ACL 2026 Papers — Page 3
Annual Meeting of the Association for Computational Linguistics · 2296 papers
Beyond Detection: Evaluating Fallacy Awareness of LLMs in Interactive Scenarios
Conghui Niu (Beijing Language and Culture University), Pengyuan Liu (Beijing Language and Culture University)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the concept of 'fallacy awareness' and constructs a large-scale Chinese interactive benchmark, ISFallacy, to evaluate the ability of LLMs to identify and resist fallacies in dynamic situations without explicit prompts.
Beyond End-to-End: Dynamic Chain Optimization for Private LLM Adaptation on the Edge
Yebo Wu (University of Macau), Li Li (University of Macau)
OptimizationFederated LearningComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose ChainFED, a federated fine-tuning method that overcomes the memory bottleneck of LLMs on edge devices through chained layer-wise freezing and training
Beyond Examples: Towards Automated Thought-level In-Context Reasoning for Large Language Models
Jinyang Wu (Tsinghua University), Jianhua Tao (Tsinghua University)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes ThoughtICR, an automated thought-level context reasoning framework that first generates high-level thinking paths using Monte Carlo Tree Search (MCTS) on a small number of seed samples, then clusters them into thinking cards. During reasoning, it dynamically matches thinking cards based on problem attributes and guides LLMs to generate answers through a verification strategy.
Beyond Experience Retrieval: Learning to Generate Utility-Optimized Structured Experience for Frozen LLMs
Xuancheng Li (Tsinghua University), Qingyao Ai (Tsinghua University)
OptimizationComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: A SEAM module is designed for frozen LLM agents, capable of generating structured experiential prompts based on instances in a single forward inference, helping the agent perform reasoning tasks more efficiently and accurately.
Beyond Explicit Refusals: Soft-Failure Attacks on Retrieval-Augmented Generation
Wentao Zhang (University of Electronic Science and Technology of China), Fuji Ren (University of Electronic Science and Technology of China)
GenerationRetrievalOptimizationAdversarial AttackData-Centric LearningTransformerLarge Language ModelPrompt EngineeringDiffusion modelContrastive LearningTextBenchmarkFinance RelatedRetrieval-Augmented Generation
🎯 What it does: To address the threat of soft failure in retrieval-augmented generation (RAG) systems, we propose a black-box evolutionary deceptive interference attack (DEJA), which can silently reduce system usability by automatically generating and injecting a single forged document, causing RAG to produce fluent but information-poor responses after retrieving the document.
Beyond Fully Random Masking: Attention-Guided Denoising and Optimization for Diffusion Language Models
Jia Deng (Renmin University of China), Ji-Rong Wen (Meituan)
OptimizationAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningDiffusion modelText
🎯 What it does: This paper proposes an attention-based denoising and optimization framework, AGDO, for the post-training tasks of diffusion large language models (dLLMs). By analyzing the internal attention structure of the model, it constructs a denoising order that conforms to natural dependency relationships, and on this basis performs supervised fine-tuning and reinforcement learning, significantly enhancing the model's reasoning and code generation capabilities.
Beyond Hard Masks: Progressive Token Evolution for Diffusion Language Models
Linhao Zhong (Zhejiang University), Chunhua Shen (Zhejiang University)
GenerationComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningDiffusion modelText
🎯 What it does: Propose EvoToken-DLM, which replaces traditional hard binary masks with soft tokens that evolve over time, enabling parallel decoding with progressive refinement;
Beyond Itinerary Planning—A Real-World Benchmark for Multi-Turn and Tool-Using Travel Tasks
Xiang Cheng (Renmin University of China), Yong Liu (Alibaba Group)
Large Language ModelAgentic AITextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the TravelBench benchmark for multi-round travel planning evaluation using real user queries, implicit preferences, and tools.
Beyond Literal Mapping: Benchmarking and Improving Non-Literal Translation Evaluation
Yanzhi Tian (Beijing Institute of Technology), Yuhang Guo (Beijing Institute of Technology)
TransformerLarge Language ModelAgentic AITextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed the non-literal translation evaluation dataset MENT and proposed the reflective proxy framework RATE to improve the reliability of translation quality assessment.
