From Regulatory Approvals to Patents: Cross-Domain Linking for Cardiovascular Device Traceability
Qingqing Yang, Moyan Li (Hong Kong University of Science and Technology (Guangzhou))
CodeDrug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextBiomedical DataElectronic Health RecordsReview/Survey PaperBenchmarkRetrieval-Augmented Generation
π― What it does: This study proposes the Bridge-MedDevKG framework, which aligns FDA-approved medical devices with their related U.S. patents across domains and constructs the first benchmark for cardiovascular device-patent alignment;
From Synthesis to Clinical Assistance: A Strategy-Aware Agent Framework for Autism Intervention based on Real Clinical Dataset
Junhong Lai (Zhejiang University), Yueming Wang (Zhejiang University)
CodeData SynthesisRecommendation SystemExplainability and InterpretabilityKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringTextBiomedical DataElectronic Health RecordsChain-of-Thought
π― What it does: Proposed the ASDAGENT framework, combining DOCTORAGENT (O-T-A-C iterative reasoning) and CHILDAGENT (probabilistic behavior simulation), to achieve high-fidelity autism intervention dialogue synthesis and clinical decision support.
Yang Yang (Sun Yat-sen University), Wenqi Ren (Sun Yat-sen University)
CodeGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelDiffusion modelContrastive LearningImageTextMultimodalityGraphBenchmarkRetrieval-Augmented Generation
π― What it does: This paper proposes the task of generating image captions for game characters, constructs the GC-Bench benchmark and the Graph-F1 evaluation metric, further builds the GC-148K large-scale dataset, and fine-tunes the G-Cap series of models on it.
G-IdiomAlign: A Gloss-Pivoted Benchmark for Cross-Lingual Idiom Alignment
Fengying Ye (University of Macau), Derek F. Wong (University of Macau)
CodeTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmarkRetrieval-Augmented Generation
π― What it does: Constructed a cross-lingual idiom alignment benchmark, G-IdiomAlign, based on English definitions, and proposed two evaluation protocols (multiple-choice and contrastive generation), providing a reproducible diagnostic framework for cross-lingual idiom equivalence.
Gained in Translation: Privileged Pairwise Judges Enhance Multilingual Reasoning
Lintang Sutawika (Carnegie Mellon University), Graham Neubig (Carnegie Mellon University)
CodeComputational EfficiencyData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextMultimodalityRetrieval-Augmented Generation
π― What it does: Propose the SP3F framework in the absence of target language data, first using translated English QA for supervised fine-tuning, and then training through self-adversarial RL and pairwise judge with restricted information.
Gap-K%: Measuring Top-1 Prediction Gap for Detecting Pretraining Data
Minseo Kwak (Yonsei University), Jaehyung Kim (Yonsei University)
CodeAnomaly DetectionExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelContrastive LearningText
π― What it does: A pre-training data detection method based on the training dynamics of autoregressive language models, called Gap-K%, is studied. It distinguishes between training set samples and untrained samples by measuring the log probability difference between the target word and the model's highest probability word.
GASim: A Graph-Accelerated Hybrid Framework for Social Simulation
Xuan Zhou (University of Science and Technology of China), Wu Liu (University of Science and Technology of China)
CodeComputational EfficiencyReinforcement Learning from Human FeedbackGraph Neural NetworkTransformerLarge Language ModelTextGraphRetrieval-Augmented Generation
π― What it does: Proposed GASim, a graph-accelerated hybrid multi-agent framework for large-scale social simulations, addressing the high latency issues caused by traditional frameworks where LLM retrieval and ABM sequential execution are used.
Generating Literature-Driven Scientific Theories at Scale
Peter Jansen (Allen Institute for Artificial Intelligence), Daniel S Weld
CodeGenerationExplainability and InterpretabilityKnowledge DistillationData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextReview/Survey PaperRetrieval-Augmented Generation
π― What it does: Developed a system called THEORIZER for automatically generating theories from a large volume of scientific literature.
