Add Batch 36c539db-b0bd-4919-86f7-fb25f74602ee data

.gitattributes

@@ -5103,3 +5103,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2025/Butterfly[[:space:]]Effects[[:space:]]in[[:space:]]Toolchains_[[:space:]]A[[:space:]]Comprehensive[[:space:]]Analysis[[:space:]]of[[:space:]]Failed[[:space:]]Parameter[[:space:]]Filling[[:space:]]in[[:space:]]LLM[[:space:]]Tool-Agent[[:space:]]Systems/8c2cf4ca-62b0-4db5-939d-f0eeca852dcb_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2025/CAARMA_[[:space:]]Class[[:space:]]Augmentation[[:space:]]with[[:space:]]Adversarial[[:space:]]Mixup[[:space:]]Regularization/17f446bb-68df-47ea-a5db-62a858fc7df0_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2025/CAC-CoT_[[:space:]]Connector-Aware[[:space:]]Compact[[:space:]]Chain-of-Thought[[:space:]]for[[:space:]]Efficient[[:space:]]Reasoning[[:space:]]Data[[:space:]]Synthesis[[:space:]]Across[[:space:]]Dual-System[[:space:]]Cognitive[[:space:]]Tasks/7627dd42-601c-4073-a98e-bd6906f5a3e2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/KoACD_[[:space:]]The[[:space:]]First[[:space:]]Korean[[:space:]]Adolescent[[:space:]]Dataset[[:space:]]for[[:space:]]Cognitive[[:space:]]Distortion[[:space:]]Analysis[[:space:]]via[[:space:]]Role-Switching[[:space:]]Multi-LLM[[:space:]]Negotiation/68de1955-5717-4baa-afad-4f38141302dd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/KoLEG_[[:space:]]On-the-Fly[[:space:]]Korean[[:space:]]Legal[[:space:]]Knowledge[[:space:]]Editing[[:space:]]with[[:space:]]Continuous[[:space:]]Retrieval/c37f6d72-5c27-4a79-8b7b-8402c8729735_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Krikri_[[:space:]]Advancing[[:space:]]Open[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]for[[:space:]]Greek/a01b7f99-74bd-4cb7-9a41-265c6435fe83_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/KurTail[[:space:]]_[[:space:]]Kurtosis-based[[:space:]]LLM[[:space:]]Quantization/3649cf39-4457-4adc-be41-8f76450f5b9e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LAGCL4Rec_[[:space:]]When[[:space:]]LLMs[[:space:]]Activate[[:space:]]Interactions[[:space:]]Potential[[:space:]]in[[:space:]]Graph[[:space:]]Contrastive[[:space:]]Learning[[:space:]]for[[:space:]]Recommendation/c5d15e62-530f-4ac3-b25e-1c7904eb8e11_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LAVa_[[:space:]]Layer-wise[[:space:]]KV[[:space:]]Cache[[:space:]]Eviction[[:space:]]with[[:space:]]Dynamic[[:space:]]Budget[[:space:]]Allocation/a94b9893-339c-481e-b5c0-29fad3e14c83_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LAWCAT_[[:space:]]Efficient[[:space:]]Distillation[[:space:]]from[[:space:]]Quadratic[[:space:]]to[[:space:]]Linear[[:space:]]Attention[[:space:]]with[[:space:]]Convolution[[:space:]]across[[:space:]]Tokens[[:space:]]for[[:space:]]Long[[:space:]]Context[[:space:]]Modeling/04580249-c454-4e92-b9e7-a09c71d71b81_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LC-Eval_[[:space:]]A[[:space:]]Bilingual[[:space:]]Multi-Task[[:space:]]Evaluation[[:space:]]Benchmark[[:space:]]for[[:space:]]Long-Context[[:space:]]Understanding/bfbb1c17-291f-491b-b8e4-17a9af50d32d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LCAN_[[:space:]]A[[:space:]]Label-Aware[[:space:]]Contrastive[[:space:]]Attention[[:space:]]Network[[:space:]]for[[:space:]]Multi-Intent[[:space:]]Recognition[[:space:]]and[[:space:]]Slot[[:space:]]Filling[[:space:]]in[[:space:]]Task-Oriented[[:space:]]Dialogue[[:space:]]Systems/82b6ed46-a9f8-45e5-85b9-e1b0bc58ecb6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LEAF_[[:space:]]Large[[:space:]]Language[[:space:]]Diffusion[[:space:]]Model[[:space:]]for[[:space:]]Time[[:space:]]Series[[:space:]]Forecasting/0e501bb2-a116-48ba-b882-1533b8584ac6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LIFTED_[[:space:]]Multimodal[[:space:]]Clinical[[:space:]]Trial[[:space:]]Outcome[[:space:]]Prediction[[:space:]]via[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]and[[:space:]]Mixture-of-Experts/f54a1869-a540-4f05-9a2b-d67b7845d24b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LLM[[:space:]]Agents[[:space:]]at[[:space:]]the[[:space:]]Roundtable_[[:space:]]A[[:space:]]Multi-Perspective[[:space:]]and[[:space:]]Dialectical[[:space:]]Reasoning[[:space:]]Framework[[:space:]]for[[:space:]]Essay[[:space:]]Scoring/d2b04270-61b4-449b-85ea-a940c525de0a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LLM[[:space:]]Agents[[:space:]]for[[:space:]]Education_[[:space:]]Advances[[:space:]]and[[:space:]]Applications/da6da90b-43b3-4ab9-b1a3-894dc7f3404a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LLM[[:space:]]Distillation[[:space:]]for[[:space:]]Efficient[[:space:]]Few-Shot[[:space:]]Multiple[[:space:]]Choice[[:space:]]Question[[:space:]]Answering/2181c27e-4e50-4d9f-bc77-914fc9e63490_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LLM[[:space:]]Jailbreak[[:space:]]Detection[[:space:]]for[[:space:]](Almost)[[:space:]]Free!/8cf1eae3-a8bc-4e25-8645-9a17e88d5d44_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LLM-Based[[:space:]]Web[[:space:]]Data[[:space:]]Collection[[:space:]]for[[:space:]]Research[[:space:]]Dataset[[:space:]]Creation/66c45d51-4a2b-43b6-8fc8-d892e7a2c1ec_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LLM-based[[:space:]]Conversational[[:space:]]Recommendation[[:space:]]Agents[[:space:]]with[[:space:]]Collaborative[[:space:]]Verbalized[[:space:]]Experience/2d2180b1-193d-457d-bf06-2dffe4474327_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LLM-based[[:space:]]Open[[:space:]]Domain[[:space:]]Planning[[:space:]]by[[:space:]]Leveraging[[:space:]]Entity-Attribute-Level[[:space:]]Domain[[:space:]]Models/a8aa13d6-ec38-43fb-973b-ea2a594438da_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LLM-empowered[[:space:]]Dynamic[[:space:]]Prompt[[:space:]]Routing[[:space:]]for[[:space:]]Vision-Language[[:space:]]Models[[:space:]]Tuning[[:space:]]under[[:space:]]Long-Tailed[[:space:]]Distributions/7dde2e52-68d6-4ce9-aa27-bd2a40b4ba65_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LLMAP_[[:space:]]LLM-Assisted[[:space:]]Multi-Objective[[:space:]]Route[[:space:]]Planning[[:space:]]with[[:space:]]User[[:space:]]Preferences/2da3c7b8-30b1-4a77-9cd0-a9c9f10a5289_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LLMEval-Med_[[:space:]]A[[:space:]]Real-world[[:space:]]Clinical[[:space:]]Benchmark[[:space:]]for[[:space:]]Medical[[:space:]]LLMs[[:space:]]with[[:space:]]Physician[[:space:]]Validation/b6bdfb4e-22d2-4204-8a7d-22da0d2a9016_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LLMs[[:space:]]Can[[:space:]]Compensate[[:space:]]for[[:space:]]Deficiencies[[:space:]]in[[:space:]]Visual[[:space:]]Representations/cabd9ffc-1921-425c-a58d-035a4caf2059_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LLMs[[:space:]]Reproduce[[:space:]]Stereotypes[[:space:]]of[[:space:]]Sexual[[:space:]]and[[:space:]]Gender[[:space:]]Minorities/64a7e54a-af0e-42a8-9b6b-6d3aa99e2fbf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LLMs[[:space:]]are[[:space:]]Privacy[[:space:]]Erasable/ca2184c7-27f0-402c-914b-36cef96d7b11_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LLMs[[:space:]]as[[:space:]]a[[:space:]]synthesis[[:space:]]between[[:space:]]symbolic[[:space:]]and[[:space:]]distributed[[:space:]]approaches[[:space:]]to[[:space:]]language/b097142c-dd5f-45d3-8137-253412550285_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LLMs[[:space:]]for[[:space:]]Bayesian[[:space:]]Optimization[[:space:]]in[[:space:]]Scientific[[:space:]]Domains_[[:space:]]Are[[:space:]]We[[:space:]]There[[:space:]]Yet_/68a53f51-4c93-4cee-91c0-ae6aaee7a1f5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LLMsPark_[[:space:]]A[[:space:]]Benchmark[[:space:]]for[[:space:]]Evaluating[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]in[[:space:]]Strategic[[:space:]]Gaming[[:space:]]Contexts/373ead25-9104-4681-be63-1442d14520f9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LLaVE_[[:space:]]Large[[:space:]]Language[[:space:]]and[[:space:]]Vision[[:space:]]Embedding[[:space:]]Models[[:space:]]with[[:space:]]Hardness-Weighted[[:space:]]Contrastive[[:space:]]Learning/922e7a46-fe5f-425f-b3f3-50246284d059_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LM2Protein_[[:space:]]A[[:space:]]Structure-to-Token[[:space:]]Protein[[:space:]]Large[[:space:]]Language[[:space:]]Model/ad2c4cd3-a6d8-4458-a4e4-3408f643f6b6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LMUNIT_[[:space:]]Fine-grained[[:space:]]Evaluation[[:space:]]with[[:space:]]Natural[[:space:]]Language[[:space:]]Unit[[:space:]]Tests/38ed9229-ba91-47ae-975f-ebe4ba9e7252_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LNE-Blocking_[[:space:]]An[[:space:]]Efficient[[:space:]]Framework[[:space:]]for[[:space:]]Contamination[[:space:]]Mitigation[[:space:]]Evaluation[[:space:]]on[[:space:]]Large[[:space:]]Language[[:space:]]Models/52f11578-f861-486c-9b02-95eb20058e70_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LOHRec_[[:space:]]Leveraging[[:space:]]Order[[:space:]]and[[:space:]]Hierarchy[[:space:]]in[[:space:]]Generative[[:space:]]Sequential[[:space:]]Recommendation/4675ed3c-1261-4ada-bcb9-efc4268a2127_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LORE_[[:space:]]Continual[[:space:]]Logit[[:space:]]Rewriting[[:space:]]Fosters[[:space:]]Faithful[[:space:]]Generation/bb496e61-57d5-406b-8a09-0d7cf2a14fa4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LRPLAN_[[:space:]]A[[:space:]]Multi-Agent[[:space:]]Collaboration[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]and[[:space:]]Reasoning[[:space:]]Models[[:space:]]for[[:space:]]Planning[[:space:]]with[[:space:]]Implicit[[:space:]]&[[:space:]]Explicit[[:space:]]Constraints/1cb8f85c-7aac-4d85-8ef9-ee0b1f1907e2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LSRL_[[:space:]]Process-Supervised[[:space:]]GRPO[[:space:]]on[[:space:]]Latent[[:space:]]Recurrent[[:space:]]States[[:space:]]Improves[[:space:]]Mathematical[[:space:]]Reasoning/7602ed67-cd08-46d4-b62d-41f6b43a1f01_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LUME_[[:space:]]LLM[[:space:]]Unlearning[[:space:]]with[[:space:]]Multitask[[:space:]]Evaluations/aa162b2d-77c0-4bf7-8d34-8d8c787995a6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LaMP-Cap_[[:space:]]Personalized[[:space:]]Figure[[:space:]]Caption[[:space:]]Generation[[:space:]]With[[:space:]]Multimodal[[:space:]]Figure[[:space:]]Profiles/5c1d4396-04e2-40e5-beca-a8246b57bb56_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LaMP-Val_[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]Empower[[:space:]]Personalized[[:space:]]Valuation[[:space:]]in[[:space:]]Auction/c2bafe0a-49cb-4e86-a1a1-f6a135f7fbf4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LangProBe_[[:space:]]a[[:space:]]Language[[:space:]]Program[[:space:]]Benchmark/7d82d569-baed-48fc-ba49-69fcf43dc919_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Language[[:space:]]Models[[:space:]]Can[[:space:]]Easily[[:space:]]Learn[[:space:]]to[[:space:]]Reason[[:space:]]from[[:space:]]Demonstrations/88e6e1a1-1749-4a89-9721-b10084475d4b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Language-Informed[[:space:]]Synthesis[[:space:]]of[[:space:]]Rational[[:space:]]Agent[[:space:]]Models[[:space:]]for[[:space:]]Grounded[[:space:]]Theory-of-Mind[[:space:]]Reasoning[[:space:]]On-the-fly/cb7a64fa-daf7-4557-bf8a-52b2062eb337_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Language-Specific[[:space:]]Layer[[:space:]]Matters_[[:space:]]Efficient[[:space:]]Multilingual[[:space:]]Enhancement[[:space:]]for[[:space:]]Large[[:space:]]Vision-Language[[:space:]]Models/409c9787-5325-473b-a730-a87698a4ce5a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Large[[:space:]]Language[[:space:]]Model[[:space:]]Agents[[:space:]]in[[:space:]]Finance_[[:space:]]A[[:space:]]Survey[[:space:]]Bridging[[:space:]]Research,[[:space:]]Practice,[[:space:]]and[[:space:]]Real-World[[:space:]]Deployment/7ee7efdf-3a25-4d8c-a980-20ffa38f4dfc_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Large[[:space:]]Language[[:space:]]Model[[:space:]]Evaluation[[:space:]]via[[:space:]]Matrix[[:space:]]Nuclear-Norm/a49984c1-dc08-4d57-b024-0b24118dccfa_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Large[[:space:]]Language[[:space:]]Models[[:space:]]Might[[:space:]]Not[[:space:]]Care[[:space:]]What[[:space:]]You[[:space:]]Are[[:space:]]Saying_[[:space:]]Prompt[[:space:]]Format[[:space:]]Beats[[:space:]]Descriptions/a6552072-b52e-4ecd-b887-ba0e30d0144f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Large[[:space:]]Language[[:space:]]Models[[:space:]]as[[:space:]]Reader[[:space:]]for[[:space:]]Bias[[:space:]]Detection/58ac6468-1afb-4e1d-98f6-62d192d3ef18_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Large[[:space:]]Language[[:space:]]Models[[:space:]]for[[:space:]]Controllable[[:space:]]Multi-property[[:space:]]Multi-objective[[:space:]]Molecule[[:space:]]Optimization/4f87c197-d0a4-4b15-9c9f-a05e83f20076_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Large[[:space:]]Language[[:space:]]Models[[:space:]]for[[:space:]]Multilingual[[:space:]]Previously[[:space:]]Fact-Checked[[:space:]]Claim[[:space:]]Detection/e125155b-ad09-4e57-bc83-230701f57dd8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Large[[:space:]]Language[[:space:]]Models[[:space:]]with[[:space:]]Temporal[[:space:]]Reasoning[[:space:]]for[[:space:]]Longitudinal[[:space:]]Clinical[[:space:]]Summarization[[:space:]]and[[:space:]]Prediction/65e6a6b1-e411-4e24-b308-716f24a8e244_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Large[[:space:]]Vision-Language[[:space:]]Model[[:space:]]Alignment[[:space:]]and[[:space:]]Misalignment_[[:space:]]A[[:space:]]Survey[[:space:]]Through[[:space:]]the[[:space:]]Lens[[:space:]]of[[:space:]]Explainability/833b86ec-0195-410c-9687-0d58b196b88b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LastingBench_[[:space:]]Defend[[:space:]]Benchmarks[[:space:]]Against[[:space:]]Knowledge[[:space:]]Leakage/40264d67-41de-4e6e-9ad3-558234efe887_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Layer[[:space:]]Duplication[[:space:]]in[[:space:]]LLMs/020bb96d-59f9-4483-a686-652a08f22c06_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Layer-Aware[[:space:]]Task[[:space:]]Arithmetic_[[:space:]]Disentangling[[:space:]]Task-Specific[[:space:]]and[[:space:]]Instruction-Following[[:space:]]Knowledge/8e09e52c-bbe3-4629-b339-83dc87325b13_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Learning[[:space:]]API[[:space:]]Functionality[[:space:]]from[[:space:]]In-Context[[:space:]]Demonstrations[[:space:]]for[[:space:]]Tool-based[[:space:]]Agents/e197dc4a-7c5d-4244-876b-7044e7283f09_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Learning[[:space:]]Is[[:space:]]Not[[:space:]]A[[:space:]]Race_[[:space:]]Improving[[:space:]]Retrieval[[:space:]]in[[:space:]]Language[[:space:]]Models[[:space:]]via[[:space:]]Equal[[:space:]]Learning/83db89fa-aac4-4e1b-8fe9-7e9b028152ed_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Learning[[:space:]]SQL[[:space:]]Like[[:space:]]a[[:space:]]Human_[[:space:]]Structure-Aware[[:space:]]Curriculum[[:space:]]Learning[[:space:]]for[[:space:]]Text-to-SQL[[:space:]]Generation/2713508d-1757-4e45-8f73-2af548bb1873_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Learning[[:space:]]Trajectories[[:space:]]of[[:space:]]Figurative[[:space:]]Language[[:space:]]for[[:space:]]Pre-Trained[[:space:]]Language[[:space:]]Models/7e8c7593-c13b-4fac-9f96-87a3058821f9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Learning[[:space:]]What[[:space:]]to[[:space:]]Remember_[[:space:]]Adaptive[[:space:]]Probabilistic[[:space:]]Memory[[:space:]]Retention[[:space:]]for[[:space:]]Memory-Efficient[[:space:]]Language[[:space:]]Models/795d51a8-5299-4296-a4ba-141033a43dba_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Learning[[:space:]]to[[:space:]]Align_[[:space:]]Addressing[[:space:]]Character[[:space:]]Frequency[[:space:]]Distribution[[:space:]]Shifts[[:space:]]in[[:space:]]Handwritten[[:space:]]Text[[:space:]]Recognition/9eace239-4c96-4d82-bbb0-23dd975fb090_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Learning[[:space:]]to[[:space:]]Describe[[:space:]]Implicit[[:space:]]Changes_[[:space:]]Noise-robust[[:space:]]Pre-training[[:space:]]for[[:space:]]Image[[:space:]]Difference[[:space:]]Captioning/d58c93f3-42ad-41a0-813a-791c432aa625_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Learning[[:space:]]to[[:space:]]Instruct_[[:space:]]Fine-Tuning[[:space:]]a[[:space:]]Task-Aware[[:space:]]Instruction[[:space:]]Optimizer[[:space:]]for[[:space:]]Black-Box[[:space:]]LLMs/a0f3c51a-0918-462e-b873-0d37d2bc1782_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Legal[[:space:]]Mathematical[[:space:]]Reasoning[[:space:]]with[[:space:]]LLMs_[[:space:]]Procedural[[:space:]]Alignment[[:space:]]through[[:space:]]Two-Stage[[:space:]]Reinforcement[[:space:]]Learning/43567c92-26ff-4c8d-a00f-752ab43c0a98_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/LegoSLM_[[:space:]]Connecting[[:space:]]LLM[[:space:]]with[[:space:]]Speech[[:space:]]Encoder[[:space:]]using[[:space:]]CTC[[:space:]]Posteriors/140442ca-4acf-4df3-aa50-552f62132b90_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2025/Lemma[[:space:]]Dilemma_[[:space:]]On[[:space:]]Lemma[[:space:]]Generation[[:space:]]Without[[:space:]]Domain-[[:space:]]or[[:space:]]Language-Specific[[:space:]]Training[[:space:]]Data/047268e9-6274-4c4b-852c-4611109e80cf_origin.pdf filter=lfs diff=lfs merge=lfs -text

2025/KoACD_ The First Korean Adolescent Dataset for Cognitive Distortion Analysis via Role-Switching Multi-LLM Negotiation/68de1955-5717-4baa-afad-4f38141302dd_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:42d68b5a849354c0cc48e740975f5e7fd0f20ec11733bea9fec0d37af4ee36ea
+size 1089275

2025/KoACD_ The First Korean Adolescent Dataset for Cognitive Distortion Analysis via Role-Switching Multi-LLM Negotiation/full.md

@@ -0,0 +1,1107 @@
+# KoACD: The First Korean Adolescent Dataset for Cognitive Distortion Analysis via Role-Switching Multi-LLM Negotiation
+
+Jun Seo Kim
+
+Hye Hyeon Kim
+
+<sup>1</sup>Department of Computer Engineering, Gachon University
+
+$^{2}$ Department of Biomedical Systems Informatics, Yonsei University
+
+kma80kjs@gachon.ac.kr
+
+hye_hyeon@yonsei.ac.kr
+
+# Abstract
+
+Cognitive distortion refers to negative thinking patterns that can lead to mental health issues like depression and anxiety in adolescents. Previous studies using natural language processing (NLP) have focused mainly on small-scale adult datasets, with limited research on adolescents. This study introduces KoACD, the first large-scale dataset of cognitive distortions in Korean adolescents, containing 108,717 instances. We applied a multi-Large Language Model (LLM) negotiation method to refine distortion classification, enabling iterative feedback and role-switching between models to reduce bias and improve label consistency. In addition, we generated synthetic data using two approaches: cognitive clarification for textual clarity and cognitive balancing for diverse distortion representation. Validation through LLMs and expert evaluations showed that while LLMs classified distortions with explicit markers, they struggled with context-dependent reasoning, where human evaluators demonstrated higher accuracy. KoACD aims to enhance future research on cognitive distortion detection. The dataset and implementation details are publicly accessible<sup>1</sup>.
+
+# 1 Introduction
+
+Negative thoughts (Pietromonaco et al., 1985) are a natural part of human cognition, often helping individuals recognize potential dangers, prepare for challenges, or engage in self-reflection. However, when these patterns become rigid and excessive, they can lead to emotional distress and contribute to mental health issues such as depression and anxiety disorders. Adolescents, in particular, may be more vulnerable to these
+
+maladaptive thought patterns due to their ongoing cognitive and emotional development.
+
+Globally, one in seven children and adolescents (about 166 million) suffers from mental illness, with  $42.9\%$  experiencing anxiety and depression (UNICEF, 2018). The rising prevalence of these conditions during adolescence has become a serious global concern. Since this stage is crucial for self-identity formation and emotional regulation (Pfeifer et al., 2018), understanding how negative thought patterns emerge and persist is essential for early intervention and prevention. In particular, depression is frequently associated with habitual negative thinking, known as cognitive distortion (Rnic et al., 2016). Adolescents experiencing these distortions may instinctively blame themselves when something goes wrong, thinking, 'I messed up again' or 'I completely failed.' These persistent negative thoughts trigger emotional distress, reinforcing cycles of depression (Chahar et al., 2020). Identifying and analyzing these patterns is essential for developing effective coping strategies. To achieve this, building a comprehensive dataset of adolescent cognitive distortions is necessary to enable more targeted and impactful mental health interventions.
+
+Our key contributions are as follows:
+
+- We present KoACD, the first large-scale dataset of cognitive distortions in Korean adolescents, filling a critical gap in non-English and youth-oriented mental health resources.  
+- We propose a multi-round negotiation framework using multiple Large Language Models (LLMs) to refine cognitive distortion classification, employing iterative role-switching to improve consistency, reduce bias, and capture context-dependent reasoning.
+
+<table><tr><td>Dataset</td><td>Language</td><td>Sample</td><td>Target</td><td>Data Source</td><td>Classification</td></tr><tr><td>Tumblr Cognitive Distortion (Simms et al., 2017)*</td><td>English</td><td>459</td><td>nonspecific</td><td>Tumblr blogs</td><td>Binary (2)</td></tr><tr><td>MH-C (Shickel et al., 2020)</td><td>English</td><td>1,164</td><td>Adult</td><td>TAO Connect</td><td>Multi-class (15)</td></tr><tr><td>MH-D (Shickel et al., 2020)</td><td>English</td><td>1,799</td><td>Adult</td><td>TAO Connect</td><td>Binary (2)</td></tr><tr><td>CrowdDist (Shickel et al., 2020)</td><td>English</td><td>7,666</td><td>Adult</td><td>Mechanical Turk</td><td>Multi-class (15)</td></tr><tr><td>Clinician-Client SMS (Tauscher et al., 2023)*</td><td>English</td><td>7,354</td><td>Adult</td><td>Clinician-Client SMS</td><td>Multi-class (5)</td></tr><tr><td>SocialCD-3K (Qi et al., 2024)</td><td>Chinese</td><td>3,407</td><td>nonspecific</td><td>Weibo</td><td>Multi-label (12)</td></tr></table>
+
+Table 1: Summary of datasets for cognitive distortion detection: The 'Sample' column indicates the number of instances, and 'Classification' specifies the type (binary, multi-class, or multi-label). *Indicates unofficially named datasets.
+
+# 2 Related Work
+
+# 2.1 Cognitive distortions Detection
+
+Cognitive distortions are closely related to negative thinking patterns and are defined as distorted ways of thinking that reinforce negative emotions (Beck, 1979). With the recent development of natural language processing (NLP), research has been actively conducted to automatically classify cognitive distortions using various datasets.
+
+Previous studies on cognitive distortion classification have primarily relied on small-scale, adult-focused, and English-language datasets. Early research utilized Linguistic Inquiry and Word Count (LIWC)-based regression models on social media posts (Simms et al., 2017), while later studies adopted deep learning models, including RNNs, CNNs, and BERT, using datasets from counseling platforms and therapist-patient conversations (Shickel et al., 2020; Tauscher et al., 2023). More recent approaches have leveraged Large Language Models (LLMs) for cognitive distortion classification, further enhancing performance and adaptability across diverse datasets (Chen et al., 2023; Qi et al., 2024).
+
+Table 1 summarizes existing datasets that deal with cognitive distortions.
+
+Despite these advancements, existing datasets remain limited in scale and predominantly focus on English-speaking adults. To address these limitations, we propose a Korean-language dataset specifically designed for adolescents, filling a crucial gap in research on cognitive distortions in younger populations.
+
+# 2.2 LLMs-Based Negotiation
+
+Attempts have been made to explore the possibility of models going beyond independent judgment and deriving more sophisticated conclusions through interaction through negotiations between LLMs.
+
+Self-Play and In-Context Learning techniques using AI feedback were applied to improve the negotiation capabilities of LLMs (Yao Fu et al., 2024), and a method for providing feedback by developing an LLM-based Assistant for Coaching nEgotiation (ACE) using negotiation data from MBA students was proposed (Ryan Shea et al., 2024). In addition, research has been conducted on applying negotiation methods to emotional analysis. It has been demonstrated that using LLM negotiation methods to interact between models can outperform the existing single-pass decision (Xiaofei Sun et al., 2024).
+
+Previous studies have shown that LLM negotiations can yield sophisticated results, but challenges remain in balancing negotiation outcomes due to fixed roles and limited structures. To overcome this, we use role-switching and multiple types of LLMs to balance the negotiation process. We also introduce multi-round negotiations to give equal consideration to the 10 cognitive distortions, ultimately arriving at the optimal conclusion. Therefore, this study aims to generate and validate data using LLM negotiation techniques.
+
+<table><tr><td>Cognitive Distortion Type</td><td>Definition</td><td>Examples</td></tr><tr><td>All-or-Nothing Thinking</td><td>Viewing situations in only two categories (e.g., perfect or failure) instead of on a spectrum.</td><td>“If I fail this test, I’m a total failure.”</td></tr><tr><td>Overgeneralization</td><td>Drawing broad conclusions from a single event or limited evidence.</td><td>“My one friend ignored me, so everyone else will hate me too.”</td></tr><tr><td>Mental Filtering</td><td>Focusing only on the negative aspects of a situation while ignoring the positive.</td><td>“I only remember my mistake though I got compliments on my presentation.”</td></tr><tr><td>Discounting the Positive</td><td>Rejecting positive experiences or compliments by insisting they don’t count.</td><td>&quot;People told me I did well, but I was just being polite.&quot;</td></tr><tr><td>Jumping to Conclusions</td><td>Predicting negative outcomes without evidence.</td><td>&quot;She didn’t text back. She must be mad at me.&quot;</td></tr><tr><td>Magnification and Minimization</td><td>Exaggerating negative or risky aspects while minimizing positive or positive aspects.</td><td>“One little mistake at work means I’m incompetent.”</td></tr><tr><td>Emotional Reasoning</td><td>Believing something must be true because you feel it strongly.</td><td>&quot;I feel worthless, so I must be worthless.&quot;</td></tr><tr><td>&quot;Should&quot; Statements</td><td>Holding rigid rules about how you or others should behave, leading to guilt or frustration.</td><td>&quot;I should always be productive; otherwise, I’m lazy.&quot;</td></tr><tr><td>Labeling</td><td>Assigning negative labels to yourself or others based on one event.</td><td>“I made a mistake, so I’m a total failure.”</td></tr><tr><td>Personalization</td><td>Blaming yourself for events outside your control or assuming excessive responsibility.</td><td>&quot;My friend looks sad, maybe I did something wrong.&quot;</td></tr></table>
+
+Table 2: Classification of cognitive distortions with definitions and examples
+
+# 3 Constructing KoACD
+
+# 3.1 Data Source and Preprocessing
+
+We crawled posts on Naver Knowledge  $\mathbf{iN}^2$ , a Q&A platform where users can post questions and receive answers, to analyze the cognitive distortions of Korean adolescents. Naver Knowledge iN is Korea's largest open Q&A platform, with over 32 million users and 800 million Q&A entries since its launch (Jang & Kim, 2024). The use of Naver Knowledge iN complies with its terms of service and copyright regulations and is permitted through prior approval or explicit permission. Since Naver Knowledge covers a wide range of age groups, we used only data from five major adolescent counseling organizations and services, covering the years 2011-2024, to focus on adolescent concerns. A total of 69,925 questions were collected, and the distribution of data sources is provided in Appendix A.
+
+A pre-processing step refined the data to align with the research purpose and excluded irrelevant questions. This included removing entries from elementary students or adults, filtering inappropriate content, deleting vague questions, and eliminating duplicates. After applying these criteria, 37,124 questions remained for analysis.
+
+Details on preprocessing and removed cases are in Appendix B.
+
+# 3.2 Definition of Cognitive Distortions
+
+Aaron Beck, a pioneer in cognitive therapy (Beck, 1979), identified 10 cognitive distortions in patients with depression and incorporated them into psychotherapy. He emphasized that reducing these distortions could alleviate stress and anxiety (Beck, 1991). We used these distortions, listed in Table 2, to classify questions reflecting the emotional struggles commonly reported by adolescents.
+
+# 3.3 Multi-LLMs Negotiation for Identifying Cognitive Distortions
+
+To effectively identify cognitive distortions, this study designed a process for deriving optimal distortions using the multi-LLM negotiation method (Xiaofei Fu et al., 2024), where relevant distortions are gradually derived through LLM interactions.
+
+This study uses a multi-LLM negotiation method based on the interaction between Google's Gemini 1.5 Flash (Team et al., 2024) and OpenAI's GPT-4o mini (OpenAI, 2024). The two models work together to identify the most accurate cognitive distortion. One model acts as the Analyzer and the other as the Evaluator. Through
+
+![](images/e0ce49f9b821e2880792e445b937ee2eaf442fb29d5bd12d89b475ffd81a58b6.jpg)  
+Figure 1: Process for identifying and evaluating cognitive distortions through negotiation
+
+their collaboration, cognitive distortions are gradually refined. Negotiation is conducted up to 5 rounds to systematically explore all 10 predefined cognitive distortions, as each round consists of two turns, evaluating one distortion per turn. This structure allows for the possibility that the same sentence may be interpreted in multiple cognitive distortions before reaching a final classification. Table 11 in Appendix E details the LLMs parameters used in this process.
+
+Here's how the roles work:
+
+- Analyzer: Identifies the most relevant cognitive distortion in a sentence and suggests sentences that match it.  
+- Evaluator: Reviews the suggestions made by the Analyzer and provides feedback on their accuracy.
+
+The prompts used for these roles are detailed in Appendix H. In each round of negotiation, the models take turns playing the roles of Analyzer and Evaluator.
+
+A round consists of two turns and proceeds in the following structure:
+
+- T1 (Initial Analysis): Identify the most relevant cognitive distortion in the sentence. (Options: one of the 10 cognitive distortions.)  
+- T1 (Evaluation): Assess whether the proposed cognitive distortion from T1 (Initial Analysis) accurately reflects the distortion present in the sentence. (Options: "Yes" or "No.") The evaluator provides a justification.)  
+- T2 (Reanalysis): If T1 (Evaluation) results in rejection, select the next most relevant cognitive distortion, excluding previously rejected options. (Options: one of the remaining cognitive distortions.)  
+- T2 (Evaluation): Determine whether the cognitive distortion from T2 (Reanalysis) is appropriate. (Options: "Yes" or "No." The evaluator provides a justification.)
+
+Each step is performed sequentially, incorporating feedback from the previous evaluation. T1 (Evaluation) assesses the distortion proposed in T1 (Initial Analysis), and T2 (Reanalysis) refines the selection based on that feedback. Similarly, T2
+
+(Evaluation) verifies the suitability of the distortion chosen in T2 (Reanalysis).
+
+Throughout the negotiation process, distortions deemed inappropriate are systematically excluded, ensuring the selection of the most fitting cognitive distortion. To maintain fairness, the models alternate roles in T2 (Reanalysis) so that both contribute equally to the negotiation.
+
+If a consensus is not reached after five rounds, the question is classified as unknown. This indicates that all cognitive distortions proposed during the negotiation process were considered inherently inappropriate.
+
+The number of turns required to identify cognitive distortions or classify the question as unknown varies from dataset to dataset. Some sentences reach conclusions early, while others require multiple turns for final classification. Details of turn counts and classification ratios are given in Appendix C. The overall structure of this negotiation process is illustrated in Figure 1.
+
+# 3.4 Independent Evaluation
+
+After the negotiation process is complete, Anthropic's Claude 3 Haiku (Anthropic, 2024) is used for an independent validation of the final cognitive distortion and its corresponding sentence. Claude 3 Haiku is not involved in the negotiation process; instead, it evaluates whether the selected cognitive distortion correctly aligns with the given sentence.
+
+During negotiation, the models assess cognitive distortions in the context of the entire original text, whereas Claude 3 Haiku determines the appropriateness based solely on the selected sentence. This additional validation step helps identify potential misclassifications and ensures that the cognitive distortion is properly connected to the sentence.
+
+Claude 3 Haiku assigns a relevance score from 1 to 3, and only cognitive distortion-sentence pairs that receive a score of 3 are used as the final-stage data for generating synthetic data. A summary of the validation score distribution is presented in Table 3. The parameters used for this validation are detailed in Table 11 (Appendix E), and the prompts used for evaluation are provided in Appendix H.
+
+# 3.5 Synthetic Data Generation
+
+The original data consists of free-form text written by adolescents, often containing spelling errors, excessive use of emoticons, or unclear wording, making it difficult to interpret.
+
+<table><tr><td>Score</td><td>Count</td><td>Proportion (%)</td></tr><tr><td>1</td><td>11</td><td>0.06</td></tr><tr><td>2</td><td>874</td><td>4.41</td></tr><tr><td>3</td><td>18,897</td><td>95.53</td></tr><tr><td>Total</td><td>19,782</td><td>100.00</td></tr></table>
+
+Table 3: Distribution of validation scores
+
+Additionally, some texts lack contextual coherence, with disjointed narratives or insufficient background information to accurately assess cognitive distortions. As a result, the data could be difficult to use as is.
+
+Furthermore, the distribution of the 10 cognitive distortion categories we propose was imbalanced, leading to a potential bias in the dataset. To address these issues, we employ two methods to generate synthetic data. The prompts used for both methods are provided in Appendix H.
+
+# 3.5.1 Cognitive Clarification of Cognitive Distortions
+
+The first approach to generating synthetic data is to identify cognitive distortions and rephrase the text in a clearer and more structured form while preserving the meaning of the original text: maintaining the emotional tone and context.
+
+We used three LLMs—Gemini 1.5 Flash, Claude 3 Haiku, and GPT-4o mini—independently to generate a wide variety of expressions from 18,897 samples, ensuring greater diversity in the generated content. The parameters of these models used for synthetic data generation are detailed as shown in Table 11 in Appendix E.
+
+# 3.5.2 Balancing Cognitive Distortions with Context-Preserved Data
+
+The second approach we adopted aimed to address the imbalance of cognitive distortions by utilizing data classified as 'Unknown' or data for which a suitable cognitive distortion could not be identified. First, we analyzed the distribution of cognitive distortions to detect which types were underrepresented. Then, synthetic data was generated by reconstructing and reorganizing 17,342 samples labeled as "Unknown", ensuring that the overall context was preserved.
+
+Table 4 summarizes the distribution of cognitive distortions produced through both the cognitive clarification and cognitive balancing methods, along with the overall total after combining both approaches.
+
+<table><tr><td>Cognitive Distortion Type</td><td>Cognitive Clarification (%)</td><td>Cognitive Balancing</td><td>Total</td></tr><tr><td>All-or-Nothing Thinking</td><td>5,949 (10.50%)</td><td>4,920 (9.46%)</td><td>10,869 (10.00%)</td></tr><tr><td>Overgeneralization</td><td>11,418 (20.14%)</td><td>0 (0.00%)</td><td>11,418 (10.50%)</td></tr><tr><td>Mental Filtering</td><td>2,763 (4.88%)</td><td>8,139 (15.64%)</td><td>10,902 (10.03%)</td></tr><tr><td>Discounting the Positive</td><td>822 (1.45%)</td><td>9,873 (18.98%)</td><td>10,695 (9.84%)</td></tr><tr><td>Jumping to Conclusions</td><td>10,479 (18.48%)</td><td>183 (0.35%)</td><td>10,662 (9.81%)</td></tr><tr><td>Magnification and Minimization</td><td>6,078 (10.72%)</td><td>4,836 (9.30%)</td><td>10,914 (10.04%)</td></tr><tr><td>Emotional Reasoning</td><td>10,842 (19.12%)</td><td>0 (0.00%)</td><td>10,842 (9.98%)</td></tr><tr><td>&quot;Should&quot; Statements</td><td>2,697 (4.76%)</td><td>7,998 (15.37%)</td><td>10,695 (9.84%)</td></tr><tr><td>Labeling</td><td>2,373 (4.19%)</td><td>8,463 (16.27%)</td><td>10,836 (9.97%)</td></tr><tr><td>Personalization</td><td>3,270 (5.77%)</td><td>7,614 (14.63%)</td><td>10,884 (10.01%)</td></tr><tr><td>Total</td><td>56,691 (100.00%)</td><td>52,026 (100.00%)</td><td>108,717 (100.00%)</td></tr></table>
+
+Table 4: Distribution of cognitive distortion types across synthetic data generation methods
+
+# 4 Validating Synthetic Data with Clustering
+
+To verify the validity of the synthetic data we created, we performed clustering based on two criteria: (1) topics that trigger negative emotions in adolescents and (2) negative emotions and symptoms outlined in the DSM-5 $^3$ , a widely used framework for assessing and diagnosing mental disorders (Lee et al., 2023).
+
+# 4.1 Topic-Based Classification of Adolescent Negative Thinking
+
+The Korea National Youth Policy Institute (NYPI $^4$ ), under the Ministry of Gender Equality and Family, categorized adolescents' concerns into five areas: (1) academic and career concerns, (2) relationships (friendships, romance, bullying), (3) physical and mental health, (4) family issues, and (5) appearance and self-image.
+
+To assess the alignment of our synthetic data on adolescent negative thinking with these categories, we applied K-means clustering, an unsupervised machine learning algorithm that partitions data into distinct groups, to keywords extracted from 37,124 adolescents' questions (Section 3.1). Specifically, K-means clustering was used to define hierarchical relationships and classifications, grouping the data into the five predefined subject clusters, each of which was assigned sub-keywords based on relevance. As a result, a dictionary with five topics and 120 keywords was created, as shown in Table 12 in Appendix F.
+
+We identified the most frequent keywords for each topic. The top topic was Academic
+
+performance and career concerns, with 99,076 times  $(36.9\%)$ , followed by Relationships (73,586,  $27.4\%$ ), Physical and mental health (71,249,  $26.5\%$ ), Family issues (20,532,  $7.6\%$ ), and Appearance and self-image (4,007,  $1.5\%$ ).
+
+# 4.2 DSM-5 based Classification of Adolescent Negative Thinking
+
+Cognitive distortions can contribute to depression, so we examined the nine categories of the DSM-5 to determine whether a significant relationship exists. To explore this, we analyzed 37,124 adolescents' questions (Section 3.1). and identified DSM-5-related word distributions using NLTK text mining $^{5}$  (3.9.1). We used NLTK to build a keyword-topic dictionary based on DSM-5 criteria, extracting synonyms for DSM-5 symptoms with WordNet, identifying related keywords, and using LDA topic modeling to derive and compare topics with the DSM-5 criteria. These distributions were then used to create dictionaries for DSM classification, resulting in nine categories and 121 keywords, as shown in Table 13 in Appendix F.
+
+For keyword mapping, we used our dataset of 108,717 synthesized data points (Section 3.5.2), allowing multiple keywords per data point. For DSM-based keyword mapping, 69,290 data points  $(63.7\%)$  were successfully mapped, with 115 unique keywords assigned 1,335,337 times. For negative emotion-triggering topic-based mapping, 103,183 data points  $(94.9\%)$  were successfully mapped, with 129 unique keywords assigned 268,450 times.
+
+Among the DSM-5 symptom categories, five
+
+![](images/13d84f8d8bd9649e7c1482846af771d79d1f1820f988e4e360e50cd4d78e0dfd.jpg)  
+Figure 2: Cluster distribution of high-frequency by (A) negative emotion-triggering topics, (B) DSM-5 symptom keywords ( $\geq$ 1,000 occurrences)
+
+![](images/0a049d4d66fca492ba13df961d783a474c7f68ecd046a71a7576b5f8827e7bc3.jpg)
+
+<table><tr><td rowspan="2">Criteria</td><td colspan="2">LLMs Evaluation</td><td colspan="2">Human Evaluation</td></tr><tr><td>Cognitive Clarification</td><td>Cognitive Balancing</td><td>Cognitive Clarification</td><td>Cognitive Balancing</td></tr><tr><td>Consistency</td><td>2.400 ± 0.232</td><td>2.105 ± 0.173</td><td>2.254</td><td>2.160</td></tr><tr><td>Accuracy</td><td>2.708 ± 0.177</td><td>2.416 ± 0.270</td><td>2.322</td><td>2.738</td></tr><tr><td>Fluency</td><td>2.655 ± 0.219</td><td>2.529 ± 0.223</td><td>2.904</td><td>2.690</td></tr></table>
+
+Table 5: Evaluation results of synthetic data by LLMs and humans: The LLMs evaluation (left) reports mean ± standard deviation scores assigned by three models, where the standard deviation represents variations across models. The human evaluation (right) presents the average scores given by two experts after cross-validation.
+
+out of nine categories appeared more than 15,000 times. The most frequent keyword was B. Loss of interest or pleasure (321,157 occurrences,  $23.8\%$ ), followed by H. Decreased concentration (25,580 occurrences,  $18.9\%$ ), A. Depressed mood (25,258 occurrences,  $18.7\%$ ), D. Insomnia or hypersomnia (24,864 occurrences,  $18.4\%$ ), and E. Psychomotor agitation or retardation (15,235 occurrences,  $11.3\%$ ).
+
+We found 34 keywords (Table 14 in Appendix F) for cognitive distortion-triggering topics and 20 (Table 15 in Appendix F) for DSM-5 categories, each with a frequency of 1,000 or more, are listed in Figure 2.
+
+The generated synthetic data mainly highlighted academic and career stress, along with social conflicts like friendships and romantic relationships, while underrepresenting appearance and self-image issues. Additionally, its cognitive distortions were closely linked to five of the nine DSM-5 depression symptom keywords.
+
+# 5 Evaluation
+
+We evaluate the quality of two types of synthetic data: data generated from clearly identified cognitive distortions (Section 3.5.1) and data generated to address cognitive distortion imbalances (Section 3.5.2). The evaluation was conducted independently using three evaluation criteria, with both LLMs and human assessments.
+
+# 5.1 Evaluation criteria
+
+We evaluated the generated synthetic data using three criteria: (1) Consistency, (2) Accuracy, and (3) Fluency. Scores ranged from 1 to 3, with 1 indicating 'inappropriate' and 3 indicating 'highly appropriate'. Consistency checked if the cognitive distortion was logically maintained between the original and synthetic data. Accuracy assessed whether the labeled cognitive distortion matched the correct classification. Fluency evaluated how natural, grammatically correct, and easy to read the sentences were. The prompts used for these criteria are provided in Appendix H.
+
+<table><tr><td>Original Distress Question</td><td>&quot;Whenever I meet my cousins, my mom asks me why I&#x27;m not tall. Even my friends, who are shorter than me, stay up past 2 a.m., and when I see them grow taller, I can only think about why I&#x27;m not tall.&quot;</td></tr><tr><td>[Type of Cognitive Distortion] Synthetic Data</td><td>[&quot;Should&quot; Statement]&quot;My mom often compares me to her cousins and says, &#x27;Why are you so short?&#x27; I don&#x27;t understand why I&#x27;m the only one who&#x27;s so short when all my friends are growing taller. Both my mom and dad are tall, but I feel like something is wrong with being short.&quot;</td></tr><tr><td>Commands from Expert</td><td>The expert provided two points in the accuracy evaluation of the cognitive distortion type, and chose mental filtering rather than &#x27;should&#x27; statement. The belief that one should be tall (&#x27;should&#x27; statement) usually comes from parents. In this article, we confirmed that the fact that the mother is not tall triggers anxiety. However, this presupposes that the individual has negative thoughts (mental filtering), as she believes she will not grow taller. While the &#x27;should&#x27; statement seems to be the main issue in literal terms, mental filtering—an error in self-judgment—is considered the primary cognitive distortion.&quot;</td></tr></table>
+
+Table 6: Expert analysis of case with synthetic data accuracy score of 2: Explanations for LLM misclassification  
+
+<table><tr><td rowspan="2">Cognitive Distortion Type</td><td colspan="3">LLMs Evaluation</td><td colspan="3">Human Evaluation</td><td colspan="3">Difference</td></tr><tr><td>Cos</td><td>Acc</td><td>Flu</td><td>Cos</td><td>Acc</td><td>Flu</td><td>Cos</td><td>Acc</td><td>Flu</td></tr><tr><td>All-or-Nothing Thinking*</td><td>2.203</td><td>2.607</td><td>2.470</td><td>2.610</td><td>2.590</td><td>2.730</td><td>0.407</td><td>0.017</td><td>0.260</td></tr><tr><td>Overgeneralization*</td><td>2.287</td><td>2.767</td><td>2.609</td><td>2.280</td><td>2.520</td><td>2.860</td><td>0.007</td><td>0.247</td><td>0.251</td></tr><tr><td>Mental Filter*</td><td>2.247</td><td>2.677</td><td>2.578</td><td>2.480</td><td>2.460</td><td>2.830</td><td>0.233</td><td>0.217</td><td>0.252</td></tr><tr><td>Discounting the Positive</td><td>2.153</td><td>2.240</td><td>2.640</td><td>2.120</td><td>2.710</td><td>2.880</td><td>0.033</td><td>0.470</td><td>0.240</td></tr><tr><td>Jumping to Conclusions</td><td>2.279</td><td>2.361</td><td>2.550</td><td>2.560</td><td>2.890</td><td>2.840</td><td>0.281</td><td>0.529</td><td>0.290</td></tr><tr><td>Magnification and Minimization*</td><td>2.212</td><td>2.531</td><td>2.625</td><td>2.330</td><td>2.100</td><td>2.730</td><td>0.118</td><td>0.431</td><td>0.105</td></tr><tr><td>Emotional Reasoning*</td><td>2.624</td><td>2.887</td><td>2.713</td><td>2.020</td><td>2.200</td><td>2.880</td><td>0.604</td><td>0.687</td><td>0.167</td></tr><tr><td>Should Statements</td><td>2.315</td><td>2.562</td><td>2.654</td><td>2.110</td><td>2.600</td><td>2.770</td><td>0.205</td><td>0.038</td><td>0.116</td></tr><tr><td>Labeling</td><td>2.309</td><td>2.563</td><td>2.499</td><td>2.380</td><td>2.700</td><td>2.770</td><td>0.071</td><td>0.137</td><td>0.271</td></tr><tr><td>Personalization</td><td>2.250</td><td>2.632</td><td>2.648</td><td>1.890</td><td>2.810</td><td>2.860</td><td>0.360</td><td>0.178</td><td>0.212</td></tr><tr><td>Total mean</td><td>2.287</td><td>2.582</td><td>2.598</td><td>2.278</td><td>2.558</td><td>2.815</td><td>0.231</td><td>0.295</td><td>0.216</td></tr></table>
+
+Table 7: Comparative evaluation of cognitive distortions by LLMs and humans: Cos (Consistency), Acc (Accuracy), and Flu (Fluency). *Types of cognitive distortions easily detected by LLMs.
+
+# 5.2 Comparison of LLMs and Human Evaluations Across Criteria
+
+To ensure objectivity, the model generating synthetic data was excluded from evaluation. Two other models independently scored the data, averaging their scores for the final result. Evaluation parameters for the three LLMs are in Table 11, Appendix E.
+
+For human evaluation, 50 or 100 synthetic samples per distortion were randomly selected, totaling 900 samples. Two psychology experts independently assessed them using the same criteria as LLM evaluation, with a Cohen's kappa of 0.78 indicating substantial agreement.
+
+Table 5 summarizes the evaluation results from both LLMs and humans, highlighting the differences between the two types of synthetic data (Section 3.5)—cognitive clarification and cognitive balancing—across the three criteria.
+
+Detailed results for each model are in Table 10 in Appendix D.
+
+Human evaluation scores were lower across all criteria except fluency, with accuracy showing the largest gap. This difference stems from LLMs' strength in detecting explicit text patterns while struggling with the implicit reasoning essential for cognitive distortion evaluation, highlighting their limitations. Table 6 provides detailed expert feedback.
+
+Regarding the two synthetic data generation methods, in the LLM evaluation, the cognitive clarification method scored 0.1 to 0.3 points higher on all criteria than the cognitive balancing method. However, in the human evaluation, only the cognitive balancing method showed higher accuracy.
+
+# 5.3 Comparison of LLM and Human Performance in Cognitive Distortion Classification
+
+To further analyze the differences between LLM-based and human evaluations, we compared the scores for each cognitive distortion. Table 7 presents comparative results, highlighting key discrepancies between the two evaluation methods. Scores were compared between LLM and human evaluations, with the higher values in bold. The 'Difference' column shows score gaps, with differences of 0.4 or greater also bolded.
+
+The average LLMs evaluation scores were 2.287 for consistency, 2.582 for accuracy, and 2.598 for fluency, while the average human evaluation scores were 2.278 for consistency, 2.558 for accuracy, and 2.815 for fluency. Fluency was higher in human evaluation, whereas consistency and accuracy showed no significant difference, though human scores were slightly lower overall. The higher fluency score in human evaluation is likely because LLMs assessed synthetically generated sentences, which were naturally structured and free of pauses.
+
+In the evaluation of cognitive distortions by type, human scores were lower than those of LLMs in some cases, particularly in accuracy. For example, scores for "Emotional Reasoning" (2.887 vs. 2.200) and "Magnification and Minimization" (2.531 vs. 2.100) showed notable differences. This discrepancy may be because LLMs excel at detecting clear linguistic patterns, such as "Should Statements," "Labeling," and "Discounting the Positive." However, human evaluation tends to be more reliable for distortions requiring inferential reasoning, such as "Mental Filtering" and "Magnification and Minimization," since these rely on deeper contextual understanding.
+
+These findings highlight that LLMs rely more on explicit linguistic patterns, whereas human evaluators consider deeper contextual reasoning, which may impact their ability to identify distortions that require implicit inference.
+
+# 6 Conclusion and Future Work
+
+We developed KoACD, a dataset of cognitive distortions in Korean adolescents, overcoming the limitations of small-scale datasets focused on English-speaking adults. KoACD offers a balanced representation of cognitive distortions through the creation of synthetic data. To our knowledge, it is the first dataset specifically designed for Korean adolescents.
+
+We introduced a multi-LLM negotiation method to improve the objectivity and accuracy of the
+
+synthetic data. By using multiple LLMs to negotiate and refine cognitive distortion labels, we minimized biases and enhanced data quality. Expert and LLM evaluations confirmed that LLMs performed well when clear linguistic cues were present, while human evaluators showed higher accuracy in context-dependent situations. Discrepancies between LLMs and human evaluations highlighted the LLMs' reliance on superficial linguistic patterns.
+
+Future work will focus on fine-tuning models with adolescent-specific data to enhance contextual understanding of cognitive distortions. Additionally, we aim to improve LLM performance by developing algorithms that better distinguish cognitive distortions, mitigating biases toward specific types and enhancing both balance and accuracy in detection.
+
+# Limitations
+
+We recognize that there are some limitations to the methods for detecting cognitive distortions and to the KoACD dataset:
+
+Cognitive Distortion Classification We assigned the most appropriate cognitive distortion to each question, but some questions may involve multiple distortions simultaneously. The boundaries between some types of distortions are blurred, making classification challenging and leading to potential discrepancies between the model and human raters. To address these issues, a multi-label classification method and more refined criteria are needed.
+
+Multi-LLMs Negotiation Methods We designed the LLMs to alternate between Analyzer and Evaluator roles, but the results can vary depending on the model used. Therefore, negotiation results with different LLMs should also be considered. Additionally, discrepancies between analysts and evaluators sometimes result in data being classified as "Unknown," even after five rounds of negotiation, due to the inability to fit the data within the ten cognitive distortion categories. Interpretation of such data is essential, and further research is needed to develop more accurate detection methodologies.
+
+LLMs and Human Evaluation While the KoACD is a large dataset, the amount of data reviewed by human raters is relatively small. Although human raters excel at considering context for accurate judgments, subjectivity in the evaluation process and inconsistency due to differing standards among raters may arise. Future research should focus on securing more human
