ACL 2026 Papers — Page 12
Annual Meeting of the Association for Computational Linguistics · 2296 papers
LazyEviction: Lagged KV Eviction with Attention Pattern Observation for Efficient Long Reasoning
Haoyue Zhang (Hong Kong University of Science and Technology), Song Guo (Hong Kong University of Science and Technology)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose a LazyEviction framework based on an observation window for lagging KV cache eviction, aiming to reduce GPU memory usage during long inference tasks.
LBLLM: Lightweight Binarization of Large Language Models via Three-Stage Distillation
Siqing Song (Chinese Academy of Sciences), Xu-Yao Zhang (Huawei)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose the LBLLM framework, which adopts a three-stage quantization (PTQ initialization → weight quantization → activation refinement) to achieve lightweight binarization of LLMs (W(1+1)A4), completing training with only 0.016B tokens on a single GPU.
LCMA-SRT: Language-Conditional Mixture-of-Experts Adapters for Joint Multilingual Speech Recognition and Translation
Nanjie Li (Xinjiang University), Wei Shi (Timekettle AI Lab)
RecognitionKnowledge DistillationTransformerMixture of ExpertsContrastive LearningTextAudio
🎯 What it does: Propose LCMA-SRT, which integrates source language conditioned and target language conditioned Mixture-of-Experts (MoE) adapters into a hierarchical neural transducer, achieving a single model for multilingual bidirectional speech recognition and translation;
LCR-RAG: Enhancing Logical Consistency in Retrieval-Augmented Generation via Neuro-symbolic Reinforcement Learning
Wenxiang Zheng (Harbin Institute of Technology), Ming Liu (Du Xiaoman Technology)
Recommendation SystemAutonomous DrivingOptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningData-Centric LearningDrug DiscoveryAI Code AssistantNeural Architecture SearchProtein Structure PredictionTransformerLarge Language ModelReinforcement LearningPrompt EngineeringMixture of ExpertsTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the LCR-RAG framework, combining a neuro-symbolic verifier with PPO reinforcement learning to explicitly optimize the logical consistency of retrieval-augmented generation (RAG) models.
LEAF: Knowledge Distillation of Text Embedding Models with Teacher-Aligned Representations
Robin Vujanic (MongoDB Research), Thomas Rückstieß (MongoDB Research)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Built the LEAF framework, which uses simple L2 error for knowledge distillation to generate student models that produce text embeddings compatible with the teacher, supporting an asymmetric architecture where documents are processed with large models and queries with small models.
League of LLMs: A Benchmark-Free Paradigm for Mutual Evaluation of Large Language Models
Qianhong Guo (National University of Defense Technology), Baosheng Wang (National University of Defense Technology)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the League of LLMs (LOL) evaluation paradigm, where multiple models generate questions, answer them, and evaluate each other, forming a benchmark-free, dynamic, and transparent evaluation loop.
Learn Like Humans: Use Meta-cognitive Reflection for Efficient Self-Improvement
Xinmeng Hou, Yang Liu (Nanyang Technological University)
Computational EfficiencyMeta LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Propose the MARS framework, which achieves efficient self-improvement within a single loop through metacognitive reflection
Learn to Relax with Large Language Models: Solving Constraint Optimization Problems via Bidirectional Coevolution
Beidan Liu (National University of Defense Technology), Quanjun Yin (National University of Defense Technology)
OptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringMixture of ExpertsTextTabularBenchmark
🎯 What it does: Propose AutoCO, an end-to-end constrained optimization method based on large language models, which automatically generates and optimizes constraint relaxation strategies, algorithmic concepts, and executable code, thus addressing the issues of fragmented feasible regions and search stagnation caused by hard constraints in constrained optimization problems (COP).
LearnerCoMPASS: Intelligent Tutoring System with Dynamic Cognitive Diagnosis and Multi-Model Path Planning
Ziji Sheng (Huazhong University of Science and Technology), Daizong Liu (Wuhan University)
Knowledge DistillationGraph Neural NetworkTransformerLarge Language ModelAgentic AIPrompt EngineeringMixture of ExpertsTextGraphTabularBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes LEARNERCOMPASS, an end-to-end intelligent tutoring system that integrates dynamic cognitive diagnosis, graph knowledge reasoning, and multi-model path planning.
Learning Diverse Responses with Prefix-Conditioned Supervised Fine-Tuning
Zhiyuan Fan (Hong Kong University of Science and Technology), Yi R. Fung (Hong Kong University of Science and Technology)
GenerationData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: Address the problem of large language models producing homogeneous and repetitive responses (diversity collapse) under the same prompt, and propose the Prefix-conditioned SFT (P-SFT) method.
Learning Faster with Better Tokens: Parameter-Efficient Vocabulary Adaptation for Specialized Text Summarization
Gunjan Balde (Indian Institute of Technology Kharagpur), Niloy Ganguly (Indian Institute of Technology Kharagpur)
Domain AdaptationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextReview/Survey PaperBenchmarkRetrieval-Augmented Generation
🎯 What it does: A parameter-efficient vocabulary adaptation framework named VOCABADAPT is proposed for the task of text summarization in specialized domains. First, tokens in the pre-trained model that are under-trained or unreachable are replaced with domain-specific tokens, and the vocabulary is expanded as needed. Subsequently, the model is fine-tuned for the domain using LoRA.
Learning from Cognition: Enhancing RL Efficiency for LLM Reasoning via Hierarchical Metacognitive Decomposition and Refinement
Zexu Sun (Renmin University of China), Xu Chen (Shenzhen Technology University)
Computational EfficiencyRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes Cog‑Rethinker, a hierarchical meta-cognitive reinforcement learning framework that improves the sample utilization and accuracy of LLMs in tasks such as mathematical reasoning by employing a two-stage rollout process involving decomposition and reflection.