GenesisFunc: Multi-Agent Data Generation for Accurate and Generalizable Function-Calling
Hao-Xiang Xu (University of Science and Technology of China), Zhen-Hua Ling (University of Science and Technology of China)
CodeGenerationData SynthesisReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIPrompt EngineeringTextSequentialRetrieval-Augmented Generation
π― What it does: Built the GENESISFUNC multi-agent data generation pipeline, automatically generating high-quality, scalable multi-turn function call training data, covering multiple tools, multiple tasks, and multiple scenarios;
GenomeQA: Benchmarking General Large Language Models for Genome Sequence Understanding
Weicai Long (Hong Kong University of Science and Technology (Guangzhou)), Yanlin Zhang (Hong Kong University of Science and Technology (Guangzhou))
CodeProtein Structure PredictionTransformerLarge Language ModelPrompt EngineeringTextSequentialBiomedical DataBenchmarkChain-of-Thought
π― What it does: Propose the GenomeQA benchmark, which designs 5,200 questions based on original DNA sequences, covering six major tasks: enhancer/promoter identification, splicing sites, classification, histone marks, transcription factor binding sites, and motif prediction;
GenPT: Beyond Self-Report for Reliable LLM Psychometrics via Generative Projective Testing
Ming Wang (Northeastern University), Yufan Sun (Northeastern University)
CodeExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelAgentic AIPrompt EngineeringDiffusion modelImageTextTabularBenchmarkRetrieval-Augmented Generation
π― What it does: Proposed and implemented the GenPT generative projective test framework to evaluate the psychological characteristics of LLM agents, addressing issues of training data contamination and social desirability bias in self-report questionnaires.
GiLT: Augmenting Transformer Language Models with Dependency Graphs
Tianyu Huang (ShanghaiTech University), Kewei Tu (ShanghaiTech University)
CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelContrastive LearningTextGraph
π― What it does: Propose a new Transformer language model called GiLT, which enhances the syntactic structure during the parsing process by utilizing dependency graphs.
GKnow: Measuring the Entanglement of Gender Bias and Factual Gender
Leonor Veloso (LMU Munich), Hinrich Schuetze
CodeExplainability and InterpretabilityTransformerLarge Language ModelTextBenchmark
π― What it does: This paper constructs the GKnow benchmark, systematically analyzing the high coupling between gender bias and factual gender information at the circuit layer and neuron layer in language models, and verifying the destructive impact of ablation on factual gender knowledge through neuron ablation experiments.
CodeComputational EfficiencyRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelContrastive LearningImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
π― What it does: Propose the Glyph framework, which renders long text into a visual page and then processes it with a vision-language model, achieving scalability for long contexts.
GMoE: Global Mixture of Experts with Logit Propagation
Geonwoo Hong (Ulsan National Institute of Science and Technology), Taehwan Kim (Ulsan National Institute of Science and Technology)
CodeComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelMixture of ExpertsText
π― What it does: Propose a GMoE architecture that achieves parameter efficiency in sparse mixture-of-experts networks through shared global experts, local experts, and a global router.
GQLBench: A Large-Scale Cross-Domain, Cross-Dialect Benchmark for NL2GQL
Yanning Su (Fudan University), Hongfeng Chai (Fudan University)
CodeData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextGraphTabularBenchmark
π― What it does: Constructed GQLBench, a cross-domain, cross-dialect, executable NL2GQL benchmark, and provided corpus based on automated migration and synthesis.
GrACE: A Generative Approach to Better Confidence Elicitation and Efficient Test-Time Scaling in Large Language Models
Zhaohan Zhang (Queen Mary University of London), Ioannis Patras (Queen Mary University of London)
CodeGenerationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented Generation
π― What it does: Proposes GrACE, a generative confidence mining method that directly outputs confidence during the generation phase, by adding a special <CNF> token to the vocabulary, allowing the model to generate this token at the end of the response, and using the similarity between its hidden state and the embedded <CNF> token to estimate confidence in real-time;
π― What it does: Propose GRAD, a framework that dynamically activates multiple RAG objectives (model expansion, domain adaptation, position debiasing) during decoding and converts them into token-level guidance for small models, avoiding the need to retrain large models for each task.
CodeOptimizationExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
π― What it does: This paper proposes an efficient and robust automatic prompt optimization framework called GMPO, which uses gradient approximation to score the importance of prompt paragraphs and combines the loss integration of multiple discriminators to guide prompt rewriting.
Graph-Based Alternatives to LLMs for Human Simulation
Joseph Suh (University of California, Berkeley), Serina Chang (University of California, Berkeley)
CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextGraphChain-of-Thought
π― What it does: Proposed a graph-based model called GEMS to simulate human behavior in closed-choice tasks, replacing large language models.