+
+evaluation data and developing more precise evaluation standards to increase reliability.
+
+# Ethical Considerations
+
+In this study, we collected publicly accessible data from Naver Knowledge iN, and users participate anonymously on the platform. We only used publicly available data in the course of our research and did not interact directly with Naver Knowledge iN users.
+
+We have identified that the data collection process may include various inappropriate topics, such as hate speech, violence, sexual content, and profanity. Accordingly, we have attempted to exclude such data as much as possible by applying strict filtering criteria including keyword-based filtering and manual inspection of sampled posts. However, we cannot completely rule out the possibility that some inappropriate content may be included in the data. We are aware of the risk that AI models may be trained on inappropriate data and produce biased or unethical results. Therefore, it is important to continuously monitor the ethical use of AI models and improve filtering techniques to address this risk.
+
+All processed and derived datasets are intended solely for research purposes and must not be used for commercial or any other non-research purposes. Their use is consistent with the original access conditions of Naver Knowledge iN.
+
+To ensure the reliability of the cognitive distortion annotations, two psychological experts—each holding a master's degree in clinical or counseling psychology and possessing over five years of relevant experience—independently reviewed and validated 900 samples from the dataset. The experts assessed the appropriateness of the assigned cognitive distortion types based on predefined guidelines established for this study.
+
+All data samples were fully anonymized prior to evaluation to eliminate any risk of personal identification. No personal information was collected from either the dataset or the expert reviewers.
+
+The experts voluntarily participated in the annotation task after being fully informed of the study's purpose and procedures. They were compensated with 300,000 KRW, an amount determined to reflect fair payment based on standard professional rates and estimated task duration.
+
+The annotation process was conducted under strict confidentiality, and all procedures adhered to ethical standards for research involving domain experts.
+
+All experts provided informed consent to participate in this study, and we explained how the annotation evaluation would be conducted and how their feedback would be used for research purposes.
+
+# Acknowledgments
+
+This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (RS-2024-00459226).
+
+# References
+
+Anthropic. 2024. The Claude 3 Model Family: Opus, Sonnet, Haiku.  
+Aaron T. Beck. 1979. Cognitive therapy and the emotional disorders. Penguin.  
+Aaron T. Beck. 1991. Cognitive therapy: A 30-year retrospective. American Psychologist, 46(4):368. https://doi.org/10.1037/0003-066X.46.4.368.  
+Zhiyu Chen, Yujie Lu, and William Yang Wang. 2023. Empowering psychotherapy with large language models: Cognitive distortion detection through diagnosis of thought prompting. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 4295-4304, Singapore. Association for Computational Linguistics. https://doi.org/10.18653/v1/2023.findings-emnlp.284.  
+Yao Fu, Hao Peng, Tushar Khot, and Mirella Lapata. 2023. Improving language model negotiation with self-play and in-context learning from AI feedback. arXiv preprint, arXiv:2305.10142. https://doi.org/10.48550/arXiv.2305.10142.  
+Gemini Team. 2024. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint, arXiv:2403.05530. https://doi.org/10.48550/arXiv.2403.05530.  
+Moonkyoung Jang and Seongcheol Kim. 2024. Key traits of top answers on Korean social Q&A platforms: Insights into user performance and entrepreneurial potential. Humanities and Social Sciences Communications, 11(1):1-13. https://doi.org/10.1057/s41599-024-03239-3.  
+Lee, Geonju, Dabin Park, and Hayoung Oh. 2023. Methodology of labeling according to 9 criteria of DSM-5. Applied Sciences, 13(10481). https://doi.org/10.3390/app131810481.  
+Saghar Chahar Mahali, Shadi Beshai, Justin R. Feeney, and Sandeep Mishra. 2020. Associations of negative cognitions, emotional regulation, and depression symptoms across four continents: International support for the cognitive model of depression. BMC Psychiatry, 20(18):1-12. https://doi.org/10.1186/s12888-019-2423-x.
+
+OpenAI. 2024. GPT-4o mini: advancing cost-efficient intelligence. https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence/  
+Jennifer H. Pfeifer and Elliot T. Berkman. 2018. The development of self and identity in adolescence: Neural evidence and implications for a value-based choice perspective on motivated behavior. *Child Development Perspectives*, 12(3):158-164. https://doi.org/10.1111/cdep.12279.  
+Paula R. Pietromonaco and Hazel Markus. 1985. The nature of negative thoughts in depression. Journal of Personality and Social Psychology, 48(3):799-807. https://doi.org/10.1037/0022-3514.48.3.799.  
+Hongzhi Qi, Qing Zhao, Jianqiang Li, Changwei Song, Wei Zhai, Dan Luo, Shuo Liu, Yi Jing Yu, Fan Wang, Huijing Zou, Bing Xiang Yang, and Guanghui Fu. 2023. Supervised learning and large language model benchmarks on mental health datasets: Cognitive distortions and suicidal risks in Chinese social media. Preprint, version 1. Available at Research Square. https://doi.org/10.21203/rs.3.rs-3523508/v1.  
+Katerina Rnic, David J. A. Dozois, and Rod A. Martin. 2016. Cognitive distortions, humor styles, and depression. *Europe's Journal of Psychology*, 12(3):348-362. https://doi.org/10.5964/ejop.v12i3.1118.  
+Ryan Shea, Aymen Kallala, Xin Lucy Liu, Michael W. Morris, and Zhou Yu. 2024. ACE: A LLM-based Negotiation Coaching System. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP 2024), pages 12720-12749, Miami, Florida, USA. Association for Computational Linguistics. https://doi.org/10.18653/v1/2024.emnlp-main.709.  
+Benjamin Shickel, Scott Siegel, Martin Heesacker, Sherry Benton, and Parisa Rashidi. 2020. Automatic detection and classification of cognitive distortions in mental health text. In Proceedings of the 2020 IEEE 20th International Conference on Bioinformatics and Bioengineering (BIBE), pages 275-280. IEEE. https://doi.org/10.1109/BIBE50027.2020.00052.  
+Taetem Simms, Christopher Ramstedt, Marc Rich, Matthew Richards, Thomas Martinez, and Christophe Giraud-Carrier. 2017. Detecting cognitive distortions through machine learning text analytics. In Proceedings of the 2017 IEEE International Conference on Healthcare Informatics (ICHI), pages 508-512. IEEE. https://doi.org/10.1109/ICHI.2017.39.  
+Xiaofei Sun, Xiaoya Li, Shengyu Zhang, Shuhe Wang, Fei Wu, Jiwei Li, Tianwei Zhang, and Guoyin Wang. 2023. Sentiment analysis through LLM negotiations.
+
+arXiv preprint, arXiv:2311.01876. https://doi.org/10.48550/arXiv.2311.01876.  
+Justin S. Tauscher, Kevin Lybarger, Xiruo Ding, Ayesha Chander, William J. Hudenko, Trevor Cohen, and Dror Ben-Zeev. 2023. Automated detection of cognitive distortions in text exchanges between clinicians and people with serious mental illness. *Psychiatric Services*, 74(4):407-410. https://doi.org/10.1176/appi.ps.202100692.  
+UNICEF. 2018. Ensuring mental health and well-being in an adolescent's formative years can foster a better transition from childhood to adulthood. UNICEF.
+
+<table><tr><td>Institution</td><td>2011 - 2015</td><td>2016 - 2020</td><td>2021 - 2024</td><td>Total</td></tr><tr><td>이성가복부.한국형upro단복지개발단
+Korea Youth Counseling &amp; Welfare Institute</td><td>17,479</td><td>19,465</td><td>15,699</td><td>52,643</td></tr><tr><td>이터지.
+Incheon Youth Support Center</td><td>1,357</td><td>1,355</td><td>-</td><td>2,712</td></tr><tr><td>울수지.
+Ulsan Youth Support Center</td><td>1,655</td><td>2,360</td><td>1,518</td><td>5,533</td></tr><tr><td>glyeonggi Youth Support Center</td><td>6,862</td><td>-</td><td>53</td><td>6,915</td></tr><tr><td>형upro단 motion상dap선拓
+Youth Mobile Counseling Center</td><td>-</td><td>407</td><td>1,715</td><td>2,122</td></tr><tr><td>Overall Total</td><td>27,353</td><td>23,587</td><td>18,985</td><td>69,925</td></tr></table>
+
+# A Distribution of Data Sources
+
+We collected 69,925 questions and answers from five major organizations and services specializing in adolescent counseling on Naver Knowledge iN. Table 8 shows the data collection status by organization and year, along with the distribution of questions collected from 2011 to 2024.
+
+# B Data Preprocessing Details
+
+The collected data was refined and pre-processed to ensure relevance. The following criteria were applied to remove misaligned data:
+
+1. Non-adolescent questions: To exclude questions written by elementary school students or adults, we applied keyword-based filtering, resulting in the removal of 14,075 questions.  
+2. Inappropriate content: A total of 7,397 questions containing inappropriate sexual content were removed to maintain alignment with the research scope.  
+3. Lack of specificity: To eliminate vague questions that hinder meaningful analysis, 9,240 questions with 15 words or fewer in the detailed worry column were deleted.  
+4. Duplicate entries: To ensure data uniqueness and prevent redundancy, 2,089 duplicate questions were removed.
+
+After applying the above criteria for pre-processing, 37,124 data points were selected and used in the study.
+
+Table 8: Distribution of Q&A and worry Q&A data by institution and year  
+
+<table><tr><td>Round</td><td>Turn 1</td><td>Turn 2</td><td>Total (Cumulative %)</td></tr><tr><td>Round 1</td><td>17,694</td><td>1,841</td><td>17,694 (51%)</td></tr><tr><td>Round 2</td><td>6,132</td><td>361</td><td>6,493 (75%)</td></tr><tr><td>Round 3</td><td>4,240</td><td>57</td><td>4,297 (87%)</td></tr><tr><td>Round 4</td><td>1,243</td><td>37</td><td>1,280 (91%)</td></tr><tr><td>Round 5</td><td>784</td><td>4,735</td><td>5,519 (100%)</td></tr></table>
+
+Table 9: Turn counts across negotiation rounds
+
+# C Changes in Cognitive Distortion Classification Through Negotiation
+
+We analyze the distribution of data based on the number of negotiation rounds required to determine cognitive distortions. Table 9 presents the count of instances finalized at each round, illustrating how much data was classified early versus how much required additional rounds. The cumulative percentage represents the proportion of data for which cognitive distortion classification was completed at each round.
+
+# D Detailed Evaluation Results of LLM-Based Assessment
+
+This appendix presents the detailed evaluation results of the LLM-based assessment for the two synthetic data generation methods: Cognitive Clarification and Cognitive Balancing. Each model's performance was assessed based on three criteria—Consistency, Accuracy, and Fluency—using independent evaluations by Gemini 1.5 Flash, GPT-4o mini, and Claude 3 Haiku, as shown in Table 10.
+
+<table><tr><td rowspan="3">Generation Method</td><td rowspan="3">Generation Model</td><td colspan="9">Evaluation model</td></tr><tr><td colspan="3">Gemini-1.5-flash</td><td colspan="3">GPT-4o mini</td><td colspan="3">Claude-3-haiku</td></tr><tr><td>Cos</td><td>Acc</td><td>Flu</td><td>Cos</td><td>Acc</td><td>Flu</td><td>Cos</td><td>Acc</td><td>Flu</td></tr><tr><td rowspan="3">Cognitive Clarification</td><td>Gemini-1.5-flash</td><td>-</td><td>-</td><td>-</td><td>2.596</td><td>2.929</td><td>2.948</td><td>2.400</td><td>2.779</td><td>2.416</td></tr><tr><td>GPT-4o mini</td><td>2.142</td><td>2.498</td><td>2.638</td><td>-</td><td>-</td><td>-</td><td>2.508</td><td>2.774</td><td>2.472</td></tr><tr><td>Claude-3-haiku</td><td>2.150</td><td>2.519</td><td>2.643</td><td>2.606</td><td>2.754</td><td>2.814</td><td>-</td><td>-</td><td>-</td></tr><tr><td rowspan="3">Cognitive Balancing</td><td>Gemini-1.5-flash</td><td>-</td><td>-</td><td>-</td><td>2.253</td><td>2.718</td><td>2.740</td><td>2.134</td><td>2.589</td><td>2.298</td></tr><tr><td>GPT-4o mini</td><td>1.882</td><td>2.090</td><td>2.515</td><td>-</td><td>-</td><td>-</td><td>2.111</td><td>2.333</td><td>2.312</td></tr><tr><td>Claude-3-haiku</td><td>1.966</td><td>2.162</td><td>2.547</td><td>2.284</td><td>2.604</td><td>2.760</td><td>-</td><td>-</td><td>-</td></tr></table>
+
+Table 10: Detailed evaluation results of synthetic data: Cos (Consistency), Acc (Accuracy), and Flu (Fluency)  
+
+<table><tr><td>Methodology</td><td>Model</td><td>Temperature</td><td>Max Tokens</td><td>Top-p</td></tr><tr><td rowspan="3">(A) Negotiation process</td><td>Gemini 1.5 Flash</td><td>0.5</td><td>1,024</td><td>0.9</td></tr><tr><td>GPT-4o mini</td><td>0.5</td><td>1,024</td><td>0.9</td></tr><tr><td>Claude-3 Haiku</td><td>0.5</td><td>1,024</td><td>0.9</td></tr><tr><td rowspan="3">(B) Synthetic data generation</td><td>Gemini 1.5 Flash</td><td>1.0</td><td>1,024</td><td>0.9</td></tr><tr><td>GPT-4o mini</td><td>1.0</td><td>1,024</td><td>0.9</td></tr><tr><td>Claude-3 Haiku</td><td>1.0</td><td>1,024</td><td>0.9</td></tr><tr><td rowspan="3">(C) Evaluation</td><td>Gemini 1.5 Flash</td><td>0.5</td><td>512</td><td>0.9</td></tr><tr><td>GPT-4o mini</td><td>0.5</td><td>512</td><td>0.9</td></tr><tr><td>Claude-3 Haiku</td><td>0.5</td><td>512</td><td>0.9</td></tr></table>
+
+Table 11: Hyperparameters for negotiation process, synthetic data generation, and evaluation
+
+# E Hyperparameters of LLMs Models
+
+We utilized Claude-3 Haiku, Gemini 1.5 Flash, and GPT-4o Mini at different stages of this study, summarizing the hyperparameters used at each step. Table 11(A) presents the hyperparameters for the negotiation process and independent evaluation, with Gemini 1.5 Flash and GPT-4o Mini used during the negotiation, and Claude-3 Haiku employed for independent evaluation. Table 11(B) summarizes the hyperparameters for synthetic data generation, while Table 11(C) outlines the hyperparameters used for evaluating synthetic data. The hyperparameters were empirically determined through multiple experiments, focusing on maximizing output diversity, ensuring reproducibility, and guaranteeing sufficient output size.
+
+# F Validating Synthetic Data with Clustering
+
+To validate the synthetic data, we conducted clustering based on two criteria: (1) topics that elicit cognitive distortion in adolescents and (2) negative emotions and symptoms from the DSM-5.
+
+To perform clustering, we first created mapping dictionaries for each criterion. Table 12 lists keywords for cognitive distortion topics in adolescents, and Table 13 shows DSM-5 depression symptom categories with related keywords.
+
+We discovered keywords mapped to each category based on the topic-based mapping (Table 14) and DSM-5 symptom-based mapping (Table 15) of the synthetic data in the mapping dictionary. Only keywords with a mapping frequency of over 1,000 were selected, and the results were checked with the keyword in English, Korean, and the mapping frequency.
+
+<table><tr><td>Topic (Korean, n)</td><td>List of Keywords (Korean)</td></tr><tr><td>Academic performance and career concerns (학现阶段에 specificor, n=31)</td><td>Academics (학), Academy (학원), Advancement to Higher Education (학교), Class (수簸), Club (동아리), College Entrance Exam (일서), College Entrance Exam (수簸), Dropping out (자회), English (영어), English Academy (영어학원), Enrollment (재hua), Exam (서형), Final Exam (기목고사), GED ( gymnasiol), Grades (성SOC), Harass/Bully (개류하다), High School (고등학교), Homework (수 stu), Math (수학), Middle School (중학교), Midterm Exam (중kan고수), Mock Exam (모고수), Private Tutoring (과와), Rank (등목), Report Card (성SOC), Retaking the College Entrance Exam (개수), Scholarship (상학원), School (학교), School Life (학교生命周期), School Record (내신), Specialized High School (류상회고), Study (공부), Teacher (선형NM)</td></tr><tr><td>Friendships, romantic relationships, and interpersonal relationships (무형, 전해, 썼이에 기재, n=19)</td><td>Acquaintance (지인), Best Friend (단rock), Boyfriend (단가 친구), Boyfriend (단본), Break up (개어지다), Bullying (울러), Close Friend (단한 친구), Confession (고백), Crush (심사랑), Dating (사려다), Friend ( 친구), Friendship ( 친구사 ), Girlfriend ( 친자 친구), Heartbreak (심은,마운의상chter), Jealousy (울복), Love (사라), Loyalty (우gly, 편리)</td></tr><tr><td>Physical and mental health (선체워, 썼이에 기재, n=39)</td><td>Alone (화자), Anxiety (울한), Binge Eating Disorder (류식물), Comfort (류로), Confidence (자신安全感), Concern (고려), Counseling (상품), Counselor (상품사), Depression (울울), Depression (울증), Domestic Violence (가형복려), Emotional Wound (상품), Exercise (운동), Fatigue (피로), Guilt (피복 ), Headache (류화), Heart/Mind (마화), Inferiority Complex (열들고�), Inner Self (내면), Insomnia (fol各个环节), Loneliness (심로울), Mental Illness (성신화), Mental Strength ( Behavioral), Obesity (비단), Psychotherapy (심리지료), Running Away (가화), School Bullying (가화 ), School Violence (가화복려), Self-esteem (자운 가 ), Sleep Disorder (수만 기재), Stress (스트래斯), Therapy ( 친류), Unconscious Mind (무의심), Violence (류려), Worry (戰態)</td></tr><tr><td>Family issues (가조목제, n=20)</td><td>Dad (어머), Divorce (이후), Domestic Conflict (가형복화), Domestic Violence (가형복려), Estrangement (소원화), Family (가후), Family Conflict (가후 갑등), Family History (가후사), Father (어머지), Financial Issues (가형복课题), Home/Family Environment (가후), Mom (어머), Mother (어머니), Neglect (방울), Older Sister (수다), Older Sister (머니), Parents (부모머), Younger Brother (남동상 ), Younger Sibling (동상 ), Younger Sister (어로상 )</td></tr><tr><td>Appearance and self-image (무모에 어머지, n=11)</td><td>Acne (어드름), Appearance (무모), Body Shape (목무), Bulking Up (버크어 ), Diet (다이어트), Facial Features (IELGE 生長 ,(Height (身高)), Makeup (메이크워 ), Muscle (고목 ), Plastic Surgery (성형), Skin (피부)</td></tr></table>
+
+Table 12: List of Keywords of Negative Emotion-Triggering Topics in Adolescents
+
+<table><tr><td>DSM-5 Depression Symptom Class (Korean, n)</td><td>List of Symptom or Emotion (Korean)</td></tr><tr><td>A. Depressed mood (울다, Depression (울다,울다), Despair (심명), Disappointment (심명), Emptiness (공허, 퍼타), Frustration (화워), Guilt (심traffam), Hard (심들다), Heartache (상ikh), Pain (고通過), Sad (수화다), Scared (무서운, 죽iates), Suffering (피로울), Tough (심들), Unhappiness (심traffam), Upset (화cona, 출상회), Worthlessness (무가지회)</td><td>Crying (울다), Depression (울다,울다), Despair (심traffam), Disappointment (심traffam), Emptiness (공허, 퍼타), Frustration (화워), Guilt (심traffam), Hard (심들다), Heartache (상ikh), Pain (고通過), Sad (수화다), Scared (무서운, 죽iates), Suffering (심traffam), Tough (심들), Unhappiness (심traffam), Upset (화cona, 출상회), Worthlessness (무가지회)</td></tr><tr><td>B. Loss of interest/pleasure (증마이에는 죽iates에의상품, n=14)</td><td>Alienated (소의), Alone (한자, 퍼로), Apathy (내용드), Bore (지류회), Bullying (마���목), Unpleasant (복테일), Ignored (무지), Indifference (무원신), Isolated (고려), Loneliness (고려울), Lonely (고려울), Meaningless (무의지회), Disinterest (증마이에 odds)</td></tr><tr><td>C. Weight loss or gain (체 weight or gain, n=9)</td><td>Appetite (식물), Binge Eating (식물), Body (물개), Diet (다이어트, 죽iates), Fat (상지다), Loss of Appetite (식물 기회), Nausea (개수 기회), Weight (체 weight)</td></tr><tr><td>D. Insomnia or hypersomnia (복만중에고해 죽iates, n=15)</td><td>Daytime Fatigue (주가지회), Hypersomnia (과다수면), Insomnia (身高, 컡지), Restless Sleep (류chts일), Sleep (수면, 컡), Sleep Deprivation (수면부종), Sleep Disorder (수면장에), Sleep Patterns (수면 편면), Sleepiness (身高), Sleeping Pills (수면제), Stress (스트래스), Worry (고려, 겨름)</td></tr><tr><td>E. Psychomotor agitation or retardation (道路上만 기재에고해 죽iates, n=13)</td><td>Anger (본단), Anxiety (身高, 컡일), Anxiety disorder (身高), Irritability (과지, 사很快就�), Nervousness (身高, 편조, 편성), Obsessive (身高), Obsessive-compulsive disorder (身高), Sensitive (에지, 편일), Tension (고려, 편성)</td></tr><tr><td>F. Fatigue (PTHROMAS, n=17)</td><td>Dejected (고려, 퍼타), Empty (공허), Exhausted (지려다, 편진), Fatigued (PTHROMAS), Helpless (무려고�), Incompetence (무能力强, 편종), Inferiority (PTHROMAS), Self-harm (PTHROMAS), Lethargy (무기고�,무기고�), Powerless (무기고�), 편이 doesn&#x27;t, Tired (PTHROMAS), Lethargy (PTHROMAS)</td></tr><tr><td>G. Inappropriate guilt (부ध patrimonial, 편고�, 편성), n=11)</td><td>Guilt (심traffam), Helplessness (무려고�), Incompetence (무能力强, 편성), Inferiority (PTHROMAS), Regret (고려), Self-blame (자ophile), Shame (수고�, 편성)</td></tr><tr><td>H. Decreased concentration (심중요고�, n=12)</td><td>Concentration (심중요,심중요), Concern (심려, 유려, 고지), Confusion (심려), Distracted (심만고�, majority), Discomfort/Inconvenience (심려,심TRAffam), Forgetfulness (심TRAffam), Judgment (심TRAffam)</td></tr><tr><td>I. Thoughts of suicide (자심사고�, n=9)</td><td>Death (심은), Desperation (심중�, 편성), Die (심다), Fear (심려고�), Self-harm (심려), Suicide (자심), Suicidal Ideation (자심울동, majority), Suicide (자심), Suicide ideation (자심울동, majority)</td></tr></table>
+
+Table 13: DSM-5 Depression Symptom related Classes and Keywords
+
+<table><tr><td>Topic</td><td>Keyword
+(Korean)</td><td>Keyword
+(English)</td><td>Count</td></tr><tr><td>Relationship</td><td>KHJUG</td><td>Friend</td><td>64,041</td></tr><tr><td>Academic</td><td>CHOHO</td><td>School</td><td>21,298</td></tr><tr><td>Mental Health</td><td>HNJAN</td><td>Alone</td><td>18,482</td></tr><tr><td>Mental Health</td><td>JEYHON</td><td>Worry</td><td>16,636</td></tr><tr><td>Academic</td><td>SINHOM</td><td>Exam</td><td>13,168</td></tr><tr><td>Academic</td><td>선목상품</td><td>Teacher</td><td>11,163</td></tr><tr><td>Academic</td><td>공부</td><td>Study</td><td>9,987</td></tr><tr><td>Mental Health</td><td>복원</td><td>Anxiety</td><td>9,716</td></tr><tr><td>Academic</td><td>성종貌</td><td>Grades</td><td>9,407</td></tr><tr><td>Family</td><td>머지</td><td>Mom</td><td>8,458</td></tr><tr><td>Mental Health</td><td>마이트</td><td>Heart/Mind</td><td>7,543</td></tr><tr><td>Academic</td><td>수CPP</td><td>Class</td><td>6,987</td></tr><tr><td>Family</td><td>부모statt侣</td><td>Parents</td><td>5,121</td></tr><tr><td>Mental Health</td><td>고만</td><td>Concern</td><td>3,966</td></tr><tr><td>Academic</td><td>고등학교</td><td>High School</td><td>3,725</td></tr><tr><td>Academic</td><td>학원</td><td>Academy</td><td>3,409</td></tr><tr><td>Academic</td><td>수CPP</td><td>Math</td><td>3,347</td></tr><tr><td>Mental Health</td><td>수CPP</td><td>Depression</td><td>2,927</td></tr><tr><td>Appearance</td><td>기</td><td>Height</td><td>2,690</td></tr><tr><td>Family</td><td>이름</td><td>Dad</td><td>2,312</td></tr><tr><td>Academic</td><td>중학교</td><td>Middle School</td><td>2,065</td></tr><tr><td>Academic</td><td>수CPP</td><td>Homework</td><td>1,992</td></tr><tr><td>Mental Health</td><td>스트래스</td><td>Stress</td><td>1,798</td></tr><tr><td>Mental Health</td><td>운동</td><td>Exercise</td><td>1,639</td></tr><tr><td>Family</td><td>가SOC</td><td>Family</td><td>1,636</td></tr><tr><td>Mental Health</td><td>상dap</td><td>Counseling</td><td>1,619</td></tr><tr><td>Academic</td><td>자회</td><td>Dropping out</td><td>1,458</td></tr><tr><td>Academic</td><td>선목상품</td><td>Report Card</td><td>1,343</td></tr><tr><td>Academic</td><td>진학</td><td>Advancement to Higher Education</td><td>1,241</td></tr></table>
+
+<table><tr><td>Relationship</td><td>화해</td><td>Reconciliation</td><td>1,148</td></tr><tr><td>Mental Health</td><td>상chter</td><td>Emotional Wound</td><td>1,125</td></tr><tr><td>Relationship</td><td>고복</td><td>Confession</td><td>1,113</td></tr><tr><td>Relationship</td><td>선empre</td><td>Senior</td><td>1,086</td></tr><tr><td>Relationship</td><td>사랑</td><td>Love</td><td>1,025</td></tr></table>
+
+Table 14: Frequency Distribution of 34 Keywords Across Topics
+
+<table><tr><td>Topic</td><td>Keyword (Korean)</td><td>Keyword (English)</td><td>Count</td></tr><tr><td>A. Depressed mood</td><td>심명</td><td>Disappointment</td><td>3,722</td></tr><tr><td>A. Depressed mood</td><td>울울</td><td>Depression</td><td>2,927</td></tr><tr><td>A. Depressed mood</td><td>/help들</td><td>Struggling</td><td>13,694</td></tr><tr><td>A. Depressed mood</td><td>/help들다</td><td>Difficult</td><td>1,332</td></tr><tr><td>B. Loss of interest/pleasure</td><td>부視</td><td>Neglect</td><td>8,226</td></tr><tr><td>B. Loss of interest/pleasure</td><td>소之乡</td><td>Alienation</td><td>2,609</td></tr><tr><td>B. Loss of interest/pleasure</td><td>회자</td><td>Alone</td><td>18,482</td></tr><tr><td>D. Insomnia or hypersomnia</td><td>戰CRY</td><td>Worry</td><td>16,636</td></tr><tr><td>D. Insomnia or hypersomnia</td><td>고ikit</td><td>Concern</td><td>3,966</td></tr><tr><td>D. Insomnia or hypersomnia</td><td>스트래斯</td><td>Stress</td><td>1,798</td></tr><tr><td>D. Insomnia or hypersomnia</td><td>증</td><td>Nap</td><td>2,357</td></tr><tr><td>E. Psychomotor agitation or retardation</td><td>相关内容</td><td>Anxiety</td><td>9,716</td></tr><tr><td>E. Psychomotor agitation or retardation</td><td>목상</td><td>Annoyance</td><td>2,868</td></tr><tr><td>F. Fatigue</td><td>부nergy</td><td>Inability</td><td>2,332</td></tr><tr><td>G. Inappropriate guilt</td><td>무既能</td><td>Incompetence</td><td>2,332</td></tr><tr><td>G. Inappropriate guilt</td><td>후회</td><td>Regret</td><td>1,451</td></tr><tr><td>H. Decreased concentration</td><td>過程</td><td>Worry</td><td>16,636</td></tr><tr><td>H. Decreased concentration</td><td>고一定能</td><td>Concern</td><td>3,966</td></tr><tr><td>H. Decreased concentration</td><td>相关内容</td><td>Discomfort</td><td>1,959</td></tr><tr><td>H. Decreased concentration</td><td>相关内容</td><td>Focus</td><td>1,971</td></tr></table>
+
+Table 15: Frequencies of 20 keywords across DSM-5 symptom categories
+
+# G Examples of Synthetic Data in KoACD
+
+In Table 16, we provide one synthetic example per cognitive distortion, totaling 10.
+
+<table><tr><td>Cognitive Distortion
+Type (English / Korean)</td><td>Sentence (English / Korean)</td></tr><tr><td>All-or-Nothing Thinking
+(심 bile사고)</td><td>&quot;I&#x27;m really bad at studying, and my grades are at the bottom. I can&#x27;t even think about college, and getting a job seems impossible too. It feels like I have no future, like I&#x27;ve completely failed.&quot; 
+(공부를 전쟁무해서 쓸이에요. entire of the whole thing that is not just the whole thing that is the whole thing that is the whole thing that is the whole thing that is the whole thing that is the whole thing that is the whole thing that is the whole thing that is the whole thing that is the whole thing that is the whole thing that is the whole thing that is the whole thing that is the whole thing that is the whole thing that is the whole thing that is the whole thing that is the whole thing that is the whole thing that is the whole thing that is the whole thing that in some way has been so much more than what it was before. 
+(雖然 흠울성 기재에요, but it&#x27;s not enough to say anything else. 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.)</td></tr><tr><td>Overgeneralization
+(과일본화)</td><td>&quot;I feel anxious because it seems like my classmates avoid talking to me. One day, I felt so left out that I cried. There have been so many times when everyone gathered and left me out. Now, I&#x27;m scared of being alone.&quot; 
+(수어는 사원을 나와의 어머기를 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기我们可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 춅고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅��以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고ISC.
+(내용품만 기재에요, but it&#x27;s not enough to say anything else. 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don’t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t known what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know what to do.) 
+(I don&#x27;t know who is the person who will be going to see you, but you are going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person Who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is she or he who will be going to see you.) 
+(I don&#x27;t know who is she or she who will be going to see you.) 
+(I don&#x27;t know who is she or she who will be going to see you.) 
+(I don&#x27;t know who is she or she who will be going to see you.) 
+(I don&#x27;t know who is she or she who will be going to see you.) 
+(I don&#x27;t know who is she or she who will be going to see you.) 
+(I don&#x27;t know who is she or she who will be going to see you.) 
+I don&#x27;t know what I should study hard for this exam, but it feels so tough every time I try, and I just want to give up. But I know I can&#x27;t, I have to keep going and work hard to get good grades.&quot; 
+(이터어는 출류를 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고ISC. 
+(I don&#x27;t know who is the person who will be going to see you, but you are going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who in the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t known who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+I don&#x27;t know what I should study hard for this exam, but it feels so tough every time I try, and I just want to give up. But I know I can&#x27;t, I have to keep going and work hard to get good grades.&quot; 
+(이터어는 출류를 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고 기不仅可以 썸다고ISC. 
+(I don&#x27;t know who is the person who will be going to see you, but you are going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who Will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is theperson who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who in the person who will be going to see you.) 
+(I don&#x27;t know who is the person Who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t know who is the person who will be going to see you.) 
+(I don&#x27;t known who is the person who will be going to see you.) 
+(I don&#x27;t know who is in the person who will be going to see you.) 
+(I don&#x27;t know who is in the person who will be going to see you.) 
+(I don&#x27;t know who is in the person who will be going to see you.) 
+(I don&#x27;t know who is in the person who will be going to see you.) 
+(I don&#x27;t know who is in the person who will be going to see you.) 
+(I don&#x27;t know who is in the person who will be going to see you.)
+(I don&#x27;t know who is in the person who will be going to see you.) 
+(I don&#x27;t know who is in the person who will be going to see you.) 
+(I don&#x27;t know who is in the person who will be going to see you.) 
+(I don&#x27;t know who is in the person who will be going to see you.) 
+(I don&#x27;t know who is in the person who will be going to see you.) 
+(I don&#x27;t know who is in the person Who will be going to see you.) 
+(I don&#x27;t know who is in the person who will be going to see you.) 
+(I don&#x27;t know who is in the person who will be going to see you.) 
+(I don&#x27;t know who is in the person who will be going to see you.) 
+(I don&#x27;t know who is in the person who will be going to see you.) 
+(I don&#x27;t know who is in the person who will be going to see you.) 
+(I Don&#x27;t know who is in the person who will be going to see you.) 
+(I don&#x27;t know who is in the person who will be going to see you.) 
+(I don&#x27;t know who is in the person who will be going to see you.) 
+(I don&#x27;t know who is in the person who will be going to see you.) 
+(I don&#x27;t know who is in the person who will be going to see you.) 
+(I don&#x27;t know who is in the person who will go to see you.) 
+(I don&#x27;t know who is in the person who will go to see you.) 
+(I don&#x27;t know who is in the person who will go to see you.) 
+(I don&#x27;t know who is in the person who will go to see you.) 
+(I don&#x27;t know who is in the person who will go to see you.) 
+(I don&#x27;t know who is in the person who will go to see you.) 
+(I don&#x27;t know who is in the person Who will be going to see you.) 
+(I don&#x27;t know who is in the person who will be going to see you.) 
+(I don&#x27;t know who is in the person who will be going to see you.) 
+(I don&#x27;t know who is in the person who will be going to see you.) 
+(I don&#x27;t know who is in the person who will be going to see you.) 
+(I don&#x27;t know where I am going to see you.) 
+(I don&#x27;t know where I am going to see you.) 
+(I don&#x27;t know where I am going to see you.) 
+(I don&#x27;t know where I am going to see you.) 
+(I don&#x27;t know where I am going to see you.) 
+(I don&#x27;t know where I am going to see you.) 
+(I don&#x27;t know where I am going to see you.) 
+(I don&#x27;t know where I am going to see you.) 
+(I don’t know where I am going to see you.) 
+(I don&#x27;t know where I am going to see you.) 
+(I don&#x27;t know where I am going to see you.) 
+(I don&#x27;t know where I am going to see you.) 
+(I don&#x27;t know where I am going to see you.) 
+(I don&#x27;t know where I am going to see you.) 
+(I don&#x27;t know where I am going to see you.) 
+(I don&#x27;t know where I am going to see you.)
+(I don&#x27;t know where I am going to see you.)
+(I don&#x27;t know where I am going to see you.)
+(I don&#x27;t know where I am going to see you.)
+(I don&#x27;t know where I am going to see you.)
+(I don&#x27;t know where I am going to see you.)
+(I don&#x27;t know where I am going to see you.)
+(I don&#x27;t know where I am going to see you.)
+(I don&#x27;t know where I am going to see you.)
+(I don&#x27;t known where I am going to see you.)
+(I don&#x27;t know where I am going to see you.)
+(I don&#x27;t know where I am going to see you.)
+(I don&#x27;t know where I am going to see you.)
+(I don&#x27;t know where I am going to see you.)
+(I don&#x27;t know where I am going to see you.)
+(I don&#x27;t know where I am going to see you.)
+(I don&#x27;t know where I am going to see you.)
+(I don&#x27;t know where I am go)
+(I don&#x27;t know where I am go)
+(I don&#x27;t know where I am go)
+(I don&#x27;t know where I am go)
+(I don&#x27;t know where I am go)
+(I don&#x27;t know where I am go)
+(I don&#x27;t know where I am go)
+(I don&#x27;t know where I am go)
+(I don&#x27;t know where I am go)
+(I don&#x27;t know where I am go)
+(I don&#x27;t know where I am go)
+(I don&#x27;t know where I am go)
+I don&#x27;t know where I am go.
+I don&#x27;t know where I am go.
+I don&#x27;t know where I am go.
+I don&#x27;t know where I am go.
+I don&#x27;t know where I am go.
+I don&#x27;t know where I am go.
+I don&#x27;t know where I am go.
+I don&#x27;t know where I am go.
+I don&#x27;t know where I am go.
+I don&#x27;t know where I am go.
+I don&#x27;t know where I am go.
+I don&#x27;t know where I am go.
+我也不知道。</td></tr></table>
+
+Table 16: Examples of Synthetic Data for Each Cognitive Distortion in KoACD
+
+# H Prompt Templates
+
+We present the prompt templates used throughout the study for various stages of cognitive distortion identification, synthetic data generation, and evaluation. These prompts were designed to ensure consistency and accuracy across different processes.
+
+To maintain conciseness, we replaced detailed descriptions and examples of cognitive distortions with the phrase 'Refer to Table 2 for a detailed explanation of each cognitive distortion.' This appendix includes Tables 17, 18, 19, 20, 21, 22, 23, and 24, which provide the full prompt templates for each stage.
+
+# Analyzer
+
+You are a psychology expert.
+
+Analyze the text below and, if a relevant cognitive distortion is present, select the most appropriate one.
+
+Choose from the following ten cognitive distortions: All-or-Nothing Thinking, Overgeneralization, Mental Filter, Discounting the
+
+Positive, Jumping to Conclusions, Magnification and Minimization, Emotional Reasoning, Should Statements, Labeling, and
+
+Personalization.
+
+{previous_cognitive_distortions} were deemed inappropriate in the previous analysis. Do not select them again under any circumstances.
+
+Previously rejected cognitive distortions: {previous_cognitive_distortions}
+
+Reason for rejection: {previous Reasons}
+
+# Since {previous_cognitive_distortions} were already deemed inappropriate:
+
+1. Do not select any of the above cognitive distortions again.  
+2. You must choose only from the remaining cognitive distortions.  
+3. If none of the remaining cognitive distortions are appropriate, respond with "Unknown."
+
+When identifying cognitive distortions, carefully refer to the definitions and examples of the ten distortions to consider a variety of cognitive distortions.
+
+When deciding on a cognitive distortion, analyze the overall context of the text rather than focusing on a single sentence.
+
+# Criteria for Responding with "Unknown":
+
+- The response requires speculation or subjective interpretation.  
+- The intent of the sentence is unclear.  
+- The speaker is not explicitly identified.  
+- The text consists only of simple emotional expressions.  
+- The text is merely a description of a situation or a question.  
+- Context from prior conversations is necessary for understanding.  
+- The text lacks value judgments or personal interpretation.  
+- The meaning is unclear without external context.  
+- The experience is described from another person's perspective.  
+- Negative emotions are present, but no specific cognitive distortion is identifiable.  
+- The text is a request for information, advice, or help.
+
+Important: If you determine "Unknown," this is a final decision, and no further analysis or reconsideration is needed. If any of the above criteria apply, immediately respond with "Unknown" without considering alternative interpretations.
+
+# Text to Analyze:
+
+{input_text}
+
+# List of Cognitive Distortions:
+
+Refer to Table 2 for a detailed explanation of each cognitive distortion.
+
+# Analysis Request:
+
+1. When determining cognitive distortions, consider the overall context.  
+2. Copy and paste all relevant sentences or paragraphs that support the selected cognitive distortion. Include at least two complete sentences.  
+3. Provide a clear explanation for selecting the sentences, ensuring a logical cause-and-effect relationship in your reasoning.  
+4. If no cognitive distortion applies, respond with "Unknown."
+
+# Output Format:
+
+- Cognitive Distortion: [Selected Cognitive Distortion]  
+- Relevant Sentences/Paragraphs: [Text]  
+- Reason for Selection: [Explanation]
+
+# Additional Output Rules:
+
+- All responses must be grammatically complete sentences.  
+- Sentences should not be cut off mid-thought.  
+- The final sentence of the response must be fully structured and complete.
+
+- Do not use Markdown formatting.  
+- When outputting [Selected Cognitive Distortion], do not select any distortions from {previous_cognitive_distortions}.
+
+Table 17: Prompt for the analyzer role in the negotiation process
+
+# Evaluator
+
+You are a psychology expert.
+
+Strictly evaluate the following cognitive distortion analysis provided by the analyzer.
+
+Refer to the cognitive distortions list for definitions and examples.
+
+# Original Text:
+
+{input_text}
+
+# Analyzer's Assessment:
+
+Cognitive Distortion: {cognitive_distortions}
+
+Relevant Sentences/Paragraphs: {related_text}
+
+Reason for Selection: {reason_text}
+
+# List of Cognitive Distortions:
+
+Refer to Table 2 for a detailed explanation of each cognitive distortion.
+
+# Evaluation Rules:
+
+1. Is the selected cognitive distortion present in the text?  
+- Assess whether the identified cognitive distortion can be reasonably inferred from the original text.  
+- Do not rely on isolated sentences; patterns must be found within the overall flow of the text.  
+2. Do the selected relevant sentences and reasoning properly support the cognitive distortion?  
+- Check whether the selected sentences accurately align with the definition and examples of the cognitive distortion.  
+- Evaluate whether the explanation logically connects the chosen sentences to the cognitive distortion.  
+- Ensure that the justification is not overly interpretative or speculative.
+
+# Judgment Criteria:
+
+- If any of the evaluation rules are violated, classify the analysis as "Inappropriate."  
+- If deemed inappropriate, clearly specify which rule was violated.  
+- If the response is "Unknown," accept it immediately.
+
+# Output Format:
+
+Evaluation Result: [Appropriate / Inappropriate]
+
+Evaluation Reason: [Detailed explanation for each rule]
+
+Conclusion:
+
+[If appropriate] "The current analysis is valid."
+
+[If inappropriate] "The cognitive distortion should be reassessed."
+
+# Additional Output Rules:
+
+- The evaluation reason must be fully structured in grammatically complete sentences.  
+- Sentences should not be cut off mid-thought.  
+- The final sentence must be a fully completed statement.  
+- Do not use Markdown formatting.
+
+Table 18: Prompt for the evaluator role in the negotiation process
+
+# Independent Evaluator
+
+You are a psychology expert.
+
+Thoroughly evaluate the appropriateness of the extracted cognitive distortion and its associated sentences/paragraphs.
+
+# Content to Evaluate:
+
+Selected Cognitive Distortion: {selected_cognitive_distortion}
+
+Relevant Sentences: {related_sentences}
+
+# Evaluation Criteria (1-3 points):
+
+1 Point: Inappropriate
+
+- The relevant sentences do not contain the identified cognitive distortion.  
+- OR the sentences are incomplete or lack clear context.
+
+2 Points: Partially Appropriate
+
+- The relevant sentences contain a cognitive distortion, but it does not match the selected one.  
+- OR another cognitive distortion would be a better fit.
+
+3 Points: Appropriate
+
+- The relevant sentences clearly demonstrate the selected cognitive distortion.  
+- The content aligns well with the definition and examples of the cognitive distortion.
+
+# Output Format:
+
+Score: [1-3 points]
+
+# Important Notes:
+
+- Only scores of 1, 2, or 3 may be used.  
+- Intermediate scores (e.g., 1.5 or 2.5) are not allowed.  
+- The evaluation rationale must be consistent with the assigned score.
+
+Table 19: Prompt for independently verifying cognitive distortion
+
+# Cognitive Clarification Method
+
+Generate a realistic fictional adolescent story based on the given cognitive distortion and reference case.
+
+When writing the fictional story, ensure that the age and content remain within the adolescent range.
+
+Consider a variety of situations that may occur both inside and outside of school.
+
+Strictly follow the output format specified below.
+
+# Input Information:
+
+Cognitive Distortion: {cognitive_distortions}
+
+Relevant Real-Life Sentence/Paragraph: {related_ssentences}
+
+Original Text: {input_text}
+
+# Story Writing Requirements:
+
+1. Length: Must be 40 words or fewer (Exceeding 40 words is strictly prohibited).  
+2. Format: [Gender/Age] --- [Story Content]  
+3. Age: Must be between 13 and 19 years old.
+
+If gender, age, or school grade is mentioned in the original text, use that information to generate [Gender/Age].
+
+(Gender: Male or Female, Middle School: 14-16 years old, High School: 17-19 years old)  
+4. Theme: Events that occur in school, home, friendships, or daily adolescent life.  
+5. Perspective: Write from a first-person point of view.  
+6. Content:
+
+Clearly establish the situation (when, where, what, how).
+
+Maintain a logical cause-and-effect relationship within the story.
+
+The narrator (first-person) should naturally exhibit cognitive distortion.
+
+# Constraints:
+
+1. The story must be inspired by the given real-life sentence, adapting it to a similar but new context.  
+2. Utilize grammatical transformations, such as active/passive voice changes and word order modifications.  
+3. Avoid starting the story with any of the following words: {used_words}  
+4. The word "today" must not be used.  
+5. Do not explicitly mention cognitive distortion terms in the story.
+
+(e.g., Do NOT use terms like "overgeneralization" or "all-or-nothing thinking.")
+
+# Output Format:
+
+[Gender/Age] --- [Generated Story]
+
+# Important Notes:
+
+Ensure that the cognitive distortion characteristics reflected in the reference sentence are incorporated into the new story in a different yet relevant context.
+
+Do NOT exceed 40 words in the generated story (Strict limit: 40 words maximum).
+
+Table 20: Prompt for cognitive clarification-based synthetic story generation
+
+# Cognitive Balancing Method
+
+Generate a realistic fictional adolescent story that reflects the characteristics of {cognitive_distortions}, based on the provided real-life example.
+
+Strictly follow the output format specified below.
+
+# Input Information:
+
+Original Text: {input_text}
+
+# Story Writing Requirements:
+
+1. Length: Must be 40 words or fewer (Exceeding 40 words is strictly prohibited).  
+2. Format: [Gender/Age] --- [Story Content]  
+3. Age: Must be between 13 and 19 years old.
+
+If gender, age, or school grade is mentioned in the original text, use that information to generate
+
+[Gender/Age].
+
+(Gender: Male or Female, Middle School: 14-16 years old, High School: 17-19 years old)
+
+4. Theme: Events that occur in school, home, friendships, or daily adolescent life.  
+5. Perspective: Write from a first-person point of view.  
+6. Content:
+
+Clearly establish the situation (when, where, what, how).
+
+Maintain a logical cause-and-effect relationship within the story.
+
+The narrator (first-person) should naturally exhibit cognitive distortion.
+
+# List of Cognitive Distortions:
+
+Refer to Table 2 for a detailed explanation of each cognitive distortion.
+
+# Current Cognitive Distortion for Story Generation:
+
+{cognitive_distortions}
+
+# Constraints:
+
+1. The story must be inspired by the given real-life sentence, adapting it to a similar but new context.  
+2. Utilize grammatical transformations, such as active/passive voice changes and word order modifications.  
+3. Avoid starting the story with any of the following words: {used_words}  
+4. The word "today" must not be used.  
+5. Do not explicitly mention cognitive distortion terms in the story.
+
+(e.g., Do NOT use terms like "overgeneralization" or "all-or-nothing thinking.")
+
+# Output Format:
+
+[Gender/Age] --- [{cognitive_distortions} Reflected Story]
+
+# Important Notes:
+
+Do NOT exceed 40 words in the generated story (Strict limit: 40 words maximum).
+
+Table 21: Prompt for cognitive balancing-based synthetic story generation
+
+# Consistency Evaluator
+
+Please evaluate the consistency of the given text and assign a single score between 1 and 3.
+
+# Input Information:
+
+Selected Cognitive Distortion: {selected_cognitive_distortion}
+
+Relevant Sentences: {related_ssentences}
+
+Generated Story: {generated_story}
+
+# Evaluation Criteria (1-3 points):
+
+Assess whether the selected cognitive distortion is neither exaggerated nor minimized and whether the original meaning of the relevant sentences is preserved while being appropriately expressed in the generated story.
+
+- Is the selected cognitive distortion accurately maintained without distortion from the relevant sentences?  
+- Has the meaning of the relevant sentences been appropriately conveyed in the generated story without excessive modification?  
+- Does the generated story logically align with the selected cognitive distortion and its context?
+
+# Scoring Guidelines:
+
+1 Point: The selected cognitive distortion or the context of the relevant sentences is significantly distorted or altered in the generated story.  
+2 Points: The selected cognitive distortion and the context of the relevant sentences are partially retained, but there are some inconsistencies or unnatural expressions.  
+3 Points: The selected cognitive distortion and the context of the relevant sentences are naturally maintained, forming a logically coherent story.
+
+# Output Format:
+
+Score: [1-3 points]
+
+# Important Notes:
+
+- Only scores of 1, 2, or 3 may be used.  
+- Intermediate scores (e.g., 1.5 or 2.5) are not allowed.  
+- The evaluation rationale must be consistent with the assigned score.
+
+Table 22: Prompt for consistency evaluation of synthetic data
+
+# Accuracy Evaluator
+
+Please evaluate the accuracy of the given text and assign a single score between 1 and 3.
+
+# Input Information:
+
+Selected Cognitive Distortion: {selected_cognitive_distortion}
+
+Relevant Sentences: {related_sentences}
+
+Generated Story: {generated_story}
+
+# Evaluation Criteria (1-3 points):
+
+Evaluate whether the generated story is correctly classified under the most relevant cognitive distortion among the ten defined categories.
+
+- Does the selected cognitive distortion correctly classify the cognitive distortion present in both the relevant sentences and the generated story?  
+- When compared to other cognitive distortions, is the selected cognitive distortion the most appropriate choice?  
+- Is there a logical consistency between the selected cognitive distortion and the way it is expressed in the generated story?
+
+# Scoring Guidelines:
+
+1 Point: The selected cognitive distortion significantly mismatches the cognitive distortion found in the relevant sentences and the generated story.  
+2 Points: The selected cognitive distortion is partially appropriate, but another cognitive distortion might be a better fit.  
+3 Points: The selected cognitive distortion is the most accurate classification of the cognitive distortion found in the relevant sentences and the generated story.
+
+# Output Format:
+
+Score: [1-3 points]
+
+# Important Notes:
+
+- Only scores of 1, 2, or 3 may be used.  
+- Intermediate scores (e.g., 1.5 or 2.5) are not allowed.  
+- The evaluation rationale must be consistent with the assigned score.
+
+Table 23: Prompt for accuracy evaluation of synthetic data
+
+# Fluency Evaluator
+
+Please evaluate the fluency of the given text and assign a single score between 1 and 3.
+
+# Input Information:
+
+Generated Story: {generated_story}
+
+# Evaluation Criteria (1-3 points):
+
+Evaluate whether the generated story is grammatically sound and maintains human-like fluency in its structure and readability.
+
+- Is the sentence structure natural and fluent?  
+- Are there any grammatical errors?  
+- Is the flow between sentences smooth, making the overall story cohesive?
+
+# Scoring Guidelines:
+
+1 Point: The text contains many grammatical errors or is highly unnatural.  
+2 Points: The text has minor grammatical issues or slightly awkward expressions but is still generally understandable.  
+3 Points: The text is grammatically correct and reads naturally with a smooth sentence structure.
+
+# Output Format:
+
+Score: [1-3 points]
+
+# Important Notes:
+
+- Only scores of 1, 2, or 3 may be used.  
+- Intermediate scores (e.g., 1.5 or 2.5) are not allowed.  
+- The evaluation rationale must be consistent with the assigned score.
+
+Table 24: Prompt for fluency evaluation of synthetic data
+
+# I Evaluation Form
+
+# Korean
+
+安國海
+
+六通韓江，
+
+本作章是清初的世纪白海之迈儿托而，古云道中，三人，自国的，
+
+本明在，
+
+明者，将之用以示其法。
+
+首则删是单据的，即当载于其后。
+
+亚司的民时国用其云，其云。
+
+$\therefore {a}_{1} = 2,{a}_{2} =  - 4$
+
+1. 1:
+
+2. 2:
+
+3. 3:
+
+把副本是说量的完整表里，
+
+103
+
+i
+
+1. 请：选择属于印尼的歌词，如：
+
+2. 2: 1
+
+3. 3: 1
+
+$③$  反删是：云的，
+
+訳音，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，，,。
+
+i
+
+1. 議：訳當在目本音人，不。
+
+2. 2:
+
+3. 请：证明当  $n$  次的  $n$  次方程组的解为  $\{x_{1}, x_{2}, x_{3}\}$ ，求证：
+
+a.
+
+音。
+
+函的自形也成，
+
+daii
+
+Table 25: Expert Evaluation Form for Cognitive Distortion (Korean)
+
+# English
+
+Hello,
+
+Thank you very much for taking the time to participate in this task.
+
+This task involves evaluating each sentence from real adolescent utterances based on both the type of cognitive distortion it contains and the overall quality of the sentence's expression.
+
+All results collected in this evaluation will be used solely for research purposes and data analysis, and will not be used for commercial purposes or distributed outside the research context.
+
+The sentences to be evaluated are provided in a separate file, and for each sentence, please make your assessment according to the following three criteria.
+
+The first criterion is Consistency. Please evaluate whether the selected cognitive distortion is neither exaggerated nor minimized and whether it is appropriately represented in the generated story while preserving the original meaning of the related sentence. Assess how well the indicated cognitive distortion type explains the sentence, using a 1-to-3 point scale.
+
+# Scoring guidelines:
+
+1 point: The selected cognitive distortion or the situation in the related sentence is significantly distorted or altered in the generated story.
+
+2 points: The selected cognitive distortion and the situation in the related sentence are partially maintained in the generated story, but there are some inconsistencies or unnatural expressions.
+
+3 points: The selected cognitive distortion and the situation in the related sentence are naturally preserved in the generated story, forming a logically consistent narrative.
+
+The second criterion is Accuracy. Please evaluate whether the generated story correctly reflects the most appropriate category among the 10 cognitive distortion types.
+
+# Scoring guidelines:
+
+1 point: The selected cognitive distortion significantly disagrees with the type present in the related sentence and generated story.
+
+2 points: The selected cognitive distortion is partially appropriate, but another distortion type might be more suitable.
+
+3 points: The selected cognitive distortion most accurately matches the cognitive distortion in the related sentence and generated story.
+
+The third criterion is Fluency. Please assess whether the sentence is grammatically natural, easy to read, and coherent in terms of meaning. In particular, consider whether the cause-problem-solution structure within the sentence is logically connected.
+
+# Scoring guidelines:
+
+1 point: The sentence is unnatural or contains many grammatical errors, making it difficult to understand.
+
+2 points: The sentence is somewhat awkward or partially lacks logical flow, but overall meaning can be conveyed.
+
+3 points: The sentence is smooth, natural, and maintains structural and logical consistency.
+
+In addition, if you believe that a sentence could be interpreted as a different type of cognitive distortion, please indicate the possibility and provide a brief explanation.
+
+If you feel any burden during the task or need to adjust your schedule, please feel free to pause or coordinate as needed.
+
+If you have any questions, do not hesitate to contact us.
+
+Once again, thank you very much for your contribution.
+
+Table 26: Expert Evaluation Form for Cognitive Distortion (English)