Learning from Contrasts: Synthesizing Reasoning Paths from Diverse Search Trajectories
Peiyang Liu (Peking University), Wei Ye (Peking University)
Data SynthesisExplainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningMixture of ExpertsContrastive LearningTextChain-of-Thought
🎯 What it does: This paper proposes the Contrastive Reasoning Path Synthesis (CRPS) framework, which uses MCTS to generate diverse search trajectories, and then generates high-quality reasoning chains through contrastive analysis, achieving a transition from filtering to synthesis;
Learning from Emptiness: De-biasing Listwise Rerankers with Content-Agnostic Probability Calibration
Hang Lv (University of Science and Technology of China), Enhong Chen (University of Science and Technology of China)
Recommendation SystemComputational EfficiencyData-Centric LearningTransformerPrompt EngineeringGenerative Adversarial NetworkContrastive LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose a training-agnostic and content-agnostic probabilistic calibration framework called CapCal, aimed at eliminating position bias in generative listwise re-ranking models.
Learning from Evolving Training Dynamics: An Entropy-Maximizing Data Curation Strategy for LLM Supervised Post-Training
Mengxiang Zhang (University of Hong Kong), Lingyuan Liu (Independent Researcher)
Knowledge DistillationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes the EVO-Curate framework, dynamically aligning sample complexity with model maturity during training under LLM supervision.
Learning from Mistakes: Negative Reasoning Samples Enhance Out-of-Domain Generalization
Tian Xueyun (Institute of Computing Technology, Chinese Academy of Sciences), Huawei Shen (Institute of Computing Technology, Chinese Academy of Sciences)
Domain AdaptationExplainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Studying negative reasoning trajectories in SFT of large models can improve OOD generalization, and a gain-based adaptive weighting scheme called GLOW is proposed.
Learning from Near-Misses: Error-Aware Contrastive Few-Shot Learning for NL2Formula
Zhihao Shuai (Hong Kong University of Science and Technology), Weikai Yang (Hong Kong University of Science and Technology)
GenerationData SynthesisAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextTabularBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes an error-aware contrastive few-shot learning framework, ECFL, designed to learn from approximate errors, thereby improving the accuracy of generating natural language to Excel formula.
Learning from the Irrecoverable: Error-Localized Policy Optimization for Tool-Integrated LLM Reasoning
Qiao Liang (Tongji University), Sheng Guo (MYbank, Ant Group)
OptimizationTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringTextChain-of-Thought
🎯 What it does: By constructing a binary search backtracking tree, the tool identifies the first unrecoverable error step during the reasoning process of the integrated inference, and on this basis, achieves fine-grained advantage allocation and adaptive trimming, thereby performing precise reinforcement learning on the LLM agent;
Learning How and What to Memorize: Cognition-Inspired Two-Stage Optimization for Evolving Memory
Derong Xu (University of Science and Technology of China), Tong Xu (University of Science and Technology of China)
Recommendation SystemOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringContrastive LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes a cognitive-inspired two-stage optimization framework called MemCoE, which learns how to organize and update the external memory of large language models, thereby achieving long-term personalized dialogue.
Learning Invariant Modality Representation for Robust Multimodal Learning from a Causal Inference Perspective
Sijie Mai (South China Normal University), Shiqin Han (South China Normal University)
Domain AdaptationAnomaly DetectionRepresentation LearningTransformerAuto EncoderContrastive LearningTextMultimodalityAudio
🎯 What it does: Proposed the CmIR framework, which enhances the robustness of multimodal sentiment computing by learning causal-invariant representations and decomposing each modality into causal-invariant and environment-specific spurious representations.
Learning More from Less: Exploiting Counterfactuals for Data-Efficient Chart Understanding
Jianzhu Bao (Nanyang Technological University), Wenya Wang (Nanyang Technological University)
OptimizationComputational EfficiencyData-Centric LearningTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Generate adversarial charts by making minor, targeted modifications to chart plotting code, and improve the performance of vision-language models on chart understanding tasks through bimodal preference optimization (text + image).
Learning to Conceal Risk: Controllable Multi-turn Red Teaming for LLMs in the Financial Domain
Gang Cheng (Bloomberg), Jun Zhuang (Boise State University)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextBenchmarkFinance RelatedRetrieval-Augmented Generation
🎯 What it does: Proposes a controllable multi-round red teaming framework, CoRT, specifically designed to test compliance risks in financial domain large language models (LLMs).
Learning to Edit Knowledge via Instruction-based Chain-of-Thought Prompting
Jinhu Fu (Beijing University of Posts and Telecommunications), Sen Su (Beijing University of Posts and Telecommunications)
Explainability and InterpretabilityKnowledge DistillationRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose CoT2Edit, a method for training LLMs to perform knowledge editing through Chain-of-Thought (CoT) reasoning, combining Retrieval-Augmented Generation (RAG), Supervised Fine-Tuning (SFT), and Group Relative Policy Optimization (GRPO);
Learning to Select: Query-Aware Adaptive Dimension Selection for Dense Retrieval
Zhanyu Wu, Zhijie Nie (Beihang University)
RetrievalKnowledge DistillationRepresentation LearningTransformerSupervised Fine-TuningPrompt EngineeringContrastive LearningText
🎯 What it does: This paper proposes a query-adaptive dimension selection method based on supervised label learning, retaining only the most important dimensions for the query during retrieval to improve the performance of dense retrieval.