Beyond Majority Voting: Towards Fine-grained and More Reliable Reward Signal for Test-Time Reinforcement Learning
Weiqin Wang, Hui Huang (Fuzhou University)
Reinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: SCOPE proposes a pseudo-label generation and reward mechanism that combines step-level confidence with dynamic subgroup partitioning in reinforcement learning during unsupervised testing.
Beyond Markovian Forgetfulness: Episodic Memory for Reasoning-Intensive Retrieval
Dohyeon Lee (Korea Advanced Institute of Science and Technology), Seung-won Hwang (Seoul National University)
RetrievalTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes a retrieval framework called EMR based on episodic memory, aimed at solving the issue of cyclic redundancy that occurs during multi-step reasoning retrieval.
Beyond Meta-Reasoning: Metacognitive Consolidation for Self-Improving LLM Reasoning
Ziqing Zhuang (Southeast University), Yulan He (King's College London)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationMeta LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed and implemented the Metacognitive Consolidation (MC2) framework, enabling large language models to accumulate metacognitive experiences across different instances and convert them into reusable meta-knowledge, thereby self-improving in subsequent reasoning.
Beyond Monolingual Assumptions: A Survey on Code-Switched NLP in the Era of Large Language Models across Modalities
Rajvee Sheth (IIT Gandhinagar), Mayank Singh (IIT Gandhinagar)
ClassificationRecognitionRecommendation SystemFederated LearningExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningData-Centric LearningMeta LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringMixture of ExpertsVision Language ModelContrastive LearningTextMultimodalityReview/Survey PaperBenchmarkRetrieval-Augmented Generation
🎯 What it does: Systematic review of code-switching NLP, covering 327 studies in the era of LLMs, 15 NLP tasks, 30+ datasets, and 80+ languages.
Beyond Noise: Characterizing Creative Potential in Unverifiable LLM Hallucinations
Yu Yan (Beijing University of Posts and Telecommunications), Zheng Hu (Beijing University of Posts and Telecommunications)
GenerationExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextMultimodalityPhysics RelatedRetrieval-Augmented Generation
🎯 What it does: Studied the creative potential in LLM-generated unverifiable hallucinations, proposing a binary classification between creative synthesis (Region A) and groundless fabrication (Region B).
Beyond Pedagogical Principles: Multi-Horizon Preference Optimization for Efficient Socratic Tutoring
Xin Shi (Zhejiang University), Yawei Luo (Zhejiang University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringDiffusion modelTextRetrieval-Augmented Generation
🎯 What it does: Proposes a multi-perspective preference optimization (MHPO) framework, which iteratively optimizes the LLM teacher agent in a dynamic teacher-student simulation environment using multi-perspective reward functions (instantaneous and trajectory-level).
Beyond Query Memorization: Large Language Model Routing with Query Decomposition and Historical Matching
Bo Lv (Tencent Hunyuan), Yue Yu (Peng Cheng Laboratory)
Computational EfficiencyData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose a LLM routing framework called DecoR based on historical log matching, avoiding query memory traps;
Beyond Ranking: Fine-Grained Diagnostics and Self-Improvement for MLLMs
Mingze Xu (Tsinghua University), Yuxing Han (Tsinghua University)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the ATOM framework for fine-grained process diagnosis of multi-modal large language models (MLLMs), and achieve model self-improvement through closed-loop self-correction.
Beyond Scaling: Measuring and Predicting the Upper Bound of Knowledge Retention in Language Model Pre-Training
Changhao Jiang (Fudan University), Xuanjing Huang (Fudan University)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Proposed an information theory-based knowledge retention metric called SMI to predict the accuracy of large language models in closed-book question answering before pre-training.