GRASPrune: Global Gating for Budgeted Structured Pruning of Large Language Models
Ziyang Wang (Beijing Institute of Technology), Jianbin Qin (Shenzhen University)
CodeComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelContrastive LearningText
π― What it does: This paper proposes a global budget structured pruning framework called GRASPrune, which jointly prunes FFN channels and KV head groups after pre-training while keeping the model weights unchanged;
GuideDog: A Real-World Egocentric Multimodal Dataset for Blind and Low-Vision Accessibility-Aware Guidance
Junhyeok Kim (Yonsei University), Youngjae Yu (Seoul National University)
CodeObject DetectionDepth EstimationRecommendation SystemAnomaly DetectionAutonomous DrivingData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: This paper proposes a scalable and assessable egocentric dataset called GUIDE DOG for blind/low vision (BLV) navigation assistance, and provides a multimodal guidance generation task and a fine-grained visual perception question answering benchmark based on BLV standards.
HAG: Hierarchical Demographic Tree-based Agent Generation for Topic-Adaptive Simulation
Rongxin Chen (State Key Laboratory of AI Safety), Huawei Shen (State Key Laboratory of AI Safety)
CodeGenerationData SynthesisTransformerLarge Language ModelWorld ModelTextBenchmarkRetrieval-Augmented Generation
π― What it does: Propose a hierarchical population distribution tree combined with retrieval and incremental generation, named the HAG framework, to achieve high-fidelity Agent generation that adapts to specific themes.
π― What it does: Constructed the HalluxAudio dataset, which covers three major audio domainsβspeech, environmental sounds, and musicβwith over 5k human-verified QA pairs, to systematically evaluate the hallucination behavior of large-scale audio-language models.
Hallucination Detection in LLMs with Topological Divergence on Attention Graphs
Alexandra Bazarova (Applied AI Institute), Alexey Zaytsev (Applied AI Institute)
CodeAnomaly DetectionExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelTextRetrieval-Augmented Generation
π― What it does: Proposes a training-free, attention map topology-based hallucination detection method called TOHA, which uses the topological divergence of attention maps, MTop-Div G(R,P), to measure the structural differences between responses and prompts, thereby determining hallucinatory outputs of LLMs in RAG scenarios.
π― What it does: This paper systematically analyzes hallucinated citations (HalluCitation) appearing in conference papers from ACL, NAACL, and EMNLP 2024-2025, and provides detection methods and their impact on the peer review process.
Hard2Verify: A Step-Level Verification Benchmark for Open-Ended Frontier Math
Shrey Pandit (Salesforce AI Research), Shafiq Joty (Salesforce AI Research)
CodeLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Proposed the Hard2Verify benchmark, using manually annotated step-by-step verification data to evaluate the step-by-step verifiers of state-of-the-art mathematical reasoning models.
Harmonizing the Past, Present, and Future: A Null-Space Constrained Region-Specific Method for Continual Learning in LLMs
Jinhui Chen (Institute of Automation, Chinese Academy of Sciences), Jun Zhao (Institute of Automation, Chinese Academy of Sciences)
CodeFederated LearningExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmarkRetrieval-Augmented Generation
π― What it does: Propose a zero-replay continual learning framework called PaRSP, which achieves a triple balance of stability, plasticity, and sustainability for large language models by leveraging task-specific parameter region activation and orthogonal projection in the zero space.
Harnessing Negative Signals: Reinforcement Distillation from Teacher Data for LLM Reasoning
Shuyao Xu (INFLY TECH COMPANY LIMITED), Yuan Qi (INFLY TECH COMPANY LIMITED)
CodeKnowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― What it does: Under the offline distillation setup, training LLMs using positive and negative reasoning trajectories to enhance reasoning performance.
HEALing Entropy Collapse: Enhancing Exploration in Few-Shot RLVR via Hybrid-Domain Entropy Dynamics Alignment
Zhanyu Liu (Xiamen University), Jinsong Su (Xiamen University)
CodeData-Centric LearningMeta LearningReinforcement Learning from Human FeedbackReinforcement LearningPrompt EngineeringTextTabularBiomedical DataBenchmarkPhysics RelatedChain-of-Thought
π― What it does: Propose the HEAL framework, combining high-quality general-domain data selection and entropy dynamics alignment (EDA) reward to address the entropy collapse problem in few-shot RLVR;
Heterogeneous Adaptive Policy Optimization: Tailoring Optimization to Every Tokenβs Nature
Zheng Liu (Peking University), Wentao Zhang (Peking University)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
π― What it does: Proposed a Heterogeneous Adaptive Policy Optimization (HAPO) algorithm, which utilizes entropy information throughout the entire RLHF training process to achieve token-level adaptive optimization, including adaptive sampling temperature, token-level advantage estimation, differential advantage redistribution, and asymmetric adaptive clipping.