Learning to Think on Hypergraph: HyperCoT for Structure-Guided N-ary Knowledge Graph Completion
Mengxue Yang (University of Chinese Academy of Sciences), Ying Li (University of Chinese Academy of Sciences)
Recommendation SystemExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningMeta LearningDrug DiscoveryAI Code AssistantReinforcement Learning from Human FeedbackNeural Architecture SearchProtein Structure PredictionGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextGraphRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes HyperCoT, a n-ary knowledge hypergraph completion framework that integrates multi-hop structural reasoning with LLM generation.
Learning Uncertainty from Sequential Internal Dispersion in Large Language Models
Ponhvoan Srey (Nanyang Technological University), Anh Tuan Luu (Nanyang Technological University)
Anomaly DetectionExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelAuto EncoderContrastive LearningTextSequential
🎯 What it does: This paper proposes a framework called Sequential Internal Variance Representation (SIVR) for detecting hallucinations and estimating uncertainty when large language models generate text.
Learning What Matters: Dynamic Dimension Selection and Aggregation for Interpretable Vision-Language Reward Modeling
Qiyuan Chen (Zhejiang University), Jian Wu (Zhejiang University)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerReinforcement LearningAgentic AIVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose the VL-MDR framework, achieving interpretable and efficient vision-language reward models through visual perception dimension selection and adaptive aggregation, and construct a preference dataset with approximately 321k samples and 21-dimensional fine-grained labels.
Learning What to Ignore: Mitigating Negative Transfer in Medical Knowledge Fusion via Clinical Task-Adaptive Selection
Xinyan Deng (South China University of Technology), Xiaorou Zheng (South China University of Technology)
Federated LearningExplainability and InterpretabilityKnowledge DistillationRepresentation LearningDrug DiscoveryRecurrent Neural NetworkGraph Neural NetworkTransformerLarge Language ModelContrastive LearningTextTabularTime SeriesBiomedical DataElectronic Health RecordsRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the TrustKE framework, which integrates LLM inference with the UMLS structure through a dual-layer knowledge graph to achieve interpretable clinical sequence prediction;
Learning While Staying Curious: Entropy-Preserving Supervised Fine-Tuning via Adaptive Self-Distillation for Large Reasoning Models
Hao Wang (City University of Hong Kong), Dapeng Wu (City University of Hong Kong)
Computational EfficiencyKnowledge DistillationRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextSequential
🎯 What it does: Propose an adaptive self-distillation method called CurioSFT, which enhances the exploration capability of large-scale reasoning models during the supervised fine-tuning phase by maintaining entropy, thereby improving performance in the subsequent reinforcement learning phase.
LEASH: Adaptive Length Penalty and Reward Shaping for Efficient Large Reasoning Model
Yanhao Li (Peking University), Guibo Luo (Peking University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: Propose LEASH, which utilizes the Lagrangian primal-dual method from reinforcement learning to achieve adaptive length penalty and reward shaping, dynamically controlling the thinking chain length of large-scale reasoning models, significantly reducing the generated length;
Leave My Images Alone: Preventing Multi-Modal Large Language Models from Analyzing Images via Visual Prompt Injection
Zedian Shao (Georgia Institute of Technology), Neil Zhenqiang Gong (Duke University)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: Propose ImageProtector, a method that inserts almost imperceptible perturbations on images using a visual prompt injection approach, enabling multimodal large language models (MLLMs) to refuse answering any queries related to the image, thereby protecting user privacy.
LeCoDe: A Benchmark Dataset for Interactive Legal Consultation Dialogue Evaluation
Weikang Yuan (Zhejiang University), Xiaozhong Liu (Worcester Polytechnic Institute)
Recommendation SystemFederated LearningExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextSequentialBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Constructed and publicly released a large-scale multi-turn legal consultation dialogue dataset called LeCoDe, and proposed a comprehensive evaluation framework to measure the performance of LLMs in two aspects: clarifying information and providing professional legal advice.
LEDOM: Reverse Language Model
Xunjian Yin (Peking University), Xiaojun Wan (Peking University)
GenerationData SynthesisExplainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Studied and released LEDOM, a large-scale purely right-to-left autoregressive language model, and systematically evaluated its differences with forward models.
Legal Judgment Prediction: A Reflection on the State of the Art
Yi Feng (Nanjing University), Vincent Ng (University of Texas at Dallas)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: Reviews the current state of the field of legal judgment prediction (LJP), proposes the vision of building interpretable and trustworthy LJP systems, and uses argument trees as the core structure to demonstrate how legal norms, precedents, and humanistic values can be embedded into the reasoning process;
LegalChainReasoner: Grounding Criminal Judicial Opinion Generation via Structured Legal Chains
Weizhe Shi (University of Auckland), Kaiqi Zhao (Harbin Institute of Technology)
GenerationTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsTextChain-of-Thought
🎯 What it does: Propose the criminal justice opinion generation task, combining legal reasoning with sentencing prediction, and design the LegalChainReasoner framework.
LegalGraphRAG: Multi-Agent Graph Retrieval-Augmented Generation for Reliable Legal Reasoning
Zerui Chen (Xiamen University), Jinsong Su (Xiamen University)
OptimizationFederated LearningSafty and PrivacyExplainability and InterpretabilityGraph Neural NetworkTransformerLarge Language ModelAgentic AITextGraphRetrieval-Augmented Generation
🎯 What it does: Proposed the LegalGraphRAG framework, which achieves reliable legal reasoning by constructing a hierarchical legal knowledge graph and introducing multiple agents (retrieval, review, adjudication).