2025/KoACD_ The First Korean Adolescent Dataset for Cognitive Distortion Analysis via Role-Switching Multi-LLM Negotiation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7a290522ad8cd6183f9a93ac4b374dbdc38f9a85f8a3aaa83d03990d056c206e
+size 2077152

2025/KoLEG_ On-the-Fly Korean Legal Knowledge Editing with Continuous Retrieval/c37f6d72-5c27-4a79-8b7b-8402c8729735_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2449cfcd44901d2101fed841aa25ae73d95663feaccc529483b2df83dfb6556a
+size 4982794

2025/KoLEG_ On-the-Fly Korean Legal Knowledge Editing with Continuous Retrieval/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2f427bcf694029740b95728b2b055f4ee3343693a3f47e1e859304003c9f9144
+size 1845037

2025/Krikri_ Advancing Open Large Language Models for Greek/a01b7f99-74bd-4cb7-9a41-265c6435fe83_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:52e47508a488b926559d02a07d5a4a4b013261ea4d728ddf033baa1fe8206a80
+size 406979

2025/Krikri_ Advancing Open Large Language Models for Greek/full.md

@@ -0,0 +1,512 @@
+# Krikri: Advancing Open Large Language Models for Greek
+
+Dimitris Roussis*, Leon Voukoutis*, Georgios Paraskevopoulos*, Sokratis Sofianopoulos*, Prokopis Prokopidis*, Vassilis Papavasileiou, Athanasios Katsamanis, Stelios Piperidis, Vassilis Katsouros*
+
+Institute for Speech and Language Processing, Athena Research Center Artemidos 6 & Epidavrou, Athens, Greece vsk@athenarc.gr
+
+# Abstract
+
+We introduce Llama-Krikri-8B, a cutting-edge Large Language Model tailored for the Greek language, built on Meta's Llama 3.1-8B. Llama-Krikri-8B has been extensively trained on high-quality Greek data to ensure superior adaptation to linguistic nuances. With 8 billion parameters, it offers advanced capabilities while maintaining efficient computational performance. Llama-Krikri-8B supports both Modern Greek and English, and is also equipped to handle polytonic text and Ancient Greek. The chat version of Llama-Krikri-8B features a multi-stage post-training pipeline, utilizing both human and synthetic instruction and preference data, by applying techniques such as MAGPIE. In addition, for evaluation, we propose three novel public benchmarks for Greek. Our evaluation on existing as well as the proposed benchmarks shows notable improvements over comparable Greek and multilingual LLMs in both natural language understanding and generation as well as code generation.
+
+# 1 Introduction
+
+Recent advancements in AI have been largely driven by the development of large-scale foundation models. Meta's Llama 3 (Grattafori et al., 2024) fostered a new generation of open models, designed for strong multilingual capabilities, code generation, reasoning, and tool use. With extended context windows, and refined training strategies, models based on Llama 3 have achieved performance comparable to proprietary systems like GPT-4. A critical aspect in this evolution is the development of multilingual and language-specific models, democratizing access to AI technologies and preserving linguistic diversity.
+
+While substantial progress has been made for widely spoken languages, low and medium resource languages remain underrepresented. Greek,
+
+in particular, has received limited attention despite its linguistic complexity, rich cultural heritage, and historical significance. Addressing this gap, we present Llama-Krikri-8B, a cutting-edge open Large Language Model tailored for the Greek language. Built on Meta's Llama 3.1-8B architecture, Llama-Krikri has been continually pretrained on a diverse, high-quality Greek corpus. This allows the model to effectively capture the syntactic and semantic nuances of Greek, while retaining the multilingual strengths of the base model. Notably, Llama-Krikri also supports English and is capable of handling polytonic and Ancient Greek texts, addressing not only contemporary but also historical forms of the language.
+
+Compared to Meltemi-7B (Voukoutis et al., 2024), the previous state-of-the-art open Greek LLM built on Mistral 7B (Jiang et al., 2023), Llama-Krikri-8B significantly increases the number of parameters, context length, and training data scale. Additionally, it features an enhanced posttraining pipeline using both human and synthetic data. Following the MAGPIE methodology (Xu et al., 2024), we generate high-quality instruction-response pairs via prompting of aligned models, and apply multiple rounds of instruction tuning and alignment using Direct Preference Optimization (DPO) (Rafailov et al., 2024). This pipeline ensures that the model produces helpful, honest, and harmless outputs.
+
+To evaluate Llama-Krikri-8B, we also introduce three novel public benchmarks specifically designed for Greek. These, alongside existing evaluation suites, show that Llama-Krikri outperforms comparable Greek and multilingual LLMs in both natural language understanding and generation, as well as code-related tasks. Moreover, it supports function calling and agentic behaviors, opening new application domains for Greek users. Llama-Krikri-8B is available under the Llama 3.1 Com
+
+munity License Agreement<sup>1</sup>.
+
+Our key contributions are:
+
+- We present Llama-Krikri-8B, a state-of-the-art open Greek foundation model based on Llama 3.1, demonstrating strong capabilities in Modern and Ancient Greek, English, and code generation, along with support for function calling and agentic behavior.  
+- We implement a rigorous post-training pipeline incorporating synthetic instruction tuning via MAGPIE and alignment through DPO.  
+- We introduce three new benchmarks for evaluating Greek LLMs, covering language understanding, generation, and code tasks.  
+- We show that Llama-Krikri-8B significantly outperforms existing open Greek and multilingual models across several domains, while supporting advanced features such as function calling.  
+- We open source the foundation<sup>2</sup> and instruct<sup>3</sup> models on HuggingFace for public use.
+
+# 2 Background and Related Work
+
+Large Language Models (LLMs) have achieved state-of-the-art performance across a wide variety of natural language processing (NLP) tasks. These models are typically trained on massive corpora dominated by English, leading to strong performance in English-language tasks but comparatively weaker capabilities in other languages (Devlin et al., 2019; Brown et al., 2020). As a result, the development of language-specific LLMs has become an active area of research, particularly for under-represented languages.
+
+One prominent strategy for developing such models is continual pretraining, where a pretrained base model is further trained on data in the target language. This approach allows researchers to leverage the general capabilities of large base models while improving performance in specific linguistic domains, without the prohibitive cost of training from scratch (Gururangan et al., 2020).
+
+Several recent projects have successfully applied continual pretraining to adapt existing models to new languages. BgGPT-GEMMA-2-27B-Instruct (Alexandrov et al., 2024) fine-tunes Google's Gemma-2 model (Riviere et al., 2024) for Bulgarian, combining over 100B tokens of Bulgarian and English data and applying techniques such as Branch-and-Merge to mitigate catastrophic forgetting. Similarly, LeoLM (LAION, 2023) and the Sabia models (Pires et al., 2023) adapt LLaMA and Mistral-based architectures for German and Portuguese, respectively, through targeted continual pretraining.
+
+For Greek, Meltemi-7B represents the first open generative LLM tailored to the language. It was developed by continually pretraining Mistral-7B on a substantial Greek corpus, followed by instruction fine-tuning. While effective, Meltemi's performance is bounded by the size and capabilities of the base model, as well as the limited post-training alignment techniques employed at the time.
+
+Beyond language adaptation, alignment of LLMs to generate helpful, harmless, and honest outputs has become increasingly central. Early approaches such as InstructGPT (Ouyang et al., 2022) and Constitutional AI (Bai et al., 2022) rely on multi-stage fine-tuning pipelines involving human feedback or rule-based constraints. More recently, DPO and data synthesis methods like MAGPIE have enabled scalable and effective instruction tuning. MAGPIE, in particular, leverages high-quality prompting of already-aligned models to generate large volumes of instruction-response pairs, demonstrating that synthetic data can rival or surpass human-curated datasets.
+
+These advancements highlight a trend toward bootstrapping high-quality training data using strong base models, especially in low-resource languages. Our work builds on this foundation by employing Llama 3.1 as a base architecture, scaling up parameter count and context length, and applying a more rigorous post-training pipeline, including instruction tuning with MAGPIE-generated data and DPO for alignment.
+
+# 3 Methodology
+
+Llama-Krikri-8B is based on the Transformer architecture (Vaswani et al., 2023), which has become the de facto standard for large language models. The model inherits its architecture from Meta's Llama 3.1-8B, leveraging the strong foundation in
+
+multilingual understanding, code generation, and reasoning provided by Llama 3.1.
+
+Adapting an LLM for the Greek language requires addressing the lack of high-quality Greek data in the massive datasets typically used to train foundation models. Even though Llama 3.1's pretraining corpus comprises trillions of tokens, it struggles to generate coherent Greek text, thus indicating that Greek data is only a tiny fraction of its training data; we should note however that there is limited information on the composition of its pretraining data. Our approach is to perform continual pretraining with Greek and parallel data to infuse the model with Greek knowledge. This training must be done carefully to avoid catastrophic forgetting (Luo et al., 2025) of the base model's prior knowledge in other languages and domains. We tackle this by (a) constructing a high-quality, large-scale Greek corpus, (b) extending and tuning the tokenizer, (c) interleaving Greek training data with datasets containing English, mathematics, code, and parallel data in languages that the model has already been trained on, (d) employing a dataset sampling schedule during training that prefers data closer to the initial llama-3.1 distribution in the beginning, while shifting closer to our true dataset distribution as training continues and (d) re-warming and re-decaying the learning rate (Ibrahim et al., 2024).
+
+Pretraining Data Collection & Cleaning: As a foundation for continual pretraining, we curated a large corpus of texts totalling approximately 91 billion tokens (after filtering and deduplication), which was upsampled to 110 billion tokens for the final pretraining mix. This corpus was constructed with a primary focus on Greek and aiming on retaining and enhancing the original model's capabilities. The distribution included 56.7 billion monolingual Greek tokens  $(62.3\%)$ , 21 billion monolingual English tokens  $(23.1\%)$ , 5.5 billion parallel data tokens  $(6.0\%)$ , and 7.8 billion math and code tokens  $(8.6\%)$ . Table 1 presents the distribution of the pretraining data mix, with more details provided in Appendix A Pretrained data mix.
+
+After corpus collection, we implemented a multi-stage preprocessing and filtering pipeline to ensure a high quality for the pretraining data. Various parts of our filtering methodology have been informed by approaches used in Voukoutis et al. (2024) and large-scale corpus creation efforts such as Zyda (Tokpanov et al., 2024). However, we have adapted
+
+these approaches to cater for the peculiarities of the Greek language. We detail the preprocessing pipelines we used in Appendix B Pretraining data cleaning pipelines.
+
+Tokenizer and Embeddings Expansion: The original Llama 3 tokenizer comprises 128,000 tokens and is inefficient for Greek texts, as it generally performs character-level tokenization for Greek. This was determined, through the approximation of the llama-3.1 tokenizer's fertility (Csaki et al., 2023), a metric of the average tokens per word produced. To determine the efficiency of the original Llama 3 tokenizer and compare with our approach, we conducted tests on diverse Greek and English corpora (each one containing 100,000 rows and totalling approximately 2M words) and calculated the difference in fertility, as can be seen in Table 2.
+
+In order to develop an optimal tokenizer for Greek which is also efficient in historical dialects of the language, as well as in critical domains (such as legal and scientific texts), we extended the Llama 3 tokenizer with 20,992 new tokens through a multistage process which encompasses curating high-quality texts and allocating new tokens across five domains. This process is especially important during model inference, as it significantly reduces the input and output token cost during model use. Furthermore, more compact representations of input text help to improve model performance. We provide detail on the steps for the tokenizer and embeddings expansion in Appendix C, Tokenizer and embeddings expansion process.
+
+Greek embeddings training: To effectively integrate newly introduced Greek tokens into the model, we implemented an initial, targeted training phase for their corresponding embeddings. By preparing the new token representations prior to full-scale pretraining, we mitigate potential disruptions to the existing model parameters.
+
+We initialized the model with Llama 3.1-8B-Base weights, freezing all but the embeddings and output-projection weights for the 20,992 new tokens, allowing their initial training without large gradient updates to the rest of the model. The dataset for this step was comprised of 5B tokens sampled from the main pre-training dataset. This short, several-thousand-step training regimen ensured a smoother integration of the new vocabulary into the model's existing knowledge representation.
+
+<table><tr><td>Subcorpus</td><td>Original Tokens (B)</td><td>Percentage</td><td>Upsampled Tokens (B)</td><td>Percentage</td></tr><tr><td>Greek</td><td>56.7</td><td>62.3%</td><td>66.1</td><td>60.0%</td></tr><tr><td>English</td><td>21.0</td><td>23.1%</td><td>25.2</td><td>22.9%</td></tr><tr><td>Parallel</td><td>5.5</td><td>6.0%</td><td>8.8</td><td>8.0%</td></tr><tr><td>Math/Code</td><td>7.8</td><td>8.6%</td><td>10.1</td><td>9.1%</td></tr><tr><td>Total</td><td>91.0</td><td>100%</td><td>110.2</td><td>100%</td></tr></table>
+
+Table 1: Composition of the pretraining corpus - original and upsampled  
+
+<table><tr><td>Tokenizer</td><td>Vocabulary Size</td><td>Fertility Greek</td><td>Fertility English</td></tr><tr><td>Mistral-7B</td><td>32,000</td><td>6.80</td><td>1.49</td></tr><tr><td>Meltemi-7B</td><td>61,362</td><td>1.52</td><td>1.44</td></tr><tr><td>Llama-3.1-8B</td><td>128,000</td><td>2.73</td><td>1.33</td></tr><tr><td>Llama-Krikri-8B</td><td>149,248</td><td>1.65</td><td>1.33</td></tr></table>
+
+Continual Pretraining Process: After embedding training, all parameters were unfrozen, and training continued on the 110B token corpus using a mixed-curriculum strategy. Training alternated between chunks of predominantly Greek text and supporting data (English, parallel, code) in a round-robin fashion. This interleaving improved Greek performance while maintaining/improving English validation perplexity, similar to findings in related work. The strategy involved curriculum learning and experience replay: starting with simpler Greek/more English, progressing to diverse Greek, and mixing in periodic English/code replays. Training was conducted on two machines, each equipped with 8 NVIDIA H200 GPUs, using DeepSpeed Zero 3 and bf16 mixed precision for  $\sim 50$  days on the 110B token dataset at 128K context length. Training used packed sequences, cosine annealing LR, AdamW optimizer, gradient clipping, and weight decay.
+
+Annealing Phase: Following pretraining, a short annealing pass used a curated 3.5B token dataset of very high-quality texts across all subcorpora. We used within-dataset normalized perplexity, calculated using KenLM (Heafield, 2011), to implement a dataset-aware fluency scoring method for document selection. To boost comprehension and reasoning, a synthetic question-answer dataset (189M tokens) was created by prompting Gemma-2-27B-IT to generate Q&A triplets with reasoning from curated documents. Annealing was tested with and without the synthetic QA data. Performance (Table 3) showed continual pretraining improved Greek (+8.7) but reduced English (-4) vs Llama-3.1.
+
+Annealing with curated data gave modest gains. Most notably, adding synthetic QA significantly improved Greek (+2.1 vs continual pretraining) and enhanced English beyond original Llama-3.1 (+0.8). Liger kernels (Hsu et al., 2024) were used for efficiency.
+
+Table 2: Tokenizer statistics for Greek and English  
+
+<table><tr><td>Training Stage</td><td>Avg. Greek</td><td>Avg. English</td></tr><tr><td>Llama-3.1-8B</td><td>48.7</td><td>66.2</td></tr><tr><td>+ Continual Pretraining</td><td>57.4</td><td>62.2</td></tr><tr><td>+ Curated Corpora</td><td>58.0</td><td>63.4</td></tr><tr><td>+ Synthetic QA Dataset</td><td>59.5</td><td>67.0</td></tr></table>
+
+Table 3: Average performance across training stages on Greek and English benchmarks
+
+Instruction Tuning and Alignment: Llama-Krikri-8B-Instruct was created by fine-tuning the base model for instruction following and dialogue. Addressing Greek data scarcity and avoiding translation artifacts, the pipeline combined data synthesis, filtering, two-stage Supervised Fine-Tuning (SFT), and Direct Preference Optimization (DPO).
+
+Data collection, synthesis, & curation involved collecting high-quality English instruction, e.g., Tulu 3 (Lambert et al., 2025) and SmolTalk (Allal et al., 2025), and preference data e.g., UltraFeedback (Cui et al., 2023). Additionally, Greek data was synthesized via translation (with post-editing), regenerating responses using LLMs (Gemma-2-27B-IT), and generating synthetic instructions directly in Greek using the MAGPIE technique. Curated corpora from annealing were reused for synthetic Q&A and dialogues. Data was scored and filtered using the Skyword-Reward-Gemma-2-27Bv0.2 (Liu et al., 2024) reward model, known for Greek accuracy. Rule-based filters ensured formatting and language verification.
+
+SFT was done in two stages ( $\sim 856\mathrm{k}$  pairs in Stage 1,  $\sim 638\mathrm{k}$  in Stage 2), with progressively higher data quality. Datasets included filtered original English, reward-model-filtered synthetic MAGPIE data (higher scores in Stage 2), translated/post-edited data (Stage 1), regenerated responses (in
+
+cluding "thinking" in Stage 2), multi-language translation data, synthetic QA, synthetic multiturn dialogues, and upsampled manual safety data. Training used linear decay LR, AdamW, and Liger kernels, masking prompt loss. SFT produced a model that followed instructions but needed alignment for helpfulness, precision, and safety.
+
+DPO provided final alignment using  $\sim 92\mathrm{k}$  preference triplets. Data included selected/high-scored original preferences, preferences from MAGPIESynthesized data (using reward model scores), preferences derived from translated data (contrasting regenerated responses or regenerated vs translated reference), and safety preferences. DPO maximized preferred response likelihood while minimizing dispreferred, using length normalization. Training used AdamW (Loshchilov and Hutter, 2017), linear decay LR, and Liger kernels. DPO significantly improved response quality, safety, and helpfulness compared to the SFT-only model. The DPO-tuned model is the final Llama-Krikri-8B-Instruct.
+
+# 4 Evaluation
+
+In this section, we present evaluation details for Llama-Krikri-8B-Base and Llama-Krikri-8B-Instruct, across six Greek and six English benchmarks. We compare our base model directly with the base model Llama-3.1-8B (Grattaftori et al., 2024) and the previous Greek state-of-the-art model Meltemi-7B-v1.5 (Voukoutis et al., 2024). Additionally, we evaluate our chat model, Llama-Krikri-8B-Instruct on three challenging English benchmarks, as well as three novelly constructed Greek benchmarks which correspond to the English ones.
+
+# 4.1 Base Model Evaluation: Krikri-8B-Base
+
+We evaluated Llama-Krikri-8B-Base against Llama-3.1-8B and Meltemi-7B-v1.5 in a few-shot setting, consistent with the Open LLM Leaderboard<sup>4</sup>.
+
+Greek Benchmarks: The evaluation was carried out on a suite of six Greek-specific benchmarks<sup>5</sup> used in Voukoutis et al. (2024), including machine-translated versions of established English datasets (ARC-Challenge Greek, HellaSwag Greek,
+
+MMLU Greek), manually post edited MT of English datasets (Truthful QA Greek), the existing Belebele Greek benchmark (Bandarkar et al., 2024) (created by native speakers with translation experience), and a novel medical QA benchmark (Medical MCQA).
+
+Results in Table 4 demonstrate substantial improvements for Greek  $(+10.8\%)$  compared to Llama-3.1-8B. Moreover, we observe that Llama-Krikri-8B-Base surpasses Meltemi-7B-v1.5 with a notable  $+11.6\%$  average improvement across all benchmarks. On MMLU Greek, Llama-Krikri-8B-Base surpasses Llama-3.1-8B and Meltemi-7B-v1.5 by  $+9.4\%$  and  $+10.8\%$  respectively, while on ARC-Challenge Greek, it achieves an accuracy of  $49.4\%$ , compared to Llama-3.1-8B's and Meltemi-7B-v1.5's  $39.9\%$  and  $40.0\%$ , respectively. Similar substantial gains are observed on the Belebele Greek dataset, where Llama-Krikri-8B-Base scores  $82.7\%$ , surpassing Meltemi-7B-v1.5 and Llama-3.1-8B by  $21.7\%$  and  $+9.9\%$ , respectively. In the Greek Medical MCQA, Llama-Krikri-8B-Base reaches  $53.8\%$ , demonstrating clear advancements over Llama-3.1-8B  $(+20.4\%)$  in a domain-specific Greek benchmark that was not translated from English.
+
+English Benchmarks For the evaluation of base models on English, we utilized six benchmarks, with five of them being the original versions of those also used for Greek: ARC-Challenge (Clark et al., 2018), Truthful QA (Lin et al., 2022), HellaSwag (Zellers et al., 2019), MMLU (Hendrycks et al., 2021), and Belebele (Bandarkar et al., 2024). Additionally, the Winogrande (Sakaguchi et al., 2021) test set was used as the sixth benchmark for English. In the results presented in Table 5 we see that our training methodology not only mitigates catastrophic forgetting effectively, but also improves average performance across all English test sets by  $+0.8\%$ .
+
+# 4.2 Chat Model Evaluation: Krikri-8B-Instruct
+
+For evaluating the capabilities of Llama-Krikri-8B-Instruct as a conversational assistant, suitable for multi-turn dialogue, instruction-following and complex coding and math queries, we used a suite of benchmarks in both English and Greek. For English, we conducted evaluations across two paths:
+
+- We submitted our model to the Open LLM Leaderboard (Fourrier et al., 2024) which au
+
+<table><tr><td>Benchmark</td><td>Meltemi-7B-v1.5</td><td>Llama-3.1-8B</td><td>Krikri-8B-Base</td></tr><tr><td>Medical MCQA EL (15-shot)</td><td>42.2</td><td>33.4</td><td>53.8</td></tr><tr><td>Belebele EL (5-shot)</td><td>61.0</td><td>72.8</td><td>82.7</td></tr><tr><td>HellaSwag EL (10-shot)</td><td>53.8</td><td>52.1</td><td>64.6</td></tr><tr><td>ARC-Challenge EL (25-shot)</td><td>40.0</td><td>39.9</td><td>49.4</td></tr><tr><td>TruthfulQA MC2 EL (0-shot)</td><td>49.0</td><td>51.1</td><td>54.2</td></tr><tr><td>MMLU EL (5-shot)</td><td>41.2</td><td>42.6</td><td>52.0</td></tr><tr><td>Average</td><td>47.9</td><td>48.7</td><td>59.5</td></tr></table>
+
+Table 4: Greek benchmark results (accuracy %) for base models.  
+
+<table><tr><td>Benchmark</td><td>Meltemi-7B-v1.5</td><td>Llama-3.1-8B</td><td>Krikri-8B-Base</td></tr><tr><td>Winogrande (5-shot)</td><td>73.4</td><td>74.6</td><td>72.6</td></tr><tr><td>Belebele EN (5-shot)</td><td>77.7</td><td>71.5</td><td>79.8</td></tr><tr><td>HellaSwag EN (10-shot)</td><td>79.6</td><td>82.0</td><td>80.7</td></tr><tr><td>ARC-Challenge EN (25-shot)</td><td>54.1</td><td>58.5</td><td>57.8</td></tr><tr><td>TruthfulQA MC2 EN (0-shot)</td><td>40.5</td><td>44.2</td><td>44.8</td></tr><tr><td>MMLU EN (5-shot)</td><td>56.9</td><td>66.2</td><td>65.1</td></tr><tr><td>Average</td><td>63.7</td><td>66.2</td><td>67.0</td></tr></table>
+
+Table 5: English benchmark results (accuracy %) for base models.
+
+tomatically evaluates models on IFEval, BBH, MATH, GPQA, MUSR, and MMLU-Pro using the Eleuther AI Language Model Evaluation Harness (Gao et al., 2021), a unified framework to test generative language models on a large number of different evaluation tasks.
+
+- We used the Arena Hard Auto v0.1 (Li et al., 2024; Chiang et al., 2024), IFEval (Zhou et al., 2023) (strict avg) and MT-Bench (Zheng et al., 2023) benchmarks. Although IFEval was already included in the Open LLM Leaderboard, we re-implemented it to enable accurate comparison with multiple models. In the evaluation of MT-Bench we used GPT-4o (2024-08-06) as the judge model, while in the evaluation of Arena Hard Auto v0.1 we used the standard approach with GPT-4-0314 as the baseline model (by default scoring  $50\%$ ) and GPT-4-1106-Preview as the judge model, while also reusing the generations and judgments already computed by the authors.
+
+For Greek, we created three novel evaluation benchmarks by translating three challenging, diverse, and widely used English benchmarks, ensuring high-quality translations through careful post-editing, validation and adaptation involving human annotators:
+
+- IFEval Greek (strict avg.): a manual translation of 541 prompts from the original Instruction-Following Evaluation benchmark (Zhou et al., 2023), featuring verifiable instructions such as "απάντησε με περισότερες από 400 λέξεις" (answer with more than 400 words) and "ανάφερε τη λέξη TN τουλαχοτοτον 3 φορές" (mention the word AI at least 3 times), designed to assess the model's ability to follow specific instructions.  
+- MT-Bench Greek: a translated version of the Multi-turn Benchmark (Zheng et al., 2023) containing 80 high-quality, multi-turn conversations across eight diverse categories (e.g., STEM, humanities, roleplay, coding, etc.), carefully post-edited to ensure natural Greek phrasing and cultural appropriateness. MT-Bench is also used to evaluate the function-calling capabilities of LLMs (Chen et al., 2025). The performance of each model is calculated using LLM-as-Judge (Zheng et al., 2023) with GPT-4o (2024-08-06) serving as the scoring model.  
+- Arena-Hard-Auto Greek: a translated version of Arena-Hard-Auto v0.1, which originates from Chatbot Arena (Chiang et al., 2024) was included in m-ArenaHard (Dang et al., 2024) after translation with Google Translate API v3. We later post-edited using Claude Son
+
+<table><tr><td>Model</td><td>IFEval EL</td><td>IFEval EN</td><td>MT-Bench EL</td><td>MT-Bench EN</td></tr><tr><td>Qwen 2.5 7B</td><td>46.2</td><td>74.8</td><td>5.83</td><td>7.87</td></tr><tr><td>EuroLLM 9B</td><td>51.3</td><td>64.5</td><td>5.98</td><td>6.27</td></tr><tr><td>Aya Expanse 8B</td><td>50.4</td><td>62.2</td><td>7.68</td><td>6.92</td></tr><tr><td>Meltemi-7B-v1.5</td><td>32.7</td><td>41.2</td><td>6.25</td><td>5.46</td></tr><tr><td>Llama-3.1-8B</td><td>45.8</td><td>75.1</td><td>6.46</td><td>7.25</td></tr><tr><td>Llama-Krikri-8B</td><td>67.5</td><td>82.4</td><td>7.96</td><td>7.21</td></tr><tr><td>Gemma 2 27B IT</td><td>63.2</td><td>75.6</td><td>8.23</td><td>8.00</td></tr><tr><td>Aya Expanse 32B</td><td>60.3</td><td>70.2</td><td>8.27</td><td>7.40</td></tr></table>
+
+net 3.5 (Anthropic, 2024) with 10-shot examples to address translation issues, particularly in coding-related prompts where some parts would best be left untranslated, as well as to retain the original style of the prompts, since some of them would be best left vaguely posed as in the original prompt. We used the version of the benchmark with style control methods for Markdown elements<sup>6</sup>. We set GPT-4o-Mini (2024-07-18) as the baseline model (by default  $50\%$  score) and GPT-4o (2024-08-06) as the judge model.
+
+As shown in Table 6, Llama-Krikri-8B-Instruct demonstrates exceptional performance across both Greek and English benchmarks, substantially outperforming not only its parent model Llama-3.1-8B-Instruct but also other competitive multilingual models in the 7-9B parameter range. It should be noted that the IFEval scores reported in this table reflect our own implementation of the benchmark, which may differ from the Open LLM Leaderboard implementation due to variations in prompt formatting and evaluation criteria. Despite these methodological differences, the relative performance comparisons remain valid within each implementation context.
+
+On IFEval Greek, Llama-Krikri-8B-Instruct achieves a remarkable  $67.5\%$  accuracy, surpassing Llama-3.1-8B-Instruct by  $+21.7\%$  and Meltemi-7B-v1.5 by  $+34.8\%$ . Notably, our 8B model even outperforms much larger models like Gemma 2 27B IT  $(+4.3\%)$  and Aya Expanse 32B  $(+7.2\%)$  on this Greek instruction-following benchmark. As regards the original English IFEval, Llama-Krikri-8B-Instruct scores  $82.4\%$ , significantly higher than all other models, including those with 3-4 times more parameters. This dramatic improvement suggests that our data synthesis instruction tuning
+
+Table 6: Greek and English evaluation results using IFEval and MT-Bench.  
+
+<table><tr><td>Task</td><td>Llama-3.1-Instr.</td><td>Krikri-8B-Instr.</td></tr><tr><td>IFEval</td><td>49.22</td><td>60.79</td></tr><tr><td>BBH</td><td>29.38</td><td>29.31</td></tr><tr><td>MATH</td><td>15.56</td><td>11.78</td></tr><tr><td>GPQA</td><td>8.72</td><td>7.05</td></tr><tr><td>MUSR</td><td>8.61</td><td>10.46</td></tr><tr><td>MMLU-PRO</td><td>31.09</td><td>25.70</td></tr><tr><td>Avg.</td><td>23.76</td><td>24.18</td></tr></table>
+
+Table 7: Comparative evaluation on English benchmarks from the Open LLM Leaderboard.
+
+approach successfully addresses the unique challenges of following instructions in Greek, where naive translations of instruction data often fail to capture language-specific nuances.
+
+For MT-Bench Greek, which evaluates multi-turn conversation quality, Llama-Krikri-8B-Instruct achieves a score of 7.96, making it the top performer amongst other models in its size class. While larger models like Gemma 2 27B IT (8.23) and Aya Expanse 32B (8.27) achieve slightly higher scores on MT-Bench Greek, the margin is surprisingly small given the substantial difference in model size. On MT-Bench English, Llama-Krikri-8B maintains competitive performance at 7.21, essentially identical with Llama-3.1-8B-Instruct (-0.04), though understandably lower than the larger Gemma 2 27B IT (-0.79) and Aya Expanse (-0.19). This demonstrates that our instruction tuning approach can achieve a very high conversational performance on Greek conversational benchmarks, while also producing a competitive model for English benchmarks with a much more compact approach (Grattafori et al., 2024).
+
+As detailed in Table 7, Llama-Krikri-8B-Instruct's official Open LLM Leaderboard submission shows an average score of  $24.18\%$  across all tests, slightly surpassing the  $23.76\%$  of Llama-3.1-8B-Instruct. The model shows particularly impress
+
+<table><tr><td>Model</td><td>ArenaHard EL</td><td>ArenaHard EN</td></tr><tr><td>Aya Expanse 8B</td><td>23.8</td><td>-</td></tr><tr><td>Llama 3.1 8B Instr.</td><td>4.0</td><td>19.7</td></tr><tr><td>Krikri 8B Instr.</td><td>31.8</td><td>35.1</td></tr><tr><td>Aya Expanse 32B</td><td>40.1</td><td>45.1</td></tr><tr><td>Gemma 2 27B IT</td><td>32.2</td><td>49.6</td></tr><tr><td>Llama 3.1 70B Instr.</td><td>27.4</td><td>53.9</td></tr><tr><td>GPT 4o Mini</td><td>50.0</td><td>65.0</td></tr></table>
+
+Table 8: Arena Hard evaluation results (\% win rate) for Greek and English.
+
+sive gains on IFEval implementation (60.79% vs. 49.22%) and MUSR (10.46% vs. 8.61%), while closely matching performance on the Big Bench Hard (BBH) benchmark (29.31% vs. 29.38%). Although Llama-Krikri performs slightly below Meta-Llama-3.1-8B-Instruct in the MMLU-PRO category (25.70% vs. 31.09%), the overall performance indicates successful retention of English capabilities during the Greek-focused continual pretraining.
+
+The results from our Arena Hard evaluations, presented in Table 8, reveal that, in the 8B parameter range, Llama-Krikri-8B-Instruct significantly outperforms its competitors, achieving a  $31.8\%$  win rate on Arena Hard Greek compared to Aya Expanse 8B's  $23.8\%$  and Llama 3.1 8B Instruct's  $4.0\%$ $(+27.8\%)$  improvement). This demonstrates the effectiveness of our Greek-focused training approach. Even more impressively, Llama-Krikri-8B-Instruct achieves a  $35.1\%$  win rate on Arena Hard English, substantially outperforming the original Llama-3.1-8B-Instruct  $(19.7\%)$  by  $+16.2\%$ , despite our focus on Greek capabilities. While Aya Expanse 32B leads on Arena Hard Greek with  $40.1\%$ , our 8B model is on par with Gemma 2 27B IT  $(31.8\%$  vs.  $32.2\%)$  and outperforms the 8.75 times larger Llama-3.1-70B-Instruct  $(27.4\%)$  by  $+4.4\%$  on the Greek evaluation data. On the original English Arena Hard, the larger models generally perform better, although it should be noted that Llama-Krikri-8B-Instruct outperforms Llama-3.1-8B-Instruct by  $+15.4\%$ $(35.1\%$  vs.  $19.7\%)$ .
+
+Please note that while all models trail behind GPT-4o-Mini (used as baseline on the Greek Arena Hard), recent research (Li et al., 2025) has shown that judge models are biased towards student models, i.e., models finetuned on distilled data from the stronger & larger teacher model which also acts as a judge. While details on the post-training data of GPT-4o-Mini are undisclosed, it would be very reasonable to assume that it has been trained -at least partly- with GPT-4 and GPT-4o serving as teacher models and, therefore, that the judges that
+
+we utilized are biased towards it compared to all other evaluated models.
+
+This performance comparison with much larger models highlights the efficiency of our approach since Llama-Krikri-8B-Instruct achieves comparable or even superior performance on Greek benchmarks compared to models with 3-4x more parameters, while maintaining strong English capabilities. This efficiency is particularly important for deployment scenarios where computational resources may be limited, demonstrating that a carefully trained smaller model can rival much larger ones for specific languages.
+
+Apart from the comparative evaluations mentioned above, we have performed zero-shot MT experiments on an Ancient-Modern Greek  $(\mathrm{grc}\leftrightarrow \mathrm{ell})$  translation dataset that includes 100 sentences of Ancient Greek texts manually translated into Modern Greek. The dataset was developed by automatically sentence aligning document pairs, with the result corrected by human annotators. Using Llama-Krikri-8B-Instruct we have observed a 54.66 BLEU score for the Ancient to Modern Greek  $(\mathrm{grc}\rightarrow \mathrm{ell})$  translation direction, with the reverse direction  $(\mathrm{ell}\rightarrow \mathrm{grc})$  being more challenging (20.41 BLEU).
+
+Finally, we have also developed a novel high-quality dataset consisting of 1200 multiple-choice questions prepared manually and published online by the Greek Supreme Council for Civil Personnel Selection. The dataset covers several domains including Law, Economics, Informatics, and Modern Greek History. The evaluation results on this dataset demonstrate a  $59.75\%$  accuracy that is similar to a  $61.25\%$  accuracy for larger Gemma3-12b.
+
+# 5 Discussion and Conclusions
+
+In this paper, we presented Llama-Krikri-8B, a new LLM that exhibits significant skills in understanding and generating Greek, while also showing highly accurate handling of text in English and historical Greek dialects. We achieved this by developing an efficient tokenizer that exhibits a low token/words fertility for Greek and by further training Llama 3.1-8B using a carefully constructed dataset that covered a wide variety of domains. In evaluation experiments on a benchmark suite comprising Greek and English datasets, we have observed
+
+that Llama-Krikri-8B performs significantly better in Greek  $(+10.8\%)$  compared to its base model, while also showing gains in English  $(+0.8\%)$ . We then created Llama-Krikri-8B-Instruct, a version designed for following instructions and engaging in helpful conversations. This involved a multi-step process that comprised synthetic data generation in a multitude of domains, fine-tuning the model and then aligning it with human preferences. Evaluations revealed that Llama-Krikri-8B-Instruct significantly outperformed Llama-3.1-8B-Instruct in both Greek  $(+21.7\%)$  and English  $(+7.3\%)$  IFEval. Our model also demonstrated highly competitive chat abilities in both languages across several benchmarks.
+
+# 6 Limitations
+
+The quality and accessibility of Greek datasets are critical to the development of Krikri. Greek opensource corpora are becoming more numerous, but they might not be as large or varied as datasets for more extensively spoken languages, like English. This may result in biases in the model's understanding of Greek, especially with regard to regional variances, dialects, and specialized fields like technical fields, law, or medicine.
+
+As an 8B parameter model, our model shows a fairly high level of Greek fluency, but it is less effective than larger-class and commercial models at reasoning and instruction following, and is more likely to experience hallucinations.
+
+In the future, our evaluation benchmarks should include more original Greek LLM datasets that are not the result of machine translation and post-editing. These datasets will help minimize the effect of machine translation on evaluation results and also better reflect the target language and culture.
+
+# 7 Risks and ethical considerations
+
+To mitigate potential risks, we took several steps to ensure the data used for training did not contain personally identifiable information, offensive, or otherwise inappropriate content. We sourced data from publicly available, licensed, or open-access datasets, ensuring compliance with their respective policies and any flagged data points were excluded. We did not collect data from private communications or data sources that could contain personally identifiable information. We have also given special care to align our model's responses with safety
+
+guidelines followed by manual reviews.
+
+However, we have not performed a systematic evaluation against LLM risks including risks related to discrimination, hate speech and exclusion; information hazards; and misinformation harms (Weidinger et al., 2022). As this was due to lack of relevant evaluation material in Greek, we want to contribute towards improving this situation in the future.
+
+We recognize that these measures are not a substitute for more thorough evaluation protocols. Moving forward, we aim to contribute towards addressing these limitations by promoting the development of Greek-language evaluation resources for LLM risks. This will enable more robust and contextually appropriate assessments of ethical risks in future models.
+
+# Acknowledgements
+
+The authors wish to thank AWS and GRNET, especially Nikiforos Botis and Panos Louridas, for their ongoing support and helping us attain the required training infrastructure. We express our sincere gratitude to all members of the Institute for Language and Speech Processing, Athena RC, for their unwavering support of this project.
+
+# References
+
+Alon Albalak, Yanai Elazar, Sang Michael Xie, Shayne Longpre, Nathan Lambert, Xinyi Wang, Niklas Muennighoff, Bairu Hou, Liangming Pan, Haewon Jeong, et al. 2024. A survey on data selection for language models. arXiv preprint arXiv:2402.16827.  
+Anton Alexandrov, Veselin Raychev, Dimitar I. Dimitrov, Ce Zhang, Martin Vechev, and Kristina Toutanova. 2024. Bggpt 1.0: Extending english-centric llms to other languages. Preprint, arXiv:2412.10893.  
+Loubna Ben Allal, Anton Lozhkov, Elie Bakouch, Gabriel Martin Blázquez, Guilherme Penedo, Lewis Tunstall, Andrés Marafioti, Hynek Kydlíček, Agustín Piqueres Lajarín, Vaibhav Srivastav, Joshua Lochner, Caleb Fahlgren, Xuan-Son Nguyen, Clémentine Fourrier, Ben Burtenshaw, Hugo Larcher, Haojun Zhao, Cyril Zakka, Mathieu Morlon, Colin Raffel, Leandro von Werra, and Thomas Wolf. 2025. Smollm2: When smol goes big - data-centric training of a small language model. Preprint, arXiv:2502.02737.  
+Duarte M Alves, José Pombal, Nuno M Guerreiro, Pedro H Martins, João Alves, Amin Farajian, Ben Peters, Ricardo Rei, Patrick Fernandes, Sweta Agrawal, et al. 2024. Tower: An open multilingual large
+
+language model for translation-related tasks. arXiv preprint arXiv:2402.17733.  
+AI Anthropic. 2024. The claude 3 model family: Opus, sonnet, haiku. Claude-3 Model Card.  
+Mikel Artetxe and Holger Schwenk. 2018. Margin-based parallel corpus mining with multilingual sentence embeddings. arXiv preprint arXiv:1811.01136.  
+Mikel Artetxe and Holger Schwenk. 2019. Massively multilingual sentence embeddings for zero-shot cross-lingual transfer and beyond. Transactions of the association for computational linguistics, 7:597-610.  
+Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. 2023. Qwen technical report. arXiv preprint arXiv:2309.16609.  
+Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, Carol Chen, Catherine Olsson, Christopher Olah, Danny Hernandez, Dawn Drain, Deep Ganguli, Dustin Li, Eli Tran-Johnson, Ethan Perez, Jamie Kerr, Jared Mueller, Jeffrey Ladish, Joshua Landau, Kamal Ndousse, Kamile Lukosuite, Liane Lovitt, Michael Sellitto, Nelson Elhage, Nicholas Schiefer, Noemi Mercado, Nova DasSarma, Robert Lasenby, Robin Larson, Sam Ringer, Scott Johnston, Shauna Kravec, Sheer El Showk, Stanislav Fort, Tamera Lanham, Timothy Telleen-Lawton, Tom Conerly, Tom Henighan, Tristan Hume, Samuel R. Bowman, Zac Hatfield-Dodds, Ben Mann, Dario Amodei, Nicholas Joseph, Sam McCandlish, Tom Brown, and Jared Kaplan. 2022. Constitutional ai: Harmlessness from ai feedback. Preprint, arXiv:2212.08073.  
+Lucas Bandarkar, Davis Liang, Benjamin Muller, Mikel Artetxe, Satya Narayan Shukla, Donald Husa, Naman Goyal, Abhinandan Krishnan, Luke Zettlemoyer, and Madian Khabsa. 2024. The belebele benchmark: a parallel reading comprehension dataset in 122 language variants. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 749-775.  
+Marta Bañón, Miquel Esplà-Gomis, Mikel L. Forcada, Cristian García-Romero, Taja Kuzman, Nikola Ljubesić, Rik van Noord, Leopoldo Pla Sempere, Gema Ramírez-Sánchez, Peter Rupnik, Vít Suchomel, Antonio Toral, Tobias van der Werff, and Jaume Zaragoza. 2022. MaCoCu: Massive collection and curation of monolingual and bilingual data: focus on under-resourced languages. In Proceedings of the 23rd Annual Conference of the European Association for Machine Translation, pages 303-304, Ghent, Belgium. European Association for Machine Translation.  
+A Broder. 1997. On the resemblance and containment of documents. In Proceedings of the Compression and Complexity of Sequences 1997, page 21.
+
+Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. Preprint, arXiv:2005.14165.  
+Nicholas Carlini, Daphne Ippolito, Matthew Jagielski, Katherine Lee, Florian Tramér, and Chiyuan Zhang. 2023. Quantifying memorization across neural language models. In The Eleventh International Conference on Learning Representations. OpenReview.  
+Ilias Chalkidis, Manos Fergadiotis, and Ion Androutopoulos. 2021. Multieurlex - a multi-lingual and multi-label legal document classification dataset for zero-shot cross-lingual transfer. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics.  
+Ilias Chalkidis, Manos Fergadiotis, Prodromos Malakasiotis, and Ion Androutsopoulos. 2019. Large-scale multi-label text classification on EU legislation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 6314-6322, Florence, Italy. Association for Computational Linguistics.  
+Yi-Chang Chen, Po-Chun Hsu, Chan-Jan Hsu, and Dashan Shiu. 2025. Enhancing function-calling capabilities in LLMs: Strategies for prompt formats, data integration, and multilingual translation. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 3: Industry Track), pages 99-111, Albuquerque, New Mexico. Association for Computational Linguistics.  
+Wei-Lin Chiang, Lianmin Zheng, Ying Sheng, Anastasios Nikolas Angelopoulos, Tianle Li, Dacheng Li, Banghua Zhu, Hao Zhang, Michael Jordan, Joseph E Gonzalez, et al. 2024. Chatbot arena: An open platform for evaluating llms by human preference. In International Conference on Machine Learning, pages 8359-8388. PMLR.  
+Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. 2018. Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge. arXiv:1803.05457v1.  
+Zoltan Csaki, Pian Pawakapan, Urmish Thakker, and Qiantong Xu. 2023. Efficiently adapting pretrained language models to new languages. arXiv preprint arXiv:2311.05741.
+
+Ganqu Cui, Lifan Yuan, Ning Ding, Guanming Yao, Wei Zhu, Yuan Ni, Guotong Xie, Zhiyuan Liu, and Maosong Sun. 2023. Ultrafeedback: Boosting language models with high-quality feedback. Preprint, arXiv:2310.01377.  
+John Dang, Shivalika Singh, Daniel D'souza, Arash Ahmadian, Alejandro Salamanca, Madeline Smith, Aidan Peppin, Sungjin Hong, Manoj Govindassamy, Terrence Zhao, Sandra Kublik, Meor Amer, Viraat Aryabumi, Jon Ander Campos, Yi-Chern Tan, Tom Kocmi, Florian Strub, Nathan Grinsztajn, Yannis Flet-Berliac, Acyr Locatelli, Hangyu Lin, Dwarak Talupuru, Bharat Venkitesh, David Cairuz, Bowen Yang, Tim Chung, Wei-Yin Ko, Sylvie Shang Shi, Amir Shukayev, Sammie Bae, Aleksandra Piktus, Roman Castagné, Felipe Cruz-Salinas, Eddie Kim, Lucas Crawhall-Stein, Adrien Morisot, Sudip Roy, Phil Blunsom, Ivan Zhang, Aidan Gomez, Nick Frosst, Marzieh Fadaee, Beyza Ermis, Ahmet Üstün, and Sara Hooker. 2024. Aya expanse: Combining research breakthroughs for a new multilingual frontier. Preprint, arXiv:2412.04261.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. Preprint, arXiv:1810.04805.  
+Tomaž Erjavec, Maciej Ogrodniczuk, Petya Osenova, Nikola Ljubesic, Kiril Simov, Andrej Pančur, Michal Rudolf, Matyáš Kopp, Starkažur Barkarson, Steingrímsson, Cagri Žoltekin, Jesse de Does, Katrien Depuydt, Tommaso Agnoloni, Giulia Venturi, María Calzada Pérez, Luciana D. de Macedo, Costanza Navarretta, Giancarlo Luxardo, Matthew Coole, Paul Rayson, Vaidas Morkevicius, Tomas Krilavicius, Roberts DarUndefinedis, Orsolya Ring, Ruben van Heusden, Maarten Marx, and Darja Fiser. 2022. The ParlaMint corpora of parliamentary proceedings. Lang. Resour. Eval., 57(1):415-448.  
+Clémentine Fourrier, Nathan Habib, Alina Lozovskaya, Konrad Szafer, and Thomas Wolf. 2024. Open llm leaderboard v2. https://huggingface.co/spaces/open-llm-leaderboard/open_11mleaderboard.  