HiEdit: Lifelong Model Editing with Hierarchical Reinforcement Learning
Yangfan Wang (Harbin Institute of Technology), Jingchi Jiang (Harbin Institute of Technology)
CodeExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelReinforcement LearningPrompt EngineeringMixture of ExpertsText
π― What it does: Propose the HiEdit framework, which utilizes hierarchical reinforcement learning to achieve lifelong editing of LMMs, automatically selecting the most relevant layers and updating parameters;
Hierarchical Policy Optimization for Simultaneous Translation of Unbounded Speech
Siqi Ouyang (Carnegie Mellon University), Lei Li (Carnegie Mellon University)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextAudio
π― What it does: This paper proposes a post-training framework called Hierarchical Policy Optimization (HPO), which uses reinforcement learning to simultaneously optimize translation quality and latency in the task of simultaneous speech translation.
CodeRetrievalRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmark
π― What it does: Propose a training-free hierarchical token prepending method (Hierarchical Token Prepending, HTP), which inserts sentence-level and document-level summary tokens into the input to construct multi-level reversed information flow, and uses mean pooling as the readout function to enhance the embedding quality of the decoder LLM.
HOPE: Hybrid Optimized Parallel Encoding with Supervised and Unsupervised Semantic Fusion for Depression Symptom Detection
Tu-Phuong Mai (VNU University of Engineering and Technology), Hoang-Quynh Le (VNU University of Engineering and Technology)
CodeClassificationRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextElectronic Health RecordsRetrieval-Augmented Generation
π― What it does: Proposed the HOPE framework, using hybrid parallel encoding for multi-label detection of depressive symptoms.
CodeData SynthesisTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextRetrieval-Augmented Generation
π― What it does: Propose the HopWeaver framework, which automatically synthesizes high-quality, multi-hop cross-document questions (bridge-type and comparison-type) from raw corpora and provides an evaluation system;
How Instruction and Reasoning Data shape Post-Training: Data Quality through the Lens of Layer-wise Gradients
Ming Li (University of Maryland), Tianyi Zhou (University of Maryland)
CodeExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Conducts gradient spectrum analysis of large language models trained on instructions and reasoning data of varying quality, investigating the impact of data quality on the dynamics of fine-tuning.
How Memory Management Impacts LLM Agents: An Empirical Study of Experience-Following Behavior
Zidi Xiong (Harvard University), Zhen Xiang (University of Minnesota-Twin Cities)
CodeExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelReinforcement LearningAgentic AITextTabularTime SeriesRetrieval-Augmented Generation
π― What it does: An empirical study on memory management (memory addition and deletion) of large language model (LLM) agents, exploring its impact on long-term performance.
How Tokenization Limits Phonological Knowledge Representation in Language Models and How to Improve Them
Disen Liao (University of Waterloo), Freda Shi (University of Waterloo)
CodeExplainability and InterpretabilityRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningText
π― What it does: Studied the impact of subword tokenization on language models' phonological knowledge (such as prosody, syllable segmentation, etc.), and proposed a lightweight IPA fine-tuning method to enhance the model's phonological reasoning ability.
HyperMem: Hypergraph Memory for Long-Term Conversations
Juwei Yue (Institute of Information Engineering Chinese Academy of Sciences), Yafeng Deng (EverMind AI)
CodeRetrievalFederated LearningSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextGraphSequentialRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Constructed the HyperMem hypergraph hierarchical memory system, which includes dialogue segmentation (episode detection), topic aggregation, fact extraction, and links topics, chapters, and facts through hyperedges to achieve explicit modeling of high-order associations; meanwhile, a coarse-to-fine retrieval process was designed to sequentially filter topics β chapters β facts.
Identifying the Periodicity of Information in Natural Language
Yulin OU, Hendrik Buschmeier (Bielefeld University)
CodeRecognitionText
π― What it does: Proposed the AutoPeriod of Surprisal (APS) algorithm, which directly identifies information periodicity in a corpus at the single-document level;
IEvoAgent: Evolving Conversational Agent based on User Implicit Feedback
Yichen Cai (Zhejiang University), Pei Chen (Zhejiang University)
CodeOptimizationRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextSequentialBenchmarkRetrieval-Augmented Generation
π― What it does: Proposed a two-stage IEvoAgent framework that utilizes user implicit feedback to achieve real-time evolution and multi-turn alignment in dialogue systems
Illusions of the Gold Standard: A Large-scale Analysis of Human Evaluation Protocols for Long-form Text Generation
Katelyn X. Mei (University of Washington), Lucy Lu Wang (University of Washington)
CodeGenerationTransformerLarge Language ModelPrompt EngineeringTextReview/Survey PaperRetrieval-Augmented Generation
π― What it does: Conducted a large-scale analysis of human evaluation protocols for long text generation in *CL conference papers from 2023 to 2025, constructing and applying 20 reportable criteria to statistically analyze and visualize manual and LLM-assisted annotation results.
CodeExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: This study systematically explores the fundamental role of Immediate Inference (II) in large language models (LLMs), constructs the first benchmark for II called IIBench, and demonstrates that II can explain and enhance the performance of various downstream reasoning tasks through zero-shot evaluation, error type analysis, causal mediation models, and training intervention experiments.
ImReasoner: Improving Memory-based Language Models for Reasoning-in-a-Haystack Tasks
Ching-Yun Ko (IBM Research), Pin-Yu Chen (IBM Research)
CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningRecurrent Neural NetworkTransformerLarge Language ModelAuto EncoderContrastive LearningTextRetrieval-Augmented GenerationChain-of-Thought
π― What it does: This paper proposes an improved memory-enhanced language model called ImReasoner, designed to address the 'needle-in-a-haystack' task in long-context reasoning, combining memory updates during inference with a staged training approach;
CodeExplainability and InterpretabilityAdversarial AttackTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
π― What it does: Proposed and verified Doublespeak: an attack method that replaces harmful keywords with harmless synonyms in context, enabling large language models' internal representations to gradually shift from harmless to harmful without altering the surface semantics of the input.
π― What it does: This paper proposes the CohTP framework, which achieves cross-document event coreference resolution by constructing and refining event time graphs, dividing events into time periods, and performing coreference inference within the same time period.
InferenceDynamics: Adaptive LLM Routing through Structured Capability and Knowledge Profiling
Haochen Shi (Hong Kong University of Science and Technology), Yangqiu Song (Hong Kong University of Science and Technology)
CodeRecommendation SystemComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextBenchmark
π― What it does: Propose an expandable LLM routing framework called InferenceDynamics, which intelligently assigns queries based on the model's capabilities and knowledge characteristics;
Inferring Events from Time Series using Language Models
Mingtian Tan (University of Virginia), Thomas Hartvigsen (University of Virginia)
CodeData SynthesisKnowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextTime SeriesBenchmarkChain-of-Thought
π― What it does: Investigated whether large language models can infer natural language events corresponding to time series, and constructed a benchmark dataset and automated task generation method based on sports matches.
Injecting Context via Situation Working Memory for Logical Reasoning with LLMs
Jieun Kim (Yonsei University), Sung-Bae Cho (Yonsei University)
CodeExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
π― What it does: Proposes SituW, a situation working memory based on a situational model, which guides the logical reasoning of large language models by dynamically constructing and updating situational memory.
Inside Out: Evolving User-Centric Core Memory Trees for Long-Term Personalized Dialogue Systems
Jihao Zhao (Renmin University of China), Zhiyu li
CodeRecommendation SystemData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextRetrieval-Augmented Generation
π― What it does: Proposes the Inside Out framework, which manages core memories in long-term personalized dialogues by constructing an evolvable PersonaTree (a hierarchical structure based on the bio-psycho-social model), and uses a lightweight MemListener (reinforcement learning based on process rewards) to convert dialogues into tree operations. Finally, it adopts an adaptive generation strategy (fast mode and agent recall) to improve response quality.
CodeOptimizationRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
π― What it does: Propose a self-supervised reinforcement learning framework that does not require external supervision, specifically aimed at improving the performance of large language models in multi-constraint instruction following tasks.
Interleaved Tool-Call Reasoning for Protein Function Understanding
Chuanliu Fan (Soochow University), Guohong Fu (Soochow University)
CodeExplainability and InterpretabilityDrug DiscoveryProtein Structure PredictionTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextBiomedical DataBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Proposed a tool-enhanced protein function reasoning agent, PFUA, which generates verifiable intermediate evidence by alternately invoking domain-specific tools (such as MMseqs2, Pfam, TMbed, etc.) during the reasoning process, thereby enabling answers to multi-dimensional questions about protein function, catalytic activity, domains, etc.
Into the Gray Zone: Domain Contexts Can Blur LLM Safety Boundaries
Ki Sen Hung (Hong Kong University of Science and Technology), Yangqiu Song (Hong Kong University of Science and Technology)
CodeSafty and PrivacyAdversarial AttackTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Studied how domain-specific contexts can lead large language models (LLMs) to exhibit gray areas on safety boundaries, proposing and evaluating the JARGON framework: leveraging security research contexts and multi-turn dialogues to generate 'academic-style' attacks, significantly improving the success rate of breaking through safety defenses.