Beyond Self-Report: Bridging the Intention-Behavior Gap in Critical Thinking Assessment via Interpretable Multi-Agent System
Zekun Li (Tsinghua University), Yu Zhang (Tsinghua University)
Explainability and InterpretabilityTransformerLarge Language ModelAgentic AITextChain-of-Thought
🎯 What it does: Propose the MASA framework, transforming static self-assessment questionnaires into multi-agent contextualized dialogue evaluations, bridging the intention-behavior gap.
Beyond Single Representations: Multi-Model Embedding Fusion for Stable Text Classification
Jiho Gwak (Kumoh National Institute of Technology), Yuchul Jung (Kumoh National Institute of Technology)
ClassificationTransformerLarge Language ModelMixture of ExpertsContrastive LearningText
🎯 What it does: This paper systematically evaluates the effects of single-layer embeddings and multi-model embedding fusion in text classification, exploring the impact of layer selection, pooling methods, and fusion strategies on performance.
Beyond Single View: A Comprehensive Benchmark for Medical Multimodal Large Language Models on Multi-Image Understanding
Dexuan Xu (Peking University), Yu Huang (Peking University)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringImageTextMultimodalityBiomedical DataBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed MedMultiBench, a medical multi-image understanding benchmark comprising 11,392 expert-verified samples, to evaluate the cross-image reasoning capabilities of multi-modal large language models (MLLMs).
Beyond Single-shot Writing: Deep Research Agents are Unreliable at Multi-turn Report Revision
Bingsen Chen (New York University), Chen Zhao (New York University)
TransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose a multi-round report revision (MR DRE) evaluation framework to investigate the reliability of deep research agents (DRA) in multi-round revisions after receiving user feedback.
Beyond Single-View Detection: A Dual-Space Reasoning Framework for Interpretable Harmful Meme Understanding
Wenqing Hou (National University of Defense Technology), Bin Zhou (National University of Defense Technology)
ClassificationExplainability and InterpretabilityTransformerPrompt EngineeringMixture of ExpertsVision Language ModelImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose the BPDMoE-Hate framework, which generates adversarial dual perspectives through a VLM, selects perspectives using Adaptive Viewpoint Gating, and performs hierarchical fusion via Dual-Space Mixture-of-Experts to achieve explainable harmful meme detection.
Beyond Static Alignment: Adaptive Arbitration for Semantic Incongruence in Semi-Supervised Multimodal Sentiment Analysis
Huicong Li (Inner Mongolia University), Wei Wu (Inner Mongolia University)
ClassificationDomain AdaptationKnowledge DistillationRepresentation LearningConvolutional Neural NetworkTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose the Adaptive Arbitration for Semantic Incongruence (A2SI) framework to address the issue of semantic inconsistency in semi-supervised multi-modal sentiment analysis.
Beyond Static Artifacts: An Evolutionary Framework for Synthetic Claim Generation
Yeqing Teng (Qilu University of Technology (Shandong Academy of Sciences)), Xiaoming Wu (Qilu University of Technology (Shandong Academy of Sciences))
GenerationData SynthesisTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes a synthetic statement generation framework based on social evolution, transforming statements from static text into traceable social dissemination processes.
Beyond Static Benchmarks: Synthesizing Harmful Content via Persona-based Simulation for Robust Evaluation
Huije Lee (Korea Advanced Institute of Science and Technology), Jong C. Park (Korea Advanced Institute of Science and Technology)
Data SynthesisSafty and PrivacyTransformerLarge Language ModelAgentic AIPrompt EngineeringText
🎯 What it does: Built and used an LLM agent based on the character dimension to generate context-aware and diverse harmful content;
Beyond Static Persona Consistency: Dynamic Persona Coherence in LLM Role-Playing
Yirui QI, Zhenyu Guan (Beihang University)
GenerationExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes a dynamic personality consistency framework that separates the identity stability layer from the adaptation layer, and realizes emotional evolution across time through the L/M/S psychological state model.
Beyond Surface Features: Advancing Medical Vision-Language Alignment via Dynamic Evidence-Guided Preference Optimization
Zixuan Huang (University of Chinese Academy of Sciences), Yefeng Zheng (Westlake University)
OptimizationExplainability and InterpretabilityData-Centric LearningReinforcement Learning from Human FeedbackTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityBiomedical DataBenchmark
🎯 What it does: Proposed a dynamic evidence-guided preference optimization framework, DEPO, to enhance image-text alignment in medical vision-language models, significantly reducing hallucinations.