+Leo Gao, Jonathan Tow, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Kyle McDonell, Niklas Muennighoff, Jason Phang, Laria Reynolds, Eric Tang, Anish Thite, Ben Wang, Kevin Wang, and Andy Zou. 2021. A framework for few-shot language model evaluation.  
+Maria Gavriilidou, Stelios Piperidis, Dimitrios Galanis, Kanella Pouli, Penny Labropoulou, Juli Bakagianni, Iro Tsiouli, Miltos Deligiannis, Athanasia Kolovou, Dimitris Gkoumas, Leon Voukoutis, and Katerina Gkirtzou. 2023. The CLARIN:EL infrastructure: Platform, Portal, K-Centre. In Selected papers from the CLARIN Annual Conference 2023.  
+Aaron Grattafori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-
+
+Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston Zhang, Aurelien Rodriguez, Austen Gregerson, Ava Spataru, Baptiste Roziere, Bethany Biron, Binh Tang, Bobbie Chern, Charlotte Caucheteux, Chaya Nayak, Chloe Bi, Chris Marra, Chris McConnell, Christian Keller, Christophe Touret, Chunyang Wu, Corinne Wong, Cristian Canton Ferrer, Cyrus Nikolaidis, Damien Alonsius, Daniel Song, Danielle Pintz, Danny Livshits, Danny Wyatt, David Esiobu, Dhruv Choudhary, Dhruv Mahajan, Diego Garcia-Olano, Diego Perino, Dieuwke Hupkes, Egor Lakomkin, Ehab AlBadawy, Elina Lobanova, Emily Dinan, Eric Michael Smith, Filip Radenovic, Francisco Guzmán, Frank Zhang, Gabriel Synnaeve, Gabrielle Lee, Georgia Lewis Anderson, Govind Thattai, Graeme Nail, Gregoire Mialon, Guan Pang, Guillem Cucurell, Hailey Nguyen, Hannah Korevaar, Hu Xu, Hugo Touvron, Iliyan Zarov, Imanol Arrieta Ibarra, Isabel Kloumann, Ishan Misra, Ivan Evtimov, Jack Zhang, Jade Copet, Jaewon Lee, Jan Geffert, Jana Vranes, Jason Park, Jay Mahadeokar, Jeet Shah, Jelmer van der Linde, Jennifer Billock, Jenny Hong, Jenya Lee, Jeremy Fu, Jianfeng Chi, Jianyu Huang, Jiawen Liu, Jie Wang, Jiecao Yu, Joanna Bitton, Joe Spisak, Jongsoo Park, Joseph Rocca, Joshua Johnstun, Joshua Saxe, Junteng Jia, Kalyan Vasuden Alwala, Karthik Prasad, Kartikeya Upasani, Kate Plawiak, Ke Li, Kenneth Heafield, Kevin Stone, Khalid El-Arini, Krithika Iyer, Kshitiz Malik, Kuenley Chiu, Kunal Bhalla, Kushal Lakhotia, Lauren Rantala-Yeary, Laurens van der Maaten, Lawrence Chen, Liang Tan, Liz Jenkins, Louis Martin Lovish Madaan Lubo Malo Lukas Blecher, Lukas Landzaat Luke de Oliveira Madeline Muzzi Mahesh Pasupuleti Mannat Singh Manohar Paluri Marcin Kardas Maria Tsimpoukelli Mathew Oldham Mathieu Rita Maya Pavlova Melanie Kambadur Mike Lewis Min Si Mitesh Kumar Singh Mona Hassan Naman Goyal Narjes Torabi Nikolay Bashlykov,Nikolay Bogoychev,Niladri Chatterji Ning Zhang,Olivier Duchenne Onur Celebi Patrick Alrassy,Pengchuan Zhang,Pengwei Li,Petar Vasic Peter Weng Prajwal Bhargava Pratik Dubal Praveen Krishnan,Punit Singh Koura Puxin Xu Qing He,Qingxiao Dong,Ragavan Srinivasan Raj Ganapathy Ramon Calderer Ricardo Silveira Cabral Robert Stojnic Roberta Raileanu Rohan Maheswari Rohit Girdhar Rohit Patel Romain Sauvestre Ronnie Polidoro Roshan Sumbaly Ross Taylor Ruan Silva Rui Hou Rui Wang Saghar Hosseini Sahana Chennabasappa Sanjay Singh Sean Bell Seohyun Sonia Kim,Sergey Edunov Shaoliang Nie Sharan Narang Sharath Rarparthy Sheng Shen Shengye Wan Shruti Bhosale Shun Zhang Simon Vandenhende Soumya Batra Spencer Whitman Sten Sootla Stephane Collot Suchin Gururangan Sydney Borodinsky Tamar Herman Tara Fowler Tarek Sheasha Thomas Georgiou Thomas Scialom Tobias Speckbacher,Todor Mihaylov Tong Xiao Ujjwal Karn,Vedanuj Goswami Vibhor Gupta Vignesh Ramanathan Viktor Kerkez Vincent Gonguet Vir
+
+ginie Do, Vish Vogeti, Vitor Albiero, Vladan Petrovic, Weiwei Chu, Wenhan Xiong, Wenyin Fu, Whitney Meers, Xavier Martinet, Xiaodong Wang, Xiaofang Wang, Xiaoqing Ellen Tan, Xide Xia, Xinfeng Xie, Xuchao Jia, Xuewei Wang, Yaelle Goldschlag, Yashesh Gaur, Yasmine Babaei, Yi Wen, Yiwen Song, Yuchen Zhang, Yue Li, Yuning Mao, Zacharie Delpierre Coudert, Zheng Yan, Zhengxing Chen, Zoe Papakipos, Aaditya Singh, Aayushi Srivastava, Abha Jain, Adam Kelsey, Adam Shajnfeld, Adithya Gangidi, Adolfo Victoria, Ahuva Goldstand, Ajay Menon, Ajay Sharma, Alex Boesenberg, Alexei Baevski, Allie Feinstein, Amanda Kallet, Amit Sangani, Amos Teo, Anam Yunus, Andrei Lupu, Andres Alvarado, Andrew Caples, Andrew Gu, Andrew Ho, Andrew Poulton, Andrew Ryan, Ankit Ramchandani, Annie Dong, Annie Franco, Anuj Goyal, Aparajita Saraf, Arkabandhu Chowdhury, Ashley Gabriel, Ashwin Bharambe, Assaf Eisenman, Azadeh Yazdan, Beau James, Ben Maurer, Benjamin Leonhardi, Bernie Huang, Beth Loyd, Beto De Paola, Bhargavi Paranjape, Bing Liu, Bo Wu, Boyu Ni, Braden Hancock, Bram Wasti, Brandon Spence, Brani Stojkovic, Brian Gamido, Britt Montalvo, Carl Parker, Carly Burton, Catalina Mejia, Ce Liu, Changhan Wang, Changkyu Kim, Chao Zhou, Chester Hu, ChingHsiang Chu, Chris Cai, Chris Tindal, Christoph Feichtenhofer, Cynthia Gao, Damon Civin, Dana Beaty Daniel Kreymer, Daniel Li, David Adkins, David Xu, Davide Testuggine, Delia David, Devi Parikh, Diana Liskovich, Didem Foss, Dingkang Wang, Duc Le,Dustin Holland Edward Dowling,Eissa Jamil Elaine Montgomery,Eleonora Presani, Emily Hahn Emily WoodEric-Tuan LeErik BrinkmanEsteban Arcaute,Evan Dunbar,Evan Smothers,Fei Sun Felix KreukFeng Tian Filippos Kokkinos Firat Ozgenel Francesco Caggioni Frank Kanayet Frank Seide,Gabriela Medina Flores,Gabriella Schwarz Gada Badeer Georgia Swee Gil Halpern Grant Herman Grigory Sizov Guangyi ZhangGuna Lakshminarayanan,Hakan Inan Hamid Shojanazeri Han Zou Hannah WangHanwen Zha Haroun Habeeb Harrison Rudolph Helen Suk Henry Aspegren Hunter Goldman Hongyuan Zhan Ibrahim Damlaj,Igor Molybog,Igor Tufanov Ilias Leontiadis Irina-Elena Veliche Itai Gat Jake WeissmanJames Geboski James Kohli Janice Lam Japhet Asher Jean-Baptiste GayaJeff MarcusJeff TangJennifer Chan Jenny Zhen Jeremy Reizenstein Jeremy Teboul Jessica Zhong Jian Jin Jingyi Yang Joe Cummings Jon Carvill Jon Shepard Jonathan McPhie Jonathan Torres Josh Ginsburg Junjie Wang Kai Wu,Kam Hou U,Karan Saxena,Kartikay Khandelwal Katayoun Zand Kathy Matosich Kaushik Veeraraghavan Kelly Michelena Keqian Li Kiran Jagadeesh Kun HuangKunal Chawla Kyle HuangLailin ChenLakshya Garg Lavender A Leandro Silva Lee Bell Lei Zhang Liangpeng Guo,Licheng Yu,Liron MoshkovichLuca Wehrstedt,Madian KhabsaManav Avalani Manish Bhatt Martynas Mankus,Matan Hasson Matthew Lennie Matthias Reso Maxim Groshev Maxim Naumov Maya Lathi Meghan Keneally Miao Liu Michael L. Seltzer,Michal Valko Michelle Restrepo,Mihir Pa
+
+tel, Mik Vyatskov, Mikayel Samvelyan, Mike Clark, Mike Macey, Mike Wang, Miquel Jubert Hermoso, Mo Metanat, Mohammad Rastegari, Munish Bansal, Nandhini Santhanam, Natascha Parks, Natasha White, Navyata Bawa, Nayan Singhal, Nick Egebo, Nicolas Usunier, Nikhil Mehta, Nikolay Pavlovich Laptev, Ning Dong, Norman Cheng, Oleg Chernoguz, Olivia Hart, Omkar Salpekar, Ozlem Kalinli, Parkin Kent, Parth Parekh, Paul Saab, Pavan Balaji, Pedro Rittner, Philip Bontrager, Pierre Roux, Piotr Dollar, Polina Zvyagina, Prashant Ratanchandani, Pritish Yuvraj, Qian Liang, Rachad Alao, Rachel Rodriguez, Rafi Ayub, Ragtoham Murthy, Raghu Nayani, Rahul Mitra, Rangaprabhu Parthasarathy, Raymond Li, Rebekkah Hogan, Robin Battey, Rocky Wang, Russ Howes, Rudy Rinott, Sachin Mehta, Sachin Siby, Sai Jayesh Bondu, Samyak Datta, Sara Chugh, Sara Hunt, Sargun Dhillon, Sasha Sidorov, Satadru Pan, Saurabh Mahajan, Saurabh Verma, Seiji Yamamoto, Sharadh Ramaswamy, Shaun Lindsay, Shaun Lindsay, Sheng Feng, Shenghao Lin, Shengxin Cindy Zha, Shishir Patil, Shiva Shankar, Shuqiang Zhang, Shuqiang Zhang, Sinong Wang, Sneha Agarwal, Soji Sajuyigbe, Soumith Chintala, Stephanie Max, Stephen Chen, Steve Kehoe, Steve Satterfield, Sudarshan Govindaprasad, Sumit Gupta, Summer Deng, Sungmin Cho, Sunny Virk, Suraj Subramanian, Sy Choudhury, Sydney Goldman, Tal Remez, Tamar Glaser, Tamara Best, Thilo Koehler, Thomas Robinson, Tianhe Li, Tianjun Zhang, Tim Matthews, Timothy Chou, Tzook Shaked, Varun Vontimitta, Victoria Ajayi, Victoria Montanez, Vijai Mohan, Vinay Satish Kumar, Vishal Mangla, Vlad Ionescu, Vlad Poenaru, Vlad Tiberiu Mihailescu, Vladimir Ivanov, Wei Li, Wenchen Wang, Wenwen Jiang, Wes Bouaziz, Will Constable, Xiaocheng Tang, Xiaojian Wu, Xiaolan Wang, Xilun Wu, Xinbo Gao, Yaniv Kleinman, Yanjun Chen, Ye Hu, Ye Jia. Ye Qi,Yenda Li,Yilin Zhang,Ying Zhang,Yossi Adi Youngjin Nam,Yu,Wang,Yu Zhao,Yuchen Hao. Yundi Qian,Yunlu Li,Yuzi He,Zach Rait,Zachary DeVito,Zef Rosnbrick,Zhaoduo Wen,Zhenyu Yang Zhiwei Zhao,and Zhiyu Ma.2024.The llama 3 herd of models.Preprint arXiv:2407.21783.
+
+Suchin Gururangan, Ana Marasovic, Swabha Swayamdipta, Kyle Lo, Iz Beltagy, Doug Downey, and Noah A. Smith. 2020. Don't stop pretraining: Adapt language models to domains and tasks. Preprint, arXiv:2004.10964.
+
+Kenneth Heafield. 2011. KenLM: Faster and smaller language model queries. In Proceedings of the Sixth Workshop on Statistical Machine Translation, pages 187-197, Edinburgh, Scotland. Association for Computational Linguistics.
+
+Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring massive multitask language understanding. In International Conference on Learning Representations.
+
+Pin-Lun Hsu, Yun Dai, Vignesh Kothapalli, Qingquan Song, Shao Tang, Siyu Zhu, Steven Shimizu, Shivam
+
+Sahni, Haowen Ning, and Yanning Chen. 2024. Liger kernel: Efficient triton kernels for llm training. arXiv preprint arXiv:2410.10989.  
+Adam Ibrahim, Benjamin Thérien, Kshitij Gupta, Mats L Richter, Quentin Anthony, Timothee Lesort, Eugene Belilovsky, and Irina Rish. 2024. Simple and scalable strategies to continually pre-train large language models. arXiv preprint arXiv:2403.08763.  
+Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. 2023. Mistral 7b. arXiv preprint arXiv:2310.06825.  
+Armand Joulin, Édouard Grave, Piotr Bojanowski, and Tomáš Mikolov. 2017. Bag of tricks for efficient text classification. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 427-431.  
+LAION. 2023. LeoLM: Igniting German-Language LLM Research. https://laion.ai/blog/leo-lm/. Accessed: (12 July 2024).  
+Nathan Lambert, Jacob Morrison, Valentina Pyatkin, Shengyi Huang, Hamish Ivison, Faeze Brahman, Lester James V. Miranda, Alisa Liu, Nouha Dziri, Shane Lyu, Yuling Gu, Saumya Malik, Victoria Graf, Jena D. Hwang, Jiangjiang Yang, Ronan Le Bras, Oyvind Tafjord, Chris Wilhelm, Luca Soldaini, Noah A. Smith, Yizhong Wang, Pradeep Dasigi, and Hannaneh Hajishirzi. 2025. Tulu 3: Pushing frontiers in open language model post-training. Preprint, arXiv:2411.15124.  
+Katherine Lee, Daphne Ippolito, Andrew Nystrom, Chiyuan Zhang, Douglas Eck, Chris Callison-Burch, and Nicholas Carlini. 2022. Deduplicating training data makes language models better. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8424-8445.  
+Jure Leskovec, Anand Rajaraman, and Jeffrey David Ullman. 2020. Mining of massive data sets. Cambridge university press.  
+Dawei Li, Renliang Sun, Yue Huang, Ming Zhong, Bohan Jiang, Jiawei Han, Xiangliang Zhang, Wei Wang, and Huan Liu. 2025. Preference leakage: A contamination problem in llm-as-a-judge. Preprint, arXiv:2502.01534.  
+Tianle Li, Wei-Lin Chiang, Evan Frick, Lisa Dunlap, Tianhao Wu, Banghua Zhu, Joseph E Gonzalez, and Ion Stoica. 2024. From crowdsourced data to high-quality benchmarks: Arena-hard and benchbuilder pipeline. arXiv preprint arXiv:2406.11939.  
+Stephanie Lin, Jacob Hilton, and Owain Evans. 2022. TruthfulQA: Measuring how models mimic human falsehoods. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics
+
+(Volume 1: Long Papers), pages 3214-3252, Dublin, Ireland. Association for Computational Linguistics.  
+Chris Yuhao Liu, Liang Zeng, Jiacai Liu, Rui Yan, Ju-jie He, Chaojie Wang, Shuicheng Yan, Yang Liu, and Yahui Zhou. 2024. Skywork-reward: Bag of tricks for reward modeling in llms. arXiv preprint arXiv:2410.18451.  
+Kyle Lo, Lucy Lu Wang, Mark Neumann, Rodney Kinney, and Daniel S Weld. 2020. S2orc: The semantic scholar open research corpus. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4969-4983.  
+Andrea Lösch, Valérie Mapelli, Khalid Choukri, Maria Giagkou, Stelios Piperidis, Prokopis Prokopidis, Vassilis Papavassiliou, Miltos Deligiannis, Aivars Berzins, Andrejs Vasiljevs, et al. 2021. Collection and Curation of Language Data within the European Language Resource Coordination (ELRC). In *Quator*.  
+Ilya Loshchilov and Frank Hutter. 2017. Fixing Weight Decay Regularization in Adam. CoRR, abs/1711.05101.  
+Anton Lozhkov, Raymond Li, Loubna Ben Allal, Federico Cassano, Joel Lamy-Poirier, Nouamane Tazi, Ao Tang, Dmytro Pykhtar, Jiawei Liu, Yuxiang Wei, et al. 2024. Starcoder 2 and the stack v2: The next generation. arXiv preprint arXiv:2402.19173.  
+Yun Luo, Zhen Yang, Fandong Meng, Yafu Li, Jie Zhou, and Yue Zhang. 2025. An empirical study of catastrophic forgetting in large language models during continual fine-tuning. *Preprint*, arXiv:2308.08747.  
+Pedro Henrique Martins, Patrick Fernandes, João Alves, Nuno M Guerreiro, Ricardo Rei, Duarte M Alves, José Pombal, Amin Farajian, Manuel Faysse, Mateusz Klimaszewski, et al. 2024. Eurollm: Multilingual language models for europe. In Proceedings of the Ninth Conference on Machine Translation, pages 1393-1409. Association for Computational Linguistics.  
+Thuat Nguyen, Chien Van Nguyen, Viet Dac Lai, Hieu Man, Nghia Trung Ngo, Franck Dernoncourt, Ryan A Rossi, and Thien Huu Nguyen. 2023. CulturaX: A cleaned, enormous, and multilingual dataset for large language models in 167 languages. arXiv preprint arXiv:2309.09400.  
+Long Ouyang, Jeff 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 Christiano, Jan Leike, and Ryan Lowe. 2022. Training language models to follow instructions with human feedback. Preprint, arXiv:2203.02155.  
+Christos Papaloukas, Ilias Chalkidis, Konstantinos Athinaios, Despina-Athanasia Pantazi, and Manolis
+
+Koubarakis. 2021. Multi-granular legal topic classification on greek legislation. In Proceedings of the Natural Legal Language Processing Workshop 2021, pages 63-75, Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Vassilis Papavassiliou, Sokratis Sofianopoulos, Prokopis Prokopidis, and Stelios Piperidis. 2018. The ILSP/ARC submission to the WMT 2018 parallel corpus filtering shared task. In Proceedings of the Third Conference on Machine Translation: Shared Task Papers, pages 928-933, Belgium, Brussels. Association for Computational Linguistics.  
+Guilherme Penedo, Hynek Kydlíček, Anton Lozhkov, Margaret Mitchell, Colin Raffel, Leandro Von Werra, Thomas Wolf, et al. 2024. The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scale. arXiv preprint arXiv:2406.17557.  
+Ramon Pires, Hugo Abonizio, Thales Sales Almeida, and Rodrigo Nogueira. 2023. Sabia: Portuguese large language models. In Intelligent Systems, pages 226-240, Cham. Springer Nature Switzerland.  
+Ye Qi, Devendra Sachan, Matthieu Felix, Sarguna Padmanabhan, and Graham Neubig. 2018. When and why are pre-trained word embeddings useful for neural machine translation? In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 529-535, New Orleans, Louisiana. Association for Computational Linguistics.  
+Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. 2024. Direct preference optimization: Your language model is secretly a reward model. Preprint, arXiv:2305.18290.  
+Ricardo Rei, Marcos Treviso, Nuno M. Guerreiro, Chrysoula Zerva, Ana C Farinha, Christine Maroti, José G. C. de Souza, Taisiya Glushkova, Duarte Alves, Luisa Coheur, Alon Lavie, and André F. T. Martins. 2022. CometKiwi: IST-unbabel 2022 submission for the quality estimation shared task. In Proceedings of the Seventh Conference on Machine Translation (WMT), pages 634-645, Abu Dhabi, United Arab Emirates (Hybrid). Association for Computational Linguistics.  
+Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Rame, Johan Ferret, Peter Liu, Pouya Tafti, Abe Friesen, Michelle Casbon, Sabela Ramos, Ravin Kumar, Charline Le Lan, Sammy Jerome, Anton Tsitsulin, Nino Vieillard, Piotr Stanczyk, Sertan Girgin, Nikola Momchev, Matt Hoffman, Shantanu Thakoor, Jean-Bastien Grill, Behnam Neyshabur, Olivier Bachem, Alanna Walton, Aliaksei Severyn, Alicia Parrish, Aliya Ahmad, Allen Hutchison, Alvin Abdagic, Amanda Carl, Amy Shen, Andy Brock, Andy Coenen, Anthony Laforge, Antonia Paterson, Ben Bastian, Bilal Piot, Bo Wu, Brandon
+
+Royal, Charlie Chen, Chintu Kumar, Chris Perry, Chris Welty, Christopher A. Choquette-Choo, Danila Sinopalnikov, David Weinberger, Dimple Vijaykumar, Dominika Rogozinska, Dustin Herbison, Elisa Bandy, Emma Wang, Eric Noland, Erica Moreira, Evan Senter, Evgenii Eltsyshv, Francesco Visin, Gabriel Rasskin, Gary Wei, Glenn Cameron, Gus Martins, Hadi Hashemi, Hanna Klimczak-Plucińska, Harleen Batra, Harsh Dhand, Ivan Nardini, Jacinda Mein, Jack Zhou, James Svensson, Jeff Stanway, Jetha Chan, Jin Peng Zhou, Joana Carrasqueira, Joana Iljazi, Jocelyn Becker, Joe Fernandez, Joost van Amersfoort, Josh Gordon, Josh Lipschultz, Josh Newlan, Ju yeong Ji, Kareem Mohamed, Kartikeya Badola, Kat Black, Katie Millican, Keelin McDonell, Kelvin Nguyen, Kiranbir Sodhia, Kish Greene, Lars Lowe Sjoesund, Lauren Usui, Laurent Sifre, Lena Heuermann, Leticia Lago, Lilly McNealus, Livio Baldini Soares, Logan Kilpatrick, Lucas Dixon, Luciano Martins, Machel Reid, Manvinder Singh, Mark Iverson, Martin Gorner, Mat Velloso, Mateo Wirth, Matt Davidow, Matt Miller, Matthew Rahtz, Matthew Watson, Meg Risdal, Mehran Kazemi, Michael Moynihan, Ming Zhang, Minsuk Kahng, Minwoo Park, Mofi Rahman, Mohit Khatwani, Natalie Dao, Nenshad Bardoliwalla, Nesh Devanathan, Neta Dumai, Nilay Chauhan, Oscar Wahltinez, Pankil Botarda, Parker Barnes, Paul Barham, Paul Michel, Pengchong Jin, Petko Georgiev, Phil Culliton, Pradeep Kuppala, Ramona Comanescu, Ramona Merhej, Reena Jana, Reza Ardeshir Rokni, Rishabh Agarwal, Ryan Mullins, Samaneh Saadat, Sara McCarthy, Sarah Cogan, Sarah Perrin, Sébastien M. R. Arnold, Sebastian Krause, Shengyang Dai, Shruti Garg, Shruti Sheth, Sue Ronstrom, Susan Chan, Timothy Jordan, Ting Yu, Tom Eccles, Tom Hennigan, Tomas Kocisky, Tulsee Doshi, Vihan Jain, Vikas Yadav, Vilobh Meshram, Vishal Dharmadhikari, Warren Barkley, Wei Wei, Wenming Ye, Woohyun Han, Woosuk Kwon, Xiang Xu, Zhe Shen, Zhitao Gong, Zichuan Wei, Victor Cotruta, Phoebe Kirk, Anand Rao, Minh Giang, Ludovic Peran, Tris Warkentin, Eli Collins, Joelle Barral, Zoubin Ghahramani, Raia Hadsell, D. Sculley Jeanine Banks Anca Dragan Slav Petrov Oriol Vinyls Jeff Dean Demis Hassabis Koray Kavukcuoglu Clement Farabet Elena Buchatskaya Sebastian Borgeaud Noah Fiedel Armand Joulin Kathleen Kenealy Robert Dadashi and Alek Andreev. 2024. Gemma 2: Improving open language models at a practical size. Preprint arXiv:2408.00118.  
+Richard J Roberts. 2001. Pubmed central: The genbank of the published literature.  
+Dimitrios Roussis and Vassilis Papavassiliou. 2022. The ARC-NKUA submission for the English-Ukrainian general machine translation shared task at WMT22. In Proceedings of the Seventh Conference on Machine Translation (WMT), pages 358-365, Abu Dhabi, United Arab Emirates (Hybrid). Association for Computational Linguistics.  
+Dimitrios Roussis, Vassilis Papavassiliou, Prokopis Prokopidis, Stelios Piperidis, and Vassilis Katsouros.
+
+2022a. SciPar: A collection of parallel corpora from scientific abstracts. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 2652-2657, Marseille, France. European Language Resources Association.  
+Dimitrios Roussis, Vassilis Papavassiliou, Sokratis Sofianopoulos, Prokopis Prokopidis, and Stelios Piperidis. 2022b. Constructing parallel corpora from COVID-19 news using MediSys metadata. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 1068–1072, Marseille, France. European Language Resources Association.  
+Dimitris Roussis, Sokratis Sofianopoulos, and Stelios Piperidis. 2024. Enhancing scientific discourse: Machine translation for the scientific domain. In Proceedings of the 25th Annual Conference of the European Association for Machine Translation (Volume 1), pages 275-285.  
+Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. 2021. Winogrande: An adversarial winograd schema challenge at scale. Communications of the ACM, 64(9):99-106.  
+VICTOR M. Sánchez-Cartagena, Marta Bañón, Sergio Ortiz-Rojas, and Gema Ramírez-Sánchez. 2018. Promptsit's submission to WMT 2018 Parallel Corpus Filtering shared task. In Proceedings of the Third Conference on Machine Translation, Volume 2: Shared Task Papers, Brussels, Belgium. Association for Computational Linguistics.  
+Liping Tang, Nikhil Ranjan, Omkar Pangarkar, Xuezhi Liang, Zhen Wang, Li An, Bhaskar Rao, Linghao Jin, Huijuan Wang, Zhoujun Cheng, Suqi Sun, Cun Mu, Victor Miller, Xuezhe Ma, Yue Peng, Zhengzhong Liu, and Eric P. Xing. 2024. Txt360: A top-quality llm pre-training dataset requires the perfect blend.  
+Jörg Tiedemann. 2012. Parallel data, tools and interfaces in OPUS. In Proceedings of the Eighth International Conference on Language Resources and Evaluation (LREC'12), pages 2214-2218, Istanbul, Turkey. European Language Resources Association (ELRA).  
+Yury Tokpanov, Beren Millidge, Paolo Glorioso, Jonathan Pilault, Adam Ibrahim, James Whittington, and Quentin Anthony. 2024. Zyda: A 1.3 T Dataset for Open Language Modeling. arXiv preprint arXiv:2406.01981.  
+Santosh Tyss, Rashid Haddad, and Matthias Grabmair. 2024. Ecthr-pcr: A dataset for precedent understanding and prior case retrieval in the european court of human rights. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 5473-5483.  
+Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2023. Attention is all you need. Preprint, arXiv:1706.03762.
+
+Leon Voukoutis, Dimitris Roussis, Georgios Paraskevopoulos, Sokratis Sofianopoulos, Prokopis Prokopidis, Vassilis Papavasileiou, Athanasios Katsamanis, Stelios Piperidis, and Vassilis Katsouros. 2024. Meltemi: The first open large language model for greek. Preprint, arXiv:2407.20743.  
+Laura Weidinger, Jonathan Uesato, Maribeth Rauh, Conor Griffin, Po-Sen Huang, John Mellor, Amelia Glaese, Myra Cheng, Borja Balle, Atoosa Kasirzadeh, Courtney Biles, Sasha Brown, Zac Kenton, Will Hawkins, Tom Stepleton, Abeba Birhane, Lisa Anne Hendricks, Laura Rimell, William Isaac, Julia Haas, Sean Legassick, Geoffrey Irving, and Jason Gabriel. 2022. Taxonomy of risks posed by language models. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, FAccT '22, page 214-229, New York, NY, USA. Association for Computing Machinery.  
+Zhangchen Xu, Fengqing Jiang, Luyao Niu, Yuntian Deng, Radha Poovendran, Yejin Choi, and Bill Yuchen Lin. 2024. Magpie: Alignment data synthesis from scratch by prompting aligned llms with nothing. Preprint, arXiv:2406.08464.  
+Jaume Zaragoza-Bernabeu, Gema Ramírez-Sánchez, Marta Bañón, and Sergio Ortiz Rojas. 2022. Bicleaner AI: Bicleaner goes neural. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 824-831, Marseille, France. European Language Resources Association.  
+Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. 2019. HellaSwag: Can a machine really finish your sentence? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4791-4800, Florence, Italy. Association for Computational Linguistics.  
+Yifan Zhang, Yifan Luo, Yang Yuan, and Andrew Chi-Chih Yao. 2024. Automathtext: Autonomous data selection with language models for mathematical texts. arXiv preprint arXiv:2402.07625.  
+Yiran Zhao, Wenxuan Zhang, Guizhen Chen, Kenji Kawaguchi, and Lidong Bing. 2024. How do large language models handle multilingualism? arXiv preprint arXiv:2402.18815.  
+Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. 2023. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595-46623.  
+Jeffrey Zhou, Tianjian Lu, Swaroop Mishra, Siddhartha Brahma, Sujoy Basu, Yi Luan, Denny Zhou, and Le Hou. 2023. Instruction-following evaluation for large language models. arXiv preprint arXiv:2311.07911.
+
+# A Pretrained data mix
+
+The pretraining data mix (Table 1) contains the following four components:
+
+1. Greek Texts (56.7B tokens): The Greek part of the dataset was sourced from publicly available resources spanning a wide range of domains and sources: Wikipedia, ELRC-SHARE (Lösch et al., 2021), EUR-LEX & MultiEUR-LEX (Chalkidis et al., 2019, 2021), MaCoCu (Banón et al., 2022), CLARIN-EL (Gavriilidou et al., 2023), EMEA<sup>9</sup>, parliamentary proceedings (Erjavec et al., 2022), <sup>10</sup>, governmental and legal documents from the Greek Government Gazette via the National Printing House<sup>11</sup>, the Permanent Greek Legislation Code – Raptarchis dataset<sup>12</sup> (Papaloukas et al., 2021), Greek School Books<sup>13</sup>, the Kallipos initiative of Greek open academic textbooks<sup>14</sup>, full texts from publicly available articles, theses, and dissertations from academic repositories and the National Documentation Center<sup>15</sup>, as well as pre-filtered resources originally compiled from the web, such as CulturaX (Nguyen et al., 2023) and CulturaY<sup>16</sup>. In addition to Modern Greek, we incorporated a significant amount of Ancient Greek texts into our training corpus from Wikisource, school books, web pages, and Project Gutenberg<sup>17</sup>, which provides freely available Ancient Greek texts, including classical literature and historical documents. By including Ancient Greek data, we ensured that Llama-Krikri-8B is able process polytonic Greek and engage with historical texts effectively. This enhances the model's utility for classical studies, historical research, and philological applications.
+
+2. English Texts (21B tokens): A subset of high-quality English data was mixed into the training corpus. This subset ensures that the model is continually trained on English data,
+
+$^{9}$ https://www.ema.europa.eu/  
+10https://www.gutenberg.org/  
+11https://et.gr/  
+$^{12}$ https://huggingface.co/datasets/AI-team-UoA/greek/legal_code  
+<sup>13</sup>https://books.edu.gr/ebooks/  
+$^{14}$ https://kallipos.gr/en/homepage/  
+15https://www.ekt.gr/en  
+$^{16}$ https://huggingface.co/datasets/ontocord/CulturaY  
+17https://www.gutenberg.org/
+
+and is drawn from sources that were also used for the Greek data, such as Wikipedia, Wikisource, Project Gutenberg (post-1900), EUR-LEX, EMEA, Greek academic repositories, etc. We also utilized additional English texts originating from abstracts and full texts of academic records found on multiple scientific repositories (Roussis et al., 2022a, 2024), ECtHR-PCR (Tyss et al., 2024), and pre-filtered datasets from TxD360 (Tang et al., 2024), like ArXiv, S2ORC (Lo et al., 2020), and PubMed Central (Roberts, 2001). By incorporating diverse and high-quality English texts, we mitigate the risk of catastrophic forgetting.
+
+3. Parallel Data (5.5B tokens): We compiled a diverse parallel corpus with language pairs covering multiple languages: to Greek, English, French, Portuguese, German, Spanish, and Italian. The decision to add parallel data which covers other European languages (i.e., German, French, Italian, Portuguese, and Spanish) is informed from the languages that have been included in the multilingual instruction tuning of the original Llama-3.1. We utilized resources such as SciPar (Roussis et al., 2022a), MediSys (Roussis et al., 2022b), MultiEUR-LEX (Chalkidis et al., 2021), Europarl, TED Talk transcripts (Qi et al., 2018), and other sources with sentence pairs such as ELRC-SHARE (Lösch et al., 2021) & OPUS (Tiedemann, 2012). Our data include parallel documents and sentence pairs randomly sampled for each translation direction, e.g., EN-EL/EL-EN and EN-DE/DE-EN, as well as augmented training examples with concatenated parallel content across multiple languages (e.g., a Greek text followed by its English, German, and Spanish translations with appropriate prompt templates). The addition of these documents has a twofold effect. It has been shown that parallel data boosts translation performance (Alves et al., 2024; Martins et al., 2024), while limited empirical evidence indicates that pretrained LLMs process multilingual queries by first translating the content into English, utilizing their English knowledge to answer the query and then translate the answer back to the original language(Zhao et al., 2024).
+
+4. Code and Math (7.8B tokens): We also integrated datasets containing text with code and mathematics, leveraging Stack Overflow $^{18}$ , Python-Edu which is a subset of the SmolLM corpus (Allal et al., 2025) originating from The Stack V2 dataset (Lozhkov et al., 2024) and having been scored with an educational code classifier, and the AutoMathText dataset (Zhang et al., 2024), which is a collection of math-related documents originating from web data, papers on arXiv, and code/notebooks on GitHub. AutoMathText has undergone an automatic selection process using Qwen-72B (Bai et al., 2023) for relevancy to the mathematical domain and the educational value of each document. Code and Mathematics data, although not specific to Greek, were included to preserve and enhance the model's ability to handle coding tasks, math problems and formal language. Maintaining these capabilities broadens the utility of Llama-Krikri beyond pure language tasks.
+
+# B Pretraining data cleaning pipelines
+
+Our filtering processes began with format standardization in order to facilitate uniform processing across multiple heterogeneous datasets. We converted all textual content from various formats (e.g., PDF, HTML, plain text, etc.) into JSONL containing both the document text and relevant metadata such as identified language, word count, and source information (including source URLs).
+
+For PDF documents such as academic records and laws, we implemented a specialized pipeline which integrated Marker $^{19}$  for extraction and conversion into Markdown files, as it exhibits strong performance for Greek texts. Subsequently, the pipeline included language identification using FastText (Joulin et al., 2017), removal of markdown artifacts, and removal of lines with characters outside Unicode ranges for Greek, Latin, and other common and scientific symbols. Furthermore, we utilized document structure metrics (Marker also extracts various structural metadata) as quality indicators, such as the ratio of tables to pages and the fraction of removed lines in disallowed scripts.
+
+Our main filtering pipeline used sequential rule-based and statistical filters to remove outlier documents across all data sources. First, we imple
+
+mented URL-based filtering by removing content from several blacklisted domains known to contain low-quality or problematic content. This was particularly effective for web-crawled datasets like CulturaX (Nguyen et al., 2023) where source metadata was available. We then applied a set of minimal content-quality filters:
+
+- Removal of documents containing multiple instances of profane or inappropriate terms from a curated list of Greek bad words  
+- Removal of short documents based on character and word counts  
+- Removal of documents containing multiple substrings like "lorem ipsum" which are indicative of content with low educational value  
+- Removal of documents containing extremely long words (>60 characters)  
+- Removal of documents with mean word length outside specified values.  
+- Removal of documents with a high fraction of non-alphanumeric characters.
+
+Parallel datasets were filtered using a different pipeline featuring various steps from previous work (Papavassiliou et al., 2018; Roussis and Papavassiliou, 2022; Roussis et al., 2024) which include: (a) rule-based filters, such as length ratio, language identification verification, and (b) model-based alignment quality scores using tools like LASER (Artetxe and Schwenk, 2018, 2019), BiCleaner AI (Zaragoza-Bernabeu et al., 2022), and CometKiwi (Rei et al., 2022).
+
+Additionally, in order to mitigate privacy concerns and protect sensitive information, we systematically identified and anonymized personally identifiable information (PII) with the use of regular expressions. In particular, we aimed to detect and replace e-mail addresses with a generic placeholder ("email@example.gr") and mask IP addresses (replacing them with 0.0.0.0).
+
+Finally, for Greek, English, and Mathematics/Code datasets we implemented intra-dataset dedduplication, as well as cross-dataset dedduplication. We utilized MinHashLSH near-deduplication (Broder, 1997; Leskovec et al., 2020) with 5-gram subsets, a MinHash signature of 128, and a Jaccard similarity threshold of 0.8, following parameter choices similar to those used in other works
+
+(Nguyen et al., 2023; Voukoutis et al., 2024). Regarding the dedduplication of parallel datasets, we followed a different approach. All sentence pairs were normalized and cleaned, by converting them to lowercase and removing digits, punctuation. Pairs were then deduplicated based on the existence of either the source or target within the same dataset, thus ensuring that no sentence can be found multiple times in each parallel dataset (Roussis and Papavassiliou, 2022; Roussis et al., 2024). It should be noted that deduplication has consistently been shown to lead to higher performance, reduced training costs, as well as reduced model memorization; thus indirectly protecting sensitive information (Lee et al., 2022; Carlini et al., 2023; Grattafori et al., 2024; Albalak et al., 2024). However, as we mentioned earlier, global dedduplication may also remove documents of high quality and actually hurt performance (Tang et al., 2024; Penedo et al., 2024). For this reason, we decided to up-sample datasets of specific sources with important content. Table 1 summarizes the composition of the filtered and deduplicated pretraining corpus. In total, our collected dataset comprises roughly 91B tokens, of which  $62.3\%$  is Greek text. For the final training curriculum, we upsampled parts of the corpus to effectively train on an equivalent of 110B tokens. Upsampling was used to give higher relative importance to certain underrepresented but valuable segments and it also leads to higher memorization of important content (Carlini et al., 2023; Tang et al., 2024). For example, we assigned a slightly higher weight to datasets with long-context documents, Wikipedia-like sources, dialogue data, multi-parallel documents, and to certain important domains, such as legal, scientific, and medical. The decision to include a significant amount of English and parallel data ( $23.1\%$  and  $6\%$  of tokens, respectively) was guided by prior work (Voukoutis et al., 2024) showing that mixed-language training can help retain the base model's general knowledge and prevent catastrophic forgetting.
+
+# C Tokenizer and embeddings expansion process
+
+The tokenizer and embeddings expansion process involved the following steps:
+
+- Data Acquisition: We acquired data by collecting sentences from high-quality sources of our pretraining mix in five domains:
+
+1. General domain which reuses a sampled portion of the data used to train the tokenizer of Meltemi (Voukoutis et al., 2024) and covers diverse domains,  
+2. Legal domain which uses legal texts extracted from the Greek Government Gazette and are available via the National Printing House $^{20}$ , as well as the Permanent Greek Legislation Code - Raptarchis dataset $^{21}$  (Papaloukas et al., 2021),  
+3. Scientific domain which uses publicly available articles, theses, and dissertations found in the National Documentation Center22,  
+4. Literature domain from public-domain books from Wikisource $^{23}$  and Project Gutenberg $^{24}$  which contain literature, poetry, and other original writings across various variants of Greek (e.g., Koine Greek, Medieval Greek, Modern Greek, etc.),  
+5. Ancient Greek which contains texts only in Ancient Greek sourced from Greek school books and various publicly available corpora.
+
+- Filtering and Preprocessing: Each dataset underwent sequential processing and filtering including language identification verification with FastText (Joulin et al., 2017), application of regular expressions to remove URLs and other anomalies, symbol-to-word ratio filtering to remove outliers, and NFC normalization. We then performed sentence-level exact dedduplication within each individual dataset. To ensure text quality, we applied fluency scoring using Monocleaner (Sánchez-Cartagena et al., 2018) which leverages a 7-gram KenLM model for Greek, and setting a score threshold of 0.3 for non-polytonic text.  
+- Creation of Training and Test Sets: For the tokenizer training and test sets creation, we sampled  $50\%$  of the sentences from each source and divided it into training and testing splits  $(80\% - 20\%)$ .
+
+<table><tr><td>Domain</td><td>Added Tokens</td><td>Llama-3.1-8B</td><td>Llama-Krikri-8B</td><td>Δ Fertility</td></tr><tr><td>General</td><td>15,000</td><td>2.65</td><td>1.59</td><td>-1.06</td></tr><tr><td>Legal</td><td>4,000</td><td>2.82</td><td>1.54</td><td>-1.28</td></tr><tr><td>Scientific</td><td>1,000</td><td>2.91</td><td>1.73</td><td>-1.18</td></tr><tr><td>Literature</td><td>500</td><td>2.90</td><td>1.89</td><td>-1.01</td></tr><tr><td>Ancient Greek</td><td>492</td><td>3.77</td><td>2.15</td><td>-1.62</td></tr><tr><td>Total</td><td>20,992</td><td>-</td><td>-</td><td>-</td></tr></table>
+
+Table 9: Domain-specific token allocation and fertility comparison
+
+- Domain-specific Token Allocation: New tokens were added sequentially for each domain until tokenizer fertility for this domain remained relatively stable, with most of the tokens being allocated to the General domain. This approach ensured that common Modern Greek patterns receive the largest coverage, while specialized terminology and older Greek variants are adequately represented.
+
+Table 2 presents the fertility of several tokenizers on the original Greek and English corpora. Note that these test sets have not been used anywhere in the domain-specific multi-stage training process and could be considered as out-of-domain for the Llama-Krikri-8B tokenizer. We evaluated the Llama-3.1-8B tokenizer (128,000 vocabulary size) and our custom-trained Llama-Krikri-8B tokenizer (149,248 vocabulary size). We observe that the Llama-3.1-8B tokenizer exhibits a fertility of 2.73 for Greek and 1.33 for English. Our Llama-Krikri-8B tokenizer demonstrates a significantly lower fertility of 1.65 for Greek, while maintaining the same low fertility of 1.33 for English as the base Llama-3.1-8B tokenizer. This indicates that our Llama-Krikri-8B tokenizer is more efficient for Greek texts compared to the standard Llama-3.1-8B tokenizer. The table also includes the Mistral and the Meltemi-7B Greek tokenizer for comparison.
+
+Furthermore, in Table 9, we list the token allocation per domain, as well as the tokenizer fertilities of Llama-3.1-8B and Llama-Krikri-8B for each of the five domains for which we created test sets during the tokenizer extension process. We can observe that tokenizer fertility has dropped significantly across all domains, with Ancient Greek benefiting from the highest reduction (-1.62) with just 492 added tokens, whereas fertility for Legal texts has almost halved (1.54 from 2.82) after the integration of a significant number of tokens (4,000).
+
+Following the development of the Llama-Krikri-8B tokenizer with 149,248 tokens, containing the initial Llama 3 reserved tokens to reach a vocab
+
+ulary size divisible by 256 for computational efficiency, we expanded the embeddings and lm head layer of our Llama 3 based model to match this vocabulary size.
+
+The newly initialized embeddings and lm head weights for each new token were initialized by averaging the embeddings (or lm head weights respectively) associated with the tokens retrieved by the initial Llama-3.1 tokenizer when fed with that token. This provided us with a smoother initial Greek embeddings and lm head training phase of the frozen model.
+
+# Ex. 1: Programming: Python Fibonacci function
+
+User:  $\Gamma \rho \alpha \psi \epsilon \mu \iota \alpha \upsilon \nu \acute{\alpha} \rho \tau \eta \eta$  Python  $\pi \sigma \upsilon \upsilon \omega \lambda \iota \zeta \iota \tau \eta \upsilon \alpha \kappa \omega \upsilon \iota \alpha$  fibonacci.
+
+Krikri: Axolouvei ng olooioon ng avapomixc ouvaptnong Python vaoov nouoyouo twv apwov Fibonacci:
+
+```python
+def fibonacci(n):
+    "" $\Upsilon \pi o \lambda o \gamma i \zeta e \iota$  tov n-oστó αριθμό Fibonacci.
+Args:
+    n (int): θειχός αχέραος. n >= 0 (0-indexed προδέγιση)
+Returns:
+    int: o n-oστός αριθμός Fibonacci
+Raises:
+    ValueError: αν n είναι αρνητιχός αριθμός
+    ""# 'Eλεγχος εγχύρότητας εἰσόδου
+if not isinstance(n, int) or n < 0:
+    raise ValueError("H είσδος πρέπεινα είναι μη αρνητιχός αχέραος")
+# Bαση αυδρομής:
+# F(0) = 0, F(1) = 1 (xλασιχοί σρισμοί)
+if n in (0, 1):
+    return n
+# Avadopouxós τύπος: F(n) = F(n-1) + F(n-2)
+return fibonacci(n-1) + fibonacci(n-2)
+Παράδειγμα χρήσης:
+# ηπολογισός μεριχών ὁρων τής
+# αχολουδίας Fibonacci
+for i in range(10):
+    print(f"F({i}) = {fibonacci(i)}")
+```
+
+# Ex. 2: Translation: Ancient to Modern Greek
+
+User: Mετάφρασε μου στα Νεα ελληνιχα το xειμενο:
+
+'Oi  $\mu \mathrm{ev}$  n Mhloi t ooaut a anxpivavto o de Aynvaiou diauouevoi n ex twyov epaoov «oov mov y e an tovtw vov boulueuatow, os nuov doxite, ta uev meaovta wov opomegaov oxapeotepa xpvte, ta de apovn tw ooueov c yuvovn dn veodex, xai Axedauovioic xai tyx n xai elpiotov dn npapaeBnaevov xai nioteuov xai opaaonoeve.
+
+Krikri: Axoiov et  $\eta$  etaappoan tou xeuévou oα Nεα Eληνixá:
+
+Oi Mhioi anavtnoov me auta ta loyia. Oi Ahoiaoi, xaoos aoxwpouoav ano tn ouhtnoj, eianv: Oomega eie o i moi nou ano autc ts ouzntneis ouunepaive t oot a to opata ooo xu ta apopata, oov exov hon oumuei. Eote uepolixia aoiodoqou xau exete evanovdeoei ts elniides os otous Axedauovic, stny tuxn xui ts npoofoies oc. Kau oo nepioootepo exete niotetse ie auta, tooo nepioootepe 0 anoyonteute.
+
+# Ex. 3: Translation to JSON Output: Greek to Multiple Languages
+
+User: Metáφρασέ μου το παρακάω κειμενο από τα νέα ελληνία στα αγγλία, στα οπανία xα στα γαλλία. Δώσε μου τις μεταφρασεις σε JSON.
+
+Σμρωνα με την σθηγια 2009/28/EK του Eupωπaixou Koivοβouliou, ως ευργεια από ανανεώμες μη έριχτές πηγες Σωρείαι Αλιλχή, Μλιαχή, εροθερμχή, γεωθερμχή, ὅροθερμχή κα ευργεια των ωκενύν, δύροιλεxtριχή, από βιομάζα, από τα ελλύμενα στούς χώρους Φειλονομιχής ταφής αέρια, α��ό αέρια μονάων επεξεργασιας λυματων κα από βιομερία.
+
+# Krikri:
+
+{
+
+"en": "According to Directive 2009/28/EC of the European Parliament, energy from renewable non-fossil sources includes wind, solar, aerothermal, geothermal, hydrothermal and ocean energy, hydropower, biomass, landfill gas, sewage treatment plant gas and biogases.",  
+"es": "De conformidad con la Directiva 2009/28/CE del Parlamento Europeo, la energia procedente de fuentes renovables no fosiles incluye la eólica, solar, aerotérmica, geotérmica, hidrotérmica y la energia del océano, hidroeléctrica, procedente de la biomasa, de los gases de vertedero, de los gases de estación depuradora de aguas residuales y de los biogás.",  
+"fr": "Conformément à la directive 2009/28/CE du Parlement européen, les énergies renouvelables non fossiles comprément l'éolien, le solaire, l'aérothermique, la géothermie, l'hydrothermie et l'énergie océanique, l'hydroélectricité, la biomasse, les gaz de décharge, les gaz de station d'épuration et les biogaz."
+
+# Ex. 4: Summarization: Basketball game
+
+User: Mnpoeic vα δημουργησεις μια περίληψη 5 ποτάσων για τον παρακατω αγώνα μπάσχει: (User provides the text of an article describing a 2025 basketball game)
+
+Krikri: O Oluuuaoxoc vixnoe tyn Mnpotoleova e 90-88 oe eva ouvapnatoxio naxydi, xapn o evaxvntpoia xaladou Phoupvie stny exvon. O Bezevxov aoi o Ixos onueiow av ano 24 xai 20 rovtoucavtiotaia touc vixntec, evw o Metou nay o xopupaoic twv Kataaavw y 16 rovtouc. H ouada touMnpctzox npapeve stny xopupn tnc Eupwlirxus meta ano autn onmuavtxn vixn.
+
+# Ex. 5: History: Population exchange between Greece and Turkey
+
+User: Evóψiε Εια παράραφο την Ελληνοτουρχίκ ανταλλαγπληθυσμών του 1923
+
+Krikri: H Eλληνοτουρχική ανταλλγή πληνθυσμών τοῦ 1923, που προβλέρθηκε αότη Συνθήπης Λωζάνης, αποτέλεσε μία μαζιχή μεταλίνηση πληνθυσμών μεταξύ Eλλάδα και Touρχίας με στόχο τήν Μογονενοποίŋη των ευνών-хρατών. Περίπου 1,5 exατομύριον ἀλληνες χριστιαύνι, xύριως αότη Συναλήθ Θράŋ, τήν Iωία, τοῦ το και τήν Kαπαδόχία, μεταχήνθηκαν στήν Eλλάδα, εύν αντίσοιχα 500.000 Mousούλáνοι (xύριως Touρχία σλλά ααι Αρόχοι και Ανίγγαύοι) μεταχήνθηκαν αότη Συναλά δστην Touρχία. Αντή συναλάγή, πάρ τι τραγίκες τής συνέπεις γία τους αὐφώπους που εκτόπιστηκαν, αποτέλεσε ευα πό τα σημαντίσερα γεγούτα που διαύρερωσαν τή συγχρονη δημοραρχίακη και ευνοτιχή συνθόση τήν δύο χρών, δόγδόντας σε μία πο μομοίγενη ευνίκη ταυτόπιτα γία κάθε υράος. H διαδύχασια αὐτή σλοχληρωύθηκε το 1924 και θεωρεῖτοι σρόσημο στέληνονοτουρχίες σχόσεις.
+
+# Ex. 6: High school biology: The process of photosynthesis
+
+User:  $\Delta$  ooe uovosn ts diabuxoas ts wtoovdoes oe ia npaparpo.
+
+Krikri: H  $\wp\omega\tau\omega\upsilon\upsilon\sigma\eta$  evou n diadixaia me tyn oioia ta pvta, xpnouonowvacs nyn lwxn evpyea, metatpeneuv to bioeido tou avpaxo (O2) xau to vepo (H2O) e yluxoicn (^6H12O6) xau oxiyovo (O2). Luyxexpueva, n xawopuuln st a pfllaa anoppov a wsc, dioana to vepo anelevepvovotac osvyovo, xau xnpouonowvts to udpoyovo ouvvertei yuxoicn ano to bioeido tou avpaxa. H npayoyne yluxoicn anotele i npny n evpeyias ria to pvto, evw to osvyovo anelevepvvetao stny atmuospaipw os nmavtio npapapoiov, unostnpicovtac tn vpiow ton vepoibw opaynuuow.
+
+# Ex. 7: Chain-of-Thought Reasoning: Math problem
+
+System: (Appropriate system prompt used to enable thinking mode)
+
+User: H Aλixη exει ενα καλάθι με μήλα. Δίνει τα 2/5 των μήλων στην αδερή της και το 1/5 στόν αδερύ της.  
+Teλικα της μένουν 6 μήλα. Πόσα μήλα εἰχε αρχία στο καλάθι;
+
+Krikri:
+
+<thinking>
+
+As  $\lambda \omega \sigma \omega \mu \varepsilon$  auto to  $\pi \rho \beta \lambda \eta \mu \alpha \beta \eta \mu \alpha \pi \rho \varsigma \beta \eta \mu \alpha$
+
+1. EoTw x o apxixoc aipuouc mJawn ou eixe n Alixn
+
+2. Γνωρίζουμε ὄτι:
+
+* Δινει 2/5 του × στην αὐδερή της  
+* Δινει 1/5 του x στον αδερφό της  
+* Metα από δλα αιτα της μένουν 6 μήλα
+
+3. Ac xavouue trv eEioomega:
+
+* Apxixα Εiε: x μnλ  
+* Tns μένουν: 6 μήλα  
+$\star \mathrm{A}\rho \alpha :x - (2 / 5)x - (1 / 5)x = 6$  
+$\star \mathrm{A}\pi \lambda \sigma \omega \tau \omega \nu \tau \alpha \varsigma :x - (3 / 5)x = 6$  
+$\star \Sigma \upsilon \delta \upsilon \alpha \zeta \upsilon \tau \varsigma \text{óp o u c:} (2 / 5)x = 6$  
+$\star \Pi o\lambda \pi \lambda \alpha \sigma i\zeta o\nu \tau \varsigma x\alpha \tau \alpha \delta \nu \mu \varepsilon \lambda \eta \mu \varepsilon 5 / 2$  
+$\star x = 15$
+
+As ean0ooue:
+
+\*Apxixαεiε15μnλα  
+* Δίνει: 2/5 * 15 = 6 μήλα στην αὐδερή της  
+* Δíveι: 1/5 * 15 = 3 μήλα στον αδερύ της  
+* Mévoyov: 15 - 6 - 3 = 6 μήλα  
+H  $\lambda \nu \sigma \eta$  eivua owoth!  
+</thinking>
+
+<output>
+
+H Aixn eixe apxixa 15 mλσoTo xαλdtnc
+
+Mnpouue va to eaiytheooue:  
+* Eδωσε 6 μήλα στην αδερρή της (2/5 του συύλου)  
+* Eδωσε 3 μήλα στον αδερρό της (1/5 του συνόλου)  
+* Étσι της ἔμειναν 6 μήλα (15 - 6 - 3 = 6)  
+</output>