Inverting the Shield: Systematically Generating Safety Tests from Policy Specifications
Xiaoyue Lu (Shenzhen Campus of Sun Yat-sen University), Jin Song Dong (National University of Singapore)
CodeSafty and PrivacyExplainability and InterpretabilityAdversarial AttackTransformerLarge Language ModelPrompt EngineeringGenerative Adversarial NetworkTextRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Propose the POLARIS framework, which utilizes the formalization of security policies to achieve safety testing of LLMs, generating traceable and comprehensive attack queries;
Investigating More Explainable and Partition-Free Compositionality Estimation for LLMs: A Rule-Generation Perspective
Ziyao Xu (Peking University), Houfeng Wang (Peking University)
CodeExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
π― What it does: This paper proposes a new perspective based on rule generation for evaluating the compositional ability of large language models (LLMs), and explicitly reveals the model's understanding of compositional properties through generated programs.
Is a Document Educational or Just Wikipedia-Style? β Pitfalls of Classifier-Based Quality Filtering
Mateusz Klimaszewski (Warsaw University of Technology), Piotr Andruszkiewicz (Warsaw University of Technology)
CodeClassificationFederated LearningSafty and PrivacyExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
π― What it does: By rewriting web content in a Wikipedia-style and using the Classifier-based Quality Filtering (CQF) model to score the text before and after rewriting, the study reveals the security vulnerabilities and biases of CQF when filtering pre-trained corpora.
Is Human-Like Text Liked by Humans? Multilingual Human Detection and Preference Against AI
Yuxia Wang (MBZUAI), Preslav Nakov (MBZUAI)
CodeClassificationRecognitionExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextMultimodalityBenchmarkRetrieval-Augmented Generation
π― What it does: This paper conducts large-scale human detection experiments across languages and domains, evaluating experts' accuracy in identifying texts generated by the latest large language models versus human texts, and exploring how prompting strategies narrow the gap and influence human preferences.
IUQ: Interrogative Uncertainty Quantification for Long-Form Large Language Model Generation
Haozhi Fan (University of Pennsylvania), Kaidi Xu (City University of Hong Kong)
CodeGenerationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Proposed a framework called IUQ based on question-answering probing for uncertainty quantification in long-text generation, used to evaluate the reliability of LLMs in long-text generation.
Joint Knowledge Base Completion and Question Answering by Combining Large Language Models and Small Language Models
Yinan Liu (Northeastern University), Bin Wang (Northeastern University)
CodeRecommendation SystemExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningAgentic AITextGraphBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Proposed the JCQL framework, which jointly completes knowledge graph completion (KBC) and knowledge-based question answering (KBQA), improving the performance of both tasks through iterative interaction between LLM and SLM.
JurisBench: A Deep Benchmark for Assessing Large Language Models in Professional Legal Practice
Ziang Chen (BIGAI), Bin Ling (Peking University)
CodeTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Constructed JurIBench, a legal evaluation benchmark focused on the entire process of civil litigation in China, with a vertical depth of analysis.
Know Your Place: Diagnosing Implicit Social Adaptation Failures in Chinese Large Language Models
Yu Tian (Inner Mongolia University), Xiangdong Su (Inner Mongolia University)
CodeExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmarkRetrieval-Augmented Generation
π― What it does: Diagnose and quantify the failure of large Chinese language models in adapting to implicit social cues during natural interactions, introduce the concept of Social Agnosia, and construct the C-ISA evaluation framework.
Knowing When Not to Answer: Lightweight KB-Aligned OOD Detection for Safe RAG
Ilias Triantafyllopoulos, JoΓ£o Sedoc (New York University)
CodeAnomaly DetectionExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextTabularSequentialBenchmarkRetrieval-Augmented Generation
π― What it does: This paper proposes a lightweight, knowledge-base-based out-of-distribution (OOD) detection method. By performing PCA on KB document embeddings, queries are projected into a low-dimensional subspace, and discrimination is conducted using geometric rules or lightweight classifiers, thereby achieving a 'when not to answer' gating mechanism in retrieval-augmented generation (RAG) systems.
Knowledge Poisoning Attacks on Medical Multi-Modal Retrieval-Augmented Generation
Peiru Yang (Tsinghua University), Tao Qi (Beijing University of Posts and Telecommunications)
CodeRetrievalAdversarial AttackData-Centric LearningTransformerLarge Language ModelPrompt EngineeringImageTextMultimodalityBiomedical DataElectronic Health RecordsRetrieval-Augmented Generation
π― What it does: Proposes a knowledge poisoning attack framework called M3Att for medical multimodal retrieval-augmented generation systems, assuming that the attacker only knows the database distribution information and does not need to query specific knowledge.