Leibniz: Theory-of-Mind Driven Neuro-Symbolic Logical Reasoning via Multi-Agent Collaboration
Yue Fan (Shanxi University), Yuanlong Wang (Shanxi University)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningMeta LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Propose a multi-agent neuro-symbolic reasoning framework called Leibniz based on the Theory-of-Mind, which utilizes evolutionary agents and reduction agents to collaboratively advance logical reasoning in a shared belief space;
LeLoRA: Learnable Low-Rank Adaptation of Large Language Models
Xiaoling Zhou (Peking University), Shikun Zhang (Peking University)
Computational EfficiencyKnowledge DistillationRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Proposes the LeLoRA framework, which dynamically generates adaptation policies (the effective rank and position of each weight matrix) using a policy network, and alternately updates the LLM and policy network through reinforcement learning to achieve learnable low-rank adaptation fine-tuning.
LENS: LLM-Enabled Narrative Synthesis for Mental Health by Aligning Multimodal Sensing with Language Models
Wenxuan Xu (Dartmouth College), Andrew Campbell (Dartmouth College)
Data SynthesisTransformerLarge Language ModelSupervised Fine-TuningTextMultimodalityTime SeriesElectronic Health Records
🎯 What it does: Proposed the LENS framework, which generates clinical narratives about mental health by aligning multimodal sensor data with language models.
LePREC: Reasoning as Classification over Structured Factors for Assessing Relevance of Legal Issues
Fanyu Wang (Monash University), Lizhen Qu (Monash University)
ClassificationExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningAuto EncoderGenerative Adversarial NetworkTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper constructs the first expert-annotated dataset of 769 Malaysian contract law case examples (LIC) and proposes a neuro-symbolic framework called LEPREC, which combines neural generation with sparse linear classification, for determining the relevance of legal questions.
Less is More: Improving LLM Reasoning with Minimal Test-Time Intervention
Zhen Yang (Hong Kong University of Science and Technology (Guangzhou)), Ying-Cong Chen (Hong Kong University of Science and Technology)
Explainability and InterpretabilityComputational EfficiencyAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the Minimal Test-Time Intervention (MTI) framework, which utilizes token entropy to locally stabilize the LLM inference process during reasoning, thereby improving inference accuracy and stability.
Less Languages, Less Tokens: An Efficient Unified Logic Cross-lingual Chain-of-Thought Reasoning Framework
Chenyuan Zhang (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the UL-XCoT framework, which utilizes a unified logical space to achieve cross-lingual chain-of-thought reasoning, significantly reducing language and token consumption while maintaining or even improving reasoning accuracy.
Leveraging Outline-Optimized Generative Interactions and Critique for Self-Refining Outlines with Reinforcement Learning
Hengwei Liu (Zhejiang University), Weiming Lu (Zhejiang University)
GenerationData SynthesisReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the LOGIC-RL framework, which transforms the teacher model's critique-revision process into a learnable reinforcement learning strategy for small models to generate high-quality long outlines.
LexRel: Benchmarking Legal Relation Extraction for Chinese Civil Cases
Yida Cai (Tsinghua University), Zhiyuan Liu (Tsinghua University)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed a fine-grained hierarchical schema for civil legal relationships in China, and generated the expert-annotated LexRel benchmark dataset based on this schema; evaluated multiple large language models (LLMs) on the legal relationship extraction task under two settings—zero-shot and relation-enhanced—using LexRel, and further validated the performance improvement achieved by incorporating the extracted legal relationships into downstream legal AI tasks (case analysis, legal consultation, and criminal compensation calculation).
LGSA: Label Geometry Structuring and Aligning for Hierarchical Text Classification
Shuai Zhang (Zhejiang University of Finance and Economics), Kecheng Huang (Emerging AI Technology Co Ltd)
ClassificationTransformerSupervised Fine-TuningPrompt EngineeringContrastive LearningText
🎯 What it does: This study proposes a two-stage label geometry structurization and alignment framework, LGSA, which explicitly constructs the geometry of labels for hierarchical text classification and enhances model performance through geometric regularization alignment.
Libra-VLA: Achieving Learning Equilibrium via Asynchronous Coarse-to-Fine Dual-System
Yifei Wei (Beihang University), Guanghui Ren (AgiBot)
Robotic IntelligenceTransformerReinforcement LearningVision Language ModelVision-Language-Action ModelDiffusion modelImageVideoTextMultimodalityBenchmark
🎯 What it does: Propose Libra-VLA, which adopts a coarse-to-fine dual-system architecture, decomposing robot actions into discrete macro intentions and continuous refined actions;
Lifting Optimized Binaries to Canonical Compiler IR via Structure-Aware Retrieval and Iterative Verification
Xiaoao Zhu (Shaanxi Normal University), Zheng Wang (Shaanxi Normal University)
OptimizationAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose the BRIDGE system, which utilizes structure-aware retrieval and iterative verification to reliably lift highly optimized binary files into standard LLVM IR.
LiGen: Active Lipid Generation via a Molecular Language Model
Ying Zhan (Southeast University), Beilun Wang (Southeast University)
Drug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningGenerative Adversarial NetworkTabularBiomedical Data
🎯 What it does: Propose the LiGen framework, which generates improved lipid nanoparticle (LNP) formulations in the latent space by leveraging a specially pre-trained lipid molecule language model, LiCore, and a multi-cell line predictor.
LightReasoner: Can Small Language Models Teach Large Language Models Reasoning?
Jingyuan Wang (University of Hong Kong), Chao Huang (University of Hong Kong)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningTextChain-of-Thought
🎯 What it does: Propose the LightReasoner framework, which automatically selects key reasoning steps by leveraging the behavioral differences between a small language model (amateur) and a large language model (expert), constructs contrastive supervision, and further performs self-distillation training on the expert.