2025/Krikri_ Advancing Open Large Language Models for Greek/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:98849cd0ad36ec03e67047cb946c5b26594fd1313c2c4fe2049846912d5e265d
+size 283515

2025/KurTail _ Kurtosis-based LLM Quantization/3649cf39-4457-4adc-be41-8f76450f5b9e_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:88ac03715d61e01af4e66bef6932f35b51a8af9247885ddfb0e8dfe5a67911e9
+size 546761

2025/KurTail _ Kurtosis-based LLM Quantization/full.md

@@ -0,0 +1,407 @@
+# KurtTail : Kurtosis-based LLM Quantization
+
+Sadegh Akhondzadeh* 1 Aleksandar Bojchevski<sup>1</sup> Evangelos Eleftheriou<sup>2</sup> Martino Dazzi<sup>2</sup>
+
+<sup>1</sup> University of Cologne, <sup>2</sup> Axelera AI
+
+[akhondzadeh, bojchevski]@cs.uni-koeln.de
+
+[evangelos.eleftheriou, martino.dazzi]@axelera.ai
+
+# Abstract
+
+One challenge of quantizing a large language model (LLM) is the presence of outliers. Outliers often make uniform quantization schemes less effective, particularly in extreme cases such as 4-bit quantization. We introduce KurTail, a new post-training quantization (PTQ) scheme that leverages Kurtosis-based rotation to mitigate outliers in the activations of LLMs. Our method optimizes Kurtosis as a measure of tailedness. This approach enables the quantization of weights, activations, and the KV cache in 4 bits. We utilize layer-wise optimization, ensuring memory efficiency. KurTail outperforms existing quantization methods, offering a  $13.3\%$  boost in MMLU accuracy and a  $15.5\%$  boost in Wiki perplexity compared to QuaRot (Ashkboos et al., 2024b). It also outperforms SpinQuant (Liu et al., 2024) with a  $2.6\%$  MMLU gain and reduces perplexity by  $2.9\%$ , all while reducing the training cost. For comparison, learning the rotation using SpinQuant for Llama3-70B requires at least four NVIDIA H100 80GB GPUs, whereas our method requires only a single GPU, making it more accessible.
+
+# 1 Introduction
+
+Large language models (LLMs) have advanced significantly in recent years, showcasing remarkable performance and capabilities. As these models grow in size and complexity, the computational cost required for their deployment and inference has increased dramatically. Furthermore, with new inference time methods (OpenAI, 2024; Guo et al., 2025), enhancing inference speed (tokens per second) is increasingly important. This has shifted the focus toward accelerating model performance while reducing memory and computational requirements. An effective method to achieve this is posttraining quantization (PTQ), which involves representing model weights and/or activations in lower
+
+numerical precisions. PTQ can significantly reduce the memory footprint and computational overhead and subsequently decrease latency and energy consumption, which are especially beneficial for inference on resource-constrained edge devices.
+
+Serving a model involves two stages of prefilling and generation. During prefilling, the model processes the input prompt and stores the internal state, known as key-value (KV) caching. During generation, tokens are produced auto-regressively. Quantizing each stage offers distinct advantages for improving inference efficiency. KV-cache quantization reduces memory requirements and accelerates data movement, which enhances the generation stage, particularly in scenarios involving long-context inference. Weight quantization, on the other hand, reduces the memory footprint independently, and when it is combined with activation quantization, it also reduces the computational demands. However, activation quantization presents challenges due to large outliers in certain channels (Dettmers et al., 2022; Xiao et al., 2023), which limits the effectiveness of uniform integer quantization as it destroys the dynamic range of the activations. While channel-wise quantization can effectively address this issue, the lack of hardware support makes it computationally expensive in practice. Several methods have been proposed to address this challenge. Dettmers et al. (2022) and Ashkboos et al. (2023) advocate for mixed-precision computation in which they store some of the channels in higher precision and less sensitive channels in lower precision to balance accuracy and efficiency. Xiao et al. (2023) introduces channel-wise scaling into the layer normalization and the weights of linear layers. Ashkboos et al. (2024b) proposed random rotation which takes the advantage of the computational invariance framework (Ashkboos et al., 2024a) to mitigate the outliers problem. We introduce KurTail - a novel approach to mitigating activation outliers by apply
+
+ing learnable rotations<sup>1</sup> to the activations similar to SpinQuant (Liu et al., 2024). KurTail focuses on reducing the tail density of activations, captured by the Kurtosis. Unlike SpinQuant which requires expensive end-to-end training of the model's loss, we prove that layer-wise optimization of our Kurtosis loss is equivalent to end-to-end training. We perform layer-wise inference to cache activations, and then optimize the rotations based on the cache independently. As a result, KurTail can be implemented in a significantly more memory-efficient manner. For instance, while SpinQuant requires at least four NVIDIA H100 80GB GPUs to compute rotations for Llama3-70B, KurTail achieves the same with just a single GPU. Despite its lower computational requirements, KurTail outperforms existing methods in terms of perplexity and zero-shot reasoning tasks. KurTail outperforms existing quantization methods with a  $13.3\%$  increase in MMLU accuracy and a  $15.5\%$  decrease in Wiki perplexity compared to QuaRot(Ashkboos et al., 2024b). It also performs better than SpinQuant(Liu et al., 2024), achieving a  $2.6\%$  increase in MMLU accuracy and a  $2.9\%$  decrease in perplexity, all while reducing the cost of training the rotation. We also theoretically shed light on why rotations are preferable to arbitrary linear transformations.
+
+# 2 Background
+
+Post Training Quantization. Previous work on post-training quantization fits into two main groups: weight-only quantization (Frantar et al., 2022; Lin et al., 2024b; Egiazarian et al., 2024; Tseng et al., 2024) and weight-activation quantization (Xiao et al., 2023; Dettmers et al., 2022; Ashkboos et al., 2024b; Liu et al., 2024). In weight only quantization, the weight are projected into a lower precision, such as 4 bits, 3 bits, or even less, and then de-quantized to higher precision before the actual computation, with all calculations still being done in high precision. Several studies (Xiao et al., 2023; Ashkboos et al., 2024b; Liu et al., 2024) attempted to introduce quantization methods for both weight and activation. They showed that uniform quantizing is impractical for large language models since they suffer from large outliers. To address this issue, Dettmers et al. (2022); Ashkboos et al. (2023); Saxena et al. (2024) proposed a mixed-precision approach for handling outliers at higher precision. Others (Xiao et al., 2023; Lin et al., 2024b) pro
+
+posed trading outliers between weights and activations by introducing a re-scaling paradigm. Tseng et al. (2024) introduced an incoherence processing method using random rotation matrices and applying vector quantization on the weights for compression, adding overhead to inference. QuaRot (Ashkboos et al., 2024b) was inspired by Tseng et al. (2024) and took advantage of the invariance framework proposed by Ashkboos et al. (2024a) introducing a rotation-based approach to compress and remove outliers from the activation space using a random Hadamard rotation. In a concurrent work, (Lin et al., 2024a) also introduced dequant which also combined a combination of permutation and rotation to mitigate the outliers. Later, SpinQuant (Liu et al., 2024) and OstQuant (Hu et al., 2025) improves the results of QuaRot (Ashkboos et al., 2024b) by optimizing some of these rotations to minimize the model loss through end-to-end training. While SpinQuant improves the results compared to QuaRot it suffers from a high computational cost for learning the rotations. We address this issue by introducing a novel loss for learning the rotations.
+
+Uniform Quantization for  $k$ -bit Precision. For a given vector  $\pmb{x}$ , uniform integer quantization reduces its continuous range of values to a finite set of discrete levels, enabling representation in lower precision. In  $k$ -bit quantization, the value range  $[\pmb{x}_{\mathrm{min}},\pmb{x}_{\mathrm{max}}]$  is divided into  $2^{k}$  equal intervals. Each element  $\pmb{x}_i$  in  $\pmb{x}$  is mapped to its closest quantization level by  $Q(x_{i}) = \mathrm{round}\left(\frac{\pmb{x}_{i} - b}{s}\right)\cdot s + b$ . Here  $s$  is the scale factor or step size and  $b$  is the shift. The values of  $s$  and  $b$  depend on the specific quantization scheme. In symmetric quantization, the range is assumed to be symmetric around zero. Therefore,  $b = 0$ , and  $s = \frac{\max(|\pmb{x}_{\mathrm{max}}|,|\pmb{x}_{\mathrm{min}}|)}{2^{k - 1} - 1}$ . In asymmetric quantization, the range is not assumed to be centered at zero and therefore,  $b = \min (\pmb {x}),\quad s = \frac{\pmb{x}_{\mathrm{max}} - \pmb{x}_{\mathrm{min}}}{2^k - 1}$ . Given  $\pmb{x}$  sampled from some distribution the expected mean-squared error (MSE) between the quantized and the original values is:
+
+$$
+\operatorname {M S E} (\boldsymbol {x}, Q) = \mathbb {E} \left[ (\boldsymbol {x} - Q (\boldsymbol {x})) ^ {2} \right] \tag {1}
+$$
+
+Definition 2.1. Quantization Sensitivity (Chmiel et al., 2020) For a given distribution and sample vector  $\pmb{x}$ , let  $\tilde{s}$  denote the optimal quantization step size where  $\tilde{s}$  minimizes the quantization error, and let  $Q_{\tilde{s}}(\pmb{x})$  represent the optimal quantizer. Quantization sensitivity  $\Gamma(\pmb{x}, \epsilon)$  is defined as the increase
+
+in the mean squared error (MSE) caused by a small perturbation  $\epsilon > 0$  in the quantization step size  $s$  around  $\tilde{s}$ , such that  $|s - \tilde{s}| = \epsilon$ . Specifically, the sensitivity is given by:
+
+$$
+\Gamma (\boldsymbol {x}, \epsilon) = | \operatorname {M S E} (\boldsymbol {x}, s) - \operatorname {M S E} (\boldsymbol {x}, \tilde {s}) | \tag {2}
+$$
+
+Theorem 2.2. (Chmiel et al., 2020) Considering  $\pmb{x}_U$  and  $\pmb{x}_N$  be continuous random variables with uniform and normal distributions. Then, for any given  $\varepsilon > 0$ , the quantization sensitivity  $\Gamma(\pmb{x}, \varepsilon)$  satisfies  $\Gamma(\pmb{x}_U, \varepsilon) < \Gamma(\pmb{x}_N, \varepsilon)$ .
+
+This theorem indicates that, compared to the typical normal distribution, the uniform distribution is more robust to changes in the quantization step size  $s$ . Therefore, it becomes apparent that there is great benefit in adjusting the distribution of the activations and weight to get closer to uniform distribution. This implies that the uniform distribution is a perfect fit for uniform quantization. It can also be shown that the optimal scaling  $\tilde{s}$  for the uniform distribution is equal to  $\tilde{s} = \frac{x_{\mathrm{max}} - x_{\mathrm{min}}}{2^k - 1}$ . Chmiel et al. (2020) also show that the optimal step size for a uniform distribution closely approximates the most robust quantization (least sensitive step size).
+
+Kurtosis. Kurtosis is a statistical measure that describes the degree of tailedness in the distribution of a dataset. It helps determine whether the data have heavy or light tails compared to a normal distribution. Mathematically, Kurtosis is defined as the standardized fourth moment of a population around its mean, and it is calculated using
+
+$$
+\kappa = \frac {\mathbb {E} [ (x - \mu) ^ {4} ]}{(\mathbb {E} [ (x - \mu) ^ {2} ]) ^ {2}} = \frac {\mu_ {4}}{\sigma^ {4}} \tag {3}
+$$
+
+where  $\mu$  is the mean,  $\mu_4$  is the fourth moment about the mean, and  $\sigma$  is the standard deviation. The Kurtosis of a normal distribution is 3.  $\kappa > 3$  is characterized by heavy tails and a sharp peak, indicating greater tail density than a normal distribution (e.g. the Laplacian distribution). We have a shift of mass from the shoulders to both the tails and the center. On the contrary,  $\kappa < 3$  is a sign of light tails and a flatter distribution (e.g. uniform or beta distribution) caused by mass moving from the tails and center to the shoulders. Banner et al. (2019) demonstrate that deep neural network weights and activations typically follow Gaussian or Laplace distributions. Furthermore, Dettmers et al. (2022) identifies the presence of extreme outliers in LLM parameters, which are critical for maintaining performance. Our key insight is that distributions with
+
+outliers exhibit high kurtosis, which measures the presence of extreme values. Therefore, by optimising the rotation to minimize the kurtosis we can bring the distribution closer to uniform. Uniform distribution is the desired distribution of the activations and weights for uniform quantization (§ 2), so we aim to move the distribution closer to uniform. Kurtosis serves two purposes: to encourage the distribution to resemble a uniform distribution, and to reduce the outliers. Our loss function is:
+
+$$
+\mathcal {L} _ {\kappa} = \frac {1}{L} \sum_ {i = 1} ^ {L} | \kappa \left(\bigoplus_ {j = 1} ^ {N} \boldsymbol {a} _ {i j}\right) - \kappa_ {u} | \tag {4}
+$$
+
+where  $\bigoplus$  denotes the concatenation of the activation of all tokens at that layer and  $\kappa_{u}$  is the Kurtosis of the uniform distribution.
+
+# 2.1 Optimality of orthogonal transformations
+
+There are two main reasons for using orthogonal transformations. First, when fusing the initial rotation  $R_{1}$ , an orthogonal transformation is required to maintain invariance with respect to RMSNorm (see § 3), as shown by (Ashkboos et al., 2024b). In principle some of the transformation (i.e  $R_{2}$ ) can be any full rank matrix. We show that the quantization error is upper bounded by its condition number which is minimized for orthogonal transformations.
+
+Lemma 2.3. The  $k$ -bit quantization error of  $\mathbf{X} \in \mathbb{R}^{N \times M}$  after a full rank transformation  $\mathbf{T}$  is
+
+$$
+\left\| \boldsymbol {X} - Q (\boldsymbol {X T}) \boldsymbol {T} ^ {- 1} \right\| _ {F} \leq \frac {\left\| \boldsymbol {X} \right\| _ {F}}{2 ^ {k - 1} - 1} \sqrt {N M} \cdot \operatorname {c n} (\boldsymbol {T})
+$$
+
+where  $\operatorname{cn}(\pmb{T})$  is the condition number of  $\pmb{T}$ .
+
+Corollary 2.4. The upper bound on the quantization error is minimized when  $\operatorname{cn}(\pmb{T}) = 1$  and  $\pmb{T}$  is (a scalar multiple of) an orthogonal matrix.
+
+Intuitively, the quantization error of the transformed activation is inversely related to the smallest singular value of  $T$ . To avoid amplifying the quantization error, it must not be smaller than one. An orthogonal transformation, where all singular values are equal to one, is well behaved.
+
+End-to-End training. KurTail can be run layerwise instead of end-to-end resulting in a computational benefit. We prove that end-to-end training and our layer-wise optimization converge to the same solution for certain families of models which include our current setting.
+
+Proposition 2.5. Let  $H_1, O_1 = f(X, W_1, R_1)$ , and  $H_2, O_2 = g(O_1; W_2, R_2)$  where  $f, g$  are parameterized by  $W_1, R_1$  and  $W_2, R_2$ . Given functional invariance of  $f$  and  $g$ , i.e.  $O_1 = O_1'$  for any  $O_1', H_1' = f(X, W_1, R_1')$  and any orthogonal  $R_1'$  (and similarly for  $g$ ), and given that the total loss is  $\mathcal{L}(R_1, R_2) = \mathcal{L}_1(H_1) + \mathcal{L}_2(H_2)$ , the independent minimization of each loss results in the same optimum as end-to-end: arg  $\min_{R_1, R_2} \mathcal{L}(R_1, R_2) = (\arg \min_{R_1} \mathcal{L}_1(R_1), \arg \min_{R_2} \mathcal{L}_2(f; R_2))$ , even though  $H_2$  implicitly depends on  $O_1$ .
+
+Proposition 2.5 indicates that optimizing  $R_{1}$  and  $R_{2}$  end-to-end is equivalent to optimizing each separately since our loss and the model architecture satisfy the assumptions. Inductively, this holds for all layers. However,
+
+for the output of the MHSA and the FFN blocks we jointly optimize  $\pmb{R}_{1}$  using the activations from all layers by summing them since  $\pmb{R}_{1}$  shared across layers/losses (Fig. 3).
+
+Quantization Sensitivity. We evaluate our method by measuring activation sensitivity both before and after applying rotations optimized with Kurtosis. We expect that after applying these rotations, the activation distribution will be closer to uniform, resulting in better quantization robustness. We empirically measure the sensitivity of the activation distribution before and after applying the rotation. We utilize the Llama3.1 8-B model and apply two rotation techniques: one using a random Hadamard transformation and another using a Kurtosis-optimized rotation. First, we compute the optimal scaling (Chmiel et al., 2020) for activation quantization and then calculate the quantization sensitivity based on Definition 2.1.
+
+![](images/1dc578d55326e75ddbb2073cfb49683591187b856b32e4c909398afa2ffae8f4.jpg)  
+Figure 1: Empirical sensitivity of the MHSA input distribution across different rotations.  $\alpha$  indicates the fraction of the optimal step size, i.e. sensitivity with step  $\alpha \cdot s$ .
+
+In Fig. 1,  $\alpha$  indicates the fraction of the optimal step size used to analyze quantization sensitivity. The results show that the random Hadamard trans
+
+![](images/c7d1e69eabf48f8a2897ada5b0ff1f380f8421784699f55e72f42b6730d2e0ce.jpg)  
+Figure 2: Input distributions of MHSA blocks in LLaMA3-8B before and after KurTail. Rotation reduces outliers and improves token-wise quantization.
+
+![](images/59ceb033cb1c75fdaf5835b89be023e58cd10965fd4513b52bcfa506dfe4bb23.jpg)
+
+formation reduces quantization sensitivity. Our Kurtosis-based method exhibits a bigger reduction in sensitivity, suggesting that it more effectively aligns the distribution with uniformity. Interestingly, we also observed that the sensitivity drop is strongest in the first layer compared to other layers for both methods. In Fig. 1 we compare layer 1 to layer 15, but this trend holds for deeper layers.
+
+Evaluation of KurTail on Channel Outliers. To demonstrate that the learned rotation by KurTail reduces the degree of tailedness in the distribution, we further visualize the inputs of multi-head self-attention (MHSA) and feed-forward network (FFN) blocks of layer 15 in Llama3-8B. In Fig. 2, we compare the input distribution once without rotation and once with KurTail learned rotation. Additionally, we highlight the maximum value for each token with a gray surface above each token. As shown, KurTail effectively mitigates outliers in activation quantization.
+
+Layer Kurtosis. We measure the kurtosis of LLaMA-3 8B representations in layers 1, 2, 15, 16, 31, and 32 before and after applying different rotations (Fig. 4). Since high kurtosis indicates heavily-tailed activations that are harder to quantize, lower values provide a good proxy for improved quantization. We find that Hadamard rotations (Quarot) reduce kurtosis across layers, SpinQuant, though not explicitly targeting kurtosis achieves further reductions, and our optimization process yields the strongest improvements, particularly in the first layer, while also giving consistent gains in later layers. These patterns align with our overall improvements over prior methods.
+
+# 3 KurTail
+
+Placement of the Rotations. Following the computational invariance theorem, as introduced by El
+
+![](images/15d0e3920314582bbff607f40ec6d665daf47ea46b3cc6a6d7b37a4a0acc1b92.jpg)  
+Figure 3: Diagram of a single-layer decoder network after applying rotations. Each block represents a computation unit. Blocks containing both blue and black indicate that the rotation is fused into the network without adding extra computation. In contrast, blocks with only the rotation signify additional computations during inference.
+
+![](images/8246bd4db26b65d2523ec76ec4a9b134888e84e7a5c7e62e001678bbe30a135d.jpg)  
+Figure 4: Kurtosis of LLaMA-3 8B layer representations. Lower values mean better quantization. The top-right panel zooms in to highlight kurtosis differences across rotation-based methods.
+
+hage et al. (2023); Ashkboos et al. (2024a) and later utilized by QuaRot and SpinQuant, we adopted a similar framework to transform the activation functions at each layer. The placement of rotations is illustrated in Fig. 3. This figure depicts a single layer of a transformer model, where each square represents a computation block. The rotations are categorized into fusible rotations  $(R_{1}$  and  $R_{2})$  and online rotations  $(R_{3}, R_{4},$  and  $R_{5})$ . Fusible rotations do not add additional computational costs during inference since they can be merged with the model's original parameters. Specifically, we apply  $R_{1}$  to the left side of the token embedding,  $W_{o}$ , and  $W_{d}$  within the MHSA and FFN blocks, respectively. The inverse of  $R_{1}$  is applied to the right side of  $W_{q}$ ,  $W_{k}$ ,  $W_{v}$  in the attention block, and  $W_{up}$ ,  $W_{gate}$  in the FFN block. Due to the residual connection, the exact same rotation must also be applied across subsequent layers (e.g.,  $XR_{1} + YR_{1}$  in one layer and  $YR_{1} + X_{2}R_{1}$  in the next). The second fusible rotation,  $R_{2}$ , is applied to the right side of  $W_{v}$ , with its inverse applied to the left side of  $W_{o}$ . This transformation improves the distribu
+
+tion of KV-caches and can vary across layers. The second group of rotations,  $R_{3}$ ,  $R_{4}$ , and  $R_{5}$ , are online which minimally increase the computational costs compared to the original model but they improve the performance. To mitigate this, we utilize random Hadamard matrices, which are computationally efficient, resulting in minimal overhead. For  $R_{3}$ , the transformation is applied after each rotational positional encoding for queries and keys. Since the transpose of any orthogonal matrix equals its inverse, there is no need to add the inverse matrix explicitly. During the computation of attention scores, the term  $Q^{T}K$  simplifies to  $Q^{T}R_{3}^{T}R_{3}K$ , effectively nullifying the impact of the rotation. For  $R_{4}$ , we introduce the transformation after applying the softmax scores to the values and add the inverse in the subsequent linear layer. Similarly,  $R_{5}$  is implemented in the FFN block using the same approach.
+
+Learning the Rotations. To discover the optimal rotations, we first run the vanilla model and store the inputs from both the MHSA and FFN blocks. Next, we create a small network consisting of a linear layer and an RMSNorm, designed to simulate the inputs of the MHSA and FFN blocks before quantization (Fig. 3). For optimization, we shuffle the stored input data from all transformer layers and both blocks and then train the rotation using Kurtosis loss. Since the optimization requires the rotations to remain within the orthogonal space, we use the Caley Adam (Li et al., 2020) optimizer to enforce this constraint. We train this small network for 100 iterations using 500 samples from the WikiText (Merit et al., 2016a) training set. In Table 9, we also did an ablation study on the different calibration size and datasets. After training, the resulting rotation is fused into the original network. For the  $R_{2}$ , we apply a similar approach, but we removed the RMSNorm and just
+
+optimize the linear layer with the Kurtosis loss.
+
+Optimization in the Orthogonal Space. As discussed in § 3, the transformation needs to be optimized in the orthogonal space to be consistent with a computational invariance framework. Therefore, we optimize all of the transformation matrices within the Stiefel Manifold (Li et al., 2020) i.e., the space of orthonormal matrices, using Caley Stochastic Gradient Descent (SGD) or Caley Adam (Li et al., 2020). For more detailed see (Li et al., 2020).
+
+Training Cost. While quantization makes the inference of large models feasible on consumer GPUs, finding the optimal rotation still requires substantial computational power. We address this by avoiding end-to-end fine-tuning. Since each multi-head attention and FFN is affected by  $R_{1}$ , end-to-end approaches like SpinQuant cannot optimize the rotation layer by layer, and directly optimizing  $R_{1}$  via gradient descent requires loading the entire model, which is memory-intensive. Although SpinQuant reduces training costs by eliminating the need to store weight gradients and states, it still requires loading the full model into GPU memory. Our approach uses layer-wise inference, which eliminates the need to load all the network weight on the GPU at once. Then we store the activations for each layer. The we optimize the rotation with a Kurtosis loss. This significantly lowers GPU requirements—at most, a single NVIDIA H100 (or A100) is needed for LLaMA 70B.
+
+# 4 Setup
+
+We developed KurTail using the Hugging Facelibrary (Wolf et al., 2019) integrated with the PyTorch framework (Paszke et al., 2019) and for evaluation we used EleutherAI evaluation framework (Gao et al., 2024b). For learning the transformation, we used 512 calibration samples for all models, except Mixteral and LLAMA 70B for which we use 256 calibration sample from the WikiText (Merit et al., 2016a) training set, each with a sequence length of 2048. For large models, we used less samples since they have more layers for which we can store the activations. For storing the activations we used layer-wise inference to reduce the GPU memory requirement. For optimizing the rotation, we use Caley Adam (Li et al., 2020) optimizer to find the rotation.
+
+For quantizing the activation, we used per-token
+
+dynamic symmetric quantization, where a single scale was applied to each row, and all values were clipped using a quantile of 0.98 in all experiments. For the KV-caches, we employed asymmetric quantization. For the Weight quantization, we use round-to-nearest (RTN), and GPTQ (Frantar et al., 2022), using per-column (or per-channel) symmetric quantization. For GPTQ quantization, we uses 128 calibration samples from the WikiText, each with a sequence length of 2048. Learning the transformation and Transforming LLAMA3-70B with KurTail on an NVIDIA H100 GPU took around one hour which compare the SpinQuant it uses significantly less memory (4 A100 GPU and 2 hours).
+
+Models. We evaluate KurTail on the LLAMA-2 (Touvron et al., 2023), LLAMA-3 (Dubey et al., 2024), Phi-model (Abdin et al., 2024) and Qwen 2.5 (Yang et al., 2024; Team, 2024) family on both language generation and zero-shot tasks. We further also target the mixture of experts model Mixtral (Jiang et al., 2024).
+
+Inference Speed-up. KurTail's contribution focuses on a novel approach to learning the rotation and given the architectural similarity with Spin-Quant and Quarot, we did not re-implement the low-level kernel for 4-bit matrix multiplication, as similar speedup results are expected. All results are based on simulated quantization; however, the real quantization will yield the same downstream performance.
+
+Evaluation Setting. To compare the performance of the model after quantization, we report the perplexity (PPL) score on the WikiText (Mery et al., 2016b) test set. While perplexity is a standard measure of language modeling performance, it may not be sufficient for evaluating the model's effectiveness after quantization. Therefore, we report the result for zero-shot reasoning as well. We assess performance using the lm-evaluation-harness (Gao et al., 2024a), testing the models on eight tasks: BoolQ (Clark et al., 2019), HellaSwag (Zellers et al., 2019), OpenBookQA(OBQA) (Mihaylov et al., 2018), PIQA (Bisk et al., 2020), SIQA (Sap et al., 2019), WinoGrande (Sakaguchi et al., 2021), ARC-Easy, and ARC-Challenge (Boratko et al., 2018) reporting the average performance across all eight tasks (0-shot), we also provide the performance on each task in § C. Additionally, to assess the model on more complex tasks, we benchmark its language comprehension and gen
+
+Table 1: Comparison of different quantization methods across various models. All the results are for 4 bit quantization for Weight/Activation/KV-cache. Weights are quantized using GPTQ.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="3">Llama-2-7b</td><td colspan="3">Llama-2-13b</td><td colspan="3">Llama-3-8b</td></tr><tr><td>Wiki (↓)</td><td>0-shot (↑)</td><td>MMLU (↑)</td><td>Wiki (↓)</td><td>0-shot (↑)</td><td>MMLU (↑)</td><td>Wiki (↓)</td><td>0-shot (↑)</td><td>MMLU (↑)</td></tr><tr><td>16-bit</td><td>5.5</td><td>64.1</td><td>42.1</td><td>4.9</td><td>66.5</td><td>52.7</td><td>6.1</td><td>67.2</td><td>63.2</td></tr><tr><td>GPTQ</td><td>9600.0s</td><td>38.9</td><td>23.8</td><td>3120.0</td><td>33.8</td><td>24.8</td><td>166.3</td><td>39.8</td><td>23.3</td></tr><tr><td>QuaRot</td><td>6.2</td><td>60.6</td><td>32.3</td><td>5.4</td><td>64.7</td><td>46.83</td><td>8.50</td><td>60.1</td><td>47.4</td></tr><tr><td>SpinQuant</td><td>6.0</td><td>61.0</td><td>34.8</td><td>5.2</td><td>64.8</td><td>47.8</td><td>7.4</td><td>63.8</td><td>56.2</td></tr><tr><td>Kurtail</td><td>5.9</td><td>61.3</td><td>32.9</td><td>5.2</td><td>65.2</td><td>49.1</td><td>7.2</td><td>64.6</td><td>57.3</td></tr><tr><td rowspan="2">Method</td><td colspan="3">Llama-3-70b</td><td colspan="3">Llama-3.2-1b</td><td colspan="3">Llama-3.2-3b</td></tr><tr><td>Wiki (↓)</td><td>0-shot (↑)</td><td>MMLU (↑)</td><td>Wiki (↓)</td><td>0-shot (↑)</td><td>MMLU (↑)</td><td>Wiki (↓)</td><td>0-shot (↑)</td><td>MMLU (↑)</td></tr><tr><td>16-bit</td><td>2.8</td><td>73.1</td><td>76.3</td><td>9.75</td><td>54.9</td><td>37.9</td><td>7.8</td><td>62.7</td><td>54.8</td></tr><tr><td>GPTQ</td><td>452.7</td><td>45.5</td><td>23.2</td><td>108.9</td><td>38.0</td><td>24.9</td><td>178.3</td><td>40.3</td><td>24.8</td></tr><tr><td>QuaRot</td><td>6.19</td><td>65.1</td><td>62.9</td><td>17.4</td><td>49.0</td><td>23.8</td><td>10.1</td><td>56.1</td><td>42.0</td></tr><tr><td>SpinQuant</td><td>6.2</td><td>65.7</td><td>59.4</td><td>13.6</td><td>48.8</td><td>25.6</td><td>9.2</td><td>57.9</td><td>44.2</td></tr><tr><td>Kurtail</td><td>4.2</td><td>70.7</td><td>73.1</td><td>12.9</td><td>50.1</td><td>27.2</td><td>9.0</td><td>59.0</td><td>47.8</td></tr></table>
+
+Table 2: Performance comparison of various models with 4 bits W/A/KV-cache quantization in common sense reasoning tasks. All the weight are quantized using GPTQ.  
+
+<table><tr><td>Model</td><td>Method</td><td>ARC-C</td><td>ARC-E</td><td>BoolQ</td><td>HellaSwag</td><td>OBQA</td><td>PIQA</td><td>SIQA</td><td>WinoGrande</td><td>AVG</td></tr><tr><td rowspan="4">Llama-2-7B</td><td>Vanilla</td><td>46.2</td><td>74.5</td><td>77.8</td><td>76.0</td><td>44.2</td><td>79.1</td><td>46.1</td><td>69.1</td><td>64.1</td></tr><tr><td>Quarot</td><td>41.6</td><td>70.6</td><td>73.2</td><td>72.1</td><td>41.2</td><td>76.9</td><td>44.0</td><td>65.2</td><td>60.6</td></tr><tr><td>SpinQuant</td><td>43.6</td><td>71.3</td><td>73.8</td><td>73.2</td><td>40.4</td><td>76.0</td><td>44.1</td><td>65.4</td><td>61.0</td></tr><tr><td>Kurtail</td><td>43.1</td><td>72.0</td><td>72.0</td><td>73.2</td><td>41.2</td><td>76.6</td><td>45.6</td><td>66.8</td><td>61.3</td></tr><tr><td rowspan="4">Llama-3-8B</td><td>Vanilla</td><td>53.4</td><td>77.8</td><td>81.4</td><td>79.2</td><td>45.0</td><td>80.8</td><td>47.2</td><td>72.6</td><td>67.2</td></tr><tr><td>Quarot</td><td>42.1</td><td>69.0</td><td>72.1</td><td>71.5</td><td>41.2</td><td>74.9</td><td>44.3</td><td>65.5</td><td>60.1</td></tr><tr><td>SpinQuant</td><td>48.0</td><td>75.4</td><td>75.8</td><td>75.4</td><td>43.8</td><td>77.5</td><td>45.0</td><td>69.2</td><td>63.8</td></tr><tr><td>Kurtail</td><td>48.2</td><td>75.4</td><td>79.2</td><td>76.4</td><td>43.6</td><td>78.4</td><td>45.8</td><td>70.0</td><td>64.6</td></tr><tr><td rowspan="4">Llama-3-70B</td><td>Vanilla</td><td>65.0</td><td>86.6</td><td>85.4</td><td>85.0</td><td>48.2</td><td>84.3</td><td>50.5</td><td>79.9</td><td>73.1</td></tr><tr><td>Quarot</td><td>53.0</td><td>74.8</td><td>81.2</td><td>77.7</td><td>42.0</td><td>78.2</td><td>45.7</td><td>68.4</td><td>65.1</td></tr><tr><td>SpinQuant</td><td>52.0</td><td>77.3</td><td>81.7</td><td>75.6</td><td>43.8</td><td>78.8</td><td>43.4</td><td>72.8</td><td>65.7</td></tr><tr><td>Kurtail</td><td>59.2</td><td>82.7</td><td>83.9</td><td>83.3</td><td>46.6</td><td>83.5</td><td>49.7</td><td>76.6</td><td>70.7</td></tr></table>
+
+eral understanding using the MMLU benchmark (Hendrycks et al., 2021) and for mathematical reasoning we utilize MathQA (Amini et al., 2019). We report the average performance in Table 1.
+
+Table 3: Performance on Phi-3-mini-4k-instruct.  
+
+<table><tr><td>Method</td><td>Wiki(↓)</td><td>0-shot(↑)</td><td>MMLU(↑)</td></tr><tr><td>16 bit</td><td>6.01</td><td>0.69</td><td>70.75</td></tr><tr><td>Quarot</td><td>8.46</td><td>0.61</td><td>56.01</td></tr><tr><td>KurTail</td><td>7.13</td><td>0.66</td><td>63.61</td></tr></table>
+
+# 5 Results
+
+To evaluate KurTail we focus on 4-bit quantization for weights, activations and KV-cache, which is a challenging bit-width for LLM quantization. Table 1 shows a summary where "0-shot" means the average performance over 8 tasks of common sense
+
+reasoning. For weight quantization we used GPTQ (Frantar et al., 2022). We also provide each task performance in Table 2 and and for all model in § C. We report the detailed performance of each tasks in § C. To demonstrate that our method outperforms previous works independently of the weight quantization technique, we acos provide results for round-to-nearest (RTN) in § C. Additionally, to show that our method is effective on LLM families beyond the LLaMA family, we present results on the Phi-3 model in Table 3 and Qwen 2.5 series of model in Table 5.
+
+For all of the result we have better perplexity in all of the models compared to previous methods. At the same time, our method is significantly better than SpinQuant and QuaRot in downstream tasks. We also provide results for math reasoning in Table 7.
+
+Comparison with Related Approaches. In Table 1, we compare KurTail with Quarot and Spin
+
+Table 4: Performance of different quantization methods on LLaMA-2 models across multiple benchmarks. WikiText reports perplexity (lower is better, ↓). Other benchmarks report accuracy (higher is better, ↑). The Avg. column is the mean of all accuracies except WikiText.  
+
+<table><tr><td>LLaMA-2 13B</td><td>Wiki(↓)</td><td>ARC-C</td><td>ARC-E</td><td>BoolQ</td><td>HellaSwag</td><td>PIQA</td><td>WinoGrande</td><td>Avg. ↑</td></tr><tr><td>Vanilla</td><td>4.9</td><td>49.2</td><td>77.5</td><td>80.6</td><td>79.4</td><td>80.5</td><td>72.1</td><td>73.2</td></tr><tr><td>Quarot</td><td>5.4</td><td>47.3</td><td>73.9</td><td>77.8</td><td>76.6</td><td>78.7</td><td>69.8</td><td>70.7</td></tr><tr><td>SpinQuant</td><td>5.2</td><td>49.0</td><td>76.3</td><td>78.2</td><td>77.1</td><td>79.3</td><td>69.5</td><td>71.6</td></tr><tr><td>DuQuant</td><td>5.3</td><td>42.1</td><td>56.2</td><td>65.7</td><td>73.9</td><td>77.2</td><td>65.4</td><td>63.4</td></tr><tr><td>OSTQuant</td><td>5.3</td><td>47.1</td><td>75.2</td><td>77.5</td><td>76.7</td><td>78.7</td><td>68.0</td><td>70.5</td></tr><tr><td>QUIK</td><td>6.8</td><td>43.7</td><td>68.0</td><td>71.3</td><td>73.3</td><td>75.7</td><td>64.6</td><td>66.1</td></tr><tr><td>KurTail</td><td>5.2</td><td>48.1</td><td>75.4</td><td>79.7</td><td>77.4</td><td>79.0</td><td>71.2</td><td>71.8</td></tr></table>
+
+Table 5: Performance on Qwen2.5-14B.  
+
+<table><tr><td>Method</td><td>Wiki(↓)</td><td>0-shot(↑)</td><td>MMLU(↑)</td></tr><tr><td>16-bit</td><td>5.3</td><td>70.6</td><td>78.6</td></tr><tr><td>RTN</td><td>2715.0</td><td>37.3</td><td>24.9</td></tr><tr><td>GPTQ</td><td>5100.3</td><td>36.9</td><td>24.8</td></tr><tr><td>SmoothQuant</td><td>1375.7</td><td>35.1</td><td>23.7</td></tr><tr><td>QUIK</td><td>10.5</td><td>57.7</td><td>56.6</td></tr><tr><td>QuaRot</td><td>6.8</td><td>67.1</td><td>70.9</td></tr><tr><td>KurTail</td><td>6.6</td><td>67.2</td><td>71.9</td></tr></table>
+
+Table 6: Performance comparison of different quantization methods for Mixtral-8x7B. All results correspond to 4-bit quantization for weights, activations, and KV-cache. RTN is used for weight quantization.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="3">Mixtral-8x7B</td></tr><tr><td>Wiki (↓)</td><td>0-shot (↑)</td><td>MMLU (↑)</td></tr><tr><td>16-bit</td><td>3.8</td><td>71.2</td><td>68.8</td></tr><tr><td>RTN</td><td>909.0</td><td>35.4</td><td>23.0</td></tr><tr><td>QuaRot</td><td>8.7</td><td>55.7</td><td>36.8</td></tr><tr><td>Kurtail</td><td>6.5</td><td>59.4</td><td>44.8</td></tr></table>
+
+Quant across different LLaMA-series models. We also include additional results on LLaMA-2 13B in Table 4 and LLaMA-2 7B in Table 15 to evaluate our method against further previous approaches. As shown in Table 4, KurTail consistently outperforms or matches the performance of other methods.
+
+Experiment on Mixture of Experts. Given the growing popularity of the Mixture of Experts (MoE) models, we also explore the idea of applying rotation within the mixture of experts. For this purpose, we utilize Mixtral (Jiang et al., 2024), which
+
+employs the exact same attention block. However, for the mixture of experts component, we apply rotation across all the experts. Table 6 presents the results for 4-bit quantization, where we used rounding to the nearest value. In principle, other quantization methods, such as GPTQ, HQQ (Badri and Shaji, 2023), and similar approaches, can also be employed to further enhance performance.
+
+Table 7: Comparison of different quantization methods across various models on MathQA. All results are reported for 4-bit quantization of W/A/KV-cache. For weight quantization, we use GPTQ.  
+
+<table><tr><td rowspan="2">Model</td><td colspan="3">MathQA Acc (%)</td></tr><tr><td>16-bit</td><td>QuaRot</td><td>KurTail</td></tr><tr><td>LLaMA-2-7B</td><td>28.24</td><td>26.70</td><td>26.77</td></tr><tr><td>LLaMA-2-13B</td><td>31.76</td><td>28.81</td><td>30.35</td></tr><tr><td>LLaMA-2-70B</td><td>38.39</td><td>33.97</td><td>35.68</td></tr><tr><td>LLaMA-3-8B</td><td>40.30</td><td>31.36</td><td>34.71</td></tr><tr><td>LLaMA-3-70B</td><td>51.79</td><td>35.54</td><td>45.76</td></tr><tr><td>LLaMA-3.2-1B</td><td>28.94</td><td>25.29</td><td>26.00</td></tr><tr><td>LLaMA-3.2-3B</td><td>34.67</td><td>30.75</td><td>30.52</td></tr><tr><td>Phi-3-mini</td><td>39.93</td><td>31.89</td><td>34.81</td></tr></table>
+
+Evaluating Mathematical Reasoning. To explore more complex reasoning tasks, we further evaluate the performance of the quantized model on tasks involving mathematical reasoning in Table 7 by reporting results on the MathQA (Amini et al., 2019) dataset. MathQA is a benchmark designed to test problem-solving and quantitative reasoning abilities. The dataset consists of real-world mathematical problems covering topics such as arithmetic, algebra, probability, and geometry.
+
+Each problem is accompanied by a natural language description, multiple-choice answers, and an annotated solution program that outlines the reasoning steps required to reach the correct answer. In Table 7, we compare KurTail with QuaRot, and the results show that KurTail outperforms QuaRot. This additional observation suggests that optimizing the rotations can also enhance performance on math reasoning tasks.
+
+Ablation Study on the Calibration Dataset. We also investigate the impact of the calibration dataset on performance. To this end, we modify the calibration data to optimize the rotation using different datasets. Specifically, we conduct experiments using PTB (Marcus et al., 1993), C4 (Raffel et al., 2020), WikiText (Merit et al., 2016b), and Alpaca (Taori et al., 2023). Additionally, we create a combined dataset by sampling equally from all four sources. For each experiment, we sample 512 instances and report the results for Llama-3.2 3B.
+
+Table 8: Performance on different calibration datasets.  
+
+<table><tr><td>Cal Dataset</td><td>Wiki(↓)</td><td>0-shot(↑)</td><td>MMLU(↑)</td></tr><tr><td>Quarot</td><td>10.1</td><td>56.1</td><td>42.0</td></tr><tr><td>Wikitext-2</td><td>9.0</td><td>59.05</td><td>47.76</td></tr><tr><td>C4</td><td>9.1</td><td>59.24</td><td>47.75</td></tr><tr><td>Alpaca</td><td>9.3</td><td>59.68</td><td>47.34</td></tr><tr><td>PTB</td><td>9.2</td><td>58.60</td><td>48.33</td></tr><tr><td>Combined</td><td>9.0</td><td>59.79</td><td>48.75</td></tr></table>
+
+Table 9: Effect of different calibration size.  
+
+<table><tr><td>Cal Size</td><td>Wiki(↓)</td><td>0-shot(↑)</td><td>MMLU(↑)</td></tr><tr><td>128</td><td>9.11</td><td>59.24</td><td>47.85</td></tr><tr><td>256</td><td>9.12</td><td>58.85</td><td>47.47</td></tr><tr><td>512</td><td>9.09</td><td>59.79</td><td>48.75</td></tr><tr><td>1024</td><td>9.08</td><td>59.43</td><td>49.02</td></tr></table>
+
+Table 8 presents the findings. Interestingly, all dataset variations outperform the non-training method Quarot. Moreover, we observe lower perplexity on WikiText when using other datasets for calibration. The best performance on the MMLU task is achieved with the PTB dataset, while the best results for common sense reasoning tasks are obtained using the Alpaca dataset. The combined dataset yields the best overall performance across all tasks while it uses the exact same number of
+
+samples (512 sentences). In Table 9, we explore different calibration sample sizes for learning the rotations and their impact on the model's performance in downstream tasks. In this study, we used our combined dataset and the Llama 3.2 3B model. As shown in Table 8, we observe a trend toward improvement as the sample size increases, although performance tends to saturate around a sample size of 512.
+
+# 6 Conclusion
+
+We introduced KurTail a novel technique for learning orthogonal transformations that rotate the activation distribution to address the outlier problem and overcomes the limitations of previous approaches. Compared to QuaRot, which uses nonlearnable rotation, and SpinQuant, which requires substantial computational resources for learning rotations, KurTail offers a more efficient and robust solution by learning a rotation that minimizes the kurtosis of the representations. We further provide theoretical insights into why layer-wise optimization yields the same results as end-to-end training, and why orthogonal transformations are a suitable choice of matrix space for learning the transformation. Finally, these results highlight KurTail's ability to deliver efficiency and high performance across large-scale language models.
+
+Limitations. In this work, we only focuses on dynamic per-token quantization for activations, which offers flexibility but does not fully exploit the potential of static tensor-wise quantization. Static quantization, which precomputes scaling factors for improved efficiency, could further optimize inference speed and memory usage. However, it requires careful calibration, which we leave for future work.
+
+# References
+
+Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, et al. 2024. Phi-3 technical report: A highly capable language model locally on your phone, 2024. URL https://arxiv.org/abs/2404.14219.  
+Aida Amini, Saadia Gabriel, Peter Lin, Rik Koncel-Kedziorski, Yejin Choi, and Hannaneh Hajishirzi. 2019. Mathqa: Towards interpretable math word problem solving with operation-based formalisms. arXiv preprint arXiv:1905.13319.  
+Saleh Ashkboos, Maximilian L Croci, Marcelo Gennari do Nascimento, Torsten Hoefler, and James Hens
+
+man. 2024a. Slicegpt: Compress large language models by deleting rows and columns. arXiv preprint arXiv:2401.15024.  
+Saleh Ashkboos, Ilia Markov, Elias Frantar, Tingxuan Zhong, Xincheng Wang, Jie Ren, Torsten Hoefler, and Dan Alistarh. 2023. Towards end-to-end 4-bit inference on generative large language models. arXiv preprint arXiv:2310.09259.  
+Saleh Ashkboos, Amirkeivan Mohtashami, Maximilian L Croci, Bo Li, Pashmina Cameron, Martin Jaggi, Dan Alistarh, Torsten Hoefler, and James Hensman. 2024b. Quarot: Outlier-free 4-bit inference in rotated llms. arXiv preprint arXiv:2404.00456.  
+Hicham Badri and Appu Shaji. 2023. Half-quadratic quantization of large machine learning models.  
+Ron Banner, Yury Nahshan, Elad Hoffer, and Daniel Soudry. 2019. Post-training 4-bit quantization of convolution networks for rapid-deployment. Preprint, arXiv:1810.05723.  
+Yonatan Bisk, Rowan Zellers, Ronan Le Bras, Jianfeng Gao, and Yejin Choi. 2020. Piqa: Reasoning about physical commonsense in natural language. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 7432-7439.  
+Michael Boratko, Harsh Padigela, Deepak Mikkilineni, Pavan Yuvraj, Rajarshi Das, Andrew McCallum, Mihai Chang, Achille Fokoue, Pavan Kapanipathi, Nicholas Mattei, et al. 2018. Arc: A machine reading comprehension dataset for reasoning over science text. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 1414-1423.  
+Brian Chmiel, Ron Banner, Gil Shomron, Yury Nahshan, Alex Bronstein, Uri Weiser, et al. 2020. Robust quantization: One model to rule them all. Advances in neural information processing systems, 33:5308-5317.  
+Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. 2019. Boolq: Exploring the surprising difficulty of natural yes/no questions. arXiv preprint arXiv:1905.10044.  
+Tim Dettmers, Mike Lewis, Younes Belkada, and Luke Zettlemoyer. 2022. Gpt3. int8(): 8-bit matrix multiplication for transformers at scale. Advances in Neural Information Processing Systems, 35:30318-30332.  
+Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783.  
+Vage Egiazarian, Andrei Panferov, Denis Kuznedelev, Elias Frantar, Artem Babenko, and Dan Alistarh. 2024. Extreme compression of large language models via additive quantization. arXiv preprint arXiv:2401.06118.
+
+Nelson Elhage, Robert Lasenby, and Christopher Olah. 2023. Privileged bases in the transformer residual stream. Transformer Circuits Thread.  
+Elias Frantar, Saleh Ashkboos, Torsten Hoefler, and Dan Alistarh. 2022. Gptq: Accurate post-training quantization for generative pre-trained transformers. arXiv preprint arXiv:2210.17323.  
+Leo Gao, Stella Biderman, Hailey Schoelkopf, Lintang Sutawika, et al. 2024a. Lessons from the trenches on reproducible evaluation of language models. arXiv preprint arXiv:2405.14782.  
+Leo Gao, Jonathan Tow, Baber Abbasi, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Alain Le Noac'h, Haonan Li, Kyle McDonell, Niklas Muennighoff, Chris Ociepa, Jason Phang, Laria Reynolds, Hailey Schoelkopf, Aviya Skowron, Lintang Sutawika, Eric Tang, Anish Thite, Ben Wang, Kevin Wang, and Andy Zou. 2024b. A framework for few-shot language model evaluation.  
+Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948.  
+Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300.  
+Xing Hu, Yuan Cheng, Dawei Yang, Zukang Xu, Zhihang Yuan, Jiangyong Yu, Chen Xu, Zhe Jiang, and Sifan Zhou. 2025. Ostquant: Refining large language model quantization with orthogonal and scaling transformations for better distribution fitting. arXiv preprint arXiv:2501.13987.  
+Albert Q. Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, Gianna Lengyel, Guillaume Bour, Guillaume Lample, Lélio Renard Lavaud, Lucile Saulnier, Marie-Anne Lachaux, Pierre Stock, Sandeep Subramanian, Sophia Yang, Szymon Antoniak, Teven Le Scao, Théophile Gervet, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. 2024. Mixtral of experts. Preprint, arXiv:2401.04088.  
+Jun Li, Li Fuxin, and Sinisa Todorovic. 2020. Efficient riemannian optimization on the stiefel manifold via the Cayley transform. arXiv preprint arXiv:2002.01113.  
+Haokun Lin, Haobo Xu, Yichen Wu, Jingzhi Cui, Yingtao Zhang, Linzhan Mou, Linqi Song, Zhenan Sun, and Ying Wei. 2024a. Duquant: Distributing outliers via dual transformation makes stronger quantized llms. Advances in Neural Information Processing Systems, 37:87766-87800.
+
+Ji Lin, Jiaming Tang, Haotian Tang, Shang Yang, Wei-Ming Chen, Wei-Chen Wang, Guangxuan Xiao, Xingyu Dang, Chuang Gan, and Song Han. 2024b. Awq: Activation-aware weight quantization for on-device llm compression and acceleration. Proceedings of Machine Learning and Systems, 6:87-100.  
+Zechun Liu, Changsheng Zhao, Igor Fedorov, Bilge Soran, Dhruv Choudhary, Raghuraman Krishnamoorthi, Vikas Chandra, Yuandong Tian, and Tijmen Blankevoort. 2024. Spinquant-llm quantization with learned rotations. arXiv preprint arXiv:2405.16406.  
+Mitchell P. Marcus, Beatrice Santorini, and Mary Ann Marcinkiewicz. 1993. Building a large annotated corpus of english: The penn treebank. Computational Linguistics, 19(2):313-330.  
+Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. 2016a. Pointer sentinel mixture models. arXiv preprint arXiv:1609.07843.  
+Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. 2016b. Pointer sentinel mixture models. arXiv preprint arXiv:1609.07843.  
+Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. 2018. Openbookqa: Fact-based open book question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 268-277.  
+OpenAI. 2024. Learning to reason with llms. https://openai.com/index/ learning-to-reason-with-llms. Accessed: 2025-01-30.  
+Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. 2019. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32.  
+Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67.  
+Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. 2021. Winogrande: An adversarial winograd schema challenge at scale. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 8732-8740.  
+Maarten Sap, Hannah Rashkin, Derek Chen, Ronan Le Bras, and Yejin Choi. 2019. Social iqa: Commonsense reasoning about social interactions. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 4463-4473.
+
+Utkarsh Saxena, Sayeh Sharify, Kaushik Roy, and Xin Wang. 2024. Resq: Mixed-precision quantization of large language models with low-rank residuals. arXiv preprint arXiv:2412.14363.  
+Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca.  
+Qwen Team. 2024. Qwen2.5: A party of foundation models.  
+Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288.  
+Albert Tseng, Jerry Chee, Qingyao Sun, Volodymyr Kuleshov, and Christopher De Sa. 2024. Quip#: Even better llm quantization with hadamard incoherence and lattice codebooks. arXiv preprint arXiv:2402.04396.  
+Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierrick Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, and Jamie Brew. 2019. Huggingface's transformers: State-of-the-art natural language processing. CoRR, abs/1910.03771.  
+Guangxuan Xiao, Ji Lin, Mickael Seznec, Hao Wu, Julien Demouth, and Song Han. 2023. Smoothquant: Accurate and efficient post-training quantization for large language models. In International Conference on Machine Learning, pages 38087-38099. PMLR.  
+An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. 2024. Qwen2 technical report. arXiv preprint arXiv:2407.10671.  
+Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. 2019. Hellaswag: Can a machine really finish your sentence? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4791-4800.
+
+# A Proofs
+
+Lemma A.1. The quantization error after transformation is bounded with  $\left\| X - Q(xT)T^{-1}\right\|_F \leq \frac{c}{2}\sqrt{NM}\left\| T^1\right\|_2\left\| T^{-1}\right\|_2 = \frac{c}{2}\sqrt{NM}\kappa(T)$ , where  $T$  is the transformation and  $c$  is a constant depends on  $X$  and  $\kappa(T)$  is the condition number of matrix  $T$ .
+
+# Proof of Lemma 2.3
+
+We aim to bound the quantization error defined as  $\left\| \boldsymbol{X} - Q(\boldsymbol{X}\boldsymbol{T})\boldsymbol{T}^{-1} \right\|_F$ . To do so, we use the fact that  $\boldsymbol{X} = \boldsymbol{X}\boldsymbol{T}\boldsymbol{T}^{-1}$ :
+
+$$
+\begin{array}{l} \mathbb {E} = \left\| \boldsymbol {X} - Q (\boldsymbol {X T}) \boldsymbol {T} ^ {- 1} \right\| _ {F} \\ = \left\| \boldsymbol {X} \boldsymbol {T} \boldsymbol {T} ^ {- 1} - Q (\boldsymbol {X} \boldsymbol {T}) \boldsymbol {T} ^ {- 1} \right\| _ {F}. \tag {5} \\ \end{array}
+$$
+
+Applying the sub-multiplicative property of the Frobenius norm (i.e.,  $\| AB\| _F\leq \| A\| _2\| B\| _F$ ), we obtain:
+
+$$
+\begin{array}{l} \left\| \boldsymbol {X} \boldsymbol {T} \boldsymbol {T} ^ {- 1} - Q (\boldsymbol {X} \boldsymbol {T}) \boldsymbol {T} ^ {- 1} \right\| _ {F} \leq \\ \left\| \boldsymbol {X} \boldsymbol {T} - Q (\boldsymbol {X} \boldsymbol {T}) \right\| _ {F} \cdot \left\| \boldsymbol {T} ^ {- 1} \right\| _ {2}. \tag {6} \\ \end{array}
+$$
+
+We now focus on bounding the quantization error term  $\| \boldsymbol{X}\boldsymbol{T} - Q(\boldsymbol{X}\boldsymbol{T})\|_{F}$ . Under uniform quantization, each entry of the quantization error is bounded by  $\frac{\Delta}{2}$ , where  $\Delta = \frac{\max_{ij} |(\boldsymbol{X}\boldsymbol{T})_{ij}|}{2^{k-1}-1}$  is the quantization step size. Therefore, the Frobenius norm can be bounded by:
+
+$$
+\left\| \boldsymbol {X} \boldsymbol {T} - Q (\boldsymbol {X} \boldsymbol {T}) \right\| _ {F} \leq \sqrt {N M} \cdot \frac {\Delta}{2}, \tag {7}
+$$
+
+where  $N$  and  $M$  are the number of rows and columns of  $\mathbf{X}\mathbf{T}$ , respectively.
+
+Combining Eq. 6 and Eq. 7, we get:
+
+$$
+\left\| \boldsymbol {X} - Q (\boldsymbol {X T}) \boldsymbol {T} ^ {- 1} \right\| _ {F} \leq \sqrt {N M} \cdot \frac {\Delta}{2} \cdot \left\| \boldsymbol {T} ^ {- 1} \right\| _ {2}. \tag {8}
+$$
+
+To bound  $\Delta$ , we use the fact that the maximum absolute value of elements in  $XT$  satisfies:
+
+$$
+\max  _ {i j} | (\boldsymbol {X T}) _ {i j} | \leq \| \boldsymbol {X T} \| _ {2} \leq \| \boldsymbol {X} \| _ {2} \cdot \| \boldsymbol {T} \| _ {2}. \tag {9}
+$$
+
+Substituting this into the expression for  $\Delta$ , we obtain:
+
+$$
+\Delta \leq \frac {\| \boldsymbol {X} \| _ {2} \cdot \| \boldsymbol {T} \| _ {2}}{2 ^ {k - 1} - 1}.
+$$
+
+Finally, substituting this into Eq. 8, we conclude:
+
+$$
+\left\| \boldsymbol {X} - Q (\boldsymbol {X T}) \boldsymbol {T} ^ {- 1} \right\| _ {F} \leq
+$$
+
+$$
+\frac {\left\| \boldsymbol {X} \right\| _ {2}}{2 ^ {k - 1} - 1} \cdot \sqrt {N M} \cdot \left\| \boldsymbol {T} \right\| _ {2} \cdot \left\| \boldsymbol {T} ^ {- 1} \right\| _ {2} = \tag {10}
+$$
+
+$$
+\frac {\| \boldsymbol {X} \| _ {2}}{2 ^ {k - 1} - 1} \cdot \sqrt {N M} \cdot \operatorname {c n} (\boldsymbol {T}),
+$$
+
+where  $\mathrm{cn}(\pmb{T}) = \| \pmb{T}\|_2\cdot \| \pmb{T}^{-1}\|_2$  denotes the condition number of  $\pmb{T}$ .
+
+Proposition A.2. Let  $H_1, O_1 = f(X, W_1, R_1)$ , and  $H_2, O_2 = g(O_1; W_2; R_2)$  where  $f, g$  are parameterized by  $W_1, R_1$  and  $W_2, R_2$ .
+
+Given functional invariance of  $f$  and  $g$ , i.e.  $\pmb{O}_1 = \pmb{O}_1'$  for any  $\pmb{O}_1', \pmb{H}_1' = f(\pmb{X}, \pmb{W}_1, \pmb{R}_1')$  and any orthogonal  $\pmb{R}_1'$  (and similarly for  $g$ ), and given that the total loss is  $\mathcal{L}(\pmb{R}_1, \pmb{R}_2) = \mathcal{L}_1(\pmb{H}_1) + \mathcal{L}_2(\pmb{H}_2)$ ,
+
+the independent minimization of each loss results in the same optimum as end-to-end:  $\arg \min_{\pmb{R}_1,\pmb{R}_2}\mathcal{L}(\pmb{R}_1,\pmb{R}_2) =$  (arg  $\min_{\pmb{R}_1}\mathcal{L}_1(\pmb {R}_1)$  , arg  $\min_{\pmb{R}_2}\mathcal{L}_2(f;\pmb {R}_2))$  even though  $\pmb{H}_{2}$  implicitly depends on  $O_1$
+
+Proof. We need to first prove that  $\mathcal{L}_2$  is independent of  $\pmb{R}_{1}$  and the same for  $\mathcal{L}_1$  from  $\pmb{R}_{2}$
+
+$\mathcal{L}_2$  is independent from  $R_{1}$  since  $f$  is invariant to orthogonal transformations,  $\pmb{O}_{1}$  is constant with respect to  $\pmb{R}_{1}$ . Since  $\pmb {H}_2 = g(\pmb {O}_1,\pmb {W}_2,\pmb {R}_2)$  depends on  $\pmb{O}_{1}$  but not directly on  $\pmb{R}_{1}$ , any change in  $\pmb{R}_{1}$  leaves  $\pmb{H}_{2}$  (and therefore  $\mathcal{L}_2$ ) unchanged:
+
+$$
+\frac {\partial \mathcal {L} _ {2}}{\partial \boldsymbol {R} _ {1}} = 0 \quad \Longrightarrow \quad \mathcal {L} _ {2} (\boldsymbol {H} _ {2}) \text {i s i n d e p e n d e n t o f} \boldsymbol {R} _ {1}.
+$$
+
+Similarly  $\mathcal{L}_1$  is independent of  $R_2$  by using the invariance of  $g$  shows that  $H_{1}$  and hence  $\mathcal{L}_1$  is independent of  $R_2$ :
+
+$$
+\frac {\partial \mathcal {L} _ {1}}{\partial \boldsymbol {R} _ {2}} = 0 \quad \Longrightarrow \quad \mathcal {L} _ {1} (\boldsymbol {H} _ {1}) \text {i s i n d e p e n d e n t o f} \boldsymbol {R} _ {2}.
+$$
+
+Now by considering that they have separable objective which means combining the independence result in
+
+$$
+\mathcal {L} (\boldsymbol {R} _ {1}, \boldsymbol {R} _ {2}) = \underbrace {\mathcal {L} _ {1} (\boldsymbol {H} _ {1})} _ {\text {d e p e n d s o n l y o n} \boldsymbol {R} _ {1}} + \underbrace {\mathcal {L} _ {2} (\boldsymbol {H} _ {2})} _ {\text {d e p e n d s o n l y o n} \boldsymbol {R} _ {2}},
+$$
+
+so  $\mathcal{L}$  is a sum of two functions, each depending on disjoint variables.
+
+Finally for any function of the form  $F(x,y) = F_{1}(x) + F_{2}(y)$  with independent feasible sets, every
+
+global minimiser is obtained by minimising each term separately:
+
+$$
+\begin{array}{l} (x ^ {\star}, y ^ {\star}) \in \operatorname * {a r g   m i n} _ {x, y} F (x, y) \\ \Longleftrightarrow x ^ {\star} \in \arg \min  _ {x} F _ {1} (x), \tag {11} \\ y ^ {\star} \in \arg \min  _ {y} F _ {2} (y). \\ \end{array}
+$$
+
+Applying this to  $\mathcal{L}(\pmb {R}_1,\pmb {R}_2)$  yields the claim.
+
+# B Evaluation of Rotation based method on outliers
+
+In dynamic per-token quantization, the maximum value of a token's vector plays a critical role in determining the quantization step size and range. Larger maximum values increase the quantization range, which results in larger quantization steps and greater precision loss. Alternatively, reducing the maximum value allows for smaller quantization steps, which result in more efficient representation of token values with minimal degradation of information. Therefore, lowering the maximum values across tokens is directly connected to overall quantization error and model performance. To evaluate how well different methods achieve this goal, we measure the success rate of our proposed method, KurTail, compared to its un-rotated counterpart (baseline vector) and an alternative rotation method, QuaRot. A "success" is defined as a case where the maximum value of a token's vector after applying a benchmark rotation method (KurTail or QuaRot) is smaller than that of the baseline vector. The success rate is defined as the percentage of tokens where the benchmarked rotated version achieves this reduction. In Table 10, we present the average success rates for LLAMA3-8B. KurTail consistently produces smaller maximum values across all layers, samples, and tokens, achieving a higher success rate compared to the baseline vector in nearly all cases. Additionally, it outperforms QuaRot in approximately  $63.29\%$  in MSHA,  $62.99\%$  in FFN on average.
+
+# C Further Evaluation
+
+In this section, we provide a more detailed evaluation of all tasks and more models. We present results for 4-bit quantization of weights, activations, and the KV-cache. Table 12 reports the performance of each MMLU task under 4-bit quantization for weights, activations, and the KV-cache. We use the GPTQ quantization algorithm for weight
+
+Table 10: The success rate of benchmark over baseline.  
+
+<table><tr><td></td><td>Baseline</td><td>Benchmark</td><td>Success Rate (%)</td></tr><tr><td rowspan="3">MHSA</td><td>Vanilla</td><td>KurTail</td><td>99.74%</td></tr><tr><td>Vanilla</td><td>QuaRot</td><td>99.43%</td></tr><tr><td>QuaRot</td><td>KurTail</td><td>63.29%</td></tr><tr><td rowspan="3">FFN</td><td>Vanilla</td><td>KurTail</td><td>99.96%</td></tr><tr><td>Vanilla</td><td>QuaRot</td><td>99.96%</td></tr><tr><td>QuaRot</td><td>KurTail</td><td>62.99%</td></tr></table>
+
+quantization in this experiment. Similarly, using the same setup, we evaluate common-sense reasoning tasks, as shown in Table 13. We report the performance of common-sense reasoning tasks using RTN quantization for weights in Table 14. Finally we also provide the further evaluation LLaMA-2 7B and Qwen 2.5 14B in Table 15, Table 17,Table 16 respectively.
+
+# D Comparison with ResQ
+
+We also compare our results with ResQ (Saxena et al., 2024), a concurrent work that uses a combination of mixed precision and Hadamard rotation (similar to Quarot) using Qwen 2.5 14B in Table 11. Although in this setting we cannot outperform ResQ, a fairer comparison is when mixed percision is combined with KurTail rotation, which is out of the context of our study. Furthermore, our method remains Pareto-optimal under 4-bit quantization.
+
+Table 11: Qwen2.5-14B: Perplexity, Reasoning Avg, and MMLU Avg for KurTail and ResQ.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="3">Qwen2.5-14B</td></tr><tr><td>Wiki (↓)</td><td>0-shot (↑)</td><td>MMLU (↑)</td></tr><tr><td>KurTail</td><td>6.6</td><td>67.2</td><td>71.9</td></tr><tr><td>ResQ</td><td>6.2</td><td>69.2</td><td>74.6</td></tr></table>
+
+Table 12: Performance comparison of different models using various methods across different domains.  
+
+<table><tr><td>Model</td><td>Method</td><td>Human</td><td>Other</td><td>STEM</td><td>S-Sci</td><td>AVG</td></tr><tr><td rowspan="4">Llama-2-7B</td><td>Vanilla</td><td>39.8</td><td>47.3</td><td>34.2</td><td>47.3</td><td>42.1</td></tr><tr><td>Quarot</td><td>31.1</td><td>35.7</td><td>29.9</td><td>34.1</td><td>32.7</td></tr><tr><td>SpinQuant</td><td>33.9</td><td>38.5</td><td>29.5</td><td>37.5</td><td>34.8</td></tr><tr><td>Kurtail</td><td>32.3</td><td>35.0</td><td>29.8</td><td>34.4</td><td>32.9</td></tr><tr><td rowspan="4">Llama-2-13B</td><td>Vanilla</td><td>47.9</td><td>59.4</td><td>42.3</td><td>61.2</td><td>52.7</td></tr><tr><td>Quarot</td><td>42.7</td><td>52.3</td><td>38.2</td><td>54.1</td><td>46.8</td></tr><tr><td>SpinQuant</td><td>43.5</td><td>53.1</td><td>39.1</td><td>55.4</td><td>47.8</td></tr><tr><td>Kurtail</td><td>45.3</td><td>54.0</td><td>40.4</td><td>56.6</td><td>49.1</td></tr><tr><td rowspan="4">Llama-3-8B</td><td>Vanilla</td><td>55.0</td><td>70.8</td><td>53.7</td><td>73.2</td><td>63.2</td></tr><tr><td>Quarot</td><td>42.1</td><td>52.9</td><td>39.8</td><td>54.9</td><td>47.4</td></tr><tr><td>SpinQuant</td><td>49.8</td><td>63.3</td><td>46.8</td><td>65.0</td><td>56.2</td></tr><tr><td>Kurtail</td><td>50.2</td><td>64.5</td><td>49.1</td><td>65.6</td><td>57.3</td></tr><tr><td rowspan="4">Llama-3-70B</td><td>Vanilla</td><td>67.7</td><td>81.5</td><td>69.2</td><td>86.7</td><td>76.3</td></tr><tr><td>Quarot</td><td>55.3</td><td>68.5</td><td>53.7</td><td>74.1</td><td>62.9</td></tr><tr><td>SpinQuant</td><td>50.7</td><td>67.0</td><td>51.9</td><td>68.1</td><td>59.4</td></tr><tr><td>Kurtail</td><td>65.2</td><td>79.1</td><td>63.9</td><td>84.2</td><td>73.1</td></tr><tr><td rowspan="4">Llama-3.2-1B</td><td>Vanilla</td><td>35.3</td><td>41.3</td><td>33.9</td><td>41.3</td><td>38.0</td></tr><tr><td>Quarot</td><td>25.4</td><td>26.9</td><td>24.4</td><td>25.4</td><td>25.5</td></tr><tr><td>SpinQuant</td><td>25.4</td><td>27.6</td><td>24.2</td><td>25.3</td><td>25.6</td></tr><tr><td>Kurtail</td><td>26.5</td><td>28.8</td><td>26.0</td><td>27.3</td><td>27.2</td></tr><tr><td rowspan="4">Llama-3.2-3B</td><td>Vanilla</td><td>49.0</td><td>63.1</td><td>45.5</td><td>62.9</td><td>55.1</td></tr><tr><td>Quarot</td><td>38.5</td><td>47.3</td><td>35.3</td><td>46.7</td><td>42.0</td></tr><tr><td>SpinQuant</td><td>37.0</td><td>49.4</td><td>39.9</td><td>50.5</td><td>44.2</td></tr><tr><td>Kurtail</td><td>44.8</td><td>53.4</td><td>39.5</td><td>53.4</td><td>47.8</td></tr></table>
+
+Table 13: Performance comparison of various models with 4 bits W/A/KV-cache quantization in common sense reasoning tasks. All the weight are quantized using GPTQ.  
+
+<table><tr><td>Model</td><td>Method</td><td>ARC-C</td><td>ARC-E</td><td>BoolQ</td><td>HellaSwag</td><td>OBQA</td><td>PIQA</td><td>SIQA</td><td>WinoGrande</td><td>AVG</td></tr><tr><td rowspan="4">Llama-2-7B</td><td>Vanilla</td><td>46.2</td><td>74.5</td><td>77.8</td><td>76.0</td><td>44.2</td><td>79.1</td><td>46.1</td><td>69.1</td><td>64.1</td></tr><tr><td>Quarot</td><td>41.6</td><td>70.6</td><td>73.2</td><td>72.1</td><td>41.2</td><td>76.9</td><td>44.0</td><td>65.2</td><td>60.6</td></tr><tr><td>SpinQuant</td><td>43.6</td><td>71.3</td><td>73.8</td><td>73.2</td><td>40.4</td><td>76.0</td><td>44.1</td><td>65.4</td><td>61.0</td></tr><tr><td>Kurtail</td><td>43.1</td><td>72.0</td><td>72.0</td><td>73.2</td><td>41.2</td><td>76.6</td><td>45.6</td><td>66.8</td><td>61.3</td></tr><tr><td rowspan="4">Llama-2-13B</td><td>Vanilla</td><td>49.2</td><td>77.5</td><td>80.6</td><td>79.4</td><td>45.2</td><td>80.5</td><td>47.4</td><td>72.1</td><td>66.5</td></tr><tr><td>Quarot</td><td>47.3</td><td>73.9</td><td>77.8</td><td>76.6</td><td>44.4</td><td>78.7</td><td>44.1</td><td>69.8</td><td>64.1</td></tr><tr><td>SpinQuant</td><td>49.0</td><td>76.3</td><td>78.2</td><td>77.1</td><td>42.8</td><td>79.3</td><td>46.3</td><td>69.5</td><td>64.8</td></tr><tr><td>Kurtail</td><td>48.1</td><td>75.4</td><td>79.7</td><td>77.4</td><td>45.0</td><td>79.0</td><td>45.6</td><td>71.2</td><td>65.2</td></tr><tr><td rowspan="4">Llama-3-8B</td><td>Vanilla</td><td>53.4</td><td>77.8</td><td>81.4</td><td>79.2</td><td>45.0</td><td>80.8</td><td>47.2</td><td>72.6</td><td>67.2</td></tr><tr><td>Quarot</td><td>42.1</td><td>69.0</td><td>72.1</td><td>71.5</td><td>41.2</td><td>74.9</td><td>44.3</td><td>65.5</td><td>60.1</td></tr><tr><td>SpinQuant</td><td>48.0</td><td>75.4</td><td>75.8</td><td>75.4</td><td>43.8</td><td>77.5</td><td>45.0</td><td>69.2</td><td>63.8</td></tr><tr><td>Kurtail</td><td>48.2</td><td>75.4</td><td>79.2</td><td>76.4</td><td>43.6</td><td>78.4</td><td>45.8</td><td>70.0</td><td>64.6</td></tr><tr><td rowspan="4">Llama-3-70B</td><td>Vanilla</td><td>65.0</td><td>86.6</td><td>85.4</td><td>85.0</td><td>48.2</td><td>84.3</td><td>50.5</td><td>79.9</td><td>73.1</td></tr><tr><td>Quarot</td><td>53.0</td><td>74.8</td><td>81.2</td><td>77.7</td><td>42.0</td><td>78.2</td><td>45.7</td><td>68.4</td><td>65.1</td></tr><tr><td>SpinQuant</td><td>52.0</td><td>77.3</td><td>81.7</td><td>75.6</td><td>43.8</td><td>78.8</td><td>43.4</td><td>72.8</td><td>65.7</td></tr><tr><td>Kurtail</td><td>59.2</td><td>82.7</td><td>83.9</td><td>83.3</td><td>46.6</td><td>83.5</td><td>49.7</td><td>76.6</td><td>70.7</td></tr><tr><td rowspan="4">Llama-3.2-1B</td><td>Vanilla</td><td>36.2</td><td>60.4</td><td>63.9</td><td>63.6</td><td>37.2</td><td>74.6</td><td>43.0</td><td>60.5</td><td>54.9</td></tr><tr><td>Quarot</td><td>30.0</td><td>51.4</td><td>59.1</td><td>54.0</td><td>34.2</td><td>66.7</td><td>39.6</td><td>57.1</td><td>49.0</td></tr><tr><td>SpinQuant</td><td>32.3</td><td>51.8</td><td>59.3</td><td>55.4</td><td>30.4</td><td>67.7</td><td>38.6</td><td>54.7</td><td>48.8</td></tr><tr><td>Kurtail</td><td>31.1</td><td>52.9</td><td>60.7</td><td>56.4</td><td>36.4</td><td>68.6</td><td>40.5</td><td>54.3</td><td>50.1</td></tr><tr><td rowspan="4">Llama-3.2-3B</td><td>Vanilla</td><td>46.0</td><td>71.7</td><td>73.2</td><td>73.6</td><td>43.0</td><td>77.5</td><td>47.0</td><td>69.7</td><td>62.7</td></tr><tr><td>Quarot</td><td>38.6</td><td>59.0</td><td>65.9</td><td>66.5</td><td>35.8</td><td>74.4</td><td>43.1</td><td>65.2</td><td>56.1</td></tr><tr><td>SpinQuant</td><td>38.9</td><td>64.8</td><td>68.0</td><td>69.1</td><td>39.4</td><td>74.9</td><td>45.1</td><td>62.9</td><td>57.9</td></tr><tr><td>Kurtail</td><td>42.2</td><td>66.7</td><td>69.8</td><td>68.8</td><td>39.8</td><td>75.6</td><td>44.8</td><td>64.6</td><td>59.0</td></tr></table>
+
+Table 14: Performance comparison of various models with 4 bits W/A/KV-cache quantization in common sense reasoning tasks. All the weights are quantized using RTN.  
+
+<table><tr><td>Model</td><td>Method</td><td>ARC-C</td><td>ARC-E</td><td>BoolQ</td><td>HellaSwag</td><td>OBQA</td><td>PIQA</td><td>SIQA</td><td>Winogrande</td><td>AVG</td></tr><tr><td rowspan="3">Llama-2-7B</td><td>Vanilla</td><td>46.2</td><td>74.5</td><td>77.8</td><td>76.0</td><td>44.2</td><td>79.1</td><td>46.1</td><td>69.1</td><td>64.1</td></tr><tr><td>Quarot</td><td>35.2</td><td>62.4</td><td>69.0</td><td>62.6</td><td>33.4</td><td>71.7</td><td>40.9</td><td>60.2</td><td>54.4</td></tr><tr><td>Kurtail</td><td>39.0</td><td>64.9</td><td>69.8</td><td>64.7</td><td>39.2</td><td>74.1</td><td>42.1</td><td>62.2</td><td>57.0</td></tr><tr><td rowspan="3">Llama-2-13B</td><td>Vanilla</td><td>49.2</td><td>77.5</td><td>80.6</td><td>79.4</td><td>45.2</td><td>80.5</td><td>47.4</td><td>72.1</td><td>66.5</td></tr><tr><td>Quarot</td><td>41.4</td><td>68.2</td><td>73.2</td><td>71.2</td><td>41.6</td><td>76.3</td><td>41.1</td><td>66.1</td><td>59.9</td></tr><tr><td>Kurtail</td><td>44.2</td><td>70.3</td><td>74.7</td><td>72.5</td><td>40.4</td><td>77.5</td><td>45.9</td><td>70.2</td><td>62.0</td></tr><tr><td rowspan="3">Llama-2-70B</td><td>Vanilla</td><td>57.4</td><td>81.1</td><td>83.8</td><td>83.8</td><td>48.8</td><td>82.8</td><td>49.2</td><td>78.0</td><td>70.6</td></tr><tr><td>Quarot</td><td>50.5</td><td>76.8</td><td>80.0</td><td>78.4</td><td>44.0</td><td>79.9</td><td>46.0</td><td>72.9</td><td>66.1</td></tr><tr><td>Kurtail</td><td>51.3</td><td>76.6</td><td>80.9</td><td>81.0</td><td>46.4</td><td>81.7</td><td>46.8</td><td>76.2</td><td>67.6</td></tr><tr><td rowspan="3">Llama-3-8B</td><td>Vanilla</td><td>53.4</td><td>77.8</td><td>81.4</td><td>79.2</td><td>45.0</td><td>80.8</td><td>47.2</td><td>72.6</td><td>67.2</td></tr><tr><td>Quarot</td><td>31.1</td><td>51.6</td><td>55.7</td><td>62.0</td><td>31.6</td><td>66.3</td><td>40.1</td><td>59.0</td><td>49.7</td></tr><tr><td>Kurtail</td><td>38.1</td><td>61.1</td><td>72.5</td><td>69.3</td><td>36.8</td><td>72.9</td><td>41.9</td><td>66.1</td><td>57.3</td></tr><tr><td rowspan="3">Llama-3-70B</td><td>Vanilla</td><td>65.0</td><td>86.6</td><td>85.4</td><td>85.0</td><td>48.2</td><td>84.3</td><td>50.5</td><td>80.0</td><td>73.1</td></tr><tr><td>Quarot</td><td>20.6</td><td>31.3</td><td>58.5</td><td>28.4</td><td>25.4</td><td>55.0</td><td>33.2</td><td>50.7</td><td>37.9</td></tr><tr><td>Kurtail</td><td>23.0</td><td>37.8</td><td>48.5</td><td>33.9</td><td>29.8</td><td>61.8</td><td>36.6</td><td>51.6</td><td>40.4</td></tr><tr><td rowspan="3">Llama-3.2-1B</td><td>Vanilla</td><td>36.2</td><td>60.4</td><td>63.9</td><td>63.6</td><td>37.2</td><td>74.6</td><td>43.0</td><td>60.5</td><td>54.9</td></tr><tr><td>Quarot</td><td>27.4</td><td>33.9</td><td>39.1</td><td>36.2</td><td>30.0</td><td>56.9</td><td>34.7</td><td>53.0</td><td>38.9</td></tr><tr><td>Kurtail</td><td>28.7</td><td>37.2</td><td>38.8</td><td>42.9</td><td>31.6</td><td>60.0</td><td>35.7</td><td>57.5</td><td>41.5</td></tr><tr><td rowspan="3">Llama-3.2-3B</td><td>Vanilla</td><td>46.0</td><td>71.7</td><td>73.2</td><td>73.6</td><td>43.0</td><td>77.5</td><td>47.0</td><td>69.7</td><td>62.7</td></tr><tr><td>Quarot</td><td>33.1</td><td>50.3</td><td>41.8</td><td>56.3</td><td>31.8</td><td>67.8</td><td>39.8</td><td>56.8</td><td>47.2</td></tr><tr><td>Kurtail</td><td>37.4</td><td>56.6</td><td>48.0</td><td>62.1</td><td>36.6</td><td>71.3</td><td>40.5</td><td>60.4</td><td>51.6</td></tr></table>
+
+Table 15: Performance of different quantization methods on LLaMA-2 models across multiple benchmarks. WikiText reports perplexity (lower is better, ↓). Other benchmarks report accuracy (higher is better, ↑). The Avg. column is the mean of all accuracies except WikiText.  
+
+<table><tr><td>LLaMA-2 7B</td><td>Wiki(↓)</td><td>ARC-C</td><td>ARC-E</td><td>BoolQ</td><td>HellaSwag</td><td>PIQA</td><td>WinoGrande</td><td>Avg. ↑</td></tr><tr><td>Vanilla</td><td>5.5</td><td>46.2</td><td>74.5</td><td>77.8</td><td>76.0</td><td>79.1</td><td>69.1</td><td>70.5</td></tr><tr><td>Quarot</td><td>6.2</td><td>41.6</td><td>70.6</td><td>73.2</td><td>72.1</td><td>76.9</td><td>65.2</td><td>66.6</td></tr><tr><td>SpinQuant</td><td>6.0</td><td>43.6</td><td>71.3</td><td>73.8</td><td>73.2</td><td>76.0</td><td>65.4</td><td>67.2</td></tr><tr><td>DuQuant</td><td>6.1</td><td>37.5</td><td>50.0</td><td>69.2</td><td>69.7</td><td>75.7</td><td>63.9</td><td>61.0</td></tr><tr><td>OSTQuant</td><td>5.9</td><td>42.9</td><td>72.5</td><td>74.7</td><td>73.1</td><td>77.4</td><td>66.8</td><td>67.9</td></tr><tr><td>QUIK</td><td>7.5</td><td>43.7</td><td>68.0</td><td>71.3</td><td>73.3</td><td>75.7</td><td>64.6</td><td>66.1</td></tr><tr><td>KurTail</td><td>5.9</td><td>43.1</td><td>72.0</td><td>72.0</td><td>73.2</td><td>76.6</td><td>66.8</td><td>67.3</td></tr></table>
+
+Table 16: Evaluation on Perplexity and Reasoning Benchmarks. Lower is better for Perplexity, higher is better for all other metrics.  
+
+<table><tr><td>Method</td><td>Wiki</td><td>ARC-c</td><td>ARC-e</td><td>BoolQ</td><td>HellaS</td><td>OBQA</td><td>PIQA</td><td>SIQA</td><td>WinoG</td><td>Avg</td></tr><tr><td>16-bit</td><td>5.3</td><td>58.8</td><td>79.4</td><td>85.4</td><td>82.9</td><td>45.4</td><td>81.9</td><td>55.3</td><td>75.8</td><td>70.6</td></tr><tr><td>RTN</td><td>2715.0</td><td>21.6</td><td>32.7</td><td>51.5</td><td>29.6</td><td>25.8</td><td>52.6</td><td>33.2</td><td>51.7</td><td>37.3</td></tr><tr><td>GPTQ</td><td>5100.3</td><td>23.8</td><td>29.1</td><td>47.7</td><td>30.1</td><td>27.6</td><td>51.3</td><td>34.6</td><td>51.2</td><td>36.9</td></tr><tr><td>SmoothQuant+</td><td>1375.7</td><td>27.0</td><td>26.3</td><td>38.0</td><td>26.8</td><td>29.2</td><td>51.6</td><td>32.4</td><td>49.3</td><td>35.1</td></tr><tr><td>QUIK</td><td>10.5</td><td>45.0</td><td>67.1</td><td>64.7</td><td>68.9</td><td>37.6</td><td>74.8</td><td>43.9</td><td>59.3</td><td>57.7</td></tr><tr><td>QuaRot</td><td>6.8</td><td>54.8</td><td>79.6</td><td>79.9</td><td>78.7</td><td>44.0</td><td>79.5</td><td>49.9</td><td>70.7</td><td>67.1</td></tr><tr><td>KurTail</td><td>6.6</td><td>55.3</td><td>80.8</td><td>81.3</td><td>79.3</td><td>42.4</td><td>79.4</td><td>50.4</td><td>69.1</td><td>67.2</td></tr></table>
+
+Table 17: MMLU Performance Across Categories. Higher is better.  
+
+<table><tr><td>Method</td><td>Humanities</td><td>Other</td><td>SocialS</td><td>STEM</td><td>Avg.</td></tr><tr><td>16-bit</td><td>69.9</td><td>81.9</td><td>86.2</td><td>76.5</td><td>78.6</td></tr><tr><td>RTN</td><td>25.3</td><td>23.2</td><td>26.0</td><td>25.3</td><td>24.9</td></tr><tr><td>GPTQ</td><td>25.1</td><td>24.7</td><td>25.1</td><td>24.3</td><td>24.8</td></tr><tr><td>SmoothQuant+</td><td>25.9</td><td>24.5</td><td>22.2</td><td>22.2</td><td>23.7</td></tr><tr><td>QUIK</td><td>48.9</td><td>61.1</td><td>64.7</td><td>51.5</td><td>56.6</td></tr><tr><td>QuaRot</td><td>60.9</td><td>75.1</td><td>80.2</td><td>67.3</td><td>70.9</td></tr><tr><td>KurTail</td><td>61.9</td><td>76.5</td><td>81.0</td><td>68.0</td><td>71.9</td></tr></table>