Knowledge-driven Augmentation and Retrieval for Integrative Temporal Adaptation
Weisi Liu (University of Memphis), Xiaolei Huang (University of Memphis)
CodeClassificationDomain AdaptationTransformerLarge Language ModelSupervised Fine-TuningTextBiomedical DataElectronic Health RecordsReview/Survey PaperRetrieval-Augmented Generation
π― What it does: To address text classification tasks with temporal drift, the KARITA framework is proposed, combining multi-dimensional drift detection, source data retrieval, and knowledge-driven augmentation to achieve adaptive temporal evolution.
CodeExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelReinforcement LearningTextRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Propose the KnowRL framework, which combines reinforcement learning (RL) with knowledge verification to provide process-level supervision during chain-of-thought reasoning, significantly reducing hallucinations.
Label Effects: Shared Heuristic Reliance in Trust Assessment by Humans and LLM-as-a-Judge
Xin Sun (National Institute of Informatics), Saku Sugawara (National Institute of Informatics)
CodeFederated LearningExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextBiomedical DataReview/Survey PaperRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Through a controlled experiment comparing source labels (Human vs AI) in health-related Q&A content, this study investigates the dependence of humans and large language models (LLM-as-a-Judge) on labels during trust evaluation, and explores the mechanisms behind label effects by combining eye-tracking with LLM internal attention/entropy analysis.
π― What it does: Designed and validated the post-hoc sparse autoencoder LANGSAE editing, which removes language identity signals from multilingual retrieval embeddings to improve cross-lingual retrieval performance.
Language Acquisition Device in Large Language Models
Masato Mita (University of Tokyo), Yohei Oseki (University of Tokyo)
CodeRepresentation LearningData-Centric LearningTransformerLarge Language ModelPrompt EngineeringText
π― What it does: Proposed the MP-STRUCT pre-pretraining framework based on the LAD (Language Acquisition Device) concept, using synthetic sequences to induce large language models to acquire structured language biases before natural language pretraining.
Language Models Learn Universal Representations of Numbers and Hereβs Why You Should Care
Michal Ε tefΓ‘nik (R&D Centre for Large Language Models, National Institute of Informatics), Pontus Stenetorp (R&D Centre for Large Language Models, National Institute of Informatics)
CodeExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextTabularBenchmark
π― What it does: This study investigates how large language models (LLMs) encode numbers, quantifies the prevalence of sinusoidal structures in different models, layers, and natural language contexts, and develops a new numerical probe based on sinusoidal parameterization (param-sin), further exploring the representation of multi-tag numbers, output tracking, and ordinal data.
Language on Demand, Knowledge at Core: Composing LLMs with Encoder-Decoder Translation Models for Extensible Multilinguality
Mengyu Bu (Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences), Yang Feng (Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences)
CodeComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningTextMultimodality
π― What it does: Built a framework called XBridge, which combines a pre-trained multilingual encoder-decoder NMT model with a large language model (LLM), utilizing the LLM as the core knowledge processing unit for English, while the external NMT handles multilingual understanding and generation;
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)
CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextBenchmarkRetrieval-Augmented Generation
π― 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.
Learning from Contrasts: Synthesizing Reasoning Paths from Diverse Search Trajectories
Peiyang Liu (Peking University), Wei Ye (Peking University)
CodeData SynthesisExplainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningMixture of ExpertsContrastive LearningTextChain-of-Thought
π― 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 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)
CodeDomain AdaptationExplainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
π― 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 the Irrecoverable: Error-Localized Policy Optimization for Tool-Integrated LLM Reasoning
Qiao Liang (Tongji University), Sheng Guo (MYbank, Ant Group)
CodeOptimizationTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringTextChain-of-Thought
π― 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)
CodeRecommendation SystemOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringContrastive LearningTextBenchmarkRetrieval-Augmented Generation
π― 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 More from Less: Exploiting Counterfactuals for Data-Efficient Chart Understanding
Jianzhu Bao (Nanyang Technological University), Wenya Wang (Nanyang Technological University)
CodeOptimizationComputational EfficiencyData-Centric LearningTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
π― 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)
CodeSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextBenchmarkFinance RelatedRetrieval-Augmented Generation
π― 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).
LeCoDe: A Benchmark Dataset for Interactive Legal Consultation Dialogue Evaluation
Weikang Yuan (Zhejiang University), Xiaozhong Liu (Worcester Polytechnic Institute)
CodeRecommendation SystemFederated LearningExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextSequentialBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― 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.