Lightweight and Faithful Visual Condition Checking in Behavior Trees via Expert-Regularized Reinforcement Learning
Hyosik Moon (University of Toronto), Eldan Cohen (University of Toronto)
Computational EfficiencyRepresentation LearningTransformerSupervised Fine-TuningReinforcement LearningMixture of ExpertsVision Language ModelImageTextMultimodality
🎯 What it does: Proposes a lightweight model that trains visual condition nodes in behavior trees using expert-regularized reinforcement learning, maintaining semantic consistency while significantly improving inference speed.
Lightweight LLM Agent Memory with Small Language Models
Jiaquan Zhang (University of Electronic Science and Technology of China), Yang Yang (University of Electronic Science and Technology of China)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Propose LightMem, a lightweight LLM agent memory system based on small language models
Limited Linguistic Diversity in Embodied AI Datasets
Selma Liliane Wanna (Los Alamos National Laboratory), Mitch Pryor (University of Texas at Austin)
Robotic IntelligenceVision-Language-Action ModelTextMultimodality
🎯 What it does: Conduct a systematic linguistic diversity audit of existing Vision-Language-Action datasets, quantifying diversity in three aspects: lexical repetition, semantic coverage, and syntactic structure.
Linear-Time and Constant-Memory Text Embeddings Based on Recurrent Language Models
Tobias Grantner (Dynatrace Research), Martin Flechl (Dynatrace Research)
Computational EfficiencyRepresentation LearningRecurrent Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextBenchmark
🎯 What it does: Studied text embedding generation based on recursive models such as Mamba2, RWKV, and xLSTM, and proposed a vertical chunking inference strategy to achieve linear time and constant memory.
Lingua-Graph: A Unified Representation of Cross-Task Common Substructures for Analytic Language Processing
Mingming Sun (BIMSA), Yunfeng Cai (BIMSA)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextGraph
🎯 What it does: Propose a unified intermediate structure, Lingua-Graph, as a shared graph representation for cross-task structural analysis, forming the representation-decision paradigm of Analytic NLP.
LinkQA: Synthesizing Diverse QA from Multiple Seeds Strongly Linked by Knowledge Points
Xuemiao Zhang (Peking University), Xunliang Cai (Meituan)
Data SynthesisKnowledge DistillationRepresentation LearningGraph Neural NetworkTransformerPrompt EngineeringDiffusion modelTextGraphRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Construct a knowledge point (KP) graph and synthesize diverse question-answer data through graph traversal and multi-source seed synthesis.
Listening Like Humans: Semantics-Guided Noise-Robust Multimodal Speech Recognition
Yan Fang (Hong Kong University of Science and Technology (Guangzhou)), Kaishun Wu (Hong Kong University of Science and Technology (Guangzhou))
RecognitionTransformerPrompt EngineeringAuto EncoderContrastive LearningImageTextMultimodalityRetrieval-Augmented GenerationAudio
🎯 What it does: Transforming noise-robust ASR into a semantics-guided speech reconstruction task, the Speech-MLM framework is proposed, combining audio, visual (spectrogram) and textual (synonymous transcription) multimodal information to achieve structural recovery and semantic reasoning under noise.
LitVISTA: A Benchmark for Narrative Orchestration in Literary Text
Mingzhe Lu (Institute of Information Engineering, Chinese Academy of Sciences), Yunpeng Li (Institute of Information Engineering, Chinese Academy of Sciences)
Representation LearningTransformerLarge Language ModelPrompt EngineeringTextGraphBenchmark
🎯 What it does: This paper proposes the VISTA Space framework and the LitVISTA benchmark to unify narrative representations of humans and models, and to evaluate the narrative composition capabilities of large models in literary texts.
LiveCLKTBench: Towards Reliable Evaluation of Cross-Lingual Knowledge Transfer in Multilingual LLMs
Pei-Fu Guo (National Taiwan University), Shou-De Lin (National Taiwan University)
TransformerLarge Language ModelPrompt EngineeringTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes LiveCLKTBench, an automated generation, leakage-free, real-knowledge-based, and sustainably updatable cross-lingual knowledge transfer evaluation benchmark.
LiveCultureBench: a Multi-Agent, Multi-Cultural Benchmark for Large Language Models in Dynamic Social Simulations
Viet Thanh Pham, Dinh Phung (Monash University)
TransformerLarge Language ModelAgentic AITextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed LIVECULTUREBENCH, a multi-cultural, multi-agent dynamic evaluation framework, where LLM agents complete daily tasks in a simulated small town while adhering to cultural norms during the process.
LiveFact: A Dynamic, Time-Aware Benchmark for LLM-Driven Fake News Detection
Cheng Xu (University College Dublin), Tahar Kechadi
TransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper introduces LiveFact, a dynamic and time-aware benchmark for evaluating the reasoning and uncertainty handling capabilities of LLMs in fake news detection.
Lizard: An Efficient Linearization Framework for Large Language Models
Chien Van Nguyen (University of Oregon), Thien Huu Nguyen (University of Oregon)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Linearizing the pre-trained Transformer LLM through a sub-quadratic cascading attention mechanism to build an efficient model with constant memory and infinite context.
LLaDA 1.5: Variance-Reduced Preference Optimization for Large Language Diffusion Models
Fengqi Zhu (Renmin University of China), Chongxuan Li (Renmin University of China)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningDiffusion modelTextBenchmark
🎯 What it does: A Variance-Reduced Preference Optimization (VRPO) framework is proposed for the alignment task of masked diffusion language models (MDM), which improves the stability of direct preference optimization (DPO) by reducing the variance of the ELBO estimation, achieving LLaDA 1.5 on LLaDA 8B Instruct.