2025/KurTail _ Kurtosis-based LLM Quantization/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:84b1565a773b3e828e744f8d5a3e98128f3591d81d317855f271f97e6704ec6d
+size 1232743

2025/LAGCL4Rec_ When LLMs Activate Interactions Potential in Graph Contrastive Learning for Recommendation/c5d15e62-530f-4ac3-b25e-1c7904eb8e11_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f467a15f7c47c08ab2d8cea7917c03edc8ad2459627d6279839fc011dfa0e950
+size 9042859

2025/LAGCL4Rec_ When LLMs Activate Interactions Potential in Graph Contrastive Learning for Recommendation/full.md

@@ -0,0 +1,1085 @@
+# LAGCL4Rec: When LLMs Activate Interactions Potential in Graph Contrastive Learning for Recommendation
+
+Leqi Zheng $^{1}$ , Chaokun Wang $^{*1}$ , Canzhi Chen $^{2}$ , Jiajun Zhang $^{3}$ , Cheng Wu $^{1}$ , Zixin Song $^{1}$ , Shanan Yan $^{1}$ , Ziyang Liu $^{1}$ , Hongwei Li $^{1}$
+
+Tsinghua University
+
+$^{2}$ Beijing Institute of Technology
+
+<sup>3</sup>University of Science and Technology of China
+
+{zhenglq24, wuc22, songzx24, ysn24, liu-zy21, lhw23}  $@$  mails.tsinghua.edu.cn
+
+chencanzhi@bit.edu.cn, zhangjiajun519@mail.ustc.edu.cn
+
+* Correspondence: chaokun@tsinghua.edu.cn
+
+# Abstract
+
+A core barrier preventing recommender systems from reaching their full potential lies in the inherent limitations of user-item interaction data: (1) Sparse user-item interactions, making it difficult to learn reliable user preferences; (2) Traditional contrastive learning methods often treat negative samples as equally hard or easy, ignoring the informative semantic difficulty during training. (3) Modern LLM-based recommender systems, on the other hand, discard all negative feedback, leading to unbalanced preference modeling. To address these issues, we propose LAGCL4Rec, a framework leveraging Large Language Models to Activate interactions in Graph Contrastive Learning for Recommendation. Our approach operates through three stages: (i) Data-Level: augmenting sparse interactions with balanced positive and negative samples using LLM-enriched profiles; (ii) Rank-Level: assessing semantic difficulty of negative samples through LLM-based grouping for fine-grained contrastive learning; and (iii) Rerank-Level: reasoning over augmented historical interactions for personalized recommendations. Theoretical analysis proves that LAGCL4Rec achieves effective information utilization with minimal computational overhead. Experiments across multiple benchmarks confirm our method consistently outperforms state-of-the-art baselines.
+
+# 1 Introduction
+
+Large Language Models (LLMs) have demonstrated remarkable capabilities in Natural Language Processing (NLP). The success of LLMs in NLP has encouraged researchers to explore their potential in recommendation tasks (Gao et al., 2023; Ren et al., 2023). Current LLM-based recommendation approaches generally fall into three categories: (a) Semantic Augmentation using LLMs: using LLMs to enhance textual descriptions of users and items to provide contextual information (Sun et al., 2025;
+
+Liu et al., 2024a); (b) Feature Extractions using LLMs: using LLMs as feature extractors to generate item and user embeddings (Hu et al., 2024; Liao et al., 2023) for downstream models; and (c) Decision Making using LLMs: directly produce recommendation lists through prompting or finetuning on foundation models (Li et al., 2023a; Bao et al., 2023).
+
+Despite these advances, existing approaches fail to fully unleash LLMs' potential for recommendation due to three critical limitations: (1) Sparse User-Item Interactions: While (a) approaches augment positive interactions, they neglect the latent potential in negative interactions, leading to an imbalanced activation of interaction data (Wang et al., 2024a; Liu et al., 2024a). (2) Ignorance of Difficulty of Negative Samples in Contrastive Learning: (b) approaches face significant efficiency bottlenecks, while traditional Graph Contrastive Learning (GCL) methods fail to activate the semantic potential in interactions by treating all negative samples uniformly (Yu et al., 2023; Lin et al., 2022). (3) Neglect of Informative Negative Feedback in LLM Prompting: By discarding negative interactions, Current (c) methods fail to harness the complete spectrum of user preference signals (Yuan et al., 2023; Bao et al., 2023). Meanwhile, most of these research using prompting on general foundation models, ignoring the domain-specific nature of recommendation.
+
+This leads us to a question: How can we systematically unlock and activate the full potential of both positive and negative interactions throughout the entire recommendation pipeline to maximize preference modeling capabilities and recommendation performance?
+
+As an answer to this question, we introduce LAGCL4Rec, a novel approach that leverages LLMs to Activate interaction potential in Graph Contrastive Learning through a progressive pipeline operating at each level of the recomm-
+
+![](images/28ae50f74a2b0760ff7701f882455a9b91fc5f2c3ac00bdc32b7c5395cedec6d.jpg)  
+Data Level  
+Positive-Only Feedback  
+Balanced Pos/Neg Activation
+
+![](images/feda46de590cc63f93a30b47d9854ed30c66dd1cfba5e15a82ac2e7f367f408d.jpg)  
+Rank Level  
+No Semantic Grouping  
+Semantic-Aware Grouping
+
+![](images/8de51e9f5df8495c3262f2d219948601acc6a8045e4b568fb0a9951364e9bd1f.jpg)  
+Rerank Level  
+Ignored Negative History  
+Full-Spectrum Reasoning  
+Figure 1: Comparison between conventional methods (left) and LAGCL4Rec (right). Data Level: Traditional methods rely on positive feedback only, while our approach activates potential in both positive and negative interactions. Rank Level: Conventional methods treat all negatives uniformly, whereas LAGCL4Rec implements semantic-aware grouping to differentiate hard/easy negatives, unlocking their potential for contrastive learning. Rerank Level: Existing approaches ignore negative interaction potential, while our method activates both positive and negative profiles through structured reasoning, enabling full-spectrum personalization.
+
+dation process. Figure 1 contrasts our proposed solutions with existing research at each level. Our contributions include:
+
+1. Data-Level Activation: We leverage LLMs to create semantically rich user and item profiles. Our approach activates both positive preferences and previously untapped negative interaction signals, transforming sparse data into a balanced representation of the complete user preference spectrum. (Section 2.1).  
+2. Rank-Level Activation: We propose a semantic grouping strategy to group negative samples based on their semantic relationships, which enables assessing their semantic difficulty. We then apply fine-grained contrastive learning to selectively emphasize informative negatives. By leveraging semantic grouping and efficient preprocessing instead of directly modeling negative interactions, our method improves computational efficiency while maintaining discriminative power (Section 2.2).  
+3. Rerank-Level Activation: We design a
+
+reranking approach that activates the full potential of historical interaction data by simultaneously leveraging both positive and negative signals. By employing LLMs to reason over these comprehensively activated interactions, we enable personalized and explainable recommendations that fully capitalize on all available preference information. (Section 2.3).
+
+4. Theoretical and Empirical Validation: We provide rigorous theoretical analysis establishing how our pipeline successfully unlocks the potential in interaction data, along with extensive experiments across diverse domains demonstrating significant performance improvements. (Section 3 and Section 4).
+
+# 2 Methodology
+
+We introduce LAGCL4Rec, a novel framework that leverages LLMs to Activate Graph Contrastive Learning through a progressive activation pipeline operating at multiple levels of the recommendation process, as illustrated in Figure 2. Our approach systematically unleashes interaction potential via three complementary mechanisms: (1) Data-Level Activation (Section 2.1), (2) Rank-Level Activation (Section 2.2) and (3) Rerank-Level Activation (Section 2.3).
+
+# 2.1 Data-Level Activation
+
+As illustrated in Figure 2 (Data-Level), the data-level augmentation process involves (1) generating profiles for user and item to enhance semantic context for downstream recommendation, and (2) augmenting the interaction sequence to create a balanced representation with both positive and negative interactions.
+
+User and Item Profile Generation. We generate semantically rich profiles for users and items using LLM:
+
+$$
+\mathcal {P} _ {u} = L L M s \left(\mathcal {S} _ {u}, \mathcal {Q} _ {u}\right), \quad \mathcal {P} _ {v} = L L M s \left(\mathcal {S} _ {v}, \mathcal {Q} _ {v}\right) \tag {1}
+$$
+
+where  $\mathcal{P}_u$  and  $\mathcal{P}_v$  are generated user and item profiles,  $S_u$  and  $S_v$  are system prompts (detailed in Appendix A.1.1 and A.1.2), and  $\mathcal{Q}_u$  and  $\mathcal{Q}_v$  are query inputs containing user/item information.
+
+Positive Interaction Augmentation. For cold-start users with less than 30 interactions, we define candidate items  $\nu_{u}^{cand}$  (items with no interactions
+
+![](images/54d3980f52f29e0f9ddfc89c04a0cec9d583fe47d27a2aa9bd85bc35b046d0a9.jpg)  
+Figure 2: LAGCL4Rec's Progressive Activation Pipeline: Data-Level generates semantically-rich profiles to activate potential in user-item interactions. Rank-Level implements semantic-aware grouping for differentiating hard/easy negatives with adaptive contrastive learning. Rerank-Level leverages both positive and negative historical interactions through structured reasoning for refined recommendations.
+
+![](images/39b6c25c7aadadb52755c73bdd30087a5af1907d457ef83d3ec1c5fc72a281aa.jpg)
+
+for user  $u$  in dataset) and sample 1024 items as  $\nu_{u}^{sample}$ . We score each item from  $\nu_{u}^{sample}$  using:
+
+$$
+s _ {u v} ^ {p o s} = L L M s (\mathcal {P} _ {u}, \mathcal {P} _ {v}, \mathcal {S} _ {p o s}) \qquad (2)
+$$
+
+where  $s_{uv}^{pos}$  is the preference score and  $S_{pos}$  is the prompt template detailed in Appendix A.1.3. We select the top 2% of items to form positive interactions  $\mathcal{R}_u^{pos}$  (likely positive preferences).
+
+Negative Interaction Augmentation. We augment negative interactions by similarly constructing  $\mathcal{V}_u^{neg - cand}$  and scoring items from candidate set:
+
+$$
+s _ {u v} ^ {n e g} = L L M s \left(\mathcal {P} _ {u}, \mathcal {P} _ {v}, \mathcal {S} _ {n e g}\right) \tag {3}
+$$
+
+where  $s_{uv}^{neg}$  is the dislike score and  $S_{neg}$  is the prompt template in Appendix A.1.4. We select the top 2% as negative interactions  $\mathcal{R}_u^{neg}$  (likely negative preferences). This creates truly representative negative interactions rather than random items.
+
+The final enhanced set combines both kinds of preference signals:
+
+$$
+\mathcal {R} _ {u} ^ {\text {e n h a n c e d}} = \mathcal {R} _ {u} ^ {\text {p o s}} \cup \mathcal {R} _ {u} ^ {\text {n e g}} \tag {4}
+$$
+
+# 2.2 Rank-Level Activation
+
+Unlike traditional sign-aware methods that directly model negative interactions with high computational overhead, we propose an efficient approach that indirectly leverages these signals through preprocessing and semantic grouping. This strategy significantly reduces computational complexity
+
+while maintaining recommendation effectiveness. Our key insight is that inferring dislikes within semantic categories (such as an orange lover disliking bananas) is harder and more informative than inferring dislikes for random objects. Therefore, we propose a contrastive learning approach based on semantic difficulty, as illustrated in Figure 2 (Rank-Level).
+
+LLM-Guided Semantic Grouping. We use LLMs to identify semantically similar users and items:
+
+$$
+\mathcal {G} _ {i} ^ {\text {u s e r}} = L L M s \left(\mathcal {P} _ {\mathcal {U} _ {i}}, \mathcal {S} _ {\text {g r o u p}}\right) \tag {5}
+$$
+
+where  $\mathcal{G}_i^{user}$  represents semantic groups of users,  $\mathcal{P}_{\mathcal{U}_i}$  is the set of user profiles, and  $S_{group}$  is the grouping prompt (Appendix A.2.1). We compute similarity between groups:
+
+$$
+\operatorname {s i m} \left(G _ {a}, G _ {b}\right) = L L M s \left(G _ {a}, G _ {b}, S _ {\text {s i m}}\right) \tag {6}
+$$
+
+where  $sim(G_a, G_b)$  quantifies group similarity and  $S_{sim}$  is detailed in Appendix A.2.2. Groups with similarity above threshold  $\tau_{sim}$  are merged to form collections  $\mathcal{G}^{user}$  and  $\mathcal{G}^{item}$ .
+
+Semantic Difficulty Assessment. Hard negatives  $(\mathcal{N}_h)$  are negative samples belonging to the same semantic group as the positive item  $i$ , defined as
+
+$$
+\mathcal {N} _ {h} (u, i) = \{j \mid j \in \mathcal {V}, (u, j) \notin \mathcal {R} ^ {\text {p o s}}, g (j) = g (i) \} \tag {7}
+$$
+
+Easy negatives  $(\mathcal{N}_e)$  are negative samples from other groups, defined as
+
+$$
+\mathcal {N} _ {e} (u, i) = \{j \mid j \in \mathcal {V}, (u, j) \notin \mathcal {R} ^ {\text {p o s}}, g (j) \neq g (i) \} \tag {8}
+$$
+
+Here,  $\mathcal{V}$  is the item set,  $\mathcal{R}^{pos}$  denotes positive interactions, and  $g(j)$  maps item  $j$  to its group. Similar definitions apply for item-user pairs.
+
+Difficulty-Aware Contrastive Learning. To enable detailed and differentiated supervision using negative samples of varying difficulty levels, we propose two difficulty-aware loss functions.
+
+For preference modeling, we propose LABPR loss, a difficulty-weighted version of BPR loss:
+
+$$
+\begin{array}{l} \mathcal {L} _ {l a b p r} = - \sum_ {(u, i, j _ {e}) \in D _ {e}} \ln \sigma (r _ {u i} - r _ {u j _ {e}}) \\ - \sum_ {(u, i, j _ {h}) \in D _ {h}} w _ {h} \cdot \ln \sigma \left(r _ {u i} - r _ {u j _ {h}}\right) \tag {9} \\ \end{array}
+$$
+
+where  $D_{e}$  and  $D_{h}$  contain easy and hard negative triplets,  $r_{ui}$  is predicted preference, and  $w_{h} > 1$  amplifies the emphasis on hard negatives.
+
+For clearer and more informative signals, we propose an adaptive-temperature InfoNCE loss for difficulty-aware contrastive learning:
+
+$$
+\mathcal {L} _ {\text {l a i n f o}} = - \sum_ {i = 1} ^ {N} \log \frac {\exp \left(\operatorname {s i m} \left(v _ {i} , v _ {i} ^ {+}\right) / \tau_ {p}\right)}{Z _ {i}} \tag {10}
+$$
+
+where  $\text{sim}(v_i, v_j)$  is cosine similarity and the denominator  $Z_i$  is:
+
+$$
+\begin{array}{l} Z _ {i} = \exp (s i m \left(v _ {i}, v _ {i} ^ {+}\right) / \tau_ {p}) + \\ \sum_ {j \neq i, g _ {j} \neq g _ {i}} \exp \left(\sin \left(v _ {i}, v _ {j}\right) / \tau_ {e}\right) + \tag {11} \\ \sum_ {j \neq i, g _ {j} = g _ {i}} w _ {h} \cdot \exp (s i m (v _ {i}, v _ {j}) / \tau_ {h}) \\ \end{array}
+$$
+
+We set  $\tau_h < \tau_e \leq \tau_p$  to enhance hard negatives, where  $\tau_p, \tau_e$ , and  $\tau_h$  are temperature parameters for positive, easy negative, and hard negative pairs. The total loss is computed as:
+
+$$
+\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {l a b p r}} + \lambda_ {c l} \cdot \left(\mathcal {L} _ {\text {l a i n f o}} ^ {\text {u s e r}} + \mathcal {L} _ {\text {l a i n f o}} ^ {\text {i t e m}}\right) + \lambda_ {\text {r e g}} \cdot \mathcal {L} _ {\text {r e g}} \tag {12}
+$$
+
+where  $\lambda_{cl}$  controls contrastive learning strength,  $\lambda_{reg}$  is regularization weight, and  $\mathcal{L}_{reg} = \sum_{\theta \in \Theta}||\theta ||_2^2$  is an L2 regularization term.
+
+# 2.3 Rerank-Level Activation
+
+To unleash the full potential of LLM in recommendation, we propose an LLM-based reasoning model as reranker. The reranker reason over the augmented historic interactions with positive and negative samples, and is optimized using a task-specific binary classification loss. as illustrated in Figure 2 (Rerank-Level).
+
+Historical Interaction Integration. For each  $u$ , we assemble an input sequence reflecting their historical interactions:
+
+$$
+\mathcal {I} _ {u} = \left\{\mathcal {P} _ {u}, \mathcal {C} _ {u}, \mathcal {R} _ {u} ^ {\text {p o s - h i s t}}, \mathcal {R} _ {u} ^ {\text {n e g - h i s t}} \right\} \tag {13}
+$$
+
+where  $\mathcal{P}_u$  denotes the comprehensive user profile generated in the data augmentation stage,  $\mathcal{C}_u$  represents the set of candidate items,  $\mathcal{R}_u^{pos - hist}$  contains the user's historical positive interactions,  $\mathcal{R}_u^{neg - hist}$  contains historical negative interactions.
+
+Training Phase: Binary Classification. To effectively leverage both positive and negative historical interactions, we formulate reranking as a binary classification for individual item relevance:
+
+$$
+p _ {u v} = f _ {\theta_ {r e r}} \left(\mathcal {P} _ {u}, v, \mathcal {R} _ {u} ^ {\text {p o s - h i s t}}, \mathcal {R} _ {u} ^ {\text {n e g - h i s t}}, \mathcal {S} _ {\text {c o t}}\right) \tag {14}
+$$
+
+where  $f_{\theta_{rer}}$  represents our reranker parameterized by  $\theta_{rer}$  (auto-regressive language model with binary classification head),  $v$  is a candidate item, and  $S_{cot}$  is the Chain-of-Thought prompt template to elicit reasoning in reranking (detailed in Appendix A.3).
+
+We train the model using binary cross-entropy loss on a dataset of user-item pairs:
+
+$$
+\begin{array}{l} \mathcal {D} _ {r e r} = \left\{\left(u, v, y _ {u v}\right) \mid u \in \mathcal {U} _ {\mathrm {t r a i n}}, v \in \mathcal {C} _ {u}, \right. \\ y _ {u v} = \mathbb {1} [ (u, v) \in \mathcal {R} _ {\text {t r a i n}} ] \}, \tag {15} \\ \end{array}
+$$
+
+where  $\mathcal{D}_{rer}$  is the training dataset, and  $y_{uv}$  is a binary indicator of whether the interaction  $(u,v)$  appears in the train set  $\mathcal{R}_{train}$ . The optimization objective is:
+
+$$
+\begin{array}{l} \mathcal{L}_{bce} = \frac{1}{|\mathcal{D}_{rer}|}\sum_{(u,v,y_{uv})\in \mathcal{D}_{rer}}[-y_{uv}\log (p_{uv}) \\ - \left(1 - y _ {u v}\right) \log \left(1 - p _ {u v}\right) ] \tag {16} \\ \end{array}
+$$
+
+$$
+\theta_ {r e r} ^ {*} = \arg \min  _ {\theta_ {r e r}} \mathcal {L} _ {b c e} \left(\theta_ {r e r}; \mathcal {D} _ {r e r}\right) \tag {17}
+$$
+
+Inference Phase: List-wise Reranking. During inference, we apply the reranking to generate a personalized recommendation result  $\mathcal{L}_u^{rerank}$  for each user. This process activates historical interaction signals through a structured reasoning process:
+
+$$
+\mathcal {L} _ {u} ^ {r e r a n k} = g _ {\theta_ {r e r} ^ {*}} \left(\mathcal {I} _ {u}, \mathcal {S} _ {c o t}\right) \tag {18}
+$$
+
+where  $g_{\theta_{re}^*}$  represents the application of our trained model to produce a reordered list of the candidate items.
+
+Score Fusion Strategy. We combine the original collaborative filtering scores with the model-generated relevance scores:
+
+$$
+s _ {u j} ^ {\text {f i n a l}} = \alpha \cdot s _ {u j} ^ {\text {o r i g}} + (1 - \alpha) \cdot s _ {u j} ^ {\text {l l m}} \tag {19}
+$$
+
+where  $s_{uj}^{orig}$  is the original score from the rank-level collaborative filtering model,  $s_{uj}^{llm}$  is the normalized score derived from the trained model, and  $\alpha \in [0,1]$  is a weighting parameter that balances these two signals. The model score is normalized using softmax:
+
+$$
+s _ {u j} ^ {l l m} = \frac {\exp \left(\tilde {s} _ {u j} ^ {l l m} / \tau\right)}{\sum_ {v _ {l} \in \mathcal {C} _ {u}} \exp \left(\tilde {s} _ {u l} ^ {l l m} / \tau\right)} \tag {20}
+$$
+
+where  $\tilde{s}_{uj}^{llm}$  is the raw relevance score assigned by the model and  $\tau$  is a temperature parameter controlling the sharpness of the distribution. The final ranked list is obtained by:
+
+$$
+\mathcal {L} _ {u} = \operatorname {a r g s o r t} _ {v _ {j} \in \mathcal {C} _ {u}} \left(s _ {u j} ^ {\text {f i n a l}}\right) \tag {21}
+$$
+
+where  $\mathcal{L}_u$  represents the ordered list of items recommended to user  $u$ .
+
+# 3 Theoretical Analysis
+
+This section provides a theoretical analysis of our pipeline, focusing on the information gain guaranteed by hard negative samples.
+
+# 3.1 Analysis of Rank-Level Activation
+
+The separate treatment of easy and hard negative interactions based on semantic grouping enhances the effectiveness of ranker training. Proposition below theoretically demonstrates this point.
+
+Proposition 1 (Hard-negative gradient dominance). Under Eqs. (9) and (10)-(11) with  $w_{h} > 1$  and  $\tau_{h} < \tau_{e} \leq \tau_{p}$ , the absolute gradient w.r.t. a hard negative similarity exceeds that of an easy negative whenever  $\text{sim}(v_{i}, v_{j_{h}}) > 0$ . Consequently, hard negatives tighten the InfoNCE mutual-information lower bound faster than easy negatives.
+
+This justifies our higher weight  $w_{h} > 1$  for hard negative interactions. Our analysis in Appendix B.2 shows that under the LABPR loss, hard negative gradients are amplified by a factor of  $w_{h}$ , ensuring faster convergence for challenging cases. Similarly, with our adaptive temperature mechanism (Appendix B.3), the model converges to a more discriminative embedding space.
+
+# 3.2 Analysis of Rerank-Level Activation
+
+For the reranking component, we establish:
+
+Proposition 2 (CoT features do not decrease task information). Let  $Y$  be the relevance label,  $X$  the base input, and  $Z = \mathrm{CoT}(X; S_{\mathrm{cot}})$ . Then
+
+$$
+I (Y; X, Z) = I (Y; X) + I (Y; Z \mid X) \geq I (Y; X), \tag {22}
+$$
+
+with strict inequality if  $I(Y;Z\mid X) > 0$ .
+
+In Appendix B.4, we prove there exists an optimal weighting parameter  $\alpha^{*} \in [0,1]$  that minimizes expected ranking error by balancing collaborative filtering and LLM signals. Furthermore, our reranking model trained with BCE loss converges to a stable solution when the candidate set distribution remains consistent between training and inference (Appendix B.5).
+
+# 4 Experiments
+
+# 4.1 Experimental Settings
+
+# 4.1.1 Datasets
+
+We evaluate LAGCL4Rec on many public recommendation datasets: ML-1M, Amazon-book, Yelp, and Steam. Following standard practices (Seo et al., 2022; Wang et al., 2024b; Chen et al., 2024), we consider ratings more than 3 as positive interactions and ratings no more than 3 as negative interactions for datasets, while all interactions in Steam are treated as positive due to absence of explicit ratings. We apply 5-core filtering and divide each dataset into training, validation, and testing sets in a 7:1:2 ratio. Table 1 summarizes the dataset statistics. We adopt the all-rank protocol for comprehensive and unbiased evaluation. Two widely used ranking metrics are employed: Recall@N and NDCG@N, where N=20 for main results.
+
+# 4.1.2 Baseline Models
+
+We compare LAGCL4Rec with representative methods from four categories:
+
+(1) CF Models: MF (Koren et al., 2009), NCF (He et al., 2017), NGCF (Wang et al., 2019),
+
+Table 1: Statistics of the experimental datasets.  
+
+<table><tr><td>Dataset</td><td>#Users</td><td>#Items</td><td>#Interactions</td><td>Density</td></tr><tr><td>ML-1M</td><td>6,040</td><td>3,952</td><td>1,000,209</td><td>4.2e-2</td></tr><tr><td>Amazon-book</td><td>11,000</td><td>9,332</td><td>120,464</td><td>1.2e-3</td></tr><tr><td>Yelp</td><td>11,091</td><td>11,010</td><td>166,620</td><td>1.4e-3</td></tr><tr><td>Steam</td><td>23,310</td><td>5,237</td><td>316,190</td><td>2.6e-3</td></tr></table>
+
+LightGCN (He et al., 2020), and SelfGNN (Liu et al., 2024b), representing classic and advanced collaborative filtering approaches.
+
+(2) Sign-aware Models: SGFormer (Wu et al., 2024b), SIGFormer (Chen et al., 2024), and NFARec (Wang et al., 2024b), state-of-the-art methods that specifically model signed interactions.  
+(3) LLM-based Models: RLMRec (Ren et al., 2023), LLM4Rerank (Gao et al., 2025), and RankGPT (Sun et al., 2023), recent approaches leveraging large language models.  
+(4) GCL-based Models: SGL (Wu et al., 2021), LightGCL (Cai et al.), and XSimGCL (Yu et al., 2023).
+
+# 4.1.3 Implementation Details
+
+We implement all models with embedding dimension 64. For all experiments, we use Claude 3.7 API. For rerank-level model training, we fine-tune Qwen-2.5 7B Instruct to perform the reranking task. Experiments were conducted on 8 NVIDIA A800 GPUs (80GB memory each).
+
+# 4.2 Overall Performance
+
+Table 2 compares different recommendation models across four datasets. Our LAGCL4Rec consistently enhances all graph contrastive learning backbones, with improvements ranging from  $6.79\%$  to  $17.42\%$  on Recall@20 and  $1.06\%$  to  $15.43\%$  on NDCG@20. The most substantial gains appear on the Yelp dataset with XSimGCL backbone (17.42% R@20 improvement), suggesting our approach excels at activating sparse signals in diverse categorical datasets.
+
+LAGCL4Rec variants outperform traditional CF methods by large margins and surpass specialized sign-aware and LLM-based approaches in most cases. XSimGCL+LAGCL4Rec achieves the best overall performance. The strong improvements in NDCG@20 for Amazon-book (consistently  $15\%$ ) indicate enhanced ranking quality for sparse datasets through our systematic activation of dormant interaction signals.
+
+# 4.3 Runtime Comparison
+
+Time Complexity Analysis The per-epoch computational complexity of LAGCL4Rec with different baseline GCL models is summarized in Table 4. Here,  $|E^{+}|$  denotes the number of positive edges,  $L$  the number of graph convolution layers,  $d$  the embedding dimension,  $M$  the total number of nodes, and  $B$  the batch size. Additional parameters include  $a$  (augmentation views in SGL) and  $q$  (singular values in LightGCL).
+
+Faster Ranker Training As shown in Figure 3, LAGCL4Rec introduces minimal per-epoch overhead (1.61-2.24% increase) while substantially reducing total training time (58.7-63.2% reduction) for all tested architectures (XSimGCL, LightGCL, and SGL). This favorable efficiency profile aligns with our theoretical complexity analysis, which predicted that our rank-level activation would add minimal computational overhead while enabling more efficient signal extraction.
+
+![](images/44bb7939fbc8854ff206cff52b74c245b05b529797e48a80becf31e8a1ccf850.jpg)  
+Figure 3: Rank-level runtime comparison between baseline GCL models and LAGCL4Rec variants (excluding rerank-level operations): (a) time per epoch showing minimal computational overhead; (b) total training time demonstrating significant efficiency improvements due to faster convergence.
+
+![](images/7221abf40240c75fffa0aa00a8e0bcfef93ecaa07527e05a1888ccea72496831.jpg)
+
+![](images/65e026b67d3af8b589cf4a337d5ccc11df3d2681448ef8885c4f32aa535489ed.jpg)  
+Figure 4: Impact of candidate set size  $(\mathcal{C}_u)$  on LAGCL4Rec performance across datasets, showing Recall@20 (left) and NDCG@20 (right).
+
+# 4.4 Ablation Study
+
+Component Analysis. We evaluate each component's contribution in LAGCL4Rec (Table 3). Results show that removing any component leads to performance degradation across all datasets. The
+
+Table 2: Performance comparison of different recommendation models across four datasets. The best results are in **bold**, the strongest baseline performances are **underlined**, and * indicates statistical significance with  $p < 0.01$ . R@20 and N@20 represent Recall@20 and NDCG@20 metrics, respectively.  
+
+<table><tr><td rowspan="2">Category</td><td>Dataset</td><td colspan="2">ML-1M</td><td colspan="2">Amazon-book</td><td colspan="2">Yelp</td><td colspan="2">Steam</td></tr><tr><td>Model</td><td>R@20</td><td>N@20</td><td>R@20</td><td>N@20</td><td>R@20</td><td>N@20</td><td>R@20</td><td>N@20</td></tr><tr><td rowspan="5">CF</td><td>MF</td><td>0.1329</td><td>0.1988</td><td>0.0446</td><td>0.0397</td><td>0.0451</td><td>0.0259</td><td>0.0387</td><td>0.0213</td></tr><tr><td>NCF</td><td>0.1501</td><td>0.2102</td><td>0.0693</td><td>0.0474</td><td>0.0637</td><td>0.0408</td><td>0.0603</td><td>0.0448</td></tr><tr><td>NGCF</td><td>0.1630</td><td>0.2185</td><td>0.0809</td><td>0.0663</td><td>0.0744</td><td>0.0395</td><td>0.0532</td><td>0.0406</td></tr><tr><td>LightGCN</td><td>0.1993</td><td>0.2632</td><td>0.0918</td><td>0.0699</td><td>0.0769</td><td>0.0512</td><td>0.0749</td><td>0.0632</td></tr><tr><td>SelfGNN</td><td>0.2565</td><td>0.2810</td><td>0.1024</td><td>0.0757</td><td>0.0795</td><td>0.0603</td><td>0.0954</td><td>0.0739</td></tr><tr><td rowspan="3">Sign-RS</td><td>SGFormer</td><td>0.1877</td><td>0.2680</td><td>0.1102</td><td>0.0703</td><td>0.0658</td><td>0.0504</td><td>0.1265</td><td>0.0803</td></tr><tr><td>SIGFormer</td><td>0.2995</td><td>0.3380</td><td>0.1428</td><td>0.1045</td><td>0.0959</td><td>0.0783</td><td>0.1448</td><td>0.0906</td></tr><tr><td>NFARec</td><td>0.2840</td><td>0.3212</td><td>0.1496</td><td>0.1032</td><td>0.1143</td><td>0.0876</td><td>0.1369</td><td>0.0884</td></tr><tr><td rowspan="3">LLM-RS</td><td>RLMRec</td><td>0.2966</td><td>0.3294</td><td>0.1543</td><td>0.0976</td><td>0.1279</td><td>0.0885</td><td>0.1467</td><td>0.0897</td></tr><tr><td>LLM4Rerank</td><td>0.3007</td><td>0.3316</td><td>0.1496</td><td>0.1033</td><td>0.1203</td><td>0.0892</td><td>0.1416</td><td>0.0916</td></tr><tr><td>RankGPT</td><td>0.2515</td><td>0.2801</td><td>0.1327</td><td>0.0852</td><td>0.1154</td><td>0.0857</td><td>0.1402</td><td>0.0893</td></tr><tr><td rowspan="10">GCL-RS</td><td>Backbone Variants</td><td colspan="8"></td></tr><tr><td>SGL base</td><td>0.2798</td><td>0.3037</td><td>0.1438</td><td>0.0904</td><td>0.1068</td><td>0.0847</td><td>0.1401</td><td>0.0881</td></tr><tr><td>LAGCL4Rec</td><td>0.2988</td><td>0.3279</td><td>0.1553</td><td>0.1041</td><td>0.1238</td><td>0.0856</td><td>0.1503</td><td>0.0927*</td></tr><tr><td>improv.</td><td>6.79%</td><td>7.97%</td><td>8.00%</td><td>15.15%</td><td>15.92%</td><td>1.06%</td><td>7.28%</td><td>5.22%</td></tr><tr><td>base LightGCL</td><td>0.273</td><td>0.3035</td><td>0.1484</td><td>0.0933</td><td>0.1092</td><td>0.0855</td><td>0.1322</td><td>0.0863</td></tr><tr><td>LAGCL4Rec</td><td>0.3035</td><td>0.3301</td><td>0.1598</td><td>0.1077*</td><td>0.1244</td><td>0.0882</td><td>0.1466</td><td>0.0899</td></tr><tr><td>improv.</td><td>11.17%</td><td>8.76%</td><td>7.68%</td><td>15.43%</td><td>13.92%</td><td>3.16%</td><td>10.89%</td><td>4.17%</td></tr><tr><td>base XSimGCL</td><td>0.2729</td><td>0.3087</td><td>0.1477</td><td>0.0925</td><td>0.1108</td><td>0.0849</td><td>0.1397</td><td>0.0876</td></tr><tr><td>LAGCL4Rec</td><td>0.3106*</td><td>0.3450*</td><td>0.1604*</td><td>0.1065</td><td>0.1301*</td><td>0.0903*</td><td>0.1507*</td><td>0.0924</td></tr><tr><td>improv.</td><td>13.81%</td><td>11.76%</td><td>8.60%</td><td>15.14%</td><td>17.42%</td><td>6.36%</td><td>7.87%</td><td>5.48%</td></tr></table>
+
+Table 3: Ablation study evaluating the contribution of each component in LAGCL4Rec. The table compares performance across four datasets when removing individual components: data-level, rank-level, and reranking-level. Checkmarks (√) indicate included components while cross marks (X) indicate removed components. R@20 and N@20 represent Recall@20 and NDCG@20, respectively.  
+
+<table><tr><td rowspan="2">Variant</td><td colspan="3">Components</td><td colspan="2">ML-1M</td><td colspan="2">Amazon-book</td><td colspan="2">Yelp</td><td colspan="2">Steam</td></tr><tr><td>data-level</td><td>rank-level</td><td>rerank-level</td><td>R@20</td><td>N@20</td><td>R@20</td><td>N@20</td><td>R@20</td><td>N@20</td><td>R@20</td><td>N@20</td></tr><tr><td>base</td><td>X</td><td>X</td><td>X</td><td>0.2729</td><td>0.3087</td><td>0.1477</td><td>0.0925</td><td>0.1108</td><td>0.0849</td><td>0.1397</td><td>0.0876</td></tr><tr><td>w/o data-level</td><td>X</td><td>✓</td><td>✓</td><td>0.2707</td><td>0.3111</td><td>0.1465</td><td>0.0944</td><td>0.1061</td><td>0.0873</td><td>0.1401</td><td>0.0899</td></tr><tr><td>w/o rank-level</td><td>✓</td><td>X</td><td>✓</td><td>0.2976</td><td>0.3345</td><td>0.1526</td><td>0.0989</td><td>0.1137</td><td>0.0884</td><td>0.1412</td><td>0.0907</td></tr><tr><td>w/o rerank-level</td><td>✓</td><td>✓</td><td>X</td><td>0.3011</td><td>0.3303</td><td>0.1581</td><td>0.0977</td><td>0.1232</td><td>0.0795</td><td>0.1479</td><td>0.0884</td></tr><tr><td>all</td><td>✓</td><td>✓</td><td>✓</td><td>0.3106</td><td>0.3450</td><td>0.1604</td><td>0.1065</td><td>0.1301</td><td>0.0903</td><td>0.1507</td><td>0.0924</td></tr></table>
+
+full model consistently achieves the best performance, with substantial gains on the Yelp dataset (17.42% improvement in R@20 compared to the base model). Notably, removing data augmentation sometimes reduces performance below the base model, highlighting the critical importance of our LLM-activated sparse interactions. These findings confirm each component's essential contribution to addressing the sparse interaction challenge.
+
+Candidate Set Size Analysis. Figure 4 illustrates how candidate set size  $(\mathcal{C}_u)$  affects reranker performance. Performance consistently improves as  $\mathcal{C}_u$  increases from 0 to 100, with significant gains in the 0 to 20 range. The Amazon-book dataset shows highest sensitivity for NDCG@20, suggesting complex domains benefit most from comprehensive candidate consideration. While Recall@20 varies across datasets, NDCG@20 shows consistent improvements, confirming that candidate diver
+
+sity maximizes the effectiveness of our interaction-history guided reranker.
+
+# 4.5 Empirical Analysis
+
+Convergence Speed Analysis. Figure 5 demonstrates LAGCL4Rec's dramatically accelerated convergence across all datasets. This rapid optimization empirically validates Proposition 3.1, which shows hard negatives tighten the MI lower bound faster than easy negatives. By differentially amplifying these high-value signals through our semantic-aware grouping strategy, our Progressive Activation Pipeline effectively enhances learning efficiency at multiple levels.
+
+Hyperparameter Analysis. Figure 6 shows sensitivity to key parameters, with optimal settings at fusion weight  $\alpha \approx 0.2$ , contrastive weight  $\lambda_{cl} = 0.15$ , and temperature ratios  $\tau_e / \tau_p = 0.45$  and  $\tau_h / \tau_p = 0.1$  across all datasets. The opti
+
+Table 4: Per-epoch time complexity of different LAGCL4Rec variants.  
+
+<table><tr><td>Component</td><td>LAGCL4Rec + SGL</td><td>LAGCL4Rec + LightGCL</td><td>LAGCL4Rec + XSimGCL</td></tr><tr><td>Baseline GCL Per-Epoch</td><td>O(2|E+|Ld + 4a|E+|Ld + 3Md)</td><td>O(2|E+|Ld + 2qMLd + 3Md)</td><td>O(2|E+|Ld + 3Md)</td></tr><tr><td>Rank-Level Addition</td><td>O(Bd)</td><td>O(Bd)</td><td>O(Bd)</td></tr><tr><td>Total Per-Epoch</td><td>O(2|E+|Ld + 4a|E+|Ld + 3Md + Bd)</td><td>O(2|E+|Ld + 2qMLd + 3Md + Bd)</td><td>O(2|E+|Ld + 3Md + Bd)</td></tr></table>
+
+![](images/a5fe115273947e147f48f928594e4f5002e33b5dbe408d97467d51900702e514.jpg)
+
+![](images/85beee07691081f5f6f9d35e562d70e62b4a3edb1343a97edda6ed2f50c38748.jpg)
+
+![](images/8fabf6f6a5aaea9b946c9704be7571d4cba632fa405d1e33a7969d37934f3434.jpg)
+
+![](images/c8a6b07f2c44042f1b7892400c6310e357c3457dd64d60178cb62a3702373c6e.jpg)
+
+![](images/b4dd40ff327675773b4e57b0e32f5076e38dc9c1b17f15f58cbf83f2ebd2b81b.jpg)  
+(a) ML-1M
+
+![](images/df4fbe4b451dcd3a53fa7e362807787fcd75a97e1e4c20e4d4b660e7cb1eed7a.jpg)  
+(b) Yelp
+
+![](images/15959efb026652551e0715db51663ac5401e21af0ffbc20d93493070312a1755.jpg)  
+(c) Steam
+
+![](images/ae96bb28c20807595bbc7f92a7d41bf5eee5fcb77f4f36495306d426b510352d.jpg)  
+(d) Amazon-book
+
+![](images/1a56d3474d4785017f75ecf33074726ba3650032c38b69bae588f7c112793d4c.jpg)  
+Figure 5: Performance across training epochs for LAGCL4Rec versus baseline models on four datasets.
+
+![](images/1d4875c82761b5db64d5ee134f7393bdbb7411cdfeb0a27a763843d208d78d58.jpg)
+
+![](images/ace67ad478862002212ea1258ffadc7cbb7cae1ff6400074d70e3bd8c2873c65.jpg)
+
+![](images/2d17844b8ef9e39d53f08f28d18b4027bcbf584bc1e1f3861e43dea1abf8b7cb.jpg)
+
+![](images/cc1c11f5f1fd54b0876d4550d17fa052cb7fa5fbe01f4ea80f82487642a331fd.jpg)  
+(a)Rerank weight  $\alpha$
+
+![](images/2142e69025ef57afef6626bfd32edf545dc380a77705ef3097d86a2f102db1c9.jpg)  
+(b) Contrastive weight  $\lambda_{cl}$
+
+![](images/d0d13e7ba74ba244678e37c620bf228b0f9a910fde5c6e9f4b64192f559f9e71.jpg)  
+(c) temp. ratio  $\tau_{e} / \tau_{p}$  
+Figure 6: Impact of key hyperparameters on ranking quality (decrease in Recall@20 & NDCG@20 relative to optimal settings).
+
+![](images/37adebfd0fe5844be78eb88f606ab77ff513646600f03d65403e5b8ac4b0093a.jpg)  
+(d) temp. ratio  $\tau_h / \tau_p$
+
+mal  $\alpha$  value confirms our theoretical prediction in Proposition 3.2 that an intermediate fusion weight balancing collaborative filtering and LLM signals yields the best performance. Similarly, the temperature ratio settings align with our analysis of how adaptive temperatures enhance discriminative power.
+
+User and Item Profiles Visualization. Figure 7 presents word clouds visualizing activated signals in user-item profiles. Across all domains, we observe the activation of both positive signals (terms like "enjoy," "appreciate") and negative signals (terms like "dislike," "negative review"), which aligns with our goal of awakening dormant preferences and illustrates how our approach captures domain-specific preference signals and activates
+
+relevant interactions.
+
+# 4.6 Case Study
+
+In Figure 8, we conducted detailed case analysis on the ML-1M dataset, examining a user with clear preferences for action/war films and aversions to horror. These bidirectional adjustments empirically validate Proposition 3.2, demonstrating how our approach effectively extracts higher information content from historical interactions compared to standard methods.
+
+# 5 Related Work
+
+Recommender systems have long been recognized as a canonical research problem and have attracted sustained attention from the academic
+
+![](images/4813e6871c239bb1f01f30b9a229eaf1b7d3a6bb5dfe3df996a5e9496d9f58ec.jpg)  
+(a) ML-1M User Profile
+
+![](images/3fee902cc0fe6e55a7c5c04f5e5b507ea1503d1431b27737b995a8c5a6e37a6e.jpg)  
+(e) Steam User Profile
+
+![](images/ec211ea0049feaa6bdf2fd0faae470f0246d0b048d306a83c247df19af9180de.jpg)  
+(b) ML-1M Item Profile
+
+![](images/4ac10b0f2308abb3d20b0e4d3ae31e1496263cf29328937ded82d2915c6dab1e.jpg)  
+(f) Steam Item Profile  
+Figure 8: Case study demonstrating LAGCL4Rec's recommendation process for a user (ID: 2814) and two candidate movies: Patton (1970, ID: 1272) and The Addiction (1995, ID: 152). Red highlights indicate activated positive preference signals, while purple highlights indicate activated negative preference signals. The right panels show score and ranking changes throughout the recommendation pipeline.
+
+![](images/f7bb3270f4b96e5793ea9ef688f7297e0f74d0851bc096da97055b45868618b8.jpg)  
+(c) Yelp User Profile
+
+![](images/d141b797a79d9d1381d2c3d31bc9d53cdaa9f5cfbb0e5456cd95920c54d9c545.jpg)  
+Figure 7: Word cloud visualizations of activated signals in user and item profiles from the datasets.
+
+![](images/deebcd1bcfba27f74d35eec887c690d45ad372f202cfd252afbe3d61fa1871c8.jpg)  
+(d) Yelp Item Profile
+
+![](images/67543aac79e764a284d471bb5df6ce29720103b098888ac1b78cbc4796e2f705.jpg)  
+(g) Amazon-book User Profile  
+(h) Amazon-book Item Profile
+
+# User Profile
+
+This user enjoys high-quality action thrillers with strong plots, particularly those featuring law enforcement, military, or espionage themes. He also appreciates well-crafted dramas, war films, and romantic comedies with intelligent humor. He dislikes horror films, pretentious art films, and comedies lacking substance.
+
+# Movie Profile
+
+Patton (1970) would appeal to history buffs, military enthusiasts, and viewers interested in character-driven biographical dramas. The film would resonate with older audiences, particularly men who appreciate war films that focus on leadership and strategy rather than just combat sequences. Viewers who enjoy complex character studies of controversial historical figures would find value in its nuanced portrayal of General Patton. Those who appreciate classic, well-crafted filmmaking with strong performances (particularly George C. Scott's Oscar-winning role) would also enjoy this film. The movie would appeal to audiences who prefer thoughtful, historically-grounded narratives over a pure action entertainment.
+
+# Movie Profile
+
+The Addiction (1995) would appeal to fans of intellectual horror, particularly those who appreciate philosophical themes explored through vampire mythology. Adult viewers interested in arthouse cinema black and white photography, and allegorical storytelling would find this film engaging. College students and academics studying philosophy, particularly existentialism, might appreciate its intellectual depth. Horror enthusiasts who prefer psychological dread over jump scares and those who enjoy vampire films that deviate from traditional tropes would be drawn to it. Viewers who appreciate director Abel Ferrara's distinctive style and fans of lead actress Lili Taylor would also find value in this unique horror experience.
+
+community (Wu et al., 2024a, 2025; Di et al., 2025a,b,c), and the advent of large language models (LLMs) (Touvron et al., 2023; Ila, 2023; Grattafori et al., 2024; Qwen et al., 2025) has infused new vitality into the field, opening up fresh opportunities for innovation in recommendation (Wang and Lim, 2023; He et al., 2023; Li et al., 2025; Liang et al., 2024; Zhang et al., 2024; Feng et al., 2024; Fan et al., 2023; Chen et al., 2023; Zhang et al., 2023; Ortega et al., 2024; Li et al., 2023b; Yang et al., 2023; Wang et al., 2023). While LLM-based approaches have shown promise, they typically overlook dormant negative signals.
+
+Graph Neural Networks (GNNs) have emerged
+
+as a highly active research frontier and are being applied across a wide spectrum of domains (Du et al., 2025a,b,c; Liu and Wang, 2025; Liu et al., 2024c), which have transformed recommender systems through their ability to model complex user-item relationships (Wu et al., 2023; Zheng et al., 2025a,b), but these methods treat all negative interactions uniformly rather than differentiating by difficulty level.
+
+Our work addresses these limitations through a progressive activation pipeline operating at multiple levels. Unlike existing sign-aware systems (Derr et al., 2018; Huang et al., 2023; Wang and Cao, 2021; Gong and Zhu, 2022; Wu et al., 2020; Zhao et al., 2018) that lack systematic activation mechanisms, we explicitly awaken dormant signals in raw data, differentially amplify them based on semantic difficulty, and leverage historical interactions through structured reasoning. While recent reranking approaches (Xiong et al., 2023; Carraro and Bridge, 2024; Zhang et al., 2025; Wang et al., 2025a,b) have explored LLM integration, they fail to provide the comprehensive activation strategy we introduce.
+
+# 6 Conclusion
+
+We introduced LAGCL4Rec, a framework that activates the untapped potential in user-item interactions through our Progressive Activation Pipeline. By integrating LLMs with graph contrastive learning via Data-Level, Rank-Level, and Rerank-Level mechanisms, we addressed key limitations in existing approaches. Experiments across diverse datasets demonstrated significant performance improvements, highlighting the value of systematically unlocking and leveraging interaction potential.
+
+# Limitations
+
+While our experimental results demonstrate the effectiveness of LAGCL4Rec across multiple public datasets, a notable limitation of our work is the absence of validation in industrial recommendation systems with real-world deployment. Industrial environments often present additional challenges including larger-scale data, more complex user-item interactions, and stricter latency requirements that may affect the practical applicability of our approach. Future work should focus on evaluating and adapting LAGCL4Rec for industrial deployment to comprehensively validate its effectiveness under production constraints.
+
+# Ethical Considerations
+
+Our research exclusively uses public benchmark datasets with properly anonymized data. The recommendation domains explored (movies, books, businesses, and games) involve no sensitive content categories. Our LLM prompting approach focuses solely on preference patterns without introducing demographic or social biases. No human subjects were involved, and our work complies with standard ethical practices in recommendation systems research.
+
+# Acknowledgments
+
+This work is supported in part by the National Natural Science Foundation of China (No. 92467203 and No. 62372264). Chaokun Wang is the corresponding author.
+
+# References
+
+2023. Llama 2: Open foundation and fine-tuned chat models. Preprint, arXiv:2307.09288.  
+Keqin Bao, Jizhi Zhang, Yupeng Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He. 2023. Tallrec: An effective and efficient tuning framework to align large language model with recommendation. arXiv preprint arXiv:2305.00447.  
+Xuheng Cai, Chao Huang, Lianghao Xia, and Xubin Ren. Lightgcl: Simple yet effective graph contrastive learning for recommendation. In The Eleventh International Conference on Learning Representations.  
+Diego Carraro and Derek Bridge. 2024. Enhancing recommendation diversity by re-ranking with large language models. ACM Transactions on Recommender Systems.
+
+Jiawei Chen, Zhi Liu, Xin Huang, Chuan Wu, Qinzheng Liu, Guangyao Jiang, Yihuai Pu, Yunqi Lei, Xiuying Chen, Xin Wang, and 1 others. 2023. When large language models meet personalization: Perspectives of challenges and opportunities. arXiv preprint arXiv:2307.16376.  
+Sirui Chen, Jiawei Chen, Sheng Zhou, Bohao Wang, Shen Han, Chanfei Su, Yuqing Yuan, and Can Wang. 2024. Sigformer: Sign-aware graph transformer for recommendation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1274-1284.  
+Tyler Derr, Yao Ma, and Jiliang Tang. 2018. Signed graph convolutional networks. In 2018 IEEE International Conference on Data Mining (ICDM), pages 929-934. IEEE.  
+Yicheng Di, Hongjian Shi, Ruhui Ma, Honghao Gao, Yuan Liu, and Weiyu Wang. 2025a. Fedrl: A reinforcement learning federated recommender system for efficient communication using reinforcement selector and hypernet generator. ACM Transactions on Recommender Systems, 4(1):1-31.  
+Yicheng Di, Hongjian Shi, Xiaoming Wang, Ruhui Ma, and Yuan Liu. 2025b. Federated recommender system based on diffusion augmentation and guided denoising. ACM Transactions on Information Systems, 43(2):1-36.  
+Yicheng Di, Xiaoming Wang, Hongjian Shi, Chongsheng Fan, Rong Zhou, Ruhui Ma, and Yuan Liu. 2025c. Personalized consumer federated recommender system using fine-grained transformation and hybrid information sharing. IEEE Transactions on Consumer Electronics.  
+Enjun Du, Xunkai Li, Tian Jin, Zhihan Zhang, RongHua Li, and Guoren Wang. 2025a. Graphmaster: Automated graph synthesis via llm agents in data-limited environments. ArXiv preprint arXiv:2504.00711v2.  
+Enjun Du, Siyi Liu, and Yongqi Zhang. 2025b. Graphoracle: A foundation model for knowledge graph reasoning. ArXiv preprint arXiv:2505.11125.  
+Enjun Du, Siyi Liu, and Yongqi Zhang. 2025c. Mixture of length and pruning experts for knowledge graphs reasoning. ArXiv preprint arXiv:2507.20498.  
+Wenqi Fan, Zhiwei Zhao, Jianping Li, Yingxu Liu, Xiaokai Mei, Yupeng Wang, Jiliang Tang, and Qiang Li. 2023. Recommender systems in the era of large language models (llms). arXiv preprint arXiv:2307.02046.  
+Shanshan Feng, Haoming Lyu, Fan Li, Zhu Sun, and Caishun Chen. 2024. Where to move next: Zero-shot generalization of llms for next poi recommendation. In 2024 IEEE Conference on Artificial Intelligence (CAI), pages 1530-1535. IEEE.
+
+Jingtong Gao, Bo Chen, Xiangyu Zhao, Weiwen Liu, Xiangyang Li, Yichao Wang, Wanyu Wang, Huifeng Guo, and Ruiming Tang. 2025. Llm4rerank: Llm-based auto-eranking framework for recommendations. In Proceedings of the ACM on Web Conference 2025, WWW '25, page 228-239, New York, NY, USA. Association for Computing Machinery.  
+Yunfan Gao, Tao Sheng, Yi Xiang, Yihong Xiong, Hao Wang, and Jia Zhang. 2023. Chat-rec: Towards interactive and explainable llms-augmented recommender system. arXiv preprint arXiv:2303.14524.  
+Shansan Gong and Kenny Q Zhu. 2022. Positive, negative and neutral: Modeling implicit feedback in session-based news recommendation. In Proceedings of the 45th international ACM SIGIR conference on research and development in information retrieval, pages 1185-1195.  
+Aaron Grattafori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, and 542 others. 2024. The llama 3 herd of models. Preprint, arXiv:2407.21783.  
+Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. 2020. Lightgen: Simplifying and powering graph convolution network for recommendation. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, pages 639-648.  
+Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. 2017. Neural collaborative filtering. In Proceedings of the 26th international conference on world wide web, pages 173-182.  
+Zhankui He, Zheng Xie, Rahul Jha, Harald Steck, Dawen Liang, Yansen Feng, Bodhisattwa Prasad Majumder, Nathan Kallus, and Julian McAuley. 2023. Large language models as zero-shot conversational recommenders. arXiv preprint.  
+Jiabin Hu, Weijun Xia, Xing Zhang, Chang Fu, Wei Wu, Zheng Huan, Anfeng Li, Zhengkun Tang, and Jianlong Zhou. 2024. Enhancing sequential recommendation via llm-based semantic embedding learning. In Companion Proceedings of the ACM on Web Conference 2024, pages 103-111.  
+Junjie Huang, Ruobing Xie, Qi Cao, Huawei Shen, Shaoliang Zhang, Feng Xia, and Xueqi Cheng. 2023. Negative can be positive: Signed graph neural networks for recommendation. Information Processing & Management, 60(4):103403.  
+Yehuda Koren, Robert Bell, and Chris Volinsky. 2009. Matrix factorization techniques for recommender systems. Computer, 42(8):30-37.  
+Jiacheng Li, Ming Wang, Jin Li, Jinmiao Fu, Xin Shen, Jingbo Shang, and Julian McAuley. 2023a. Text is
+
+all you need: Learning language representations for sequential recommendation. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD '23, page 1258-1267, New York, NY, USA. Association for Computing Machinery.  
+Xiangyu Li, Chen Chen, Xiangyu Zhao, Yicui Zhang, and Chao Xing. 2023b. E4srec: An elegant effective efficient extensible solution of large language models for sequential recommendation. arXiv preprint arXiv:2312.02443.  
+Yunzhe Li, Junting Wang, Hari Sundaram, and Zhining Liu. 2025. A zero-shot generalization framework for llm-driven cross-domain sequential recommendation. arXiv preprint arXiv:2501.19232.  
+Yueqing Liang, Liangwei Yang, Chen Wang, Xiongxiao Xu, Philip S Yu, and Kai Shu. 2024. Taxonomy-guided zero-shot recommendations with llms. arXiv preprint arXiv:2406.14043.  
+Junling Liao, Sichun Li, Ze Yang, Jonathan Wu, Yang Yuan, Xing Wang, and Xiangnan He. 2023. Llara: Aligning large language models with sequential recommenders. arXiv preprint arXiv:2312.02445.  
+Zihan Lin, Changxin Tian, Yupeng Hou, and Wayne Xin Zhao. 2022. Improving graph collaborative filtering with neighborhood-enriched contrastive learning. In Proceedings of the ACM web conference 2022, pages 2320-2329.  
+Qijiong Liu, Nuo Chen, Tetsuya Sakai, and Xiao-Ming Wu. 2024a. Once: Boosting content-based recommendation with both open- and closed-source large language models. WSDM '24, page 452-461, New York, NY, USA. Association for Computing Machinery.  
+Yuxi Liu, Lianghao Xia, and Chao Huang. 2024b. Self-gnn: Self-supervised graph neural networks for sequential recommendation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1609-1618.  
+Ziyang Liu and Chaokun Wang. 2025. Terdy: Temporal relation dynamics through frequency decomposition for temporal knowledge graph completion. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 9611-9622.  
+Ziyang Liu, Chaokun Wang, Hao Feng, and Ziyang Chen. 2024c. Efficient unsupervised graph embedding with attributed graph reduction and dual-level loss. IEEE Transactions on Knowledge and Data Engineering.  
+Gustavo Mendonça Ortega, Rodrigo Ferrari de Souza, and Marcelo Garcia Manzato. 2024. Evaluating zero-shot large language models recommenders on popularity bias and unfairness: A comparative approach to
+
+traditional algorithms. In Simposio Brasileiro de Sistemas Multimédia e Web (WebMedia), pages 45-48. SBC.  
+Qwen, :: An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, and 25 others. 2025. Qwen2.5 technical report. Preprint, arXiv:2412.15115.  
+Xuran Ren, Wenjie Wei, Longhui Xia, Lianghao Su, Sensen Cheng, Jiantao Wang, Dawei Yin, and Chao Huang. 2023. Representation learning with large language models for recommendation. arXiv preprint arXiv:2310.15950.  
+Changwon Seo, Kyeong-Joong Jeong, Sungsu Lim, and Won-Yong Shin. 2022. Siren: Sign-aware recommendation using graph neural networks. IEEE Transactions on Neural Networks and Learning Systems, 35(4):4729-4743.  
+Weiwei Sun, Lingyong Yan, Xinyu Ma, Pengjie Ren, Dawei Yin, and Zhaochun Ren. 2023. Is chatgpt good at search? investigating large language models as re-ranking agent. arXiv preprint arXiv:2304.09542.  
+Yuqi Sun, Qidong Liu, Haiping Zhu, and Feng Tian. 2025. Llmser: Enhancing sequential recommendation via llm-based data augmentation. Preprint, arXiv:2503.12547.  
+Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, and 1 others. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971.  
+Chen Wang, Xiaokai Wei, Yexi Jiang, Frank Ong, Kevin Gao, Xiao Yu, Zheng Hui, Se-eun Yoon, Philip Yu, and Michelle Gong. 2025a. Solving the content gap in roblox game recommendations: Llm-based profile generation and reranking. arXiv preprint arXiv:2502.06802.  
+Chen Wang, Mingdai Yang, Zhiwei Liu, Pan Li, Linsey Pang, Qingsong Wen, and Philip Yu. 2025b. Automating personalization: Prompt optimization for recommendation reranking. arXiv preprint arXiv:2504.03965.  
+Jianling Wang, Haokai Lu, James Caverlee, Ed Chi, and Minmin Chen. 2024a. Large language models as data augmenters for cold-start item recommendation. Preprint, arXiv:2402.11724.  
+Lei Wang and Ee-Peng Lim. 2023. Zero-shot next-item recommendation using large pretrained language models. arXiv preprint arXiv:2304.03153.
+
+Wei Wang and Longbing Cao. 2021. Interactive sequential basket recommendation by learning basket couplings and positive/negative feedback. ACM Transactions on Information Systems (TOIS), 39(3):1-26.  
+Xiang Wang, Xiangnan He, Meng Wang, Fuli Feng, and Tat-Seng Chua. 2019. Neural graph collaborative filtering. In Proceedings of the 42nd international ACM SIGIR conference on Research and development in Information Retrieval, pages 165-174.  
+Xinfeng Wang, Fumiyo Fukumoto, Jin Cui, Yoshimi Suzuki, and Dongjin Yu. 2024b. Nfarec: A negative feedback-aware recommender model. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 935-945.  
+Yunfan Wang, Zeyu Liu, Jiangxia Zhang, Wei Yao, Stefan Heinecke, and Philip S Yu. 2023. Drdt: Dynamic reflection with divergent thinking for lmbased sequential recommendation. arXiv preprint arXiv:2312.11336.  
+Cheng Wu, Shaoyun Shi, Chaokun Wang, Ziyang Liu, Wang Peng, Wenjin Wu, Dongying Kong, Han Li, and Kun Gai. 2024a. Enhancing recommendation accuracy and diversity with box embedding: A universal framework. In Proceedings of the ACM Web Conference 2024, pages 3756-3766.  
+Cheng Wu, Liang Su, Chaokun Wang, Shaoyun Shi, Ziqian Zhang, Ziyang Liu, Wang Peng, Wenjin Wu, and Peng Jiang. 2025. Learning multiple user distributions for recommendation via guided conditional diffusion. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 12845-12853.  
+Cheng Wu, Chaokun Wang, Jingcao Xu, Ziwei Fang, Tiankai Gu, Changping Wang, Yang Song, Kai Zheng, Xiaowei Wang, and Guorui Zhou. 2023. Instant representation learning for recommendation over large dynamic graphs. In 2023 IEEE 39th International Conference on Data Engineering (ICDE), pages 82-95. IEEE.  
+Chuhan Wu, Fangzhao Wu, Yongfeng Huang, and Xing Xie. 2020. Neural news recommendation with negative feedback. CCF Transactions on Pervasive Computing and Interaction, 2:178-188.  
+Jiancan Wu, Xiang Wang, Fuli Feng, Xiangnan He, Liang Chen, Jianxun Lian, and Xing Xie. 2021. Self-supervised graph learning for recommendation. In Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval, pages 726-735.  
+Qitian Wu, Wentao Zhao, Chenxiao Yang, Hengrui Zhang, Fan Nie, Haitian Jiang, Yatao Bian, and Junchi Yan. 2024b. Simplifying and empowering transformers for large-graph representations. Advances in Neural Information Processing Systems, 36.
+
+Jing Xiong, Zixuan Li, Chuanyang Zheng, Zhijiang Guo, Yichun Yin, Enze Xie, Zhicheng Yang, Qingxing Cao, Haiming Wang, Xiongwei Han, and 1 others. 2023. Dq-lore: Dual queries with low rank approximation re-ranking for in-context learning. arXiv preprint arXiv:2310.02954.  
+Ze Yang, Juntao Wu, Yue Luo, Junhong Zhang, Yang Yuan, Aizhen Zhang, Xing Wang, and Xiangnan He. 2023. Large language model can interpret latent space of sequential recommender. arXiv preprint arXiv:2310.20487.  
+Junliang Yu, Xin Xia, Tong Chen, Lizhen Cui, Nguyen Quoc Viet Hung, and Hongzhi Yin. 2023. Xsimgcl: Towards extremely simple graph contrastive learning for recommendation. IEEE Transactions on Knowledge and Data Engineering, 36(2):913-926.  
+Zheng Yuan, Fajie Yuan, Yu Song, Youhua Li, Junchen Fu, Fei Yang, Yunzhu Pan, and Yongxin Ni. 2023. Where to go next for recommender systems? idvs. modality-based recommender models revisited. SIGIR '23, New York, NY, USA. Association for Computing Machinery.  
+Haobo Zhang, Qiannan Zhu, and Zhicheng Dou. 2025. Enhancing reranking for recommendation with llms through user preference retrieval. In Proceedings of the 31st International Conference on Computational Linguistics, pages 658-671.  
+Jizhi Zhang, Keqin Bao, Yupeng Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He. 2023. Is chatgpt fair for recommendation? evaluating fairness in large language model recommendation. arXiv preprint arXiv:2305.07609.  
+Xuan Zhang, Chunyu Wei, Ruyu Yan, Yushun Fan, and Zhixuan Jia. 2024. Large language model ranker with graph reasoning for zero-shot recommendation. In International Conference on Artificial Neural Networks, pages 356-370. Springer.  
+Xiangyu Zhao, Liang Zhang, Zhuoye Ding, Long Xia, Jiliang Tang, and Dawei Yin. 2018. Recommendations with negative feedback via pairwise deep reinforcement learning. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 1040-1048.  
+Leqi Zheng, Chaokun Wang, Ziyang Liu, Canzhi Chen, Cheng Wu, and Hongwei Li. 2025a. Balancing self-presentation and self-hiding for exposure-aware recommendation based on graph contrastive learning. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2027-2037.  
+Leqi Zheng, Chaokun Wang, Zixin Song, Cheng Wu, Shannan Yan, Jiajun Zhang, and Ziyang Liu. 2025b. Negative feedback really matters: Signed dual-channel graph contrastive learning framework for recommendation. In The Thirty-ninth Annual Conference on Neural Information Processing Systems.
+
+# A Progressive Activation Pipeline Prompts
+
+This section provides comprehensive details of the prompts used in our Progressive Activation Pipeline. These prompts systematically activate dormant signals across the recommendation process, with particular focus on sparse interactions.
+
+# A.1 Data-Level Activation Prompts
+
+# A.1.1 User Profile Generation Prompt
+
+This prompt generates comprehensive user profiles based on demographic information and ratings. It instructs the LLM to analyze both positive preferences (ratings 4-5) and negative preferences (ratings 1-3), explicitly activating dormant negative signals. Figure 9 shows the system instruction, Figure 10 demonstrates a sample input, and Figure 11 illustrates the resulting profile.
+
+# A.1.2 Item Profile Generation Prompt
+
+This prompt analyses items and their potential appeal to different user demographics. Figure 12 shows the system instruction, Figure 13 illustrates a sample input, and Figure 14 displays the resulting profile.
+
+# A.1.3 Sparse Positive Interaction Activation Prompt
+
+This prompt activates high-quality positive interactions for cold-start users. Figure 15 illustrates the prompt that guides the LLM through a systematic evaluation process to identify and activate relevant positive interactions.
+
+# A.1.4 Sparse Negative Interaction Activation Prompt
+
+This prompt explicitly activates sparse negative preferences. Figure 16 guides the LLM to analyze patterns of aversion in user profiles and assign dislike scores with high discrimination, activating typically overlooked negative signals.
+
+# A.2 Rank-Level Activation Prompts
+
+# A.2.1 User/Item Grouping Prompt
+
+This prompt implements a differentiated activation strategy based on interaction difficulty. Figure 17 shows how the LLM partitions users into semantically similar groups, enabling identification of hard and easy negative interactions requiring different activation levels.
+
+# A.2.2 Group Similarity Assessment Prompt
+
+This prompt assesses similarity between groups to refine the differentiated activation strategy. Figure 18 illustrates the prompt that establishes the optimal grouping structure for differentiating between easy and hard negative interactions.
+
+# A.3 Rerank-Level Activation Prompts
+
+This prompt activates historical user interactions during reranking. Figure 19 presents the structured Chain-of-Thought reasoning process that enables the LLM to systematically analyze preferences, compare items, and produce personalized rankings that effectively incorporate activated signals.
+
+# B Detailed Theoretical Proofs
+
+# B.1 Relationship Between Lemmas and Propositions
+
+The theoretical analysis in Section 3 presents concise propositions that establish the effectiveness of our Progressive Activation Pipeline. These propositions are supported by a set of technical lemmas presented below, which provide more detailed mathematical properties and convergence guarantees.
+
+Specifically, the first proposition on information gain from hard negative samples is supported by Lemma 1 and Lemma 2, which establish the gradient scaling properties and temperature-adaptive convergence behavior, respectively. The second proposition on Chain-of-Thought reasoning process is supported by Lemma 3 and Lemma 4, which prove the optimal weighting parameter existence and the convergence of our reranking approach. Together, these lemmas provide the mathematical foundation for our theoretical claims and explain why the Progressive Activation Pipeline effectively addresses the sparse interaction challenge in recommendation systems.
+
+Lemma 1 (Difficulty-Based Gradient Scaling). Under the LABPR loss with  $w_{h} > 1$ , the expected gradient magnitude for hard negative samples is proportionally larger than for easy ones, ensuring accelerated learning for challenging cases.
+
+Lemma 2 (Temperature-Adaptive Convergence). With adaptive temperature parameters  $\tau_h < \tau_e \leq \tau_p$  in our InfoNCE loss, the model converges to a more discriminative embedding space compared to uniform temperature settings.
+
+Lemma 3 (Score Fusion Optimality). There exists an optimal weighting parameter  $\alpha^{*} \in [0,1]$  that
+
+minimizes the expected ranking error:
+
+$$
+\alpha^ {*} = \arg \min  _ {\alpha \in [ 0, 1 ]} \mathbb {E} _ {(u, v) \sim \mathcal {D}} \left[ L \left(s _ {u v} ^ {f i n a l}, y _ {u v}\right) \right] \tag {23}
+$$
+
+Lemma 4 (Reranking Convergence). The reranking model trained with BCE loss converges to a stable solution when the candidate set distribution remains consistent between training and inference.
+
+# B.2 Detailed Proof of Lemma 1: Difficulty-Based Gradient Scale
+
+Considering the LABPR loss function 9, for a model parameterized by  $\theta$ , the gradient with respect to parameters is:
+
+$$
+\begin{array}{l} \frac {\partial \mathcal {L} _ {l a b p r}}{\partial \theta} = (24) \\ - \sum_ {(u, i, j _ {e}) \in D _ {e}} \frac {\sigma^ {\prime} (r _ {u i} - r _ {u j _ {e}})}{\sigma (r _ {u i} - r _ {u j _ {e}})} \cdot \frac {\partial r _ {u i} - r _ {u j _ {e}}}{\partial \theta} \\ - \sum_ {(u, i, j _ {h}) \in D _ {h}} w _ {h} \cdot \frac {\sigma^ {\prime} \left(r _ {u i} - r _ {u j _ {h}}\right)}{\sigma \left(r _ {u i} - r _ {u j _ {h}}\right)} \cdot \frac {\partial r _ {u i} - \partial r _ {u j _ {h}}}{\partial \theta} (25) \\ \end{array}
+$$
+
+Since  $\sigma(x) = \frac{1}{1 + e^{-x}}$ , we have  $\sigma'(x) = \sigma(x) \cdot (1 - \sigma(x))$ . Therefore:
+
+$$
+\frac {\sigma^ {\prime} (x)}{\sigma (x)} = 1 - \sigma (x) \tag {26}
+$$
+
+The gradient becomes:
+
+$$
+\begin{array}{l} \frac {\partial \mathcal {L} _ {l a b p r}}{\partial \theta} = (27) \\ - \sum_ {(u, i, j _ {e}) \in D _ {e}} (1 - \sigma (r _ {u i} - r _ {u j _ {e}})) \cdot \frac {\partial r _ {u i} - r _ {u j _ {e}}}{\partial \theta} \\ - \sum_ {(u, i, j _ {h}) \in D _ {h}} w _ {h} \cdot (1 - \sigma \left(r _ {u i} - r _ {u j _ {h}}\right)) \cdot \frac {\partial r _ {u i} - \partial r _ {u j _ {h}}}{\partial \theta} (28) \\ \end{array}
+$$
+
+For any pair of examples, one from  $D_{e}$  and one from  $D_{h}$  with identical  $\sigma(r_{ui} - r_{uj})$  values, the ratio of their gradient contributions is exactly  $w_{h}$ . Since  $w_{h} > 1$ , the gradient contribution from hard negatives is proportionally larger, leading to faster adjustment of parameters to distinguish these challenging cases.
+
+Furthermore, assuming that the distribution of  $(r_{ui} - r_{uj})$  is similar for both easy and hard negative sets at initialization, the expected gradient magnitude ratio between hard and easy negatives is:
+
+$$
+\frac {\mathbb {E} _ {(u , i , j _ {h}) \in D _ {h}} [ \| \nabla_ {\theta} \mathcal {L} _ {(u , i , j _ {h})} \| ]}{\mathbb {E} _ {(u , i , j _ {e}) \in D _ {e}} [ \| \nabla_ {\theta} \mathcal {L} _ {(u , i , j _ {e})} \| ]} \approx w _ {h} \tag {29}
+$$
+
+This ensures that the model allocates more learning capacity to hard negative examples, which is essential for developing fine-grained preference discrimination.
+
+To establish the convergence implications, the overall loss function is as eq 12. Under standard gradient descent optimization with learning rate  $\eta$ , the parameter update rule is:
+
+$$
+\theta_ {t + 1} = \theta_ {t} - \eta \cdot \nabla_ {\theta} \mathcal {L} _ {\text {t o t a l}} \tag {30}
+$$
+
+The component of this update attributable to the LABPR loss for hard negatives is approximately  $w_{h}$  times larger than for easy negatives, assuming similar initial conditions. This leads to faster convergence for hard negative cases, which is precisely what we desire for effective contrastive learning in sparse interaction scenarios.
+
+# B.3 Detailed Proof of Lemma 2:
+
+# Temperature-Adaptive Convergence
+
+Considering the activation-enhanced InfoNCE loss with adaptive temperature parameters in Eq 10 and Eq 11, let's denote  $d_{ij} = \text{sim}(v_i, v_j)$  for simplicity. For any embedding pair  $(v_i, v_j)$ , the contribution to the loss through the temperature parameter is determined by  $\exp(d_{ij} / \tau)$ .
+
+With  $\tau_h < \tau_e$ , for any similarity value  $d_{ij}$ , we have:
+
+$$
+\exp \left(d _ {i j} / \tau_ {h}\right) > \exp \left(d _ {i j} / \tau_ {e}\right) \tag {31}
+$$
+
+assuming  $d_{ij} > 0$  (which holds for cosine similarity of normalized embeddings with some positive correlation).
+
+The gradient of the loss with respect to the similarity  $d_{ij}$  for a hard negative is:
+
+$$
+\frac {\partial \mathcal {L} _ {\text {l a i n f o}}}{\partial d _ {i j}} = - \frac {1}{Z _ {i}} \cdot w _ {h} \cdot \frac {1}{\tau_ {h}} \cdot \exp \left(d _ {i j} / \tau_ {h}\right) \tag {32}
+$$
+
+and for an easy negative:
+
+$$
+\frac {\partial \mathcal {L} _ {\text {l a i n f o}}}{\partial d _ {i j}} = - \frac {1}{Z _ {i}} \cdot \frac {1}{\tau_ {e}} \cdot \exp \left(d _ {i j} / \tau_ {e}\right) \tag {33}
+$$
+
+For a given similarity value  $d_{ij}$ , the ratio of gradient magnitudes is:
+
+$$
+\begin{array}{l} \frac {\left| \frac {\partial \mathcal {L} _ {l a i n f o}}{\partial d _ {i j}} \right| _ {h}}{\left| \frac {\partial \mathcal {L} _ {l a i n f o}}{\partial d _ {i j}} \right| _ {e}} = w _ {h} \cdot \frac {\tau_ {e}}{\tau_ {h}} \cdot \frac {\exp \left(d _ {i j} / \tau_ {h}\right)}{\exp \left(d _ {i j} / \tau_ {e}\right)} (34) \\ = w _ {h} \cdot \frac {\tau_ {e}}{\tau_ {h}} \cdot \exp \left(d _ {i j} \left(\frac {1}{\tau_ {h}} - \frac {1}{\tau_ {e}}\right)\right) (35) \\ \end{array}
+$$
+
+Since  $\tau_h < \tau_e$ , we have  $\frac{1}{\tau_h} > \frac{1}{\tau_e}$ , so  $\exp \left(d_{ij}\left(\frac{1}{\tau_h} - \frac{1}{\tau_e}\right)\right) > 1$  for positive  $d_{ij}$ . Additionally,  $\frac{\tau_e}{\tau_h} > 1$ . Therefore:
+
+$$
+\frac {\left| \frac {\partial \mathcal {L} _ {\text {l a i n f o}}}{\partial d _ {i j}} \right| _ {h}}{\left| \frac {\partial \mathcal {L} _ {\text {l a i n f o}}}{\partial d _ {i j}} \right| _ {e}} > w _ {h} > 1 \tag {36}
+$$
+
+This demonstrates that the gradient magnitude for hard negatives is substantially larger than for easy negatives, leading to more aggressive optimization for challenging cases.
+
+To establish convergence to a more discriminative embedding space, we need to analyze the equilibrium state of the optimization. Let us denote the embeddings after  $t$  iterations of gradient descent as  $v_{i}^{(t)}$ . The update rule for embeddings is:
+
+$$
+v _ {i} ^ {(t + 1)} = v _ {i} ^ {(t)} - \eta \cdot \nabla_ {v _ {i}} \mathcal {L} _ {\text {l a i n f o}} \tag {37}
+$$
+
+The gradient with respect to the embedding  $v_{i}$  depends on the similarities with all other embeddings and their contributions to the loss. Due to the larger gradient magnitudes for hard negatives (as established above), the embedding vectors will move more quickly to reduce similarities with hard negatives compared to easy negatives.
+
+At convergence, the embeddings reach a state where the gradients approach zero. Due to the differential treatment of hard and easy negatives, this equilibrium state will have:
+
+$$
+\mathbb {E} _ {j \in \mathcal {N} _ {h} (i)} [ \operatorname {s i m} \left(v _ {i}, v _ {j}\right) ] <   \mathbb {E} _ {j \in \mathcal {N} _ {e} (i)} [ \operatorname {s i m} \left(v _ {i}, v _ {j}\right) ] \tag {38}
+$$
+
+This indicates that hard negatives are pushed further away in the embedding space than easy negatives, creating a more discriminative representation where semantically similar but distinct items are well-separated. This property is crucial for fine-grained recommendation capabilities, especially in sparse interaction scenarios.
+
+# B.4 Detailed Proof of Lemma 3: Score Fusion Optimality
+
+We establish the existence of an optimal weighting parameter  $\alpha^{*}$  that minimizes the expected ranking error. The activation-weighted score is defined as:
+
+$$
+s _ {u v} ^ {\text {f i n a l}} = \alpha \cdot s _ {u v} ^ {\text {o r i g}} + (1 - \alpha) \cdot s _ {u v} ^ {\text {l l m}} \tag {39}
+$$
+
+Let  $L(s_{uv}^{final}, y_{uv})$  be a ranking loss function measuring the discrepancy between the predicted score  $s_{uv}^{final}$  and the ground truth relevance  $y_{uv}$ . Common choices include pairwise ranking loss or listwise ranking loss.
+
+Define the expected loss as:
+
+$$
+\mathcal {R} (\alpha) = \mathbb {E} _ {(u, v) \sim \mathcal {D}} [ L (s _ {u v} ^ {\text {f i n a l}}, y _ {u v}) ] \tag {40}
+$$
+
+We need to prove there exists an  $\alpha^{*}\in [0,1]$  that minimizes  $\mathcal{R}(\alpha)$ .
+
+First, the continuity of  $\mathcal{R}(\alpha)$  with respect to  $\alpha$  can be established through the following observations:  $s_{uv}^{final}$  is a continuous function of  $\alpha$ , the loss function  $L(s_{uv}^{final}, y_{uv})$  is typically continuous in its first argument for common ranking losses, and the expectation operator preserves continuity.
+
+To analyze the behavior of  $\mathcal{R}(\alpha)$ , we examine the extreme cases:
+
+$$
+\mathcal {R} (0) = \mathbb {E} _ {(u, v) \sim \mathcal {D}} [ L \left(s _ {u v} ^ {l l m}, y _ {u v}\right) ] \tag {41}
+$$
+
+(using only LLM scores)
+
+$$
+\mathcal {R} (1) = \mathbb {E} _ {(u, v) \sim \mathcal {D}} [ L \left(s _ {u v} ^ {\text {o r i g}}, y _ {u v}\right) ] \tag {42}
+$$
+
+(using only CF scores)
+
+If either  $\mathcal{R}(0) \leq \mathcal{R}(\alpha)$  for all  $\alpha \in [0,1]$  or  $\mathcal{R}(1) \leq \mathcal{R}(\alpha)$  for all  $\alpha \in [0,1]$ , then  $\alpha^{*}$  is trivially 0 or 1, respectively.
+
+Otherwise, by the extreme value theorem, since  $\mathcal{R}(\alpha)$  is continuous over the compact interval [0, 1], there exists an  $\alpha^{*}\in [0,1]$  such that  $\mathcal{R}(\alpha^{*})\leq \mathcal{R}(\alpha)$  for all  $\alpha \in [0,1]$ .
+
+To prove that  $\alpha^{*} \in (0,1)$  is likely in practice (rather than at the extremes), we can construct a scenario where the original and LLM scores have complementary strengths. Let  $\mathcal{U}_1$  and  $\mathcal{U}_2$  partition the user space such that:
+
+$$
+\begin{array}{l} \mathbb {E} _ {(u, v) \sim \mathcal {D}, u \in \mathcal {U} _ {1}} \left[ L \left(s _ {u v} ^ {\text {o r i g}}, y _ {u v}\right) \right] <   \\ \mathbb {E} _ {(u, v) \sim \mathcal {D}, u \in \mathcal {U} _ {1}} \left[ L \left(s _ {u v} ^ {l l m}, y _ {u v}\right) \right] \\ \end{array}
+$$
+
+$$
+\begin{array}{l} \mathbb {E} _ {(u, v) \sim \mathcal {D}, u \in \mathcal {U} _ {2}} [ L (s _ {u v} ^ {\text {o r i g}}, y _ {u v}) ] > \tag {44} \\ \mathbb {E} _ {(u, v) \sim \mathcal {D}, u \in \mathcal {U} _ {2}} \left[ L \left(s _ {u v} ^ {l l m}, y _ {u v}\right) \right] \\ \end{array}
+$$
+
+That is, collaborative filtering scores perform better for users in  $\mathcal{U}_1$ , while LLM scores perform better for users in  $\mathcal{U}_2$ .
+
+In this case, neither  $\alpha = 0$  nor  $\alpha = 1$  would be optimal across all users. There exists an intermediate value  $\alpha^{*} \in (0,1)$  that minimizes the overall expected loss by balancing the strengths of both score sources.
+
+To find this optimal value analytically, we can differentiate  $\mathcal{R}(\alpha)$  with respect to  $\alpha$  and set it to zero:
+
+$$
+\begin{array}{l} \frac {d \mathcal {R} (\alpha)}{d \alpha} = \mathbb {E} _ {(u, v) \sim \mathcal {D}} \left[ \frac {\partial L (s _ {u v} ^ {f i n a l} , y _ {u v})}{\partial s _ {u v} ^ {f i n a l}} \cdot \frac {\partial s _ {u v} ^ {f i n a l}}{\partial \alpha} \right] \\ = 0 \tag {45} \\ \end{array}
+$$
+
+Since  $\frac{\partial s_{uv}^{final}}{\partial\alpha} = s_{uv}^{orig} - s_{uv}^{llm}$ , we have:
+
+$$
+\mathbb {E} _ {(u, v) \sim \mathcal {D}} \left[ \frac {\partial L \left(s _ {u v} ^ {\text {f i n a l}} , y _ {u v}\right)}{\partial s _ {u v} ^ {\text {f i n a l}}} \cdot \left(s _ {u v} ^ {\text {o r i g}} - s _ {u v} ^ {\text {l l m}}\right) \right] = 0 \tag {46}
+$$
+
+This equation has a solution  $\alpha^{*} \in (0,1)$  when the original and LLM scores have complementary strengths across different regions of the user-item space. In practical recommendation scenarios, this is often the case, as collaborative filtering excels at capturing collaborative patterns while LLM-based models capture semantic relationships.
+
+# B.5 Detailed Proof of Lemma 4: Reranking Model Convergence
+
+We analyze the convergence properties of the reranking model trained with binary cross-entropy (BCE) loss. The BCE loss is defined as:
+
+$$
+\begin{array}{l} \mathcal{L}_{bce} = \frac{1}{|\mathcal{D}_{rer}|}\sum_{(u,v,y_{uv})\in \mathcal{D}_{rer}}[-y_{uv}\log (p_{uv}) \\ - \left(1 - y _ {u v}\right) \log \left(1 - p _ {u v}\right) ] \tag {47} \\ \end{array}
+$$
+
+where  $p_{uv}$  is the model's predicted probability that user  $u$  would interact with item  $v$ .
+
+First, we establish that BCE loss is convex with respect to model predictions. For a single example  $(u,v,y_{uv})$ , the loss is:
+
+$$
+\ell \left(p _ {u v}, y _ {u v}\right) = - y _ {u v} \log \left(p _ {u v}\right) - \left(1 - y _ {u v}\right) \log \left(1 - p _ {u v}\right) \tag {48}
+$$
+
+The second derivative with respect to  $p_{uv}$  is:
+
+$$
+\frac {\partial^ {2} \ell}{\partial p _ {u v} ^ {2}} = \frac {y _ {u v}}{p _ {u v} ^ {2}} + \frac {1 - y _ {u v}}{(1 - p _ {u v}) ^ {2}} \tag {49}
+$$
+
+Since  $y_{uv} \in \{0,1\}$  and  $p_{uv} \in (0,1)$  in practice, the second derivative is always positive, confirming that BCE loss is strictly convex in the model predictions.
+
+For the reranking model with parameters  $\theta_{rer}$ , assuming it maps inputs to probabilities via a function  $f_{\theta_{rer}}$ , the optimization problem is:
+
+$$
+\theta_ {r e r} ^ {*} = \arg \min  _ {\theta_ {r e r}} \mathcal {L} _ {b c e} \left(\theta_ {r e r}; \mathcal {D} _ {r e r}\right) \tag {50}
+$$
+
+When the model has sufficient capacity and the mapping from parameters to predictions is continuous, gradient-based optimization of convex loss functions converges to a global minimum under standard conditions (e.g., appropriate learning rate scheduling, sufficient iterations).
+
+Now, we analyze the stability of the solution when the candidate set distribution remains consistent between training and inference. Let  $\mathcal{D}_{train}$  be the distribution of examples in the training set and  $\mathcal{D}_{inf}$  be the distribution during inference.
+
+The training objective is:
+
+$$
+\mathbb {E} _ {(u, v, y _ {u v}) \sim \mathcal {D} _ {\text {t r a i n}}} [ \ell (f _ {\theta_ {r e r}} (u, v), y _ {u v}) ] \tag {51}
+$$
+
+The optimal parameters under this objective are:
+
+$$
+\theta_ {r e r} ^ {*} = \arg \min  _ {\theta_ {r e r}} \mathbb {E} _ {(u, v, y _ {u v}) \sim \mathcal {D} _ {t r a i n}} [ \ell (f _ {\theta_ {r e r}} (u, v), y _ {u v}) ] \tag {52}
+$$
+
+At inference time, we want these parameters to perform well under  $\mathcal{D}_{inf}$ . The performance gap can be bounded by:
+
+$$
+\begin{array}{l} \left| \mathbb {E} _ {(u, v, y _ {u v}) \sim \mathcal {D} _ {i n f}} \left[ \ell \left(f _ {\theta_ {r e r} ^ {*}} (u, v), y _ {u v}\right) \right] \right. \\ - \mathbb {E} _ {(u, v, y _ {u v}) \sim \mathcal {D} _ {t r a i n}} [ \ell (f _ {\theta_ {r e r} ^ {*}} (u, v), y _ {u v}) ] | \\ \leq T V \left(\mathcal {D} _ {\text {i n f}}, \mathcal {D} _ {\text {t r a i n}}\right) \cdot C \tag {53} \\ \end{array}
+$$
+
+where  $TV$  is the total variation distance between distributions and  $C$  is a constant related to the maximum possible loss value.
+
+When the candidate set distribution remains consistent between training and inference,  $TV(\mathcal{D}_{inf},\mathcal{D}_{train})$  is small, ensuring that the model's performance generalizes well from training to inference.
+
+Furthermore, since the collaborative filtering model is fixed during reranking optimization, any
+
+instability in the overall system would have to originate from the reranking model itself. By ensuring stable convergence of the reranking model through proper regularization and training procedures, the overall system maintains stability.
+
+To analyze the convergence rate, we can use standard results from convex optimization. For a strongly convex loss function with parameter  $\mu$  and Lipschitz continuous gradients with constant  $L$ , gradient descent with learning rate  $\eta = \frac{1}{L}$  achieves linear convergence:
+
+$$
+\begin{array}{l} \mathcal {L} _ {b c e} \left(\theta_ {t}\right) - \mathcal {L} _ {b c e} \left(\theta^ {*}\right) \\ \leq \left(1 - \frac {\mu}{L}\right) ^ {t} \cdot \left[ \mathcal {L} _ {b c e} \left(\theta_ {0}\right) - \mathcal {L} _ {b c e} \left(\theta^ {*}\right) \right] \tag {54} \\ \end{array}
+$$
+
+where  $\theta_{t}$  is the parameter vector at iteration  $t$ , and  $\theta^{*}$  is the optimal parameter vector.
+
+In practice, while the BCE loss itself is convex in model predictions, the overall loss landscape may not be strongly convex in model parameters due to the non-linear mapping from parameters to predictions in neural networks. However, empirical evidence suggests that well-regularized models with appropriate architectures converge reliably to good solutions, even if they are local rather than global minima.
+
+The stability of the reranking process is further enhanced by the activation-weighted score computation, which combines the collaborative filtering scores with the LLM-generated scores as eq 19. This weighted combination acts as a form of ensemble learning, which is known to improve stability and generalization. Even if one component (either the collaborative filtering model or the LLM reranking model) experiences instability, the weighted combination can mitigate its effects on the final recommendations.
+
+![](images/5ded4cd5659c1c9e6c1a9fc3221e02903ea4931c00243f519bccb9284edd9300.jpg)  
+Figure 9: User Profile Generation Instruct
+
+![](images/50c7260e28ceb95a4ce491a566646b3f7eaf8d892387744735fac904456684c1.jpg)  
+Figure 10:User Profile Prompt Input
+
+![](images/147a54a86e3c60155895606bbce8479b3e7eee5a7d441788b5cb8e42d01e104b.jpg)  
+Figure 11:User Profile Prompt Output
+
+User Profile Prompt Output
+
+Figure 12: Item Profile Generation Instruct  
+```txt
+You will serve as an assistant to help me summarize which types of users would enjoy a specific movie. Here are the instructions:   
+1. I will provide you with information in the form of a JSON string that describes the movie:   
+{ "movie_id": "unique identifier for the movie", title": "movie title with release year", genres": "pipe-separated list of genres" } 2. Use your knowledge about this movie to analyze what types of viewers might enjoy it. REQUIREMENTS:   
+1. Please provide your analysis in JSON format, following this structure: { "movie_id": "the movie ID from the input data", "summariation": "A summarization of what types of users would enjoy this movie", "reasoning": "briefly explain your reasoning for the summarization" }   
+2. If you don't have enough information about the movie or cannot recognize it, please still provide a reasonable estimation based on the genres and title.   
+3. Please ensure that the "summariation" is no longer than 100 words.   
+4. Please ensure that the "reasoning" is no longer than 200 words.   
+5. Do not provide any other text outside the JSON string. Item Profile Generation Instruct
+```
+
+Figure 13: Item Profile Prompt Input  
+```txt
+Here is the movie data to analyze:  
+[  
+    {"movie_id":"1", "title":"Toy Story (1995)", "genres":"Animation|Children's|Comedy"},
+    ... ]
+```
+
+Item Profile Prompt Input  
+Figure 14: Item Profile Prompt Output  
+```txt
+"movie_id": "1", "summmarization": "Toy Story (1995) would appeal to a diverse audience, with children ages 4-12 forming the primary demographic. Families seeking wholesome entertainment would enjoy its universal themes of friendship, loyalty, and growing up. Adults, particularly those with nostalgia for childhood toys or appreciation for Pixar's pioneering animation, would find deeper emotional resonance and humor. Animation enthusiasts would value its groundbreaking technical achievements as the first fully computer-animated feature film. The movie's blend of slapstick humor for children and clever wit for adults makes it ideal for multi-generational viewing. Fans of imaginative storytelling that anthropomorphizes everyday objects would connect with its premise of toys coming to life." "reasoning": "As Pixar's debut feature and the first fully computer-animated film, Toy Story revolutionized animation while delivering a story with multi-generational appeal. The genre combination of Animation, Children's, and Comedy indicates its primary target audience (children) while suggesting broader appeal through humor. The film balances child-friendly adventure with sophisticated emotional themes and clever dialogue that adults appreciate. Its innovative animation would attract film enthusiasts and technology fans. The universal themes of friendship, jealousy, and belonging transcend age barriers, while its portrayal of toys from different eras creates nostalgia for various generations. The film's G rating, adventurous plot, and emotional depth make it accessible to young viewers while engaging enough for adults, explaining why it became a beloved family classic that launched a successful franchise."   
+}, Item Profile Prompt Output
+```
+
+```txt
+You are a recommendation expert. Score how likely this user would enjoy each movie (1-10 scale).   
+USER PROFILE:   
+{json.dumps(user_profile,indent  $\vDash$  2)}   
+ITEM PROFILE:   
+{json.dumps(itemProfiles_list,indent  $= 2$  ）}   
+ANALYTICAL PROCESS:   
+STEP 1: Analyze user profile: Identify top genres, creators, and time periods user prefers.   
+STEP 2: Score each item on key factors (1-10): Genre match, Creator match, Content match.   
+STEP 3: Assign final score with high discrimination: Use FULL 1-10 range with meaningful differentiation.   
+-9-10: Perfect match (rare,  $<  5\%$  of recommendations)   
+-7-8: Strong match with minor misalignments   
+-5-6: Moderate match with notable mismatches   
+-3-4: Weak match, significant incompatibility   
+-1-2: Poor match, fundamental misalignment   
+REQUIREMENTS:   
+1. Please provide your analysis in JSON format, following this structure: { "ratings": [ { "movie_id": "movie_id_value", "like_score": "score_value", "reasoning": "Detailed explanation of scores" } ]   
+}   
+2.Use precise, highly differentiated scores   
+3. Ensure movie_id matches exactly as provided   
+IMPORTANT: Return ONLY the JSON object directly, with no additional text, comments, or markdown formatting before or after it. Positive Interaction Generation Instruct
+```
+
+Figure 15: Cold-Start Positive Interaction Activation Instruct
+
+You are an anti-recommendation expert. Score how likely this user would DISLIKE each movie (1-10 scale).
+
+USER PROFILE:
+
+{json.dumps(user_profile,indent=2)}
+
+ITEM PROFILE
+
+{json.dumps(item_profiles_list,indent  $\equiv$  2)}
+
+ANALYTICAL PROCESS
+
+STEP 1: Analyze user profile: Identify genres, creators, and time periods user clearly avoids or dislikes and viewing patterns that suggest strong aversions.
+
+STEP 2: Score each movie on key mismatch factors (1-10): Genre match, Creator match, Content match.
+
+STEP 3: Assign final dislike score with high discrimination: Use FULL 1-10 range with meaningful differentiation.
+
+- 9-10: Complete mismatch (high confidence user would dislike)
+
+- 7-8: Strong mismatch with few redeeming qualities
+
+- 5-6: Moderate mismatch with some compatibility
+
+-3-4: Minor mismatch, mostly compatible
+
+- 1-2: Minimal mismatch, likely enjoyable
+
+REQUIREMENTS:
+
+1. Please provide your analysis in JSON format, following this structure:
+
+```jsonl
+{ "ratings": [ { "movie_id": "movie_id_value", "dislike_score": "score_value", "reasoning": "Detailed explanation of scores" } ]
+```
+
+2. Use precise, highly differentiated scores
+
+3. Ensure movie_id matches exactly as provided
+
+IMPORTANT: Return ONLY the JSON object directly, with no additional text, comments, or markdown formatting before or after it.
+
+# Negative Interaction Generation Instruct
+
+Figure 16: Sparse Negative Interaction Activation Instruct
+
+You will be given {len(groups)} group profiles, each containing information about a group's movie preferences and the reasoning behind these preferences. Your task is to calculate an estimated similarity score between two groups.
+
+INPUT
+
+Group 1
+
+- ID: {group1['group_id']}
+
+- Name: {group1.get('group_name', 'Unnamed Group')}
+
+- Key Characteristics: {json.dumps(group1.get('key-characteristics', [.]))}
+
+- Reasoning: {group1.get('reasoning', '')}
+
+Group 2:
+
+- ID: {group2['group_id']}
+
+- Name: {group2.get('group_name', 'Unnamed Group')}
+
+- Key Characteristics: {json.dumps(group2.get('key-characteristics', [...])}
+
+- Reasoning: {group2.get('reasoning', '')}
+
+ANALYTICAL PROCESS
+
+STEP 1: Analyze the key characteristics of both groups. Identify shared characteristics between both groups.
+
+STEP 2: Analyze the overall themes, interests, and behaviors represented in each group.
+
+STEP 3: Calculate an estimated similarity score (0-100) based on these criteria:
+
+- 0-20: Fundamentally different groups with almost no overlapping characteristics
+
+- 21-40: Mostly different with a few minor similarities
+
+- 41-60: Moderate similarity with significant  $\alpha$
+
+- 61-80: Substantial similarity with some notable differences
+
+- 81-100: Extremely similar groups with only minimal differences
+
+STEP 4: Provide clear justification for your similarity score, explaining specific similarities and differences.
+
+REQUIREMENTS:
+
+1. Please provide your analysis in JSON format, following this structure:
+
+$\left\{  {1,2,3}\right\}$
+
+"similarity_score": [your calculated score between 0-100],
+
+"similarity Explanation": "Detailed explanation of similarities and differences..."
+
+}
+
+2. All property names and string values must use double quotes.
+
+IMPORTANT: Do not include any additional text, markdown formatting, code blocks, or explanations outside the JSON.
+
+# Group Merge Instruct
+
+Figure 17: Group Merge Instruct
+
+You will be given {len/users}) user profiles, each containing information about a user's movie preferences and the reasoning behind these preferences.
+
+Your task is to:
+
+1. Create logical groups of users with similar preferences and tastes  
+2. Name and characterize each group  
+3. Assign each user to their most appropriate group  
+4. IMPORTANT: Each user must be assigned to EXACTLY ONE group  
+5. CRITICAL: Only use valid user_ids exactly as provided in the input user profiles
+
+USER PROFILE
+
+{json.dumps/users,indent  $\equiv$  2)}
+
+REQUIREMENTS:
+
+1. Please provide your analysis in JSON format, following this structure:
+
+{"groups":[
+
+"group_id": "group_1",
+
+"group_name": "Descriptive name for the group".
+
+"key-characteristics": ["characteristic 1", "characteristic 2", ...],
+
+"members": ["user_id_1", "user_id_3", ...].
+
+"reasoning": "Explanation of what defines this group"
+
+#
+
+1
+
+2. You MUST use the exact user_ids provided in the input - do not modify, reformat, or make up user_ids  
+3. Every user_id in your response must exactly match one from the provided user profiles  
+4. Every user must be assigned to exactly one group  
+5. User_ids are case-sensitive strings - preserve their exact format  
+6. Your ENTIRE response must be a valid JSON object with the exact structure specified above.  
+IMPORTANT: Do not include any additional text, markdown formatting, code blocks, or explanations outside the JSON.
+
+Group Partition Instruct
+
+Figure 18: Group Partition Instruct
+
+You are a Precision Reranking Engine.
+
+Your sole task is to reorder the provided list of candidates strictly and exclusively based on the user preferences detailed below.
+
+USER INFORMATION
+
+User Information: User {user_id} with explicit preferences: {summarization}
+
+CANDIDATE ITEMS
+
+Candidate Items List: {item_id}
+
+Original Ranking:{item_id}
+
+USER INTERACTION HISTORY
+
+Items the user previously liked (strong relevance signals): Item {pos_id}: {summarization}
+
+Items the user previously disliked (negative signals): Item {neg_id}: {summarization}
+
+CANDIDATE ITEM DETAILS
+
+Item {item_id} {item_signal}: {summmarization}
+
+RANKING REASONING PROCESS
+
+Follow this reasoning process:
+
+Step 1: Prioritize items that closely match the user's historical preferences and explicit likes.
+
+Step 2: Demote items similar to those the user has historically disliked
+
+Step 3: Make minimal adjustments to the original ranking, changing positions only when you have high confidence.
+
+RANKING GUIDELINES
+
+1. Promote the most relevant items to top positions  
+2. Make conservative adjustments - only move items when necessary  
+3. The original ranking is usually correct - only adjust with clear evidence
+
+[TRAINING PHASE ONLY]
+
+For this single-item evaluation:  
+- If this item should be ranked highly based on user preferences, classify as relevant (1)  
+- If this item should be ranked lower, classify as less relevant (0)
+
+Rerank Prompt Instruct
+
+Figure 19: Chain-of-Thought Activation Process Instruct

2025/LAGCL4Rec_ When LLMs Activate Interactions Potential in Graph Contrastive Learning for Recommendation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d9c192fdc6c4ca3d97f6b2319554f1f96ef5ccf222df9d3db412310a76ce7d61
+size 1429024

2025/LAVa_ Layer-wise KV Cache Eviction with Dynamic Budget Allocation/a94b9893-339c-481e-b5c0-29fad3e14c83_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:56e8811c4b852ccfb09dd804558790adac25446475ec4e9b07bd98321bc62589
+size 1949242

2025/LAVa_ Layer-wise KV Cache Eviction with Dynamic Budget Allocation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f72386925d908b35838565795124e7d0865e7e7050c1baf0d7bbf2555fbc0557
+size 2125549

2025/LAWCAT_ Efficient Distillation from Quadratic to Linear Attention with Convolution across Tokens for Long Context Modeling/04580249-c454-4e92-b9e7-a09c71d71b81_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5151c7850ce891f504e2aee71f63dd239a1b667701fb8a65ca67ed2e0906b42a
+size 2490168

2025/LAWCAT_ Efficient Distillation from Quadratic to Linear Attention with Convolution across Tokens for Long Context Modeling/full.md