Xunjian Yin (Peking University), Xiaojun Wan (Peking University)
CodeGenerationData SynthesisExplainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelContrastive LearningText
π― 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.
π― 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.
LightReasoner: Can Small Language Models Teach Large Language Models Reasoning?
Jingyuan Wang (University of Hong Kong), Chao Huang (University of Hong Kong)
CodeComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningTextChain-of-Thought
π― 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)
CodeComputational EfficiencyRepresentation LearningTransformerSupervised Fine-TuningReinforcement LearningMixture of ExpertsVision Language ModelImageTextMultimodality
π― 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.
Lingua-Graph: A Unified Representation of Cross-Task Common Substructures for Analytic Language Processing
Mingming Sun (BIMSA), Yunfeng Cai (BIMSA)
CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextGraph
π― 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.
LiveFact: A Dynamic, Time-Aware Benchmark for LLM-Driven Fake News Detection
Cheng Xu (University College Dublin), Tahar Kechadi
CodeTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― 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.
Alessandro Corona Mendozza (University of Copenhagen), Anders SΓΈgaard (University of Copenhagen)
CodeExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
π― 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 Safety From Within: Detecting Harmful Content with Internal Representations
Difan Jiao (University of Toronto), Ashton Anderson (University of Toronto)
CodeSafty and PrivacyComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
π― 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.
LOGICAL-COMMONSENSEQA: A Benchmark for Logical Commonsense Reasoning
Obed Junias (University of Colorado Boulder), Maria Leonor Pacheco (University of Colorado Boulder)
CodeTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
π― 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.
Long Context Modeling with Ranked Memory-Augmented Retrieval
Ghadir Alselwi (University of New South Wales), Imran Razzak (Mohamed Bin Zayed University of Artificial Intelligence)
CodeRetrievalComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextRetrieval-Augmented Generation
π― 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)
CodeComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
π― 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.
Look Within or Beyond? A Theoretical Comparison Between Parameter-Efficient and Full Fine-Tuning
YongKang Liu, Hinrich Schuetze
CodeFederated LearningExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextReview/Survey PaperBenchmark
π― What it does: Studied and compared the theoretical and empirical performance of parameter-efficient fine-tuning (PEFT) and full-parameter fine-tuning (FFT), focusing on representation capability, robustness, and the impact on data/parameter gains.
CodeAutonomous DrivingOptimizationData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringDiffusion modelTextBenchmarkRetrieval-Augmented Generation
π― What it does: LoopTool proposes a closed-loop adaptive data generation and model training framework to enhance the tool calling capabilities of large language models.
LoRA on the Go: Instance-level Dynamic LoRA Selection and Merging
Seungeon Lee (MPI-SWS), Krishna P. Gummadi (MPI-SWS)
CodeComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextMultimodalityBenchmark
π― What it does: Propose the LOGO framework, which dynamically selects and fuses LoRA adapters for each instance during inference without requiring training.
CodeTransformerLarge Language ModelTextMultimodalityReview/Survey Paper
π― What it does: This paper, through a review of the training mechanisms of multilingual large language models (LLMs), model self-destruction (model collapse), and related research in linguistics, computer vision, and machine translation, proposes the role of LLMs in 'natural selection' and discusses their potential threat to the dilution of language tails (rare words, low-probability structures) and the loss of linguistic diversity. It also calls for considering linguistic diversity and expressive richness in the evaluation and training of LLMs.
Lost in the Mix: Evaluating LLM Understanding of Code-Switched Text
Amr Mohamed (MBZUAI), Guokan Shang (MBZUAI)
CodeExplainability and InterpretabilityRepresentation LearningAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextMultimodality
π― What it does: Systematically evaluate the understanding ability of large language models (LLMs) on code-switching (CSW) texts, construct a controllable code-switching test set based on linguistic theories, and compare the performance of different models and switching methods.
Lost in Translation, and Found: Detecting and Interpreting Translation Effects
Shira Wein (University of South Florida), Maria Leonor Pacheco (Proof School)
CodeClassificationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper performs binary classification to determine whether English text is translated text, constructing a high-precision model for detecting translated discourse.
LVLMs and Humans Ground Differently in Referential Communication
Peter Zeng (Stony Brook University), Owen Rambow (Stony Brook University)
CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodality
π― What it does: In an experiment, researchers had humans engage in referential dialogues with large vision-language models (LVLMs) and among humans to complete multi-round object matching tasks, investigating the establishment of a common ground.