LLM Agents in Law: Taxonomy, Applications, and Challenges
Shuang Liu (Carnegie Mellon University), Mengnan Du (Chinese University of Hong Kong Shenzhen)
Recommendation SystemAutonomous DrivingOptimizationFederated LearningSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningAdversarial AttackHyperparameter SearchData-Centric LearningRobotic IntelligenceMeta LearningDrug DiscoveryAI Code AssistantReinforcement Learning from Human FeedbackNeural Architecture SearchProtein Structure PredictionTransformerLarge Language ModelAgentic AIPrompt EngineeringTextReview/Survey PaperBenchmarkFinance RelatedRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: A systematic review of the current state of large language model (LLM) agents in the legal domain, proposing migration paths for agent technologies, application classifications, evaluation methods, and identifying existing challenges and future research directions.
LLM Beliefs Are in Their Heads
Alessandro Corona Mendozza (University of Copenhagen), Anders Søgaard (University of Copenhagen)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: Investigated whether the internal representations of decoder-only language models possess characteristics similar to beliefs, based on four criteria from Herrmann & Levinstein (2025) (accuracy, functionality, coherence, and consistency) through probing and intervention experiments.
LLM Reasoning as Trajectories: Step-Specific Representation Geometry and Correctness Signals
Lihao Sun (Microsoft), Saravan Rajmohan (Microsoft)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper studies the trajectory structure in the representation space during the chain-of-thought (CoT) generation process of large language models (LLMs), revealing step-specific subspaces and correctness signals, and proposes an intervention method during inference based on this trajectory.
LLM Safety From Within: Detecting Harmful Content with Internal Representations
Difan Jiao (University of Toronto), Ashton Anderson (University of Toronto)
Safty and PrivacyComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: Proposed a lightweight content safety detection framework called SIREN, which utilizes safety-related neurons within the internal layers of large language models for detection, rather than relying solely on terminal layer outputs.
LLM-as-Scheduler: Agentic Workflow Dynamic Scheduling
Dawei Xiang (University of Connecticut), Wei Zhang (University of Connecticut)
OptimizationComputational EfficiencyAI Code AssistantTransformerLarge Language ModelAgentic AITextChain-of-Thought
🎯 What it does: Propose a two-stage cascading scheduler (LAS) based on lightweight gating and large language models, dynamically deciding whether to continue executing multi-agent workflows based on query difficulty, significantly reducing token consumption and latency.
LLM-Based Multi-Agent Systems for Clinical Workflows: A Survey of AI Hospitals
Zonghai Yao (University of Massachusetts), Hong yu
Drug DiscoveryTransformerLarge Language ModelAgentic AITextBiomedical DataElectronic Health RecordsReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: Review and systematize the design and evaluation framework of LLM-driven multi-agent clinical workflows (AI hospital)
LLM-Based Multi-Task Bangla Hate Speech Detection: Type, Severity, and Target
Md Arid Hasan (University of Toronto), Syed Ishtiaque Ahmed (University of Toronto)
ClassificationRecognitionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: This study constructs the first Bangla multilingual hate speech dataset, BanglaMultiHate, and systematically evaluates the performance of classical machine learning models, monolingual pre-trained models BanglaBERT, and various large language models (LLMs) under zero-shot, few-shot, and LoRA fine-tuning settings on this dataset, exploring detection across three dimensions: type, severity, and target.
LLM-ForcedAligner: A Non-Autoregressive and Accurate LLM-Based Forced Aligner for Multilingual and Long-Form Speech
Bingshen Mu (Northwestern Polytechnical University), Lei Xie (Northwestern Polytechnical University)
RecognitionTransformerLarge Language ModelTextAudio
🎯 What it does: Propose a non-autoregressive forced alignment framework based on large language models, LLM-ForcedAligner, which treats forced alignment as a slot-filling task, directly predicting the start and end time indices of words/characters.
LLM-Generated Text May Harm Your Retrieval! A Robust Detection Strategy for Retrieval-Augmented Generation
Zhaoheng Huang (Renmin University of China), Zhicheng Dou (Renmin University of China)
Data SynthesisRetrievalTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: This paper investigates the role of LLM text detection in retrieval-augmented generation (RAG) systems, proposes a RAG-aware data augmentation strategy (RAD), and verifies its effectiveness through iterative corpus evolution experiments.
LLM-MC-Affect: LLM-Based Monte Carlo Modeling of Affective Trajectories and Latent Ambiguity for Interpersonal Dynamic Insight
Yu-Zheng Lin (University of Arizona), Pratik Satam (University of Arizona)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelTextTime Series
🎯 What it does: A probability-based affect trajectory framework, LLM‑MC‑Affect, was constructed based on Monte Carlo sampling of large language models (LLMs), aiming to capture the continuous distribution of affect from dialog text and quantify ambiguity;
LLM-Powered Benchmark Factory: Reliable, Generic, and Efficient
Peiwen Yuan (Beijing Institute of Technology), Kan Li (Beijing Institute of Technology)
TransformerLarge Language ModelPrompt EngineeringDiffusion modelContrastive LearningTextBenchmark
🎯 What it does: A general, reliable, and efficient benchmark generator based on large language models (LLM), called BENCHMAKER, was constructed, and a complete ten-dimensional evaluation framework was proposed to automatically evaluate the generated benchmarks from multiple dimensions.
LLM-SLM Collaborative Framework of Idiomatic Expression Generation
Hui Gao (Tianjin University), Liuxian Ge (Tianjin University)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Propose Auto-IDEA, a method that automatically generates a Chinese idiom polishing dataset annotated with Locate-Then-Polish (LTP) through iterative collaboration between a Large Language Model (LLM) and a Small Language Model (SLM);
LLM-VA: Resolving the Jailbreak-Overrefusal Trade-off via Vector Alignment
Haonan Zhang (Zhejiang University), Wenhai Wang (Griffith University)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Improve the safety alignment of large language models by proposing a method called LLM-VA, which aligns vectors without fine-tuning or modifying the model architecture, simultaneously reducing the occurrence of jailbreak attacks and excessive rejection.