@@ -0,0 +1,396 @@
+# LAWCAT: Efficient Distillation from Quadratic to Linear Attention with Convolution across Tokens for Long Context Modeling
+
+Zeyu Liu $^{1}$ , Souvik Kundu $^{2}$ , Lianghao Jiang $^{1}$ , Anni Li $^{1}$ , Srikanth Ronanki $^{3}$ , Sravan Bodapati $^{3}$ , Gourav Datta $^{4}$ , Peter A. Beerel $^{1}$
+
+<sup>1</sup>University of Southern California, USA <sup>2</sup>Intel Labs, San Diego, USA
+
+$^{3}$ Amazon AGI, USA,  $^{4}$ Case Western Reserve University, USA
+
+Correspondence: {liuzeyu, pabeerel} @usc.edu
+
+# Abstract
+
+Although transformer architectures have achieved state-of-the-art performance across diverse domains, their quadratic computational complexity with respect to sequence length remains a significant bottleneck, particularly for latency-sensitive long-context applications. While recent linear-complexity alternatives are increasingly powerful, effectively training them from scratch is still resource-intensive. To overcome these limitations, we propose LAWCAT (Linear Attention with Convolution Across Time), a novel linearization framework designed to efficiently transfer the capabilities of pre-trained transformers into a performant linear attention architecture. LAWCAT integrates causal Conv1D layers to enhance local dependency modeling and employs normalized gated linear attention to improve generalization across varying context lengths. Our comprehensive evaluations demonstrate that, distilling Mistral-7B with only 1K-length sequences yields over  $90\%$  passkey retrieval accuracy up to 22K tokens, significantly extending its effective context window. Similarly, Llama3.2-1B LAWCAT variant achieves competitive performance on S-NIAH 1&2&3 tasks (1K-8K context length) and BABILong benchmark (QA2&QA3, 0K-16K context length), requiring less than  $0.1\%$  pre-training tokens compared with pre-training models. Furthermore, LAWCAT exhibits faster prefetch speeds than FlashAttention-2 for sequences exceeding 8K tokens. LAWCAT thus provides an efficient pathway to high-performance, long-context linear models suitable for edge deployment, reducing reliance on extensive long-sequence training data and computational resources.
+
+# 1 Introduction
+
+The widespread adoption and significant success of transformer-based architectures have driven remarkable advancements across various natural
+
+language processing (NLP) (DeepSeek-AI, 2025; Grattafori et al., 2024; Team, 2024), reasoning (Chen et al., 2025), and computer vision (CV) tasks (Yao et al., 2024; Ramachandran et al., 2025). Transformers leverage self-attention mechanisms to effectively model long-range dependencies, achieving state-of-the-art (SOTA) performance on diverse benchmarks (Fourrier et al., 2024). However, their quadratic computational complexity with respect to sequence length remains a critical bottleneck, limiting their applicability to latency-sensitive edge based applications that are constrained in terms of training data and cost.
+
+To address this limitation, modern recurrent models have emerged as efficient alternatives, approximating or reformulating attention mechanisms to achieve linear complexity. Recent studies, such as Mamba-2 (Dao and Gu, 2024), Gated linear attention (GLA) (Yang et al., 2023), and Hierarchically gated linear RNN2 (HGRN2) (Qin et al., 2024), demonstrate the feasibility of state update mechanisms that substantially reduce computational costs while maintaining competitive performance. Despite these advances, training these models from scratch often requires substantial computational resources and extensive data, hindering rapid development and experimentation.
+
+On the other hand, knowledge distillation, a powerful technique for transferring knowledge from a large, pretrained teacher model to a smaller and more efficient student model, offers an effective strategy to alleviate these challenges. Recently, some research like MambaInLlama (Wang et al., 2024) and LoLCATs (Zhang et al., 2025) directly distill transformer models into recurrent architectures. However, MambaInLlama still requires extensive training data ( $\sim$ 20B tokens) and retains  $50\%$  self-attention layers to achieve strong performance up to 38K tokens, and LoLCATs exhibits limited generalization to sequences significantly longer than those seen during distillation.
+
+To address this problem, we introduce LAWCAT (Linear Attention with Convolution Across Time), a novel distillation framework that efficiently transfers and enhances the rich representational capabilities of pre-trained transformer models into modern linear attention models. LAWCAT is particularly well-suited for edge applications that require long-context modeling under tight computational constraints with limited training data. LAWCAT integrates a depth-separable convolution module, consisting of a depth-wise convolution layer followed by a linear layer (similar to MobileNet (Howard, 2017), but we use causal Conv1D across tokens), and a gated linear attention with normalization to effectively bridge the gap between quadratic and linear attention mechanisms. The main contributions of this paper are summarized as follows:
+
+- We proposed the integration of a causal Conv1D layer to enhance the capacity of linear attention mechanisms to model local dependencies and introduce the normalization for the gated linear attention to improve generalization to long contexts.  
+- We present LAWCAT, a framework leveraging these components to efficiently distill high-performing linear attention models. Specifically, we show that our Mistral v0.1 7B LAWCAT variant achieves over  $90\%$  accuracy on the passkey retrieval task with context lengths up to 22K tokens, significantly exceeding the original model's 8K limit. In addition, our Llama3.2 1B model LAWCAT variant achieves comparable performance on more complex retrieval tasks (NIAH1-3) and reasoning tasks (QA2&3 from BABILong benchmark) across various sequence lengths.  
+- We demonstrate the efficiency of our distillation approach, requiring significantly less data (less than  $0.1\%$  of the original pretraining tokens, and less than 1/3 of the original sequence length) to adapt a pretrained transformer into a high-performing linear attention model. Additionally, we show that the resulting LAWCAT model exhibits faster prefetch-stage processing speeds compared to FlashAttention-2 for input sequences longer than 8K tokens.
+
+# 2 Background and Related Work
+
+Self-Attention The majority of the powerful LLM models like Llama 3 (Grattafori et al., 2024), Mis
+
+tral (Jiang et al., 2023), and Phi 4 (Abdin et al., 2024) adopt the multi-head scaled dot-product softmax attention (SDPA). Given a query  $\mathbf{q}_t\in \mathbb{R}^{d_q}$  keys  $\{\mathbf{k}_i\}_{i = 1}^t\in \mathbb{R}^{d_k}$  , and values  $\{\mathbf{v}_i\}_{i = 1}^t\in \mathbb{R}^{d_v}$ $t\le N$  , a single-head SDPA computes output  $\mathbf{o}_t$  as (omitting the scaling factor  $\sqrt{d}$  for simplicity):
+
+$$
+\mathbf {o} _ {t} = \sum_ {i = 1} ^ {t} \frac {\exp \left(\mathbf {q} _ {t} \mathbf {k} _ {i} ^ {T}\right)}{\sum_ {j = 1} ^ {t} \exp \left(\mathbf {q} _ {t} \mathbf {k} _ {j} ^ {T}\right)} \mathbf {v} _ {i} \tag {1}
+$$
+
+This can be equivalently expressed as  $\mathbf{o}_t = \sum_{i=1}^t w_{ti}\mathbf{v}_i$ , where  $w_{ti}$  denotes the attention weight assigned to value  $\mathbf{v}_i$  at output position  $t$ . The softmax in the numerator/denominator ensures context-dependent normalization - each weight  $w_{ti}$  depends not only on the query-key similarity  $\mathbf{q}_t\mathbf{k}_i^T$  but also on all other similarities  $\{\mathbf{q}_t\mathbf{k}_j^T\}_{j \leq t}$ . This coupling gives softmax attention a powerful ability to focus on a few relevant tokens while normalizing out less relevant ones, but also yields the  $O(N^2)$  computation complexity with respect to the sequence length  $N$ .
+
+Kernel-Based Linear Attention To mitigate this quadratic computation complexity, many linear attention methods aim to find a kernel function whose structure approximates attention scores through inner products of transformed query and key representations. Assuming the kernel  $\kappa$  is positive semi-definite (PSD), then there exists a feature mapping  $\phi$  satisfying:  $\kappa (\mathbf{q},\mathbf{k}) = \phi (\mathbf{q})\phi (\mathbf{k})^T$  Thus, the attention output at position  $t$  can be computed as:
+
+$$
+\begin{array}{l} \mathbf {o} _ {t} = \sum_ {i = 1} ^ {t} \frac {\phi (\mathbf {q} _ {t}) \phi (\mathbf {k} _ {i}) ^ {T}}{\sum_ {j = 1} ^ {t} \phi (\mathbf {q} _ {t}) \phi (\mathbf {k} _ {j}) ^ {T}} \mathbf {v} _ {i} \tag {2} \\ = \frac {\phi (\mathbf {q} _ {t}) \sum_ {i = 1} ^ {t} \phi (\mathbf {k} _ {i}) ^ {T} \mathbf {v} _ {i}}{\phi (\mathbf {q} _ {t}) \sum_ {j = 1} ^ {t} \phi (\mathbf {k} _ {j}) ^ {T}} \\ \end{array}
+$$
+
+Similar to the Eq. 1, we can simplify is as  $\mathbf{o}_t = \sum_{i=1}^t \tilde{w}_{ti} \mathbf{v}_i$ , and  $\tilde{w}_{ti}$  can be regarded as the attention weights of the linear attention. This is analogous to softmax attention but uses  $\phi(\mathbf{q}) \phi(\mathbf{k})^T$  to approxiamte  $\exp(\mathbf{q} \mathbf{k}^T)$  for all queries/keys, at least in expectation. If  $\phi$  can perfectly realize exp as an inner product in a higher-dimensional feature space, then the linear attention exactly equals softmax attention. In practice,  $\phi$  might be a simple nonlinear function such as  $1 + \mathrm{ELU}$  (Katharopoulos et al., 2020), or a positive random feature map as in the Performer (Choromanski et al., 2020) that aims to fit the exponential. Such linear attention schemes run in  $O(N)$  time, but still suffer a noticeable performance gap between transformer models.
+
+Recent Modern Linear Attention Recent linear attention variants dispense with softmax normalization and utilize identity feature maps  $\phi (x) = x$  yielding the formulation  $\mathbf{o}_t = \mathbf{q}_t\mathbf{S}_t$  , where the KV state  $\mathbf{o}_t = \sum_{i = 1}^t\mathbf{k}_i^T\mathbf{v}_i$  is computed recursively as  $\mathbf{S}_t = \mathbf{S}_{t - 1} + \mathbf{k}_t^T\mathbf{v}_t$  . Instead of directly approximating softmax function, these models focus on enriching the state update process itself. Most models propose state transitions that take the form
+
+$$
+\mathbf {S} _ {t} = g \left(\mathbf {S} _ {t - 1}, \mathbf {x} _ {t}\right) + f \left(\mathbf {k} _ {t} ^ {T} \mathbf {v} _ {t}, \mathbf {x} _ {t}\right) \tag {3}
+$$
+
+The function  $g(\cdot, \cdot)$  often implements input-dependent gating or forgetting, controlling the influence of the historical state  $\mathbf{S}_{t-1}$  (Yang et al., 2023; Peng et al., 2024). The function  $f(\cdot, \cdot)$  defines how the current key-value outer product contributes to the state update (Dao and Gu, 2024; Yang et al., 2024; Azizi et al., 2025; Ye et al., 2025). Both functions can accept the input  $\mathbf{x}_t$  to make them input-dependent. These approaches leverage the efficiency of linear recurrence while introducing sophisticated mechanisms for state evolution and information flow, resulting in increasingly powerful sequence modeling capabilities. However, pre-training demands significant computational resources hindering widespread adoption.
+
+Linearization of Attention To reduce the training costs, recent efforts try to leverage pre-trained transformer-based LLMs to obtain computationally efficient models with linear complexity. Hedgehog (Zhang et al., 2024) introduced a learned linear feature mapping  $(\phi)$  to approximate softmax attention. Expanding on this approach, LoL-CATs (Zhang et al., 2025) integrated sliding-window attention to improve performance further. However, these methods still struggle to maintain strong performance on tasks involving significantly longer contexts than seen during distillation.
+
+Another approach, MambaInLlama (Wang et al., 2024), applied progressive distillation along with supervised fine-tuning (Kim and Rush, 2016; Ro et al., 2025) and directed preference optimization (Rafailov et al., 2023) to distill Mamba models from Llama3 models. Despite achieving strong performance at extended lengths up to 38K tokens, it still retains  $50\%$  of self-attention and requires substantial training data ( $\sim$ 20B tokens). SUPRA (Mercat et al., 2024) pursued a different strategy, modifying LLaMA3 models into architectures similar to RetNet (Sun et al., 2023), but still required extensive additional training ( $\sim$ 100B tokens) to extend
+
+effective context handling to 32K tokens. Thus, SUPRA does not fully resolve the substantial overhead associated with long-context model training.
+
+# 3 Algorithmic Motivation
+
+A fundamental distinction between softmax attention and its linear approximations lies in the normalization mechanism and the role of the exponential function. Softmax attention weights,  $\{w_{ti} \propto \exp(\mathbf{q}_t \mathbf{k}_i^T)\}$ , are normalized by a sum over all keys for a given query. This inherently creates a competitive dynamic: the exp function sharply amplifies scores for keys highly similar to the query relative to others, effectively focusing attention mass on the most relevant tokens (Vaswani et al., 2023). Although self-attention is renowned for modeling global dependencies, much of its effectiveness also stems from capturing local context (e.g., adjacent words or neighboring pixels). The softmax mechanism excels here as its exponential weighting amplifies locally coherent patterns (like increasing the relevance of "New York" when queried with "York"), effectively capturing correlations abundant in natural sequences. Empirical evidence (Han et al., 2024), including the disproportionate performance drop observed when masking local versus random tokens, further demonstrates that softmax models leverage local context more effectively than typical linear variants.
+
+In contrast, linear attention often employs feature maps  $\phi$  applied independently to queries and keys, yielding attention weights  $\tilde{w}_{ti} \propto \phi(\mathbf{q}_t) \phi(\mathbf{k}_i)^T$ . The normalization term  $\sum_{j} \phi(\mathbf{q}_t) \phi(\mathbf{k}_j)^T$  simply aggregates transformed key features without the strong competitive effect of the softmax denominator. This can lead to limitations such as attention dilution or oversmoothing, where attention distributions become less peaked (Qin et al., 2022). More critically, this formulation struggles to replicate the strong local modeling capabilities inherent in softmax attention (Han et al., 2024). This weaker local modeling is a key drawback of naive linear attention. Meanwhile, the potentially low-rank nature of the  $\phi(\mathbf{Q}) \phi(\mathbf{K})^T$  kernel can make it insensitive to crucial local variations and contextual nuances. Two distinct queries  $q$  and  $q'$ , even if originating from different local contexts, might yield similar  $\phi(q) \approx \phi(q')$  and thus nearly identical attention patterns, a failure mode less likely in softmax due to its context-sensitive normalization.
+
+# 4 Proposed Method
+
+# 4.1 Integration of Causal Conv1D Layer
+
+To address this deficiency in capturing local structure, we propose to integrate Conv1D layers to make the function  $\phi$  more context-sensitive. In particular, as shown in Fig. 1, our LAWCAT adds a causal depthwise convolution layer with kernel size of  $r + 1$  (In practice, we set kernel size as 4) for the query and key, respectively, to introduce an inductive bias for locality into the attention approximation. Since this convolution layer is causal, it only uses the tokens before the current token to calculate the convolution. Similar to the KV cache mechanism, we can cache the query and key from the previous tokens, but the size of the cache is limited by the kernel size. After the Conv1D layer, similar to the LoLCATs (Zhang et al., 2025), we use a linear layer with nonlinearity to project the query and key but we make the query and key to share the same linear projection, which makes the approximation align with the Eq. 2 which use the same function  $\phi$  for the query and value.
+
+A causal Conv1D with kernel size of  $r + 1$  will compute each transformed query  $\tilde{\mathbf{q}}_t = f_{\theta}([{\bf q}_{t - r},\dots,{\bf q}_t,])$  as a function of a local window of the original queries (and similarly  $\tilde{\mathbf{k}}_i$  from neighboring keys). The convolution is linear (a depthwise 1D convolution is essentially a weighted moving average of the inputs in a local window) and we do not include nonlinearities between this Conv1D layer and the subsequent linear layer. In the simplest case, let  $\tilde{\mathbf{q}}_t = \sum_{\delta = 0}^r w_\delta {\bf q}_{t - \delta}$  with some learnable weights  $w_{0\dots r}$  (and similarly  $\tilde{\mathbf{k}}_i$ ). This operation mixes local token information into the queries and keys before they interact. Intuitively, the convolution can be seen as smoothing and enriching the representations of  $Q$  and  $K$  with their neighbors' features. As a result, the kernel function  $\phi$  (now applied on  $\tilde{\mathbf{q}}_t$  and  $\tilde{\mathbf{k}}_i$ ) is no longer a purely pointwise function - it has implicit awareness of nearby tokens.
+
+The similarity between  $\tilde{\mathbf{q}}_t$  and  $\tilde{\mathbf{k}}_i$  in the convaugmented space includes contributions from keys in the neighborhood of  $i$  as well. In effect, a key  $\mathbf{k}_i$  will receive a high weight  $\tilde{w}_{ti}$  not only if it aligns with the query, but also if its previous neighbors  $\mathbf{k}_{i - 1},\mathbf{k}_{i - 2},\ldots$  align with the query. This property can help the linear attention mimic what softmax does: softmax will give a group of similar, adjacent keys a collectively high weight if the query matches that group. The Conv1D introduces an
+
+![](images/2daad0697ae00e64274c431131cf807ce6743241b6fd97053058b79db9eb2fdd.jpg)  
+Figure 1: The overall structure of LAWCAT. We use Casual Conv1D with a kernel size of  $r + 1$  for visualization, and use  $\mathbf{x}_t$  to represent the  $t$ -th token of input  $\mathbf{x}$ .
+
+implicit group-aware effect as neighboring keys support each other's relevance. Conversely, if a key  $\mathbf{k}_i$  is an outlier among irrelevant neighbors,  $\tilde{\mathbf{k}}_i$  may be diluted by its neighbors (lowering  $\phi (\tilde{\mathbf{k}}_i)$ 's dot with  $\phi (\tilde{\mathbf{q}}_t)$ ), which is analogous to how an isolated token might not stand out as strongly under softmax normalization if its similarity is not much larger than others. In summary, the convolution provides a localized smoothing and sharpening of attention scores: smoothing because it aggregates local information mitigating random fluctuations, and sharpening because a coherent local pattern (several similar tokens in a window) will amplify each token's effective feature.
+
+The advantages of the Conv1D layer over the sliding window attention On the other hand, LoL-CATs (Zhang et al., 2025) proposed to use a sliding window attention to improve the overall performance. Specifically, for the tokens within the window, they will calculate the softmax attention score, and for the tokens outside the window, they will use the linear attention score, and use the weighted sum as the final attention score. However, this strategy does not attempt to improve the linear attention component's handling of local interactions, so retrieving information located far beyond the window relies solely on the capabilities of the linear attention component. If the linear attention part fails to accurately capture these long-range dependencies or lacks the necessary precision, the model's ability to retrieve distant information will be significantly impaired, leading to performance degradation as the distance to the target information increases.
+
+In contrast, our LAWCAT produces one cohesive attention output that naturally balances local and global information which gets rid of the difficulties
+
+of balancing the linear and softmax attention. Additionally, by blending neighboring tokens' features, LAWCAT can reduce the risk of query confusion (non-injectivity) and increase the effective rank of the attention, leading to a better approximation of full attention (Fan et al., 2024). Moreover, LAW-CAT encodes a prior that nearby tokens are related, which can improve generalization on sequences with local structure helping combat the tendency of full attention to overfit noise.
+
+# 4.2 GLA with Normalization
+
+Another key design choice in our architecture is the adoption of Gated Linear Attention (GLA), which adds additional linear layers followed by a sigmoid activation to enable input-dependent gating. Unlike Yang et al. (2023), who omit the normalization term in Eq. 2, we find that retaining this normalization is critical for both training stability and final performance, particularly in the context of distillation from a standard Transformer model.
+
+Let  $\mathbf{G}_t$  denote the forget gate and  $\mathbf{S}_t$  represent the key-value (KV) state of the linear attention mechanism at time step  $t$ . For notational simplicity, we denote the projected query and key of the  $i$ -th token, after Conv1D and linear projection, as  $\dot{\mathbf{q}}_i$  and  $\dot{\mathbf{k}}_i$ , respectively. Then, the KV state update equation in GLA can be described as:
+
+$$
+\mathbf {S} _ {t} = \mathbf {G} _ {t} \odot \mathbf {S} _ {t - 1} + \dot {\mathbf {k}} _ {t} ^ {T} \mathbf {v} _ {t} \tag {4}
+$$
+
+where  $\odot$  is the Hadamard product. Letting the  $\mathbf{S}_0 = 0$ , we can rewrite the above equation as:
+
+$$
+\mathbf {S} _ {t} = \sum_ {i = 1} ^ {t} \left[ \left(\bigodot_ {z = i + 1} ^ {t} \mathbf {G} _ {z}\right) \odot \left(\dot {\mathbf {k}} _ {i} ^ {T} \mathbf {v} _ {i}\right) \right] \tag {5}
+$$
+
+and we use  $\mathbf{J}_{i,t}$  to denote the consecutive Hadamard products when  $i + 1\leq t$  ..
+
+$$
+\mathbf {J} _ {i, t} = \bigotimes_ {z = i + 1} ^ {t} \mathbf {G} _ {z} = \mathbf {G} _ {t} \odot \mathbf {G} _ {t - 1} \odot \dots \odot \mathbf {G} _ {i + 1} \tag {6}
+$$
+
+otherwise, we set  $\mathbf{J}_{i,t} = \mathbf{1}$ . Finally, we get the expression of the KV state as Eq. 7 and the attention output as Eq. 8:
+
+$$
+\mathbf {S} _ {t} = \sum_ {i = 1} ^ {t} \left[ \mathbf {J} _ {i, t} \odot \left(\dot {\mathbf {k}} _ {i} ^ {T} \mathbf {v} _ {i}\right) \right] \tag {7}
+$$
+
+$$
+\mathbf {o} _ {t} = \frac {\dot {\mathbf {q}} _ {t} \sum_ {i = 1} ^ {t} \left[ \mathbf {J} _ {i , t} \odot \left(\dot {\mathbf {k}} _ {i} ^ {T} \mathbf {v} _ {i}\right) \right]}{\dot {\mathbf {q}} _ {t} \sum_ {j = 1} ^ {t} \left[ \mathbf {J} _ {j , t} \odot \left(\dot {\mathbf {k}} _ {j} ^ {T} \mathbf {1}\right) \right] _ {\text {m e a n (1)}}} \tag {8}
+$$
+
+Note, since  $\mathbf{G}_i\in \mathbb{R}^{d_k\times d_v}$ , we also have  $\mathbf{J}_{i,t}\in$ $\mathbb{R}^{d_k\times d_v}$ , but  $\mathbf{k}_i\in \mathbb{R}^{1\times d_k}$ , it makes it not straightforward to calculate the normalization of the GLA. To address this, we let the  $\mathbf{k}_i^T$  times  $\mathbf{1}_{1\times d_v}$  first, which actually repeat the  $\mathbf{k}_i^T$  for  $d_v$  times along the second dimension, then follow the normal GLA operation. Before the multiplication with  $\mathbf{q}_t$ , we take the mean value of the  $\mathbf{J}_{i + 1,t}\odot (\mathbf{k}_i^T\mathbf{1})$  along the second dimension to reduce the shape from  $d_{k}\times d_{v}$  to  $d_{k}\times 1$ . In this way, we normalize the GLA with a similar definition as in Eq. 2.
+
+# 5 Experimental Results
+
+# 5.1 Setup
+
+Our training followed the two-stage methodology of LoLCATs (Zhang et al., 2025): an initial distillation stage employing only layer-wise MSE loss, followed by LoRA fine-tuning (Hu et al., 2022). In total, for Passkey Retrieval (Mohtashami and Jaggi, 2023), S-NIAH (Hsieh et al., 2024), and BA-BILong (Kuratov et al., 2024) tasks, we use 93M, 31M, and 103M tokens, respectively, with maximum input lengths of 1162, 1223, and 1300 tokens. More details regarding training datasets and training configurations are provided in Appendix A.1 and A.2. Standard pre-trained models typically leverage over 100B tokens to attain competitive performance. Compared with them, our approach requires less than  $0.1\%$  of such pre-training tokens.
+
+# 5.2 Results on Passkey Retrieval Task
+
+As shown in Fig. 2, the pre-trained LLaMA3 8B (Grattafori et al., 2024) maintains perfect accuracy from 1K to 8K tokens, but drops to zero beyond 9K, exceeding its training context length. In contrast, our method extends the effective context window: even when distilled and fine-tuned solely on 1K-length sequences, the resulting model retains competitive performance up to 12K tokens. Similarly, for the pre-trained LLaMA3.2 1B model, which performs reliably up to 32K tokens, our approach preserves strong retrieval accuracy up to 7K tokens. Compared to the LoLCATs method, whose effectiveness is limited to 1K for LLaMA3 8B and 1-2K for LLaMA3.2 1B, our method demonstrates substantially superior preservation of long-context retrieval capabilities.
+
+We also evaluate our approach on the Mistral v0.1 and v0.3 7B models (Jiang et al., 2023), which fail to achieve  $100\%$  accuracy even within 8K, likely due to their limited optimization for retrieval
+
+![](images/a731c4ad86f73087e458eb1bab829a7d38c4926ce4beb75227a1630cd2ea73f6.jpg)  
+Figure 2: The comparison between the pre-trained transformer model and our converted linear attention model.
+
+![](images/e856a81c4f39b384a81bf5ecdafbcd9a440e8728972e9d9d74319a0da069fb7b.jpg)
+
+<table><tr><td></td><td colspan="4">S-NIAH-1(pass-key retrieval)</td><td colspan="4">S-NIAH-2(number in haystack)</td><td colspan="3">S-NIAH-3(uuid in haystack)</td></tr><tr><td></td><td>1K</td><td>2K</td><td>4K</td><td>8K</td><td>1K</td><td>2K</td><td>4K</td><td>8K</td><td>1K</td><td>2K</td><td>4K</td></tr><tr><td>Llama3.2 1B Instruct</td><td>100</td><td>100</td><td>100</td><td>100</td><td>100</td><td>100</td><td>96</td><td>100</td><td>100</td><td>100</td><td>100</td></tr><tr><td>DeltaNet-1.3B</td><td>97.4</td><td>96.8</td><td>99.0</td><td>98.8</td><td>98.4</td><td>45.6</td><td>18.6</td><td>14.4</td><td>85.2</td><td>47.0</td><td>22.4</td></tr><tr><td>Mamba2-1.3B</td><td>99.2</td><td>98.8</td><td>65.4</td><td>30.4</td><td>99.4</td><td>98.8</td><td>56.2</td><td>17.0</td><td>64.4</td><td>47.6</td><td>4.6</td></tr><tr><td>Gated DeltaNet-1.3B</td><td>98.4</td><td>88.4</td><td>91.4</td><td>91.8</td><td>100</td><td>99.8</td><td>92.2</td><td>29.6</td><td>86.6</td><td>84.2</td><td>27.6</td></tr><tr><td>Distilled Model</td><td colspan="11">Pre-trained Model: Llama3.2-1B-Instruct</td></tr><tr><td>LoLCATs</td><td>100</td><td>84</td><td>0</td><td>0</td><td>84</td><td>44</td><td>0</td><td>0</td><td>72</td><td>24</td><td>0</td></tr><tr><td>Ours</td><td>100</td><td>100</td><td>100</td><td>80</td><td>100</td><td>96</td><td>88</td><td>48</td><td>56</td><td>44</td><td>24</td></tr><tr><td>LoLCATs</td><td colspan="11">Pre-trained Model: Llama3-8B-Instruct</td></tr><tr><td>Ours</td><td>100</td><td>4</td><td>0</td><td>0</td><td>92</td><td>4</td><td>0</td><td>0</td><td>84</td><td>24</td><td>0</td></tr><tr><td>LoLCATs</td><td>100</td><td>100</td><td>96</td><td>80</td><td>100</td><td>92</td><td>84</td><td>32</td><td>80</td><td>88</td><td>60</td></tr><tr><td colspan="12">Pre-trained Model: Mistral v0.1 7B</td></tr><tr><td>LoLCATs</td><td>100</td><td>100</td><td>100</td><td>100</td><td>100</td><td>96</td><td>64</td><td>0</td><td>72</td><td>72</td><td>16</td></tr><tr><td colspan="12">Pre-trained Model: Mistral v0.3 7B</td></tr><tr><td>LoLCATs</td><td>100</td><td>100</td><td>100</td><td>100</td><td>100</td><td>88</td><td>60</td><td>16</td><td>68</td><td>24</td><td>12</td></tr></table>
+
+Table 1: Performance comparison across different models on S-NIAH 1, 2, and 3 tasks.
+
+tasks. Despite this, our linearized Mistral v0.3 7B model retains good accuracy up to 15K tokens, and the linearized Mistral v0.1 7B model extends this further, achieving over  $90\%$  accuracy up to 22K tokens. In contrast, the LoLCATs-distilled Mistral models reach peak performance at only 1K tokens and rapidly degrade to zero beyond 2K tokens. These results highlight the robustness and generalizability of our method in preserving long-context capabilities across various model architectures.
+
+# 5.3 Results on S-NIAH Benchmark
+
+Beyond the Passkey Retrieval task, we evaluate our method on the more challenging S-NIAH 1-3 task from the RULER benchmark (Hsieh et al., 2024). For a comprehensive comparison, we include results from several SOTA pre-trained recurrent models, DeltaNet (Yang et al., 2024), Mamba2 (Dao and Gu, 2024), and Gated DeltaNet (Yang et al., 2025). All pre-trained models are trained on 100B tokens, and the training length is 4K. And the results are sourced from Yang et al. (2025).
+
+As shown in Table 1, on the S-NIAH 1 task,
+
+our model achieves  $100\%$  accuracy from 1K to 4K tokens, surpassing all listed SOTA recurrent models. Even at 8K, it remains competitive, despite being trained on sequences more than  $3 \times$  shorter (1223 vs. 4000 tokens) and using only  $0.03\%$  of the data (31M vs. 100B tokens). On the S-NIAH 2 task, LAWCAT achieves comparable performance up to 4K and maintains a strong  $48\%$  accuracy at 8K—substantially outperforming other models and demonstrating robust generalization to longer contexts. While performance on the more complex S-NIAH 3 task lags slightly behind others, we hypothesize this is due to the limited diversity in our training data. LoLCATs only maintain good results around 1K for all S-NIAH tasks, which is consistent with the results on the Passkey Retrieval task.
+
+The results related to the Llama3 8B and Mistral series models demonstrate similar trends; our model demonstrates superior robustness to increasing input lengths, with notably smaller performance drop compared to LoLCATs and other SOTA pre-trained recurrent models.
+
+![](images/84d5805194de9e9d7e6841a43de8b4e7a0dbee62c924ff7c44f179ce44ebb943.jpg)  
+Figure 3: Comparison of accuracy on QA2&3 from BABILong benchmark between Llama 3.2 1B (top), LoLCATs (middle), and LAWCAT (bottom).
+
+# 5.4 Results on BABILong benchmark
+
+Besides retrieval capabilities, we assessed LAW-CAT's performance on complex reasoning tasks using the BABILong benchmark (Kuratov et al., 2024), specifically focusing on multi-hop reasoning tasks, QA2 (two supporting facts) and QA3 (three supporting facts). The results, presented in Figure 3, demonstrate a significant advantage for our distilled and fine-tuned LAWCAT variant compared with the original Llama3.2 1B model. Across the evaluated context lengths (0K to 32K tokens), LAWCAT consistently maintains accuracy levels more than double those of the baseline Llama3.2 1B model for both QA2 and QA3 tasks.
+
+This notable enhancement in multi-hop reasoning performance underscores the robustness of the LAWCAT framework, particularly highlighting its capability to overcome inherent limitations observed in the foundational pretrained model. Moreover, compared with LoLCATs, which failed in this task, the results demonstrate LAWCAT's capacity not merely for transferring existing knowledge but also for significantly enhancing complex reasoning skills within extensive contexts. We also evaluate the Llama3 8B and Mistral series models on the BABILong benchmark, more details are shown in Appendix A.5.
+
+# 5.5 Efficiency Comparison
+
+To evaluate computational performance, we benchmarked the pre-fill latency of our Llama3.2 1B LAWCAT model against a Llama3.2 1B model incorporating FlashAttention-2 (FA2) (Dao, 2024) and the LoLCATs variant, using one NVIDIA RTX A6000 GPU. For a comprehensive comparison, we also include results from SOTA recurrent models, specifically GLA 1.3B and Mamba 1 1.3B.
+
+As illustrated in Fig. 4, LAWCAT incurs slightly higher prefetch latency than LLaMA3.2 1B with FA2 for sequences shorter than 8K tokens. Beyond this point, however, LAWCAT achieves significantly lower latency, with the gap widening as sequence length increases. In contrast, the naive implementation of LoLCATs (i.e., LoLCATs without the ThunderKittens kernel (HazyResearch, 2024)) exhibits substantially higher latency and greater GPU memory consumption.
+
+When compared to the GLA 1.3B model, LAW-CAT's inclusion of a Conv1D and linear layer, and GLA normalization contributes to a slightly higher latency. Notably, LAWCAT consistently outperforms the Mamba 1 1.3B model, which also incorporates a Conv1D layer, across all evaluated sequence lengths.
+
+![](images/154a26f8e7ede6b32d4f97ab8d4bc21b8f48ca9ee596693085cd6cf39362ddf7.jpg)  
+Figure 4: Comparison of prefetch-stage latency among 5 different models. Note, LoLCATs runs out of memory for sequence lengths exceeding 8K tokens.
+
+# 5.6 Ablation Study and Discussion
+
+We conducted a comprehensive ablation study to systematically evaluate the contribution of each component in the proposed LAWCAT framework and discuss the effect of prevalent techniques such as Rotary Position Embeddings (RoPE) and Stochastic Weight Averaging (SWA). Unless otherwise specified, all experiments are based on models distilled and fine-tuned from the pre-trained LLaMA3.2-1B using training data with a sequence length of  $\sim 1\mathrm{K}$  tokens. The ablation analysis is structured around the following key questions:
+
+Do we need the GLA normalization and Conv1D across tokens? To assess the contribution of key components within our proposed LAWCAT framework, we conduct an ablation study focusing on the GLA normalization and Conv1D layer.
+
+<table><tr><td></td><td>1K</td><td>2K ~ 6K</td><td>7K</td><td>8K</td><td>9K</td><td>10K</td><td>11K</td></tr><tr><td>NN</td><td>0.9</td><td>0</td><td>0</td><td>0</td><td>0</td><td>0</td><td>0</td></tr><tr><td>NC</td><td>1.0</td><td>1.0</td><td>1.0</td><td>0.8</td><td>0.9</td><td>0.9</td><td>0.4</td></tr><tr><td>O</td><td>1.0</td><td>1.0</td><td>0.9</td><td>0.9</td><td>0.8</td><td>0.5</td><td>0.2</td></tr></table>
+
+As presented in Table 2, removing GLA normalization significantly degrades performance on passkey retrieval tasks involving longer contexts, although the model maintains reasonable accuracy at the 1K-length scale. This finding supports our hypothesis that normalization is crucial for generalization. By constraining the GLA attention output to remain a normalized weighted sum of the values  $(\mathbf{v}_i)$  during layer-wise distillation (via MSE loss minimization against self-attention outputs), the normalization term mitigates overfitting to shorter sequence lengths and enhances the model's ability to preserve performance as context length increases.
+
+On the other hand, the results in Table 2 suggest that the Conv1D layer offers limited benefit for the relatively simpler passkey retrieval task. We attribute this to the nature of the task, where much of the context contains redundant or useless information; applying Conv1D over such sequences does not necessarily facilitate the extraction of more salient features for identifying and memorizing the passkey. However, its importance becomes evident in more complex scenarios. Table 3 reveals a substantial performance gap on the NIAH task between the full LAWCAT model and its variant without the Conv1D layer. Notably, without Conv1D, LAWCAT fails to retrieve the correct UUID in all tested cases. This stark difference underscores the critical role of the Conv1D layer in capturing dependencies and memorizing information across tokens, particularly when dealing with longer sequences and more intricate information retrieval demands.
+
+Table 2: Accuracy on the passkey retrieval task from 1K to 11K. "NN" means removing the GLA normalization from LAWCAT, "NC" means removing the Conv1D layer, and "O" means the complete LAWCAT model.  
+
+<table><tr><td></td><td colspan="4">S-NIAH-1</td><td colspan="4">S-NIAH-2</td><td colspan="4">S-NIAH-3</td></tr><tr><td></td><td>1K</td><td>2K</td><td>4K</td><td>8K</td><td>1K</td><td>2K</td><td>4K</td><td>8K</td><td>1K</td><td>2K</td><td>4K</td><td></td></tr><tr><td>NC</td><td>96</td><td>96</td><td>96</td><td>8</td><td>80</td><td>80</td><td>80</td><td>36</td><td>0</td><td>0</td><td>0</td><td></td></tr><tr><td>O</td><td>100</td><td>100</td><td>100</td><td>80</td><td>100</td><td>96</td><td>88</td><td>48</td><td>56</td><td>44</td><td>24</td><td></td></tr></table>
+
+Table 3: Accuracy on S-NIAH 1&2&3 tasks. "NC" means removing the Conv1D layer from LAWCAT, and "O" means the complete LAWCAT model.
+
+Do we need the RoPE and SWA? The necessity of explicit positional encodings like RoPE in linear attention is an open question, with varied adoption across models. Our ablation study (Table 5 in Appendix A.3) on LoLCATs and LAWCAT models addresses this. For LoLCATs, removing RoPE decreased performance on shorter contexts (1K-3K tokens) but slightly improved it on longer ones (4K-8K). This effect was more stark for our LAWCAT: the variant without RoPE maintained high accuracy up to 8K tokens, whereas the RoPE-equipped variant's performance collapsed beyond 3K tokens. These findings suggest that RoPE, while potentially aiding short-sequence modeling, introduces biases detrimental to long-context generalization, supporting the idea that recurrent formulations can inherently capture sequence order.
+
+We also investigated the role of SWA (Table 7 in Appendix A.3). On long-context passkey retrieval, removing SWA significantly impaired LoLCATs, confirming its reliance on this mechanism. Conversely, LAWCAT performed better without SWA, especially at longer contexts; we hypothesize this is due to the difficulty in balancing contributions from the GLA and SWA, particularly when generalizing from shorter training to longer evaluation sequences. Additional results and analysis on the impact of SWA on standard short-context benchmarks (LM Eval, Table 7) are presented in Appendix A.3.
+
+# 6 Conclusions
+
+We introduced LAWCAT (Linear Attention with Convolution Across Time), a novel and efficient distillation framework that strategically integrates a causal Conv1D layer to bolster local dependency modeling and employs a normalized gated linear attention mechanism for robust generalization across varying sequence lengths. Our comprehensive experiments validate the effectiveness of LAWCAT. Distilling models like Mistral v0.1 7B and Llama3.2 1B yielded linear attention variants capable of remarkable long-context performance, achieving over  $90\%$  accuracy on passkey retrieval tasks up to 22K tokens and matching SOTA recurrent models on NIAH benchmarks, all while using minimal training data and shorter sequence lengths during distillation. Furthermore, the resulting LAWCAT models demonstrate practical efficiency gains, exhibiting faster prefetch latency than highly optimized FlashAttention-2 implementation.
+
+tations for sequences exceeding 8K tokens.
+
+LAWCAT offers a practical and resource-efficient pathway for developing and deploying long-context models by leveraging existing pretrained transformers. This approach helps democratize access to moderately long-sequence modeling capabilities, particularly beneficial for resource-constrained edge environments or rapid prototyping. Furthermore, due to its causal and linear nature, LAWCAT is particularly attractive for streaming inference, opening new opportunities for low-latency, on-device reasoning.
+
+# Limitations
+
+While LAWCAT demonstrates strong long-context performance, on some complex benchmarks like S-NIAH-3 (1K) as shown in Table 1 and LM Evaluation Harness tasks (Gao et al., 2024) as shown in Table 7, purely distilled linear attention models may not fully match the performance of similarly sized transformer models or the linear attention model with sliding window attention (SWA). Although incorporating the SWA can bridge this gap on shorter sequences, as observed in our experiments, it potentially compromises the model's effectiveness on tasks requiring very long context generalization. This presents an interesting avenue for future research: developing dynamic mechanisms that can adaptively balance the global reach of linear attention with the local strength often captured by windowed approaches, perhaps conditioned on input characteristics or sequence length. Alternatively, exploring the integration of more sophisticated linear attention designs like Mixture-of-Memories (Du et al., 2025) might further mitigate the performance gap on standard benchmarks, although this could introduce new optimization challenges. Addressing this trade-off between benchmark performance and long-context fidelity remains a key direction for advancing efficient attention mechanisms.
+
+# Acknowledgments
+
+This work was supported in part by Amazon through the 2024-2025 Amazon ML fellowship.
+
+# References
+
+Marah Abdin, Jyoti Aneja, Harkirat Behl, Sébastien Bubeck, Ronen Eldan, Suriya Gunasekar, Michael Harrison, Russell J Hewett, Mojan Javaheripi, Piero Kauffmann, and 1 others. 2024. Phi-4 technical report. arXiv preprint arXiv:2412.08905.  
+Seyedarmin Azizi, Souvik Kundu, Mohammad Erfan Sadeghi, and Massoud Pedram. 2025. Mambaextend: A training-free approach to improve long context extension of mamba. In The Thirteenth International Conference on Learning Representations.  
+Runjin Chen, Zhenyu Zhang, Junyuan Hong, Souvik Kundu, and Zhangyang Wang. 2025. Seal: Steerable reasoning calibration of large language models for free. Conference on Language Modeling.  
+Krzysztof Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamas Sarlos, Peter Hawkins, Jared Davis, Afroz Mohiud
+
+din, Lukasz Kaiser, and 1 others. 2020. Rethinking attention with performers. arXiv preprint arXiv:2009.14794.  
+Tri Dao. 2024. FlashAttention-2: Faster attention with better parallelism and work partitioning. In International Conference on Learning Representations (ICLR).  
+Tri Dao and Albert Gu. 2024. Transformers are SSMs: Generalized models and efficient algorithms through structured state space duality. In International Conference on Machine Learning (ICML).  
+DeepSeek-AI. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. Preprint, arXiv:2501.12948.  
+Jusen Du, Weigao Sun, Disen Lan, Jiaxi Hu, and Yu Cheng. 2025. Mom: Linear sequence modeling with mixture-of-memories. arXiv preprint arXiv:2502.13685.  
+Qihang Fan, Huaibo Huang, and Ran He. 2024. Breaking the low-rank dilemma of linear attention. arXiv preprint arXiv:2411.07635.  
+Clémentine Fourrier, Nathan Habib, Alina Lozovskaya, Konrad Szafer, and Thomas Wolf. 2024. Open llm leaderboard v2. https://huggingface.co/spaces/open-llm-leaderboard/open_11m_1leaderboard.  
+Leo Gao, Jonathan Tow, Baber Abbasi, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Alain Le Noac'h, Haonan Li, Kyle McDonell, Niklas Muennighoff, Chris Ociepa, Jason Phang, Laria Reynolds, Hailey Schoelkopf, Aviya Skowron, Lintang Sutawika, and 5 others. 2024. The language model evaluation harness.  
+Saurav Goel. 2023. Paul graham essay collection dataset. https://huggingface.co/datasets/sgoel9/paul_graham Essays. Accessed: 2025-05-05.  
+Aaron Grattaftori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad AlDahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, and 542 others. 2024. The llama 3 herd of models. Preprint, arXiv:2407.21783.  
+Albert Gu and Tri Dao. 2023. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752.  
+Dongchen Han, Yifan Pu, Zhuofan Xia, Yizeng Han, Xuran Pan, Xiu Li, Jiwen Lu, Shiji Song, and Gao Huang. 2024. Bridging the divide: Reconsidering softmax and linear attention. Advances in Neural Information Processing Systems, 37:79221-79245.
+
+HazyResearch. 2024. ThunderKittens: HazyResearch/ThunderKittens. https://github.com/HazyResearch/ThunderKittens. Accessed: 2025-05-12.  
+Andrew G Howard. 2017. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861.  
+Cheng-Ping Hsieh, Simeng Sun, Samuel Kriman, Shantanu Acharya, Dima Rekesh, Fei Jia, Yang Zhang, and Boris Ginsburg. 2024. Ruler: What's the real context size of your long-context language models? arXiv preprint arXiv:2404.06654.  
+Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, and 1 others. 2022. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3.  
+Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. 2023. Mistral 7b. Preprint, arXiv:2310.06825.  
+Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. 2020. Transformers are rnns: Fast autoregressive transformers with linear attention. In International conference on machine learning, pages 5156-5165. PMLR.  
+Yoon Kim and Alexander M. Rush. 2016. Sequence-level knowledge distillation. Preprint, arXiv:1606.07947.  
+Wojciech Krysciński, Nazneen Rajani, Divyansh Agarwal, Caiming Xiong, and Dragomir Radev. 2021. Booksum: A collection of datasets for long-form narrative summarization.  
+Yuri Kuratov, Aydar Bulatov, Petr Anokhin, Ivan Rodkin, Dmitry Sorokin, Artyom Sorokin, and Mikhail Burtsev. 2024. Babilong: Testing the limits of llms with long context reasoning-in-a-haystack. In Advances in Neural Information Processing Systems, volume 37, pages 106519-106554. Curran Associates, Inc.  
+Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In International Conference on Learning Representations.  
+Jean Mercat, Igor Vasiljevic, Sedrick Scott Keh, Kushal Arora, Achal Dave, Adrien Gaidon, and Thomas Kollar. 2024. Linearizing large language models. In First Conference on Language Modeling.  
+Amirkeivan Mohtashami and Martin Jaggi. 2023. Random-access infinite context length for transformers. In Thirty-seventh Conference on Neural Information Processing Systems.
+
+Bo Peng, Daniel Goldstein, Quentin Gregory Anthony, Alon Albalak, Eric Alcaide, Stella Biderman, Eugene Cheah, Teddy Ferdinan, Kranthi Kiran GV, Haowen Hou, Satyapriya Krishna, Ronald McClelland Jr., Niklas Muennighoff, Fares Obeid, Atsushi Saito, Guangyu Song, Haoqin Tu, Ruichong Zhang, Bingchen Zhao, and 3 others. 2024. Eagle and finch: RWKV with matrix-valued states and dynamic recurrence. In First Conference on Language Modeling.  
+Zhen Qin, Xiaodong Han, Weixuan Sun, Dongxu Li, Lingpeng Kong, Nick Barnes, and Yiran Zhong. 2022. The devil in linear transformer. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 7025-7041, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Zhen Qin, Dong Li, Weigao Sun, Weixuan Sun, Xuyang Shen, Xiaodong Han, Yunshen Wei, Baohong Lv, Xiao Luo, Yu Qiao, and 1 others. 2023. Transnormerllm: A faster and better large language model with improved transnormer. arXiv preprint arXiv:2307.14995.  
+Zhen Qin, Songlin Yang, Weixuan Sun, Xuyang Shen, Dong Li, Weigao Sun, and Yiran Zhong. 2024. Hgrn2: Gated linear rnns with state expansion.  
+Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36:53728-53741.  
+Akshat Ramachandran, Mingyu Lee, Huan Xu, Souvik Kundu, and Tushar Krishna. 2025. Ouromamba: A data-free quantization framework for vision mamba models. International Conference on Computer Vision.  
+Yeonju Ro, Zhenyu Zhang, Souvik Kundu, Zhangyang Wang, and Aditya Akella. 2025. On-the-fly adaptive distillation of transformer to dual-state linear attention for long-context llm serving. In *Forty-second International Conference on Machine Learning*.  
+Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. 2024. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063.  
+Yutao Sun, Li Dong, Shaohan Huang, Shuming Ma, Yuqing Xia, Jilong Xue, Jianyong Wang, and Furu Wei. 2023. Retentive network: A successor to transformer for large language models. arXiv preprint arXiv:2307.08621.  
+Qwen Team. 2024. Qwen2.5: A party of foundation models.  
+Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2023. Attention is all you need. Preprint, arXiv:1706.03762.
+
+Junxiong Wang, Daniele Paliotta, Avner May, Alexander M Rush, and Tri Dao. 2024. The mamba in the llama: Distilling and accelerating hybrid models. In *The Thirty-eighth Annual Conference on Neural Information Processing Systems*.  
+Songlin Yang, Jan Kautz, and Ali Hatamizadeh. 2025. Gated delta networks: Improving mamba2 with delta rule. In The Thirteenth International Conference on Learning Representations.  
+Songlin Yang, Bailin Wang, Yikang Shen, Rameswar Panda, and Yoon Kim. 2023. Gated linear attention transformers with hardware-efficient training. arXiv preprint arXiv:2312.06635.  
+Songlin Yang, Bailin Wang, Yu Zhang, Yikang Shen, and Yoon Kim. 2024. Parallelizing linear transformers with the delta rule over sequence length. In *The Thirty-eighth Annual Conference on Neural Information Processing Systems*.  
+Songlin Yang and Yu Zhang. 2024. Fla: A triton-based library for hardware-efficient implementations of linear attention mechanism.  
+Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, Qianyu Chen, Huarong Zhou, Zhensheng Zou, Haoye Zhang, Shengding Hu, Zhi Zheng, Jie Zhou, Jie Cai, Xu Han, and 4 others. 2024. Minicpmv: A gpt-4v level mllm on your phone. arXiv preprint 2408.01800.  
+Zhifan Ye, Zheng Wang, Kejing Xia, Jihoon Hong, Leshu Li, Lexington Whalen, Cheng Wan, Yonggan Fu, Yingyan Celine Lin, and Souvik Kundu. 2025. Lamb: A training-free method to enhance the long-context understanding of ssms via attention-guided token filtering. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 1200-1209.  
+Michael Zhang, Simran Arora, Rahul Chalamala, Benjamin Frederick Spector, Alan Wu, Krithik Ramesh, Aaryan Singhal, and Christopher Re. 2025. LoL-CATs: On low-rank linearizing of large language models. In The Thirteenth International Conference on Learning Representations.  
+Michael Zhang, Kush Bhatia, Hermann Kumbong, and Christopher Re. 2024. The hedgehog & the porcupine: Expressive linear attentions with softmax mimicry. In *The Twelfth International Conference on Learning Representations*.
+
+# A Appendix
+
+# A.1 Training Dataset
+
+All datasets are in English. Passkey Retrieval We use the same format as (Mohtashami and Jaggi, 2023) to generate 10K samples to build the training dataset and 1K samples for validation. For the tokenizer of the Llama3.2 model, the length of each sample is 1162 tokens, and for the tokenizer of the Mistral model, the length is 1171 tokens.
+
+S-NIAH 1&2&3 We use the same format as (Hsieh et al., 2024) to generate 8456 samples to build the training dataset and 1K samples for validation. We used the BookSum dataset (Kryscinski et al., 2021) as the contextual 'haystack', different from the Paul Graham Essay Collection Dataset (Goel, 2023) in evaluation. For the tokenizer of the Llama3.2 model, the maximum length is 1231, the minimum length is 675, the average length is 785.50, the mode length is 768, and the std is 49. We also found that adding additional 1K data with the S-NIAH-1 format can improve the overall performance. For these 1K data, the maximum length is 1223, the minimum length is 1222, the average length is 1222.92, the mode length is 1223, and the std is 0.27.
+
+BABILong QA2&QA3 For the BABILong benchmark (Kuratov et al., 2024), training and validation data were generated on-the-fly per the original setup. Specifically, there are 10K samples for the training data of QA2 and 9868 samples for the training data of QA3. Different from the curriculum learning strategy in the original paper, we simplified the training by fine-tuning solely on sequences of length 1300 with a mixed format of QA2 and QA3.
+
+# A.2 Training Configuration
+
+All experiments were implemented using the Flash Linear Attention framework (Yang and Zhang, 2024). For the Passkey Retrieval task, we distilled and fine-tuned the LLaMA3.2 1B model over 4 epochs each, totaling  $1162 \times 10,000 \times 4 \times 2 = 92.96M$  training tokens. For larger models (LLaMA3 8B, Mistral v0.1 and v0.3 7B), distillation and fine-tuning were performed for 2 epochs each, resulting in  $1162 \times 10,000 \times 2 \times 2 = 92.96M$  tokens for LLaMA and  $1171 \times 10,000 \times 2 \times 2 = 93.68M$  tokens for Mistral. Similarly, for the NIAH tasks, we distilled and fine-tuned for 2 epochs, yielding approximately  $(785.5 \times 8456 + 1222.9 \times 1000) \times 2 \times 2 \approx 31.46M$  tokens. For
+
+the BABILong tasks (QA2 and QA3), datasets were generated dynamically; thus, exact token counts varied. With a maximum sequence length of 1300 tokens, the total training tokens were under  $1300 \times (10,000 + 9,868) \times 2 \times 2 \approx 103.31M$ .
+
+For the Passkey Retrieval task with Llama3.2 1B model, we use the cosine learning rate schedule, and for other models and other tasks, we all use the ReduceLROOnPlateau, which reduces the learning rate when the validation loss has stopped reducing. For all experiments, we use the Adamw optimizer (Loshchilov and Hutter, 2019) with an initial learning rate of 0.1 for distillation, 0.0001 for fine-tuning. For the LoRA fine-tuning, the rank is 8 and the alpha is 16, and we only add LoRA module for the linear layer for the query, key, value and output.
+
+We keep the similar training hyperparameters as Zhang et al. (2025). Specifically, for the Passkey Retrieval task with the LLaMA3.2-1B model, we adopt a cosine learning rate schedule, while for all other models and tasks, we employ the ReduceLROnPlateau scheduler, which decreases the learning rate when the validation loss plateaus. All experiments use the AdamW optimizer (Loshchilov and Hutter, 2019), with an initial learning rate of 0.1 during distillation and 1e-4 for fine-tuning. During LoRA fine-tuning, we set the rank to 8 and the scaling factor  $(\alpha)$  to 16, applying LoRA modules exclusively to the linear projections of the query, key, value, and output layers in the attention block. For all experiments, we run 3 times using different random seeds and report the results from the run achieving the best overall performance.
+
+# A.3 Other Ablation Study
+
+Do we need to share the linear and Conv layers? In Table 4, we analyze four distinct configurations regarding parameter sharing between linear and convolutional layers. Our empirical results demonstrate that optimal performance is achieved when allowing  $Q$  and  $K$  to independently process contextual information through separate convolutional layers while sharing the subsequent linear projection. This architecture aligns with the theoretical foundation in Equation 2, where both query and key utilize the same transformation function  $\phi(\cdot)$ . The separation of convolutional processing is intuitively justified, as queries and keys serve distinct roles in attention mechanisms and therefore benefit from specialized feature
+
+extraction pathways when incorporating previous token information. However, the shared linear projection enforces consistency in the embedding space where similarity is computed, maintaining the mathematical elegance of the linear attention formulation.
+
+<table><tr><td>C</td><td>L</td><td>1K</td><td>2K</td><td>3K</td><td>4K</td><td>5K</td><td>6K</td><td>7K</td><td>8K</td><td>9K</td></tr><tr><td>✓</td><td>✓</td><td>1.0</td><td>1.0</td><td>1.0</td><td>0.9</td><td>0.8</td><td>0.5</td><td>0.1</td><td>0.1</td><td>0</td></tr><tr><td>✓</td><td>×</td><td>0.9</td><td>1.0</td><td>0.3</td><td>0</td><td>0.2</td><td>0</td><td>0</td><td>0</td><td>0</td></tr><tr><td>×</td><td>×</td><td>1.0</td><td>1.0</td><td>1.0</td><td>0.8</td><td>0.8</td><td>0.6</td><td>0.3</td><td>0.1</td><td>0</td></tr><tr><td>×</td><td>✓</td><td>1.0</td><td>1.0</td><td>1.0</td><td>1.0</td><td>1.0</td><td>1.0</td><td>0.9</td><td>0.9</td><td>0.8</td></tr></table>
+
+Do we need to add RoPE to linear attention? Modern linear attention models exhibit diverse strategies for handling positional information. Some architectures, like GLA (Yang et al., 2023) and Gated DeltaNet (Yang et al., 2025), often omit explicit position embeddings, relying on their recurrent formulation to implicitly capture sequence order. Conversely, others, including TransNormer-LLM (Qin et al., 2023) and Retentive Networks (Sun et al., 2023), explicitly integrate mechanisms like rotary position embeddings (RoPE) (Su et al., 2024) or its variants. This divergence raises a question regarding knowledge distillation: when transferring knowledge from standard Transformers (which typically use position embeddings) to linear attention models, is it beneficial to retain these explicit positional signals? Ablation study presented in Table 5 addresses this for the LoL-CATs and LAWCAT. The results consistently indicate that incorporating RoPE actually impairs generalization performance on long-context tasks.
+
+For LoLCATs, removing RoPE decreases performance on shorter contexts (1K-3K tokens) but slightly improves performance on longer contexts (4K-8K tokens). This effect is more pronounced in our LAWCAT architecture, where the variant without RoPE maintains high accuracy form 1K to 8K context lengths, while the RoPE variant's performance drops to zero beyond 3K tokens.
+
+These results indicate that position embeddings constitute a critical factor limiting generalization to extended context lengths. The fixed positional encoding scheme optimized for training context appears to introduce inappropriate biases when
+
+Table 4: Accuracy on the passkey retrieval task from 1K to 9K. "C" and "L" mean if sharing the Conv1D or linear projection between query and key, "√" means yes, and "×" means no.  
+
+<table><tr><td>M</td><td>PE</td><td>1K</td><td>2K</td><td>3K</td><td>4K</td><td>5K</td><td>6K</td><td>7K</td><td>8K</td></tr><tr><td>LoL-</td><td>✓</td><td>0.9</td><td>0.8</td><td>0.3</td><td>0</td><td>0</td><td>0</td><td>0</td><td>0</td></tr><tr><td>CATs</td><td>×</td><td>0.4</td><td>0.2</td><td>0.1</td><td>0.2</td><td>0.2</td><td>0.1</td><td>0.2</td><td>0.2</td></tr><tr><td>LAW-</td><td>✓</td><td>1</td><td>0.8</td><td>0</td><td>0</td><td>0</td><td>0</td><td>0</td><td>0</td></tr><tr><td>CAT</td><td>×</td><td>1.0</td><td>1.0</td><td>1.0</td><td>1.0</td><td>1.0</td><td>1.0</td><td>0.9</td><td>0.9</td></tr></table>
+
+Table 5: The test accuracy on the passkey retrieval task from 1K to 8K. "M" means model name, and "PE" means if using the RoPE for linear attention. "√" means yes, and "×" means no.
+
+applied to longer sequences. This insight aligns with recent research suggesting that recurrent computation inherently captures sequential dependencies without requiring explicit positional information (Gu and Dao, 2023).
+
+Do we need sliding window attention? Prior work posited SWA as a key component for enhancing linear attention model performance (Zhang et al., 2025). To rigorously evaluate this claim, we conducted ablation studies on both the original LoL-CATs model and our proposed LAWCAT architecture, examining the impact of SWA across different task types and context lengths.
+
+Our investigation first focused on long-context retrieval tasks. As shown in Table 5, row 3, removing SWA significantly degrades the performance of the LoLCATs model, suggesting its reliance on SWA for effective passkey retrieval. Conversely, incorporating SWA into our LAWCAT model adversely affected performance, particularly as context length increased. We hypothesize that this stems from the difficulty in optimally balancing the contributions of GLA and SWA. Furthermore, achieving a balance that generalizes from shorter training sequences to longer evaluation sequences appears challenging. This suggests it is difficult to balance the weights between linear attention and softmax attention, and even if the model can make their combination fit well on training data, it may be hard to generalize it to longer context lengths. For example, the ideal weights of LA and SWA for 1K-length contexts may not be good for 8K-length contexts.
+
+We further assessed the effect of SWA on standard short-context benchmarks using LM Evaluation Harness tasks (LM Eval) (Gao et al., 2024) (Table 7). When SWA is removed, LAWCAT substantially outperforms LoLCATs, consistent with our findings on long-context tasks. However, adding SWA (window size 64) significantly boosts per
+
+<table><tr><td>M</td><td>SWA</td><td>1K</td><td>2K</td><td>3K</td><td>4K</td><td>5K</td><td>6K</td><td>7K</td><td>8K</td></tr><tr><td>LoL-</td><td>✓</td><td>0.9</td><td>0.8</td><td>0.3</td><td>0</td><td>0</td><td>0</td><td>0</td><td>0</td></tr><tr><td>CATs</td><td>×</td><td>0.1</td><td>0.0</td><td>0</td><td>0</td><td>0</td><td>0</td><td>0</td><td>0</td></tr><tr><td>LAW-</td><td>✓</td><td>0.5</td><td>0.5</td><td>0.2</td><td>0.1</td><td>0</td><td>0</td><td>0</td><td>0</td></tr><tr><td>CAT</td><td>×</td><td>1.0</td><td>1.0</td><td>1.0</td><td>1.0</td><td>1.0</td><td>1.0</td><td>0.9</td><td>0.9</td></tr></table>
+
+Table 6: The test accuracy on the passkey retrieval task from 1K to 8K. "M" means model name, and "SWA" means if using the sliding window attention. "√" means yes, and "×" means no.  
+
+<table><tr><td>Model</td><td>PI</td><td>AE</td><td>AC</td><td>HS</td><td>WG</td><td>Avg.</td></tr><tr><td colspan="7">Pre-trained Transformer Model</td></tr><tr><td>Llama 3.2 1B</td><td>74.4</td><td>65.5</td><td>35.8</td><td>63.7</td><td>60.5</td><td>60.0</td></tr><tr><td>Llama 3.2 1B*</td><td>73.2</td><td>64.5</td><td>35.6</td><td>40.3</td><td>59.9</td><td>54.7</td></tr><tr><td colspan="7">Pre-trained Recurrent Model</td></tr><tr><td>DeltaNet 1.3B</td><td>71.2</td><td>57.2</td><td>28.3</td><td>50.2</td><td>53.6</td><td>52.1</td></tr><tr><td>GLA 1.3B</td><td>71.8</td><td>57.2</td><td>26.6</td><td>49.8</td><td>53.9</td><td>51.9</td></tr><tr><td>Mamba 2 1.3B</td><td>73.2</td><td>64.3</td><td>33.3</td><td>59.9</td><td>60.9</td><td>58.3</td></tr><tr><td colspan="7">Distilled Model</td></tr><tr><td>LoLCATs</td><td>59.4</td><td>41.2</td><td>23.8</td><td>31.4</td><td>50.2</td><td>41.2</td></tr><tr><td>LoLCATs*</td><td>74.6</td><td>63.0</td><td>35.1</td><td>63.7</td><td>61.5</td><td>59.6</td></tr><tr><td>Ours</td><td>71.9</td><td>65.0</td><td>36.3</td><td>53.1</td><td>52.8</td><td>55.8</td></tr><tr><td>Ours*</td><td>75.3</td><td>65.2</td><td>35.1</td><td>62.0</td><td>62.0</td><td>59.9</td></tr></table>
+
+Table 7: Comparison of various models on 5 tasks from LM EVAL benchmark. Model name with * indicates the model add sliding window attention with size of 64. Specially, Llama 3.2 1B * means the Llama 3.2 1B model only uses SWA.
+
+formance for both models, bringing them to comparable levels (59.6 vs. 59.9 average score). For context, we include results from relevant baselines: the standard Llama 3.2 1B transformer and several SOTA recurrent models. Notably, a modified Llama 3.2 1B using only SWA (denoted Llama 3.2  $1\mathrm{B}^{\ast}$ ) achieves competitive performance (54.7 avg.) compared to the original (60.0 avg.), albeit with a marked drop on HellaSwag. This suggests that SAW alone possesses considerable capability. The strong results of LoLCATs may therefore be primarily attributed to the inherent effectiveness of SWA, with the linear component potentially offering further gains on specific tasks like HellaSwag.
+
+In conclusion, our ablation studies reveal a nuanced role for SWA. While its integration can substantially enhance performance on short-context tasks, potentially contingent on successful optimization of the hybrid attention mechanism, it appears detrimental to generalization capability in long-context scenarios.
+
+Ablation study on the rank of the gate function in GLA GLA models (Yang et al., 2023) utilize 2 consecutive low-rank linear layers to implement the gate function for efficiency, with the original models adopting a rank of 16. We investigate the impact of this rank hyperparameter in our LAWCAT models, distilled from Llama3.2 1B Instruct (Grattafori et al., 2024), on complex S-NIAH retrieval tasks: S-NIAH-1 (pass-key), S-NIAH-2 (number), and the particularly demanding S-NIAH-3 (UUID retrieval).
+
+As shown in Table 8, LAWCAT models with rank 16 performed well on S-NIAH-1 and S-NIAH-2 but struggled significantly on S-NIAH-3 (e.g.,  $12\%$  at 1K,  $0\%$  at 2K). This indicates that a rank of 16 limits the forget gate's capacity for recalling long, complex items like UUIDs. Increasing the rank to 32 substantially improved S-NIAH-3 performance (e.g., to  $56\%$  at 1K and  $44\%$  at 2K) while maintaining strong performance on S-NIAH-1 and S-NIAH-2. This suggests rank 32 offers a better balance of capacity for complex retrieval.
+
+Further rank increases to 64 yielded mixed results; while improving some S-NIAH-1/2 scores at longer contexts (e.g., S-NIAH-1 8K at  $92\%$ ; S-NIAH-2 8K at  $64\%$ ), it did not consistently improve S-NIAH-3 over rank 32 ( $36\%$  vs  $56\%$  at 1K). Ranks 128 and Full Rank (FR) generally led to performance degradation across tasks, particularly at longer contexts (e.g., rank 128 scored  $36\%$  on S-NIAH-1 8K). We hypothesize this non-monotonic trend stems from higher ranks introducing optimization challenges due to increased parameters and a greater risk of overfitting, which can impair the forget gate's ability to learn generalizable retention patterns.
+
+In conclusion, these findings indicate an optimal gate function rank—around 32 for our 1B parameter LAWCAT models—that balances model capacity for complex retrieval against optimization difficulties and overfitting.
+
+# A.4 The visualization of Attention Scores
+
+To qualitatively assess the attention mechanisms, we visualize attention score heatmaps from a representative head on an example from the S-NIAH-3 task.
+
+Figure 5 illustrates an attention head primarily focused on local contextual information. A comparative analysis of the heatmaps reveals that while the Transformer, LoLCATs, and LAWCAT models
+
+<table><tr><td></td><td colspan="4">S-NIAH-1(pass-key retrieval)</td><td colspan="4">S-NIAH-2 (number in haystack)</td><td colspan="3">S-NIAH-3 (uuid in haystack)</td></tr><tr><td>Rank</td><td>1K</td><td>2K</td><td>4K</td><td>8K</td><td>1K</td><td>2K</td><td>4K</td><td>8K</td><td>1K</td><td>2K</td><td>4K</td></tr><tr><td>16</td><td>100</td><td>100</td><td>96</td><td>80</td><td>100</td><td>96</td><td>92</td><td>52</td><td>12</td><td>0</td><td>4</td></tr><tr><td>32</td><td>100</td><td>100</td><td>100</td><td>80</td><td>100</td><td>96</td><td>88</td><td>48</td><td>56</td><td>44</td><td>24</td></tr><tr><td>64</td><td>100</td><td>92</td><td>100</td><td>92</td><td>100</td><td>96</td><td>96</td><td>64</td><td>36</td><td>32</td><td>20</td></tr><tr><td>128</td><td>100</td><td>92</td><td>88</td><td>36</td><td>92</td><td>72</td><td>72</td><td>4</td><td>20</td><td>24</td><td>12</td></tr><tr><td>FR</td><td>96</td><td>96</td><td>80</td><td>12</td><td>96</td><td>72</td><td>36</td><td>4</td><td>24</td><td>28</td><td>12</td></tr></table>
+
+Table 8: Performance comparison across different models on S-NIAH 1, 2, and 3 tasks.
+
+exhibit broadly similar local attention patterns, notable differences emerge. The LoLCATs model displays a tendency for its attention to be diffused by more distant preceding tokens, potentially diluting its focus. In contrast, LAWCAT maintains a sharper concentration on the immediately relevant local context, particularly the current token, mirroring the behavior of the standard softmax Transformer more closely. This suggests that the integrated Conv1D layer in LAWCAT effectively strengthens local bias. By design, the convolutional operation aggregates information from adjacent tokens, allowing the model to consolidate the importance of the local neighborhood onto the current token's representation, thereby achieving a more focused attention distribution akin to softmax attention.
+
+Further insights are provided by Figure 6, which depicts an attention head responsible for identifying the relationship between an answer and the "needle" (the target information to be retrieved). In this scenario, LoLCATs tends to allocate attention more broadly across the sequence, a characteristic that can hinder precise identification of the crucial answer-needle relationship. LAWCAT, however, mitigates this diffusion, exhibiting a more localized attention pattern that closely resembles the Transformer's focused engagement on the relevant segments. This closer alignment in attention scores between LAWCAT and the softmax Transformer not only correlates with LAWCAT's superior performance on retrieval tasks but also empirically validates the efficacy of the Conv1D layer in fostering a beneficial local inductive bias, enabling more precise attention allocation.
+
+# A.5 More Results on BABILong
+
+We also added the rustles on BABILong of the Llama 3 8B Instruct and the LAWCAT variants, as shown in Fig. 7, our LAWCAT variants consistently outperform the baselines, particularly at longer context lengths, mirroring the findings from experi
+
+![](images/1d55082f0570a891b2bc451611bfe8e8872024c3c528e05217b3ea0fc3757ce0.jpg)
+
+![](images/8f4e70c3281cb2b3fa4a49f6ed95fb2873be9c0d26e0e7c8673dd7ab0bb6e48d.jpg)
+
+![](images/056520ccde6d9918e6ba0420799ab5178990d1bb49a81ea3ab35e7ae5cf2e4cc.jpg)  
+Figure 5: Visualization of attention scores from layer 15, head 5 across three models: Transformer (top), LoL-CATs (middle), and LAWCAT (bottom). Each row presents three attention maps: needle-to-needle (left), answer-to-needle (center), and answer-to-answer (right)
+
+![](images/44b99f81cf0a3fc26b066ea08aa521b46f2118d129d242c3d667b0fa673833d2.jpg)
+
+![](images/0c2df3003fa8db6dc9031960b0c0ece89fbbfc6af3ac82af0cb86a3a646283df.jpg)
+
+![](images/b4b61281c0dae728459c420904e3a2be95ce356d7949c3aa9494661a8a60f3e6.jpg)  
+Figure 6: Visualization of attention scores from layer 15, head 22 across three models: Transformer (top), LoLCATs (middle), and LAWCAT (bottom). Each row presents three attention maps: needle-to-needle (left), answer-to-needle (center), and answer-to-answer (right)
+
+![](images/ce5b9b338aab74af7ce175f31f8dbfa05d77ac6d5228a76972fbcb382b90f277.jpg)  
+Figure 7: Comparison of accuracy on QA1&QA2&QA3 from BABILong benchmark between Llama 3 8B Instruct (left) and LAWCAT (right).
+
+![](images/afc9b84a8f11e270691eb26d3d39be50302c9359632ed7f9ef03980ead27356f.jpg)
+
+![](images/4bd3c30caeade6db50e22dbb2087910094f0d80fc4591555432a572a3d81c01e.jpg)  
+Figure 8: Comparison of accuracy on QA1&QA2&QA3 from BABILong benchmark between Mistral v0.3 7B (left) and LAWCAT (right).
+
+![](images/be62f480d1f76b7200d6654edcbbf9d863893724da4a93236be2bd0940a8b2e5.jpg)
+
+ments with the smaller LLaMA 3.2 1B model. As shown in Fig. 8, compared to the original model, the distilled Mistral v0.3 7B model exhibits significantly improved performance on inputs longer than 8k tokens. However, it suffers from noticeable performance degradation on shorter context lengths, especially the 0k setting, which does not contain any "haystack" content. We hypothesize that this issue arises because Mistral v0.3 is a pretrained base model that is highly sensitive to prompt formatting, particularly after limited fine-tuning. We observed that, instead of producing incorrect answers, the model often outputs nothing for 0k inputs.
+
+<table><tr><td>Task</td><td>0K</td><td>1K</td><td>2K</td><td>4K</td><td>8K</td><td>16K</td><td>32K</td></tr><tr><td colspan="8">Pre-trained Transformer Model</td></tr><tr><td>QA2</td><td>44</td><td>52</td><td>47</td><td>35</td><td>14</td><td>5</td><td>2</td></tr><tr><td>QA3</td><td>36</td><td>34</td><td>35</td><td>23</td><td>27</td><td>23</td><td>16</td></tr><tr><td colspan="8">LAWCAT</td></tr><tr><td>QA2</td><td>94</td><td>97</td><td>95</td><td>90</td><td>82</td><td>32</td><td>8</td></tr><tr><td>QA3</td><td>71</td><td>73</td><td>71</td><td>68</td><td>44</td><td>18</td><td>9</td></tr></table>
+
+Therefore, we added the results of Mistral v0.1 Instruct 7B and Mistral v0.3 Instruct 7B models on the QA2 and QA3 tasks of the BABILong benchmark, as shown in Table 9 and Table 10 respectively. As expected, applying LAWCAT to the models af
+
+Table 9: Results of Mistral v0.1 Instruct 7B on BABI-Long benchmark (QA2&QA3)  
+
+<table><tr><td>Task</td><td>0K</td><td>1K</td><td>2K</td><td>4K</td><td>8K</td><td>16K</td><td>32K</td></tr><tr><td colspan="8">Pre-trained Transformer Model</td></tr><tr><td>QA2</td><td>48</td><td>41</td><td>31</td><td>19</td><td>3</td><td>2</td><td>1</td></tr><tr><td>QA3</td><td>36</td><td>32</td><td>36</td><td>25</td><td>22</td><td>18</td><td>19</td></tr><tr><td colspan="8">LAWCAT</td></tr><tr><td>QA2</td><td>87</td><td>92</td><td>95</td><td>96</td><td>96</td><td>72</td><td>43</td></tr><tr><td>QA3</td><td>59</td><td>55</td><td>66</td><td>72</td><td>69</td><td>64</td><td>43</td></tr></table>
+
+Table 10: Results of Mistral v0.3 Instruct 7B on BABI-Long benchmark (QA2&QA3)
+
+ter instruction tuning will result in more stable performance, and the results of the Mistral models are consistent with those of the Llama model (although there are slight accuracy drops for the Mistral v0.1 7B Instruct LAWCAT variant model on QA3 when the context length is longer than  $16\mathrm{k}$ ).
+
+Overall, we believe these results further validate the effectiveness of our LAWCAT on various model architectures and tasks.

