (https://nips.cc/public/subjects/CodeSubmissionPolicy) for more details.  
+- While we encourage the release of code and data, we understand that this might not be possible, so "No" is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark).  
+- The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https ://nips.cc/public/subjects/CodeSubmissionPolicy) for more details.  
+- The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc.  
+- The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why.  
+- At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable).  
+- Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted.
+
+# 6. Experimental Setting/Details
+
+Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results?
+
+Answer: [Yes]
+
+Justification: [TODO]
+
+Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them.  
+- The full details can be provided either with the code, in appendix, or as supplemental material.
+
+# 7. Experiment Statistical Significance
+
+Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments?
+
+Answer: [No]
+
+Justification: the paper provides evaluation results and error analysis.
+
+Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper.  
+- The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions).  
+- The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.)  
+- The assumptions made should be given (e.g., Normally distributed errors).  
+- It should be clear whether the error bar is the standard deviation or the standard error of the mean.
+
+- It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a  $96\%$  CI, if the hypothesis of Normality of errors is not verified.  
+- For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates).  
+- If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text.
+
+# 8. Experiments Compute Resources
+
+Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments?
+
+Answer: [Yes]
+
+Justification: [TODO]
+
+Guidelines:
+
+- The answer NA means that the paper does not include experiments.  
+- The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage.  
+- The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute.  
+- The paper should disclose whether the full research project required more computer than the experiments reported in the paper (e.g., preliminary or failed experiments that didn't make it into the paper).
+
+# 9. Code Of Ethics
+
+Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines?
+
+Answer: [Yes]
+
+Justification: [TODO]
+
+Guidelines:
+
+- The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics.  
+- If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics.  
+- The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction).
+
+# 10. Broader Impacts
+
+Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed?
+
+Answer: [Yes]
+
+Justification: [TODO]
+
+Guidelines:
+
+- The answer NA means that there is no societal impact of the work performed.  
+- If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact.  
+- Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations.  
+- The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to
+
+generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster.
+
+- The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology.  
+- If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML).
+
+# 11. Safeguards
+
+Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)?
+
+Answer:[Yes]
+
+Justification: [TODO]
+
+Guidelines:
+
+- The answer NA means that the paper poses no such risks.  
+- Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters.  
+- Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images.  
+- We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort.
+
+# 12. Licenses for existing assets
+
+Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected?
+
+Answer: [Yes]
+
+Justification: [TODO]
+
+Guidelines:
+
+- The answer NA means that the paper does not use existing assets.  
+- The authors should cite the original paper that produced the code package or dataset.  
+- The authors should state which version of the asset is used and, if possible, include a URL.  
+- The name of the license (e.g., CC-BY 4.0) should be included for each asset.  
+- For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided.  
+- If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset.  
+- For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided.  
+- If this information is not available online, the authors are encouraged to reach out to the asset's creators.
+
+# 13. New Assets
+
+Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets?
+
+Answer:[Yes]
+
+Justification: [TODO]
+
+Guidelines:
+
+- The answer NA means that the paper does not release new assets.  
+- Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc.  
+- The paper should discuss whether and how consent was obtained from people whose asset is used.  
+- At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file.
+
+# 14. Crowdsourcing and Research with Human Subjects
+
+Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)?
+
+Answer: [NA]
+
+Justification: the paper does not involve crowdsourcing nor research with human subjects.
+
+Guidelines:
+
+- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects.  
+- Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper.  
+- According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector.
+
+# 15. Institutional Review Board (IRB) Approvals or Equivalent for Research with Human Subjects
+
+Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained?
+
+Answer: [NA]
+
+Justification: the paper does not involve crowdsourcing nor research with human subjects.
+
+Guidelines:
+
+- The answer NA means that the paper does not involve crowdsourcing nor research with human subjects.  
+- Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper.  
+- We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution.  
+- For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review.

2024/Boosting the Potential of Large Language Models with an Intelligent Information Assistant/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:75dffcab662ea421e1c6a25689203cf867cf99475b3fbfea91835c3c86ae5a64
+size 985668

2024/Boosting the Transferability of Adversarial Attack on Vision Transformer with Adaptive Token Tuning/ba47e4ae-9806-497d-a191-2bcd53dfb393_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5a2b5ba8b02007c393e4faeaadb0d7cfa731b9c6194598e32f8a9062a853e434
+size 15983864

2024/Boosting the Transferability of Adversarial Attack on Vision Transformer with Adaptive Token Tuning/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fd0b5505527760b417a6416ee1fa35c919c9566424c05788344a09bec1bf45c6
+size 2087325