LLM-XTM: Enhancing Cross-Lingual Topic Models with Large Language Models
Minh Chu Xuan (Hanoi University of Science and Technology), Trung Le (Monash University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose the LLM-XTM framework, which introduces LLM into existing cross-lingual topic models to refine topic keywords and align documents at the document level, thereby improving topic coherence and interpretability.
LLM4Cell: Taxonomy and Evaluation of LLM and Agentic Models for Single-Cell Biology
Sajib Acharjee Dip (Virginia Tech), Liqing Zhang (Fralin Biomedical Research Institute at VTC)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningData-Centric LearningDrug DiscoveryTransformerLarge Language ModelAgentic AIMixture of ExpertsTextMultimodalityTabularBiomedical DataReview/Survey PaperBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper constructs the LLM4Cell framework, conducting a unified literature review, model classification, task mapping, and evaluation of 58 large language models and their agent-based systems in the field of single-cell biology, covering various data modalities such as RNA, ATAC, multi-omics, spatial, and perturbation.
LLMEval-Fair: A Large-Scale Longitudinal Study on Robust and Fair Evaluation of Large Language Models
Ming Zhang (Fudan University), Xuanjing Huang (Fudan University)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed LLMEval-Fair, a dynamic evaluation framework based on a 220k private graduate-level question bank, dynamic sampling, dual-layer anti-cheating mechanisms, and LLM-as-Judge, and conducted longitudinal evaluations of nearly 60 LLMs over 30 months.
LLMs (Almost) Never Abstain Under Medical Uncertainty
Alessio Cocchieri (University of Bologna), Gianluca Moro (University of Bologna)
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextMultimodalityBiomedical DataElectronic Health RecordsBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Created the MedQAbstain benchmark, reconstructing medical multiple-choice questions to allow models to choose "I abstain" when uncertain, and systematically evaluated LLMs' abstention behavior under medical uncertainty.
LLMs as Knowledge Graph Refiners: Mitigating Factual Inconsistencies in Generative Knowledge Extraction
Donghyun Kim (Yonsei University), Chanhee Lee (Samsung Securities)
Explainability and InterpretabilityKnowledge DistillationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextGraph
🎯 What it does: Propose the GraphRefine framework, which performs post-hoc fine-grained refinement on the knowledge generated by generative knowledge extraction (GKE), executing four operations—KEEP, DELETE, FIX, and REWRITE—on each triplet;
LLMs Enable Bag-of-Texts Representations for Short-Text Clustering
I-Fan Lin (Leiden University), Suzan Verberne (Leiden University)
Representation LearningData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes LeBoT, a short text clustering method that does not require training or annotation, generating Bag-of-Texts vectors directly through LLM to represent text;
LLMs in Sarcasm Detection? It’s elementary! (Or is it?)
Priyanshu Mahato (IISER Kolkata), Kripabandhu Ghosh (IISER Kolkata)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: By comparing the performance of LLMs in detecting sarcasm on synthetic, naturally generated, and real human-generated data, the study reveals a catastrophic generalization gap, where models perform almost perfectly on synthetic data but make random guesses on real data. This phenomenon is explained through mechanisms such as hidden state geometry, entropy analysis, membership inference, and causal interventions, highlighting the root cause of shortcut learning.
LLMs Underperform Graph-Based Parsers on Supervised Relation Extraction for Complex Graphs
Paolo Gajo (University of Bologna), Alberto Barrón-Cedeño (University of Bologna)
Graph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextGraph
🎯 What it does: This paper systematically compares the performance of four large language models (LLM) with lightweight graph structure parsers on the relation extraction (RE) task, focusing on the impact of the complexity of language graphs in text on model performance.
LLMSurgeon: Diagnosing Data Mixture of Large Language Models
Yaxin Luo (Mohamed bin Zayed University of Artificial Intelligence), Zhiqiang Shen (Mohamed bin Zayed University of Artificial Intelligence)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelContrastive LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the LLMSurgeon framework, which infers the mixture ratio of pre-training data from the generated text of LLMs through a label migration inverse inference method, enabling post-audit of digital DNA.
Localized Low-Rank Adaptation within Clustered Parameter Subspaces
Jiahao Xiong (East China Normal University), Kai Zhang (East China Normal University)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: In parameter-efficient fine-tuning of large language models, we propose dividing the update matrix into local subspace blocks, and performing low-rank adaptation independently within each block, thereby achieving more fine-grained and efficient weight updates.
Locate and Explain: Joint Multimodal Emotion Cause Extraction and Summarization in Conversation
Jikun Wan (Soochow University), Guohong Fu (Soochow University)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelVideoTextMultimodalityAudio
🎯 What it does: Proposes a new task called MECES, which simultaneously localizes and interprets the emotional causes of conversational emotions, and constructs the MECESD dataset based on the ABC theory
Location Not Found: Exposing Implicit Local and Global Biases in Multilingual LLMs
Guy Mor-Lan (Google Research), Reut Tsarfaty (Bar Ilan University)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Constructed the LOCQA benchmark to evaluate implicit biases in multilingual LLMs when handling 'regional ambiguity' issues, and conducted zero-shot evaluations on 32 models.