2025/LAWCAT_ Efficient Distillation from Quadratic to Linear Attention with Convolution across Tokens for Long Context Modeling/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5b33f47e8d0418c2a80e771ef5b3278bb4af44501ef10311aa2290404c35c6a9
+size 621811

2025/LC-Eval_ A Bilingual Multi-Task Evaluation Benchmark for Long-Context Understanding/bfbb1c17-291f-491b-b8e4-17a9af50d32d_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2feae1603c43f227ccc60ae333668dafabe3909acf77be5a0965b90d4f8ec313
+size 1085385

2025/LC-Eval_ A Bilingual Multi-Task Evaluation Benchmark for Long-Context Understanding/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:085df3e82ba79372c4fd977d8cca840e4c814aea8b91cd827229062d2056570b
+size 2393228

2025/LCAN_ A Label-Aware Contrastive Attention Network for Multi-Intent Recognition and Slot Filling in Task-Oriented Dialogue Systems/82b6ed46-a9f8-45e5-85b9-e1b0bc58ecb6_content_list.json

@@ -0,0 +1,2039 @@
+[
+    {
+        "type": "text",
+        "text": "LCAN: A Label-Aware Contrastive Attention Network for Multi-Intent Recognition and Slot Filling in Task-Oriented Dialogue Systems",
+        "text_level": 1,
+        "bbox": [
+            129,
+            89,
+            870,
+            130
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Shuli Zhang $^{1}$ , Zhiqiang You $^{1}$ , Xiaoxiang Qi $^{1}$ , Peng Liu $^{2}$ , Gaode Wu $^{3}$ , Kan Xia $^{3}$ , Shenguang Huang $^{3}$ ,",
+        "bbox": [
+            156,
+            149,
+            845,
+            185
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$ Hunan University,  $^{2}$ Guangdong University of Technology,  $^{3}$ Ningbo Port Information Communication Co., Ltd.",
+        "bbox": [
+            258,
+            185,
+            736,
+            219
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            260,
+            339,
+            275
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Multi-intent utterances processing remains a persistent challenge due to intricate intent-slot dependencies and semantic ambiguities. Traditional methods struggle to model these complex interactions, particularly when handling overlapping slot structures across multiple intents. This paper introduces a label-aware contrastive attention network (LCAN), a joint modeling approach for multi-intent recognition and slot filling in task-oriented dialogue systems. LCAN addresses this issue by integrating label-aware attention and contrastive learning strategies, improving semantic understanding and generalization in multi-intent scenarios. Extensive experiments on the MixATIS and MixSNIPS datasets demonstrate LCAN's superiority over existing models, achieving improved intent recognition and slot filling performance, particularly in handling overlapping or complex semantic structures in multi-intent settings.",
+        "bbox": [
+            141,
+            288,
+            460,
+            575
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            588,
+            260,
+            604
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Task-oriented dialogue (TOD) systems are critical for applications like virtual assistants and customer service automation. A key challenge lies in multi-intent scenarios where single utterances express multiple intents and slots, requiring precise identification for optimal user experience (Wang et al., 2021).",
+        "bbox": [
+            112,
+            614,
+            489,
+            726
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Traditional methods (Xu and Sarikaya, 2013; Zhang and Wang, 2016) treated intent recognition and slot filling as separate tasks. Recent studies reveal their intrinsic connections, driving adoption of joint modeling to enable information sharing, reduce error propagation, and improve efficiency.",
+        "bbox": [
+            112,
+            727,
+            489,
+            824
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Early joint models combined CRFs with CNNs (Xu and Sarikaya, 2013) or GRUs (Zhang and Wang, 2016), later enhanced by attention mechanisms (Liu and Lane, 2016). Pre-trained models revolutionized the field, with Joint BERT (Chen et al., 2019) leveraging BERT embeddings. Subse",
+        "bbox": [
+            112,
+            824,
+            489,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "quent innovations include syntax-enhanced Transformers (Wang et al., 2021) and bilinear attention for parallel intent-slot extraction (Chen et al., 2022a).",
+        "bbox": [
+            507,
+            261,
+            884,
+            324
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Despite advances, multi-intent handling remains challenging. As shown in Figure 1, utterances like \"What does ewr stand for, what airlines fly from burbank to denver and airports in New York?\" contain multiple intents with distinct slot types. Current models often fail to accurately disentangle and align these overlapping semantic components.",
+        "bbox": [
+            507,
+            325,
+            884,
+            439
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/a0906b834e3ad0a666be20e5d00d4c1c3544cde52114367da6a623e45f530398.jpg",
+        "image_caption": [
+            "Figure 1: An example of utterance with multiple intents and slots."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            519,
+            455,
+            875,
+            491
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "As dialogue systems increasingly handle multi- intent scenarios in practical applications, researchers have focused on effective identification and alignment of multiple intents and slots in complex interactions. Kim et al. (2017b) first explored multi-intent recognition in spoken language understanding and introduced joint multi-intent modeling. Gangadharaiah (2019) proposed a joint framework with slot-gating mechanisms, though semantic ambiguity persisted due to single-vector guidance for multiple slot fillings.",
+        "bbox": [
+            507,
+            550,
+            884,
+            727
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To address this limitation, Qin et al. (2020) developed a Graph Attention Network (GAT)-based interaction model establishing fine-grained intent-slot mappings. However, its autoregressive structure constrained bidirectional information capture. Subsequently, Qin et al. (2021a) introduced a global-local graph framework to model complex intent-slot and slot-slot relations, enhancing joint modeling performance.",
+        "bbox": [
+            507,
+            728,
+            884,
+            872
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Xing and Tsang (2022) proposed a heterogeneous graph with mutual \"slot-to-intent\" and \"intent-to-slot\" guidance, coupled with label-aware",
+        "bbox": [
+            507,
+            873,
+            882,
+            921
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "25603",
+        "bbox": [
+            473,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2025, pages 25603-25612 November 4-9, 2025 ©2025 Association for Computational Linguistics",
+        "bbox": [
+            208,
+            945,
+            788,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "decoding for enhanced interaction. Cheng et al. (2023a) designed the SSRN model with explicit intent scope identification and auxiliary tasks to mitigate error propagation. Tu et al. (2023) introduced BiSLU, combining supervised contrastive learning with self-distillation for bidirectional multitask guidance. Cheng et al. (2023b) further enhanced intent-slot synergy through multi-granular contrastive learning.",
+        "bbox": [
+            112,
+            84,
+            489,
+            229
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Recent advances include Chen et al. (2024)'s two-stage contrastive learning with data augmentation for shared intent relations, and Ma et al. (2024)'s generative task unification using prompt-enhanced fusion. Zhuang et al. (2024) proposed interpretable joint modeling via cross-task information maximization.",
+        "bbox": [
+            112,
+            235,
+            489,
+            348
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Despite progress, significant challenges remain in multi-intent scenarios: (1) Complex intent-slot interactions hinder accurate detection and filling, and (2) Semantic ambiguity persists when intents share slots or exhibit unclear boundaries.",
+        "bbox": [
+            112,
+            355,
+            489,
+            434
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We propose LCAN (Label-Aware Contrastive Attention Network), a novel joint model integrating label-aware attention with contrastive learning to address these issues. LCAN enhances semantic understanding through contrastive intent-slot alignment while employing label-aware attention to reduce input ambiguity. Our framework enables effective information sharing between tasks while minimizing semantic interference.",
+        "bbox": [
+            112,
+            441,
+            489,
+            586
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Experiments on MixATIS and MixSNIPS datasets demonstrate LCAN's superiority over state-of-the-art baselines, particularly in complex multi-intent scenarios. Ablation studies confirm the critical role of contrastive learning, with component-wise analysis revealing performance contributions.",
+        "bbox": [
+            112,
+            592,
+            489,
+            705
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we make the following contributions: (1) We propose LCAN, a joint model that combines label-aware attention and contrastive learning to improve multi-intent recognition and slot filling in task-oriented dialogue systems. (2) We demonstrate the effectiveness of LCAN through experiments on two standard benchmark datasets, MixATIS and MixSNIPS, and compare its performance with existing state-of-the-art models. (3) We conduct ablation studies to analyze the contribution of each component of LCAN and show that the contrastive learning modules play a key role in enhancing the model's performance.",
+        "bbox": [
+            112,
+            712,
+            489,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related work",
+        "text_level": 1,
+        "bbox": [
+            509,
+            83,
+            662,
+            98
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 Multi-Intent NLU",
+        "text_level": 1,
+        "bbox": [
+            509,
+            109,
+            697,
+            124
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In comparison to single-intent detection (Wu et al., 2020), multi-intent detection is more common in real-world scenarios, especially in real-time dialogue systems. Early research on multi-intent detection (Kim et al., 2017a; Gangadharajaiah, 2019) attempted to apply traditional methods based on Convolutional Neural Networks or Recurrent Neural Networks, which were effective for capturing sequential patterns in input data. However, these models struggled with the complexity and ambiguity inherent in handling multiple intents simultaneously.",
+        "bbox": [
+            507,
+            131,
+            884,
+            323
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To address these challenges, Qin et al. (2020) proposed the Adaptive Graph-Interactive Framework, which incorporates graph-based interactions to model relationships between intents and slots more effectively. This approach was further extended to a non-autoregressive model by Qin et al. (2021a), improving the efficiency of multi-intent detection while reducing the limitations imposed by autoregressive structures. The non-autoregressive model offers a significant advantage in terms of computational efficiency, enabling real-time processing for multi-intent tasks.",
+        "bbox": [
+            507,
+            325,
+            884,
+            517
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In another approach, Cai et al. (2022) extends the JointBERT framework, which was initially designed for single-intent tasks, to handle multi-intent scenarios. It effectively addresses the shared-intent problem, where multiple intents may share common slot labels, by explicitly mapping the relationship between slots and their corresponding intents.",
+        "bbox": [
+            507,
+            518,
+            884,
+            631
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Recognizing the importance of accurately predicting the number of intents in a given utterance, Chen et al. (2022c) developed a novel threshold-free framework. This framework first predicts the number of intents present in the input before proceeding to predict the specific intents. This approach provides a more efficient and reliable way to handle multi-intent tasks, as it eliminates the need for predefined thresholds, which can vary across different datasets or domains.",
+        "bbox": [
+            507,
+            631,
+            884,
+            791
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2 Contrastive Learning",
+        "text_level": 1,
+        "bbox": [
+            507,
+            803,
+            726,
+            819
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Contrastive Learning has gained widespread application in Natural Language Understanding tasks due to the challenges of data scarcity and the diversity of expressions in real-world language. Recent studies have demonstrated the effectiveness of CL in enhancing NLU performance (Gunel et al., 2020;",
+        "bbox": [
+            507,
+            825,
+            884,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "25604",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Hou et al., 2021; Yehudai et al., 2023) Specifically, for multi-intent NLU tasks, Vulic et al. (2022) proposed a strategy that adapts a general sentence encoder into a task-specific one for multi-intent data by leveraging contrastive learning. This approach improves the encoder's ability to distinguish between multiple intents in a single input.",
+        "bbox": [
+            112,
+            84,
+            489,
+            197
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In a similar vein, Tu et al. (2023) introduced a novel bidirectional joint model trained using supervised contrastive learning and self-distillation. This model effectively utilizes both intent and slot features, allowing them to complement each other, and improving the overall performance of multi-intent NLU tasks. The integration of supervised CL with self-distillation enables the model to learn richer representations, enhancing its ability to handle complex multi-intent scenarios.",
+        "bbox": [
+            112,
+            198,
+            489,
+            357
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Model Design",
+        "text_level": 1,
+        "bbox": [
+            112,
+            374,
+            265,
+            390
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The LCAN model is designed around a label-aware attention mechanism, which enhances semantic discrimination through contrastive supervision. It further introduces a collaborative attention module to enable bidirectional interaction between intents and slots for efficient joint modeling. As illustrated in Figure 2, the overall architecture consists of four main components: a text encoder, a label-aware attention mechanism, a contrastive learning module, and intent-slot decoders.",
+        "bbox": [
+            112,
+            401,
+            489,
+            561
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/680638dd5fe9bf475582577069d4b7f6794a1b885e40686a7714d91a1f6229f9.jpg",
+        "image_caption": [
+            "Figure 2: Illustration of the architecture of our joint model LCAN."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            122,
+            577,
+            480,
+            772
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Text Encoder",
+        "text_level": 1,
+        "bbox": [
+            112,
+            850,
+            265,
+            864
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Following previous work (Qin et al., 2020; Song et al., 2022), we employ a shared encoder and task-specific encoders.",
+        "bbox": [
+            112,
+            873,
+            489,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1.1 Shared Encoder",
+        "text_level": 1,
+        "bbox": [
+            507,
+            84,
+            695,
+            98
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We first use a shared encoder to extract general features. The shared encoder is based on RoBERTa to obtain contextual representations for each word. For an input sentence  $\\mathbf{x} = \\{x_{1}, x_{2}, \\ldots, x_{n}\\}$ , where  $n$  is the number of words, the shared encoder produces the following context-aware word embeddings  $\\mathbf{H} = \\{\\mathbf{h}_1, \\mathbf{h}_2, \\ldots, \\mathbf{h}_n\\}$ , where each  $\\mathbf{h}_i$  is the contextualized embedding of word  $x_i$ :",
+        "bbox": [
+            507,
+            104,
+            884,
+            233
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {H} = \\operatorname {R o B E R T a} (\\mathbf {x}) \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            633,
+            247,
+            882,
+            261
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1.2 Task-Specific Encoders",
+        "text_level": 1,
+        "bbox": [
+            507,
+            275,
+            752,
+            290
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To refine the representations for intent recognition and slot filling, we introduce two independent task-specific BiLSTM encoders. The intent-specific encoder captures global semantic features, while the slot-specific encoder focuses on fine-grained word-level features.",
+        "bbox": [
+            507,
+            294,
+            882,
+            388
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The input sequence  $\\mathbf{H}$  is passed through the Intent BiLSTM to get the refined representation for intent recognition, denoted as  $\\mathbf{H}_{\\mathrm{intent}}$ , and through the Slot BiLSTM to get the refined representation for slot filling, denoted as  $\\mathbf{H}_{\\mathrm{slot}}$ :",
+        "bbox": [
+            507,
+            392,
+            882,
+            472
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {H} _ {\\text {i n t e n t}} = \\operatorname {B i L S T M} _ {\\text {i n t e n t}} (\\mathbf {H}) \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            611,
+            486,
+            882,
+            501
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {H} _ {\\text {s l o t}} = \\operatorname {B i L S T M} _ {\\text {s l o t}} (\\mathbf {H}) \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            618,
+            519,
+            882,
+            533
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 Label-Aware Attention",
+        "text_level": 1,
+        "bbox": [
+            507,
+            548,
+            737,
+            562
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To extract label-specific contextual representations from the input sequence, LCAN introduces a multi-level label attention mechanism for both intent and slot labels. This module learns individual attention distributions for each label, generating semantic vectors that serve as prior knowledge for joint modeling.",
+        "bbox": [
+            507,
+            570,
+            882,
+            681
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2.1 Intent Label Attention",
+        "text_level": 1,
+        "bbox": [
+            507,
+            693,
+            746,
+            707
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The goal of intent label attention is to obtain the semantic representation of each token in the input sentence for every potential intent label. Initially, the model uses a specialized encoder to extract token-level representations  $\\mathbf{H}_{\\mathrm{int} \\mathrm{ent}} \\in \\mathbb{R}^{d_{e} \\times n}$ , where  $n$  is the number of tokens, and  $d_{e}$  is the embedding dimension. A query matrix  $\\mathbf{Q}^{\\mathrm{I}} \\in \\mathbb{R}^{|L^{\\mathrm{I}}| \\times d_{a}}$ , where  $|L^{\\mathrm{I}}|$  is the number of intent labels, is initialized with label embeddings. This query matrix is projected into the attention space, and attention weights are computed as:",
+        "bbox": [
+            507,
+            713,
+            882,
+            889
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {A} ^ {\\mathrm {I}} = \\operatorname {s o f t m a x} \\left(\\mathbf {Q} ^ {\\mathrm {I}} \\times \\tanh  \\left(\\mathbf {D} ^ {\\mathrm {I}} \\times \\mathbf {H} _ {\\text {i n t e n t}}\\right) \\in \\mathbb {R} ^ {| \\mathrm {L} ^ {\\mathrm {I}} | \\times n} \\right. \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            521,
+            903,
+            882,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "25605",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where, the softmax function normalizes across the row dimension, ensuring that each label's weights sum to one, and  $\\mathbf{D}^{\\mathrm{I}}\\in \\mathbb{R}^{d_a\\times d_e}$ . The sentence-level representation for each intent label is calculated via a weighted sum of token embeddings:",
+        "bbox": [
+            112,
+            84,
+            485,
+            164
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {V} ^ {\\mathrm {I}} = \\mathbf {H} _ {\\text {I n t e n t}} \\times \\left(\\mathbf {A} ^ {\\mathrm {I}}\\right) ^ {T} \\in \\mathbb {R} ^ {d _ {e} \\times | \\mathrm {L} ^ {\\mathrm {I}} |} \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            189,
+            174,
+            487,
+            199
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2.2 Slot Label Attention",
+        "text_level": 1,
+        "bbox": [
+            112,
+            206,
+            334,
+            221
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In contrast to intent labels, slot labels usually have more complex hierarchical structures, we introduce a multi-layer slot label attention mechanism. At each layer  $\\mathbf{H}_{\\mathrm{slot}} \\in \\mathbb{R}^{d_e \\times n}$ , the model extracts token-level representations  $\\mathbf{Q}^{\\mathrm{S},k} \\in \\mathbb{R}^{\\lfloor \\mathrm{L}^{\\mathrm{S},k} \\rfloor \\times d_a}$  from a slot-specific encoder. A query matrix  $\\mathbf{D}^{\\mathrm{S},k} \\in \\mathbb{R}^{d_a \\times d_e}$  is used to calculate attention weights, which are projected into the attention space as follows:",
+        "bbox": [
+            112,
+            225,
+            487,
+            355
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {A} ^ {\\mathrm {S}, k} = \\operatorname {s o f t m a x} \\left(\\mathbf {Q} ^ {\\mathrm {S}, k} \\times \\tanh  \\left(\\mathbf {D} ^ {\\mathrm {S}, k} \\times \\mathbf {H} _ {\\text {s l o t}}\\right) \\quad (6) \\right.\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            152,
+            367,
+            487,
+            385
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "After calculating the attention for each slot label at level  $k$ , the model computes the sentence representation for all slot labels at that level as:",
+        "bbox": [
+            112,
+            390,
+            489,
+            437
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {V} ^ {\\mathrm {S}, k} = \\mathbf {H} _ {\\text {S l o t}} \\times \\left(\\mathbf {A} ^ {\\mathrm {S}, k}\\right) ^ {T} \\in \\mathbb {R} ^ {d _ {e} \\times | \\mathrm {L} ^ {\\mathrm {S}, k} |} \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            173,
+            448,
+            487,
+            472
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2.3 Fine-Grained Semantic Enhancement",
+        "text_level": 1,
+        "bbox": [
+            112,
+            480,
+            470,
+            494
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To fully leverage the semantic information from slot labels at higher layers, the model enhances the fine-grained slot label representations by incorporating semantic information from coarser labels. Specifically, the model applies a feed-forward network followed by a sigmoid activation function to estimate the prediction probability for each coarse-level slot label:",
+        "bbox": [
+            112,
+            500,
+            489,
+            626
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\np _ {j} ^ {\\mathrm {S}, k - 1} = \\sigma \\left(\\mathbf {w} _ {j} ^ {\\mathrm {S}, k - 1} \\cdot \\mathbf {v} _ {j} ^ {\\mathrm {S}, k - 1}\\right) \\tag {8}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            203,
+            640,
+            487,
+            659
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $w_{j}^{\\mathrm{S},k - 1}$  represents the weight for each coarse-level slot label class. These probabilities are concatenated and projected into a unified semantic vector:",
+        "bbox": [
+            112,
+            661,
+            487,
+            725
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {Z} ^ {\\mathrm {S}, k - 1} = \\mathbf {Z} ^ {\\mathrm {S}, k - 1} \\cdot \\mathbf {p} ^ {\\mathrm {S}, k - 1} \\in \\mathbb {R} ^ {d _ {p}} \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            191,
+            740,
+            487,
+            757
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "This vector is then concatenated with each fine-grained slot label vector to form an enhanced fine-grained label vector:",
+        "bbox": [
+            112,
+            763,
+            487,
+            810
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {v} _ {j} ^ {\\mathrm {S}, k} \\leftarrow \\mathbf {v} _ {j} ^ {\\mathrm {S}, k} \\oplus \\mathbf {z} ^ {\\mathrm {S}, k - 1} \\tag {10}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            213,
+            822,
+            487,
+            841
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Finally, the updated slot label representation at the  $k - th$  layer is:",
+        "bbox": [
+            112,
+            846,
+            487,
+            877
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {V} ^ {\\mathrm {S}, k} = \\left[ \\mathbf {v} _ {1} ^ {\\mathrm {S}, k}, \\mathbf {v} _ {2} ^ {\\mathrm {S}, k}, \\dots , \\mathbf {v} _ {| \\mathrm {L} ^ {\\mathrm {S}}, k |} ^ {\\mathrm {S}, k} \\right] \\in \\mathbb {R} ^ {\\left(d _ {e} + d _ {p}\\right) \\times | \\mathrm {L} ^ {\\mathrm {S}, k} |} \\tag {11}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            129,
+            887,
+            487,
+            923
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Through this hierarchical enhancement mechanism, the model refines the fine-grained slot label representations by integrating information from coarser slot labels, which improves the overall performance of slot filling tasks.",
+        "bbox": [
+            507,
+            84,
+            884,
+            164
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2.4 Intent-Slot Co-attention",
+        "text_level": 1,
+        "bbox": [
+            507,
+            179,
+            757,
+            193
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We propose an intent-slot co-attention mechanism that enables fine-grained bidirectional interaction via label-specific representations.",
+        "bbox": [
+            507,
+            200,
+            880,
+            247
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The module takes as input a set of label-specific representations generated by the label attention layer, including the semantic representation matrix of intent labels  $\\mathbf{Q}_1 = \\mathbf{V}^{\\mathrm{I}}\\in \\mathbb{R}^{d_e\\times |\\mathrm{L}^{\\mathrm{I}}|}$ , each layer of slot labels is represented by semantic embeddings",
+        "bbox": [
+            507,
+            250,
+            880,
+            329
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {Q} _ {2} = \\mathbf {V} ^ {\\mathrm {S}, 1}, \\mathbf {Q} _ {3} = \\mathbf {V} ^ {\\mathrm {S}, 2}, \\ldots , \\mathbf {Q} _ {l + 1} = \\mathbf {V} ^ {\\mathrm {S}, l},\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            517,
+            330,
+            853,
+            347
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "along with a soft slot embedding matrix for each token  $\\mathbf{Q}_{l + 2} = \\mathbf{S}\\in \\mathbb{R}^{d_s\\times n}$ . These inputs collectively form a semantic hierarchical sequence for modeling label-level interactions via the co-attention mechanism.",
+        "bbox": [
+            507,
+            347,
+            880,
+            426
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Since the input embeddings may differ in dimensionality, all representations are first projected into a unified semantic space to enable cross-layer interaction. To achieve this, we introduce separate forward and backward projection matrices for each input  $\\mathbf{Q}_t\\in \\mathbb{R}^{d_t\\times m_t}$ , mapping them into a shared representation space of dimension  $d$ :",
+        "bbox": [
+            507,
+            429,
+            882,
+            541
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\overrightarrow {\\mathbf {Q}} _ {t} = \\overrightarrow {\\mathbf {W}} _ {t} \\mathbf {Q} _ {t}, \\overleftarrow {\\mathbf {Q}} _ {t} = \\overleftarrow {\\mathbf {W}} _ {t} \\mathbf {Q} _ {t} \\tag {12}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            603,
+            554,
+            880,
+            571
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\overleftrightarrow{\\mathbf{Q}}_t$  and  $\\overleftarrow{\\mathbf{Q}}_t$  serve as the basis for semantic propagation: the former supports forward propagation (slot-to-intent), while the latter is used for backward propagation (intent-to-slot). At each step, a bilinear attention module computes a correlation matrix between adjacent layers:",
+        "bbox": [
+            507,
+            575,
+            880,
+            675
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {C} _ {t} = \\mathbf {Q} _ {t - 1} ^ {t} \\mathbf {X} _ {t} \\mathbf {Q} _ {t} \\in \\mathbb {R} ^ {m _ {t - 1} \\times m _ {t}} \\tag {13}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            591,
+            690,
+            880,
+            706
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathbf{X}_t\\in \\mathbb{R}^{d_{t - 1}\\times d_t}$  is a learnable attention weight matrix capturing inter-label similarity. Semantic propagation is then performed in both directions. For forward propagation (slot-to-intent), the update is:",
+        "bbox": [
+            507,
+            709,
+            880,
+            790
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\vec {\\mathbf {H}} _ {t} = \\left\\{ \\begin{array}{l l} \\tanh  \\left(\\overrightarrow {\\mathbf {Q}} _ {t - 1} \\mathbf {C} _ {t} + \\overrightarrow {\\mathbf {Q}} _ {t}\\right), & t = 2 \\\\ \\tanh  \\left(\\overrightarrow {\\mathbf {H}} _ {t - 1} \\mathbf {C} _ {t} + \\overrightarrow {\\mathbf {Q}} _ {t}\\right), & t > 2 \\end{array} \\right. \\tag {14}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            563,
+            802,
+            880,
+            840
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Conversely, backward propagation (intent-to-slot) is defined as:",
+        "bbox": [
+            507,
+            843,
+            882,
+            873
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\overleftarrow {\\mathbf {H}} _ {t} = \\left\\{ \\begin{array}{l l} \\tanh  \\left(\\overleftarrow {\\mathbf {Q}} _ {t + 1} \\mathbf {C} _ {t + 1} ^ {T} + \\overleftarrow {\\mathbf {Q}} _ {t}\\right), & t = l + 1 \\\\ \\tanh  \\left(\\overleftarrow {\\mathbf {H}} _ {t + 1} \\mathbf {C} _ {t + 1} ^ {T} + \\overleftarrow {\\mathbf {Q}} _ {t}\\right), & t <   l + 1 \\end{array} \\right. \\tag {15}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            532,
+            885,
+            880,
+            923
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "25606",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "This enables intent-level semantic cues to flow back to the slot layer, supporting more accurate slot predictions through joint label awareness. The dualpath propagation forms a closed loop of semantic communication, improving contextual understanding across labels.",
+        "bbox": [
+            112,
+            84,
+            487,
+            180
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "After propagation, the model extracts final label representations for decoding. Specifically, the first-layer output from backward propagation is used as the intent representation, while the final output of the forward path is used for slot decoding:",
+        "bbox": [
+            112,
+            181,
+            489,
+            261
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {R} _ {\\mathrm {I}} ^ {f i n a l} = \\overline {{\\mathbf {H}}} _ {1} \\in \\mathbb {R} ^ {d \\times | \\mathrm {L} ^ {\\mathrm {I}} |} \\tag {16}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            220,
+            273,
+            487,
+            292
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {R} _ {\\mathrm {S}} ^ {f i n a l} = \\overline {{\\mathbf {H}}} _ {l + 2} \\in \\mathbb {R} ^ {d \\times n} \\tag {17}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            220,
+            305,
+            487,
+            322
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3 Contrastive Learning",
+        "text_level": 1,
+        "bbox": [
+            112,
+            332,
+            329,
+            348
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "While label-aware attention captures sentence-level semantics, semantic overlap persists among similar labels in multi-intent/slot scenarios. LCAN addresses this through a label-aware contrastive learning module (Figure 3), structuring the embedding space via contrastive supervision. Following supervised contrastive principles, it creates positive pairs and negative pairs, enforcing intra-class compactness and inter-class separation. The framework extends to three semantic levels: intent labels, slot labels, and cross-task intent-slot pairs, establishing multi-granular contrastive supervision that enhances discriminative power while mitigating feature confusion.",
+        "bbox": [
+            112,
+            353,
+            489,
+            576
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/38027248dbff2ee607b58bc7b09cc43951d09e50b26c06e74137455bfa09e20f.jpg",
+        "image_caption": [
+            "Figure 3: Diagram of the Contrastive Learning Module."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            127,
+            596,
+            478,
+            720
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3.1 Intent Label Contrastive Learning",
+        "text_level": 1,
+        "bbox": [
+            112,
+            775,
+            445,
+            790
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Intent detection is a multi-label classification task at the sentence level. Semantic similarity between intent labels is common—especially when multiple intents co-occur within a user query. LCAN first extracts per-label semantic vectors via label-aware attention:",
+        "bbox": [
+            112,
+            794,
+            487,
+            888
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbf {V} ^ {\\mathrm {I}} = \\left[ \\mathbf {v} _ {1} ^ {\\mathrm {I}}, \\mathbf {v} _ {2} ^ {\\mathrm {I}}, \\dots , \\mathbf {v} _ {| \\mathrm {L} ^ {\\mathrm {I}} |} ^ {\\mathrm {I}} \\right] \\in \\mathbb {R} ^ {d _ {e} \\times | \\mathrm {L} ^ {\\mathrm {I}} |} \\tag {18}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            184,
+            902,
+            487,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Positive pairs are sampled from embeddings of the same intent label across different sentences, while negative pairs are drawn from different labels. The model is trained using the InfoNCE loss:",
+        "bbox": [
+            507,
+            84,
+            884,
+            149
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {i} - \\mathrm {c l}} = - \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(v , v ^ {+}\\right) / \\tau\\right)}{\\Sigma_ {v ^ {-} \\in \\mathcal {N}} \\exp \\left(\\operatorname {s i m} \\left(v , v ^ {-}\\right) / \\tau\\right)} \\tag {19}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            539,
+            159,
+            880,
+            189
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $sim()$  denotes cosine similarity,  $\\tau$  is a temperature parameter, and  $\\mathcal{N}$  is the set of negatives. This loss pulls together embeddings of the same intent and pushes apart others, enhancing robustness in multi-label scenarios.",
+        "bbox": [
+            507,
+            190,
+            884,
+            269
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3.2 Slot Label Contrastive Learning",
+        "text_level": 1,
+        "bbox": [
+            507,
+            281,
+            823,
+            297
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Slot filling involves a larger label space and finer semantic granularity. Overlapping meanings among hierarchical slot labels further complicate representation learning. LCAN generates multi-layer slot label embeddings  $\\mathbf{V}^{\\mathrm{S},k}\\in \\mathbb{R}^{d_k\\times |\\mathrm{L}^{\\mathrm{S},k}|}$  via hierarchical attention computed following Equation 11.",
+        "bbox": [
+            507,
+            300,
+            882,
+            411
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "For each layer  $k$ , contrastive learning is applied similarly:",
+        "bbox": [
+            507,
+            414,
+            880,
+            445
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\ell_ {i} = \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(v _ {i} , v _ {i} ^ {+}\\right) / \\tau\\right)}{\\sum_ {v _ {j} ^ {-} \\in \\mathcal {N}} \\exp \\left(\\operatorname {s i m} \\left(v _ {i} , v _ {j} ^ {-}\\right) / \\tau\\right)} \\tag {20}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            566,
+            455,
+            880,
+            491
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s}, k} = - \\frac {1}{| \\mathrm {L} ^ {\\mathrm {S} , k} |} \\sum_ {i = 1} ^ {| \\mathrm {L} ^ {\\mathrm {S}, k} |} \\ell_ {i} \\tag {21}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            616,
+            502,
+            880,
+            539
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "This contrastive loss encourages the model to differentiate between semantically similar slot labels.",
+        "bbox": [
+            507,
+            548,
+            882,
+            594
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s} - \\mathrm {c l}} = \\Sigma_ {k = 1} ^ {K} \\mathcal {L} _ {\\mathrm {s}, k} \\tag {22}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            633,
+            609,
+            880,
+            626
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3.3 Cross-Task Contrastive Learning",
+        "text_level": 1,
+        "bbox": [
+            507,
+            639,
+            830,
+            653
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Intent and slot labels are closely correlated. To align semantic spaces across tasks, LCAN incorporates cross-task contrastive learning, enforcing similarity between semantically aligned intent-slot label pairs. Given an intent embedding  $\\mathbf{v}_i^{\\mathrm{I}}$  and a related slot embedding  $\\mathbf{v}_j^{\\mathrm{S},k}$ , a contrastive loss is defined as:",
+        "bbox": [
+            507,
+            658,
+            882,
+            771
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {c r o s s} = - \\sum_ {i, j \\in \\mathcal {P}} \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(\\mathbf {v} _ {i} ^ {\\mathrm {I}} , \\mathbf {v} _ {j} ^ {\\mathrm {S} , k}\\right) / \\tau\\right)}{\\sum_ {i , j ^ {-} \\in \\mathcal {N}} \\exp \\left(\\operatorname {s i m} \\left(\\mathbf {v} _ {i} ^ {\\mathrm {I}}, \\mathbf {v} _ {j ^ {-}} ^ {\\mathrm {S}, k}\\right) / \\tau\\right)} \\tag {23}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            515,
+            783,
+            880,
+            840
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.4 Intent and Slot Decoders",
+        "text_level": 1,
+        "bbox": [
+            507,
+            851,
+            749,
+            865
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The final stage of the model involves decoding the intent and slot labels. Both the intent and slot decoders are based on a multi-label classification",
+        "bbox": [
+            507,
+            873,
+            880,
+            919
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "25607",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "approach for intent recognition and a sequence labeling approach for slot filling.",
+        "bbox": [
+            112,
+            84,
+            489,
+            116
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "3.4.1 Intent Decoder",
+        "text_level": 1,
+        "bbox": [
+            112,
+            128,
+            294,
+            142
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Intent detection is modeled as a multi-label classification task, where a single utterance may correspond to multiple semantic intents. To handle this, LCAN employs a sigmoid-based multi-label classifier. To train the model, a binary cross-entropy loss is computed independently for each label and summed across all labels:",
+        "bbox": [
+            112,
+            148,
+            487,
+            261
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {i} = - \\sum_ {j = 1} ^ {\\lfloor \\mathrm {L} ^ {\\mathrm {I}} \\rfloor} \\left[ y _ {j} ^ {\\mathrm {I}} \\log \\left(\\hat {y} _ {j} ^ {\\mathrm {I}}\\right) + \\left(1 - y _ {j} ^ {\\mathrm {I}}\\right) \\log \\left(1 - \\hat {y} _ {j} ^ {\\mathrm {I}}\\right) \\right] \\tag {24}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            127,
+            272,
+            487,
+            310
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "where  $y_{j}^{\\mathrm{I}}$  denotes the ground-truth label for the  $i - th$  intent. This objective supports robust learning for multi-intent utterances.",
+        "bbox": [
+            112,
+            313,
+            485,
+            362
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "3.4.2 Slot Decoder",
+        "text_level": 1,
+        "bbox": [
+            112,
+            375,
+            275,
+            388
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Slot filling is framed as a sequence labeling task, where each token in the utterance is assigned a corresponding slot label. To account for contextual dependencies among slot tags, LCAN adopts a hybrid decoding approach using a linear classifier followed by a Conditional Random Field.",
+        "bbox": [
+            112,
+            395,
+            487,
+            491
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To model label transitions and enforce sequential consistency, a CRF layer is applied over the emission scores. Given a sequence of slot labels  $y^{\\mathrm{S}} = y_1^{\\mathrm{S}},y_2^{\\mathrm{S}},\\dots,y_n^{\\mathrm{S}}$ , the sequence-level score is defined as:",
+        "bbox": [
+            112,
+            492,
+            489,
+            571
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {s c o r e} \\left(x, y ^ {\\mathrm {S}}\\right) = \\Sigma_ {i = 1} ^ {n} \\left(\\operatorname {e m i s s i o n} _ {i, y _ {i} ^ {\\mathrm {S}}} + T _ {y _ {i - 1} ^ {\\mathrm {S}}, y _ {i} ^ {\\mathrm {S}}}\\right) \\tag {25}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            129,
+            583,
+            487,
+            607
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "where  $emission_{i,y_i^s}$  denotes the transition score from the previous label to the current label.",
+        "bbox": [
+            112,
+            608,
+            485,
+            639
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The CRF is trained by maximizing the log-likelihood of the correct label sequence, which is equivalent to minimizing the negative log-likelihood:",
+        "bbox": [
+            112,
+            640,
+            489,
+            703
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {s} = - \\operatorname {s c o r e} (\\mathbf {x}, y ^ {\\mathrm {S}}) + \\log \\Sigma_ {\\hat {\\mathbf {y}} \\in \\mathcal {Y}} \\exp (s c o r e (\\mathbf {x}, \\hat {\\mathbf {y}})) \\tag {26}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            122,
+            719,
+            487,
+            736
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The denominator sums over all possible label sequences and is computed efficiently via the forward algorithm. During inference, the Viterbi algorithm is used to find the most probable label sequence.",
+        "bbox": [
+            112,
+            741,
+            487,
+            806
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "3.5 Overall Training Objective",
+        "text_level": 1,
+        "bbox": [
+            112,
+            819,
+            371,
+            834
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The LCAN model is trained using a joint objective function that combines two major components: 1) Task-specific supervised losses, including a multi-label classification loss for intent detection and a sequence labeling loss for slot filling; 2) Label-aware",
+        "bbox": [
+            112,
+            841,
+            489,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "contrastive losses, designed to enhance semantic separability among intent and slot representations.",
+        "bbox": [
+            507,
+            84,
+            880,
+            116
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The overall loss function is defined as:",
+        "bbox": [
+            526,
+            117,
+            815,
+            131
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {i} + \\lambda_ {0} \\mathcal {L} _ {s} + \\lambda_ {1} \\mathcal {L} _ {i - c l} + \\lambda_ {2} \\mathcal {L} _ {s - c l} + \\lambda_ {3} \\mathcal {L} _ {\\text {c r o s s}} \\tag {27}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            519,
+            142,
+            880,
+            168
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "This flexible multi-objective formulation enables the model to balance structural supervision and semantic discriminability, improving performance and generalization across both tasks.",
+        "bbox": [
+            507,
+            171,
+            880,
+            234
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4 Experiments Setup",
+        "text_level": 1,
+        "bbox": [
+            507,
+            247,
+            709,
+            263
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1 Datasets",
+        "text_level": 1,
+        "bbox": [
+            507,
+            273,
+            621,
+            286
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The experiments are conducted on two standard benchmark datasets: MixATIS (Qin et al., 2020) and MixSNIPS (Qin et al., 2020). MixATIS is a dataset primarily focusing on the airline travel domain, containing multiple intents like flight booking, flight inquiry, and time-based queries. The MixSNIPS dataset, on the other hand, covers multiple domains such as weather, music, restaurants, and movies, with a more conversational and informal style of language.",
+        "bbox": [
+            507,
+            294,
+            884,
+            455
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2 Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            507,
+            466,
+            741,
+            481
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We use RoBERTa-base as the shared encoder to extract general semantic features from input sequences. Two task-specific encoders based on BiLSTM are employed: one for intent recognition and one for slot filling. The label-aware attention mechanism is introduced to refine the interaction between intent and slot labels. Additionally, the model incorporates a contrastive learning strategy to enhance semantic differentiation between similar or overlapping labels. The AdamW optimizer (Loshchilov and Hutter, 2017) is used for model training with an initial learning rate of 0.001. The batch size is set to 32, and the maximum number of training epochs is 50. Early stopping is applied to prevent overfitting. We perform hyperparameter tuning using random search on the validation set.",
+        "bbox": [
+            507,
+            487,
+            882,
+            759
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3 Baselines",
+        "text_level": 1,
+        "bbox": [
+            507,
+            771,
+            628,
+            785
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To comprehensively evaluate the effectiveness of the proposed LCAN model on multi-intent detection and slot filling tasks, we conducted systematic comparative experiments on MixATIS and MixSNIPS. The evaluation includes several representative baseline models, ranging from traditional joint modeling approaches to recent frameworks enhanced by graph structures and attenuated by the proposed approach.",
+        "bbox": [
+            507,
+            791,
+            884,
+            921
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "25608",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "tion mechanisms:(1)GL-GIN (Qin et al., 2021b): A non-autoregressive model with global-local graph interaction for efficient intent-slot coordination.(2)SDJN (Chen et al., 2022b): Integrates self-distillation and multi-instance learning to enhance word-level intent-slot alignment.(3)Coguiding Net (Xing and Tsang, 2022): Employs heterogeneous graph attention and contrastive learning for bidirectional task interaction.(4)TFMN (Chen et al., 2022c): Multi-task framework with auxiliary intent count prediction, eliminating threshold dependencies.(5)BiSLU (Tu et al., 2023): Combines supervised contrastive learning with self-distillation for explicit intent-slot collaboration.(6)MISCA (Pham et al., 2023): Joint attention architecture enabling implicit intent-slot information exchange without graph construction.(7)UniMIS (Yin et al., 2024): Models the multi-intent SLU as a multi-view intent-slot interaction.",
+        "bbox": [
+            112,
+            84,
+            492,
+            391
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5 Results and Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            403,
+            329,
+            420
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.1 Main Results",
+        "text_level": 1,
+        "bbox": [
+            112,
+            431,
+            265,
+            445
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Experimental results are presented in Table 1. On both the MixATIS and MixSNIPS datasets, the proposed LCAN model demonstrates a clear overall performance advantage in joint multi-intent modeling tasks.",
+        "bbox": [
+            112,
+            453,
+            489,
+            533
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "On the MixATIS dataset, LCAN achieves a joint accuracy of  $59.5\\%$ , outperforming MISCA and BiSLU by  $1.4\\%$  and  $8.0\\%$ , respectively. This confirms its effectiveness in handling complex hierarchical slot labels and diverse intents in the airline domain. Notably, LCAN achieves an F1-score of  $92.4\\%$  on slot filling,  $3.0\\%$  higher than BiSLU, indicating that its label-aware attention mechanism effectively captures fine-grained slot dependencies. Although its intent accuracy is slightly lower than BiSLU, the incorporation of cross-task contrastive learning strengthens semantic alignment between intents and slots, leading to an improvement in overall joint accuracy.",
+        "bbox": [
+            112,
+            533,
+            489,
+            759
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "On the open-domain MixSNIPS dataset, LCAN achieves strong intent detection performance, benefiting from contrastive learning to alleviate semantic overlap among multiple intents and to enhance global consistency across tasks. While its slot F1-score is slightly lower than BiSLU, LCAN still achieves superior joint accuracy, highlighting its advantage in optimizing intent-slot interactions. Compared to earlier models, LCAN improves joint accuracy by  $7.8\\%$  over Co-guiding Net, further val",
+        "bbox": [
+            112,
+            760,
+            489,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "idating its robustness in multi-domain, multi-intent scenarios.",
+        "bbox": [
+            507,
+            84,
+            882,
+            115
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/2c5c1c5c3aec55dc96912cdc91f6f4fc879e97e6fcdf477e1629ff8e0757e2b9.jpg",
+        "table_caption": [
+            "Table 1: Obtained results"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Model</td><td colspan=\"3\">MixATIS</td><td colspan=\"3\">MixSNIPS</td></tr><tr><td>Intent (Acc.)</td><td>Slot (F1)</td><td>Overall (Acc.)</td><td>Intent (Acc.)</td><td>Slot (F1)</td><td>Overall (Acc.)</td></tr><tr><td>GL-GIN</td><td>76.3</td><td>88.3</td><td>43.5</td><td>95.6</td><td>94.9</td><td>75.4</td></tr><tr><td>SDJN</td><td>77.1</td><td>88.2</td><td>44.6</td><td>96.5</td><td>94.4</td><td>75.7</td></tr><tr><td>Co-guiding</td><td>79.1</td><td>89.8</td><td>51.3</td><td>97.7</td><td>95.1</td><td>77.5</td></tr><tr><td>TFMN</td><td>79.8</td><td>88.0</td><td>50.2</td><td>97.7</td><td>96.4</td><td>84.7</td></tr><tr><td>BiSLU</td><td>81.5</td><td>89.4</td><td>51.5</td><td>97.8</td><td>97.2</td><td>85.4</td></tr><tr><td>MISCA</td><td>80.6</td><td>90.0</td><td>58.1</td><td>97.6</td><td>96.1</td><td>83.1</td></tr><tr><td>Uni-MIS</td><td>78.5</td><td>88.3</td><td>52.5</td><td>97.2</td><td>96.4</td><td>83.4</td></tr><tr><td>LCAN (Ours)</td><td>81.3</td><td>92.4</td><td>59.5</td><td>98.7</td><td>96.9</td><td>85.3</td></tr></table>",
+        "bbox": [
+            519,
+            149,
+            873,
+            331
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.2 Ablation Results",
+        "text_level": 1,
+        "bbox": [
+            507,
+            357,
+            687,
+            370
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To evaluate the contribution of different contrastive learning modules, we conduct an ablation study by removing various contrastive learning modules. The results are shown in Table 2, and the ablated variants include: w/o Intent CL, w/o Slot CL, w/o Cross CL, and w/o CL.",
+        "bbox": [
+            507,
+            376,
+            884,
+            472
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/f4a4f51f8063be268236a5a513c6bcd524decccb2d2f1ed686a91106d7ab93d1.jpg",
+        "table_caption": [
+            "Table 2: Ablation Study on Contrastive Learning Module"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Variant</td><td colspan=\"3\">MixATIS</td><td colspan=\"3\">MixSNIPS</td></tr><tr><td>Intent (Acc.)</td><td>Slot (F1)</td><td>Overall (Acc.)</td><td>Intent (Acc.)</td><td>Slot (F1)</td><td>Overall (Acc.)</td></tr><tr><td>w/o CL</td><td>80.6</td><td>90.0</td><td>58.3</td><td>97.6</td><td>96.1</td><td>83.1</td></tr><tr><td>w/o Intent CL</td><td>79.9</td><td>91.8</td><td>58.9</td><td>97.7</td><td>96.5</td><td>84.8</td></tr><tr><td>w/o Slot CL</td><td>80.5</td><td>89.6</td><td>58.7</td><td>98.5</td><td>94.9</td><td>84.3</td></tr><tr><td>w/o Cross CL</td><td>81.6</td><td>91.7</td><td>58.8</td><td>98.8</td><td>95.9</td><td>84.8</td></tr><tr><td>LCAN</td><td>81.3</td><td>92.4</td><td>59.5</td><td>98.7</td><td>96.9</td><td>85.3</td></tr></table>",
+        "bbox": [
+            521,
+            521,
+            870,
+            682
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Removing all contrastive learning modules leads to a decrease of  $1.2\\%$  and  $2.2\\%$  in joint accuracy on MixATIS and MixSNIPS, respectively. This confirms the effectiveness of contrastive learning in enhancing semantic consistency between intents and slots, thereby improving overall performance. On MixATIS, intent accuracy drops  $1.4\\%$ , indicating its primary focus on intent representation. For MixSNIPS, it proves vital in open-domain intent-dominant scenarios. Causes significant slot F1 declines, highlighting its importance for fine-grained slot recognition. The  $0.8\\%$  intent accuracy drop on MixATIS suggests improved slot distinction indirectly benefits intent inference. Removal increases",
+        "bbox": [
+            507,
+            696,
+            884,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "25609",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "MixATIS intent accuracy due to reduced overfitting but lowers slot F1 and joint accuracy, demonstrating its balance between task interaction and semantic alignment. On MixSNIPS, joint accuracy drops  $0.5\\%$  despite minor intent gains. In summary, the intent, slot, and cross contrastive learning modules contribute to the LCAN model from global intent optimization, local slot representation, and semantic coordination, respectively, leading to improved multi-intent joint modeling.",
+        "bbox": [
+            112,
+            84,
+            489,
+            243
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "To evaluate the contributions of label-aware attention structure, we design three simplified model variants by removing specific components of the label-aware attention structure. The ablated variants include: w/o Co-attention, w/o Slot Label Attention, and w/o Intent Label Attention.",
+        "bbox": [
+            112,
+            244,
+            489,
+            341
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/475e295f662cd0cbbe291052c35417e6ed32ef8aa21d8996a70c1be4d97456b3.jpg",
+        "table_caption": [
+            "Table 3: Ablation Study on Label Attention Mechanism"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Variant</td><td colspan=\"3\">MixATIS</td><td colspan=\"3\">MixSNIPS</td></tr><tr><td>Intent (Acc.)</td><td>Slot (F1)</td><td>Overall (Acc.)</td><td>Intent (Acc.)</td><td>Slot (F1)</td><td>Overall (Acc.)</td></tr><tr><td>w/o Co-Attn</td><td>79.9</td><td>89.2</td><td>50.5</td><td>96.9</td><td>95.9</td><td>80.1</td></tr><tr><td>w/o SL Attn</td><td>79.6</td><td>91.2</td><td>56.1</td><td>98.7</td><td>96.1</td><td>81.4</td></tr><tr><td>w/o IL Attn</td><td>78.7</td><td>91.6</td><td>55.3</td><td>97.2</td><td>96.2</td><td>82.0</td></tr><tr><td>LCAN</td><td>81.3</td><td>92.4</td><td>59.5</td><td>98.7</td><td>96.9</td><td>85.3</td></tr></table>",
+        "bbox": [
+            131,
+            379,
+            470,
+            519
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "As shown in Table 3, removing the co-attention mechanism leads to decrease of  $9.0\\%$  and  $5.2\\%$  in joint accuracy on MixATIS and MixSNIPS, respectively. This confirms the effectiveness of co-attention mechanism in modeling semantic coordination and promoting collaborative decoding between the intents and entities. The absence of SL Attn prevents the model from establishing fine-grained semantic alignments between slot labels and the input context. This leads to a clear accuracy drop, especially on MixSNIPS (nearly  $4\\%$ ). Furthermore, the removal of the IL Attn leads to a substantial drop in intent classification accuracy, this suggests that relying solely on task-specific encoder outputs is insufficient to capture the semantic diversity of multiple intents.",
+        "bbox": [
+            112,
+            533,
+            489,
+            791
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.3 Case Study",
+        "text_level": 1,
+        "bbox": [
+            112,
+            803,
+            248,
+            819
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "To further illustrate the effectiveness of LCAN, we conduct a case study using a representative multi- intent example from the MixATIS dataset: \"Check flights from Beijing to Shanghai and book a ticket\". Traditional models exhibit blurred intent boundaries at conjunctions, while RoBERTa-base fails",
+        "bbox": [
+            112,
+            824,
+            489,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "to disambiguate slots near \"and\". As shown in Table 4, BiSLU improves intent separation but retains coarse slot granularity. This demonstrates LCAN's superiority in disentangling interconnected intents and maintaining slot-label consistency.",
+        "bbox": [
+            507,
+            84,
+            884,
+            165
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/e098d4daf7ef00c9333c2d8d0153fdbbb2ee6b0a0536ab0bb369e5180b229e6f.jpg",
+        "table_caption": [
+            "Table 4: Recognition Results for a Multi-Intent Example"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>Recognized In- \ntents</td><td>Slot Filling Results</td></tr><tr><td>RoBERTa-base</td><td>CheckFlight, \nBookTicket</td><td>from: Beijing; to: \nShanghai and book; \nticket: [Unrecognized]</td></tr><tr><td>BiSLU</td><td>CheckFlight, \nBookTicket</td><td>from: Beijing; to: \nShanghai; ticket: ticket</td></tr><tr><td>LCAN</td><td>CheckFlight, \nBookTicket</td><td>from: Beijing; to: \nShanghai; ticket: book \na ticket</td></tr></table>",
+        "bbox": [
+            514,
+            200,
+            885,
+            355
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 4 shows the recognition results of three models under this input. While all models successfully detect the two intents, there are clear differences in slot filling accuracy and boundary clarity. RoBERTa-base fails to distinguish between the intents around the conjunction, leading to incomplete slot extraction. BiSLU improves intent separation through bidirectional learning and contrastive strategies but still shows coarse-grained slot interpretation. LCAN, with its label-aware mechanism, accurately aligns slot information with each intent, demonstrating superior semantic parsing in multi-intent scenarios.",
+        "bbox": [
+            507,
+            370,
+            884,
+            577
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            590,
+            640,
+            606
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We propose LCAN, a novel joint model addressing multi-intent recognition and slot filling challenges in task-oriented dialogue systems. LCAN integrates label-aware attention for precise input-label alignment and contrastive learning to mitigate semantic ambiguity among overlapping labels. Evaluations on MixATIS and MixSNIPS datasets demonstrate LCAN's superiority over state-of-the-art models, particularly in complex multi-intent scenarios requiring nuanced intent-slot interactions. Ablation studies confirm the critical roles of contrastive learning and label-aware attention. While LCAN achieves robust performance, future work could explore hierarchical attention mechanisms and optimized contrastive strategies for improved generalization to unseen, highly diverse dialogues. The framework establishes a scalable foundation for advancing joint modeling in multi-intent dialogue systems.",
+        "bbox": [
+            507,
+            615,
+            884,
+            921
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "25610",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            220,
+            98
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "It should also be emphasized that LCAN integration of label-aware attention mechanisms and contrastive learning modules may substantially increase both parameter count and computational complexity. More critically, the efficacy of contrastive learning is contingent upon the design of positive/negative sample pairs—where semantic overlaps between intents and slots in multi-intent scenarios can induce sample selection bias. Furthermore, the label-level focus of contrastive learning might neglect finer-grained semantic distinctions at the word or phrase level. Despite LCAN consistent outperformance of existing models, we posit that incorporating fine-grained semantic similarity computation could yield additional performance gains.",
+        "bbox": [
+            115,
+            111,
+            489,
+            369
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgments",
+        "text_level": 1,
+        "bbox": [
+            114,
+            384,
+            278,
+            401
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This research was supported by the National Natural Science Foundation of China (62074055, 62174038, 62374047), and State Key Lab of Processors, Institute of Computing Technology, CAS (CLQ 202407).",
+        "bbox": [
+            112,
+            412,
+            489,
+            492
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            521,
+            213,
+            537
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Fengyu Cai, Wanhao Zhou, Fei Mi, and Boi Faltings. 2022. Slim: Explicit slot-intent mapping with bert for joint multi-intent detection and slot filling. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7607-7611. IEEE.",
+            "Dongsheng Chen, Zhiqi Huang, Xian Wu, Shen Ge, and Yuexian Zou. 2022a. Towards joint intent detection and slot filling via higher-order attention. In *IJCAI*, pages 4072-4078.",
+            "Guanhua Chen, Yutong Yao, Derek F Wong, and Lidia S Chao. 2024. A two-stage prediction-aware contrastive learning framework for multi-intent nlu. arXiv preprint arXiv:2405.02925.",
+            "Lisong Chen, Peilin Zhou, and Yuexian Zou. 2022b. Joint multiple intent detection and slot filling via self-distillation. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7612-7616. IEEE.",
+            "Lisung Chen, Nuo Chen, Yuexian Zou, Yong Wang, and Xinzhong Sun. 2022c. A transformer-based threshold-free framework for multi-intent nlu. In Proceedings of the 29th International Conference on Computational Linguistics, pages 7187-7192."
+        ],
+        "bbox": [
+            115,
+            546,
+            489,
+            920
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Qian Chen, Zhu Zhuo, and Wen Wang. 2019. Bert for joint intent classification and slot filling. arXiv preprint arXiv:1902.10909.",
+            "Lizhi Cheng, Wenmian Yang, and Weijia Jia. 2023a. A scope sensitive and result attentive model for multi-intent spoken language understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 12691-12699.",
+            "Xuxin Cheng, Wanshi Xu, Zhihong Zhu, Hongxiang Li, and Yuexian Zou. 2023b. Towards spoken language understanding via multi-level multi-grained contrastive learning. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, pages 326-336.",
+            "Rashmi Gangadharaiah. 2019. Joint multiple intent detection and slot labeling for goal-oriented dialog.",
+            "Belize Gunel, Jingfei Du, Alexis Conneau, and Ves Stoyanov. 2020. Supervised contrastive learning for pretrained language model fine-tuning. arXiv preprint arXiv:2011.01403.",
+            "Yutai Hou, Yongkui Lai, Cheng Chen, Wanxiang Che, and Ting Liu. 2021. Learning to bridge metric spaces: Few-shot joint learning of intent detection and slot filling. arXiv preprint arXiv:2106.07343.",
+            "Byeongchang Kim, Seonghan Ryu, and Gary Geunbae Lee. 2017a. Two-stage multi-intent detection for spoken language understanding. *Multimedia Tools and Applications*, 76:11377-11390.",
+            "Young-Bum Kim, Sungjin Lee, and Karl Stratos. 2017b. Onenet: Joint domain, intent, slot prediction for spoken language understanding. In 2017 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), pages 547-553. IEEE.",
+            "Bing Liu and Ian Lane. 2016. Attention-based recurrent neural network models for joint intent detection and slot filling. arXiv preprint arXiv:1609.01454.",
+            "Ilya Loshchilov and Frank Hutter. 2017. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101.",
+            "Zhiyuan Ma, Jiwei Qin, Meiqi Pan, Song Tang, Jinpeng Mi, and Dan Liu. 2024. Promoting unified generative framework with descriptive prompts for joint multi-intent detection and slot filling. *Electronics*, 13(6):1087.",
+            "Thinh Pham, Chi Tran, and Dat Quoc Nguyen. 2023. Misca: A joint model for multiple intent detection and slot filling with intent-slot co-attention. arXiv preprint arXiv:2312.05741.",
+            "Libo Qin, Tailu Liu, Wanxiang Che, Bingbing Kang, Sendong Zhao, and Ting Liu. 2021a. A co-interactive transformer for joint slot filling and intent detection. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 8193-8197. IEEE."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            920
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "25611",
+        "bbox": [
+            475,
+            928,
+            522,
+            941
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Libo Qin, Fuxuan Wei, Tianbao Xie, Xiao Xu, Wanxiang Che, and Ting Liu. 2021b. Gl-gin: Fast and accurate non-autoregressive model for joint multiple intent detection and slot filling. arXiv preprint arXiv:2106.01925.",
+            "Libo Qin, Xiao Xu, Wanxiang Che, and Ting Liu. 2020. Agif: An adaptive graph-interactive framework for joint multiple intent detection and slot filling. arXiv preprint arXiv:2004.10087.",
+            "Mengxiao Song, Bowen Yu, Li Quangang, Wang Yubin, Tingwen Liu, and Hongbo Xu. 2022. Enhancing joint multiple intent detection and slot filling with global intent-slot co-occurrence. In Proceedings of the 2022 conference on empirical methods in natural language processing, pages 7967-7977.",
+            "Nguyen Anh Tu, Hoang Thi Thu Uyen, Tu Minh Phuong, and Ngo Xuan Bach. 2023. Joint multiple intent detection and slot filling with supervised contrastive learning and self-distillation. In ECAI 2023, pages 2370-2377. IOS Press.",
+            "Ivan Vulic, Inigo Casanueva, Georgios Spithourakis, Avishek Mondal, Tsung-Hsien Wen, and Paweł Budzianowski. 2022. Multi-label intent detection via contrastive task specialization of sentence encoders. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 7544-7559.",
+            "Jixuan Wang, Kai Wei, Martin Radfar, Weiwei Zhang, and Clement Chung. 2021. Encoding syntactic knowledge in transformer encoder for intent detection and slot filling. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 13943-13951.",
+            "Di Wu, Liang Ding, Fan Lu, and Jian Xie. 2020. Slotrefine: A fast non-autoregressive model for joint intent detection and slot filling. arXiv preprint arXiv:2010.02693.",
+            "Bowen Xing and Ivor W Tsang. 2022. Co-guiding net: Achieving mutual guidances between multiple intent detection and slot filling via heterogeneous semantics-label graphs. arXiv preprint arXiv:2210.10375.",
+            "Puyang Xu and Ruhi Sarikaya. 2013. Convolutional neural network based triangular crf for joint intent detection and slot filling. In 2013 IEEE workshop on automatic speech recognition and understanding, pages 78-83. IEEE.",
+            "Asaf Yehudai, Matan Vetzler, Yosi Mass, Koren Lazar, Doron Cohen, and Boaz Carmeli. 2023. Qaid: Question answering inspired few-shot intent detection. arXiv preprint arXiv:2303.01593.",
+            "Shangjian Yin, Peijie Huang, and Yuhong Xu. 2024. Uni-mis: United multiple intent spoken language understanding via multi-view intent-slot interaction. In AAAI, pages 19395-19403."
+        ],
+        "bbox": [
+            115,
+            85,
+            485,
+            919
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Xiaodong Zhang and Houfeng Wang. 2016. A joint model of intent determination and slot filling for spoken language understanding. In *IJCAI*, volume 16, pages 2993–2999.",
+            "Xianwei Zhuang, Xuxin Cheng, and Yuexian Zou. 2024. Towards explainable joint models via information theory for multiple intent detection and slot filling. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 19786-19794."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            214
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "25612",
+        "bbox": [
+            475,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 9
+    }
+]