Locket: Robust Feature-Locking Technique for Language Models
Lipeng He (University of Waterloo), N. Asokan (University of Waterloo)
Federated LearningSafty and PrivacyComputational EfficiencyAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Designed and implemented a feature-locking technique called LOCKET, based on adapters, for implementing a paid unlocking scheme in large language models.
Locomo-Plus: Beyond-Factual Cognitive Memory Evaluation Framework for LLM Agents
Yifei Li (Xi'an Jiaotong University), Jun Liu (Xi'an Jiaotong University)
TransformerLarge Language ModelAgentic AIPrompt EngineeringTextSequentialBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the LoCoMo-Plus benchmark to evaluate the cognitive memory capability of LLM agents in long-term conversations regarding implicit constraints (such as user state, goals, values, etc.).
Logic Matters in Lightweight Hallucination Classification for RAG System
Ningyuan Yang (Duke Kunshan University), Kaizhu Huang (Duke Kunshan University)
ClassificationRetrievalExplainability and InterpretabilityComputational EfficiencyGraph Neural NetworkTransformerContrastive LearningTextGraphRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose a lightweight, modular RAG hallucination detection framework that aggregates cross-document semantic evidence using a graph structure and then uses an NLI model to determine if the answer is real.
Logical Consistency as a Bridge: Improving LLM Hallucination Detection via Label Constraint Modeling between Responses and Self-Judgments
Hao Mi (Institute of Computing Technology Chinese Academy of Sciences), Juan Cao (Institute of Computing Technology Chinese Academy of Sciences)
Anomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: Propose a hallucination detection framework called LaaB, which bridges the micro-level internal features of LLMs with macro-level self-assessment judgments by leveraging logical consistency, and jointly learns by mapping self-assessment labels back into the feature space through meta-evaluation.
Logical Phase Transitions: Understanding Collapse in LLM Logical Reasoning
Xinglang Zhang (Huazhong University of Science and Technology), Zikai Song (Huazhong University of Science and Technology)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This paper investigates whether the performance of large language models on logical reasoning tasks undergoes a sharp decline, akin to a phase transition, as logical complexity increases gradually, and proposes the Neuro-Symbolic Curriculum Tuning (NSCT) framework, which enhances the robustness of high-complexity reasoning by aligning neural and symbolic components and employing adaptive curriculum learning.
LOGICAL-COMMONSENSEQA: A Benchmark for Logical Commonsense Reasoning
Obed Junias (University of Colorado Boulder), Maria Leonor Pacheco (University of Colorado Boulder)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Propose the LOGICAL-COMMONSENSEQA benchmark, which transforms common sense reasoning into a logical combination (AND/OR/NEITHER) of two atomic statements for judgment.
LogicEval: A Systematic Framework for Evaluating Automated Repair Techniques for Logical Vulnerabilities in Real-World Software
Syed Md Mukit Rashid (Pennsylvania State University), Syed Rafiul Hussain (Pennsylvania State University)
Safty and PrivacyExplainability and InterpretabilityAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper constructs a systematic framework called LogicEval for evaluating Automatic Vulnerability Repair (AVR) techniques on logical vulnerabilities, and creates the first logical vulnerability dataset called LogicDS based on real-world security vulnerabilities; this framework is used to uniformly evaluate traditional, learning-based, and LLM-driven repair methods.
LogicPoison: Logical Attacks on Graph Retrieval-Augmented Generation
Yilin Xiao (Hong Kong Polytechnic University), Xiao Huang (Hong Kong Polytechnic University)
Explainability and InterpretabilityComputational EfficiencyAdversarial AttackData-Centric LearningGraph Neural NetworkTransformerPrompt EngineeringTextGraphRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the LOGICPOISON framework, which implicitly disrupts the logical topology of GraphRAG by performing type-preserving cyclic permutation of entities during the knowledge graph construction process, leading the model to generate erroneous answers;
LogosKG: Hardware-Optimized Scalable and Interpretable Knowledge Graph Retrieval
He Cheng (University of Colorado Anschutz), Yanjun Gao (University of Colorado Anschutz)
RetrievalExplainability and InterpretabilityComputational EfficiencyGraph Neural NetworkPrompt EngineeringGraphBiomedical DataRetrieval-Augmented Generation
🎯 What it does: Designed and implemented the LOGOSKG system, achieving hardware-optimized, scalable, and interpretable multi-hop knowledge graph retrieval.
LOKA: Conflict-Aware LLM Knowledge Update with Adaptive Knowledge Memory
Binchi Zhang (University of Virginia), Haifeng Chen (NEC Laboratories America)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Proposed the LOKA framework, aiming to simultaneously address knowledge learning and unlearning in large language models, achieving knowledge updates without modifying the original model weights by utilizing assignable external memory units.
Long Context Modeling with Ranked Memory-Augmented Retrieval
Ghadir Alselwi (University of New South Wales), Imran Razzak (Mohamed Bin Zayed University of Artificial Intelligence)
RetrievalComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes the ERMar framework, which improves context retention and reasoning quality in long-text generation through a dynamic ranking-based memory retrieval mechanism.
Long-Chain Reasoning Distillation via Adaptive Prefix Alignment
Zhenghao Liu (Northeastern University), Maosong Sun (Tsinghua University)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: By adaptively clipping and aligning the long-chain reasoning process generated by the teacher model into a learnable prefix, and then allowing the student model to perform complete reasoning on this prefix, the reasoning ability of the small model is distilled.
Long-Context Reasoning Through Proxy-Based Chain-of-Thought Tuning
Miao Li (University of Edinburgh), Mirella Lapata (University of Edinburgh)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose a two-stage training framework called ProxyCoT, which first learns chained reasoning in short proxy contexts and then transfers it to complete long contexts, thereby improving the performance of LLMs on long-context reasoning tasks.