Beyond Surface-Level Detection: Towards Cognitive-Driven Defense Against Jailbreak Attacks via Meta-Operations Reasoning
Rui Pu (Beijing University of Posts and Telecommunications), Xi Zhang (Beijing University of Posts and Telecommunications)
Safty and PrivacyExplainability and InterpretabilityAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: Propose a cognition-driven defense framework, CDD, which identifies and counters jailbreak attacks through meta-operations and structured reasoning chains.
Beyond the Context Window: Scaling Agentic RL via End-to-end Optimized Context Compression
Miao Lu (ByteDance Seed), Jiecao Chen (ByteDance Seed)
OptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextSequential
🎯 What it does: Fine-tune large language models (LLMs) with reinforcement learning in long-sequence, multi-round tool usage scenarios, and introduce a context management mechanism based on LLM-generated summaries to compress historical records, overcoming the limitations of fixed context window sizes.
Beyond the Crowd: LLM-Augmented Community Notes for Governing Health Misinformation
Jiaying Wu (National University of Singapore), Min-Yen Kan (National University of Singapore)
Recommendation SystemAnomaly DetectionExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextBiomedical DataBenchmarkRetrieval-Augmented Generation
🎯 What it does: Leverage large language models to enhance the generation and evaluation of community notes, alleviating the delay and scarcity of health information misguidance;
Beyond the Final Actor: Modeling the Dual Roles of Creator and Editor for Fine-Grained LLM-Generated Text Detection
Yang Li (Institute of Computing Technology, Chinese Academy of Sciences), Juan Cao (Institute of Computing Technology, Chinese Academy of Sciences)
ClassificationGraph Neural NetworkTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Propose a four-class LLM text detection method based on RACE
Beyond the Last Frame: Process-aware Evaluation for Generative Video Reasoning
Yifan Li (Renmin University of China), Minghui Qiu (ByteDance)
GenerationData SynthesisPrompt EngineeringVision Language ModelDiffusion modelScore-based ModelImageVideoTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes a process-based evaluation paradigm, constructing the VIPER benchmark and the POC@r metric to assess the consistency between the process and results of generative video reasoning models.
Beyond the Leaderboard: Rethinking Medical Benchmarks for Large Language Models
Wenxuan Wang (Renmin University of China), Linlin Shen (Shenzhen University)
Large Language ModelTextElectronic Health RecordsReview/Survey PaperBenchmark
🎯 What it does: Proposed and implemented MedCheck — a lifecycle-based framework containing 46 medical-specific evaluation criteria, used to systematically assess 56 medical LLM benchmarks; through this framework, systematic issues such as benchmark design flaws, data integrity risks, and insufficient safety evaluation were revealed.
Beyond the Panorama: Training-Free Hierarchical Perception-Reasoning for Fine-Grained Vision in MLLMs
Xiaoyang Yi (Nankai University), Jian Zhang (Nankai University)
RecognitionExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityChain-of-Thought
🎯 What it does: Proposes a training-agnostic hierarchical perception-reasoning framework called HiPerson, enhancing the performance of multi-modal large language models on fine-grained visual tasks.
Beyond Timestamps: Bridging Forward and Backward Reasoning in Temporal Numerical and Relational Understanding
Xinying Qian (Nankai University), Jeff Z. Pan (University of Edinburgh)
TransformerLarge Language ModelPrompt EngineeringTextGraphTime SeriesBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the TNR benchmark, using a bidirectional evaluation framework to simultaneously assess temporal numerical and relational reasoning.
Beyond Transcripts: A Renewed Perspective on Audio Chaptering
Fabian Retkowski (Karlsruhe Institute of Technology), Alexander Waibel (Karlsruhe Institute of Technology)
SegmentationTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextMultimodalityAudio
🎯 What it does: This paper systematically studies the task of audio chapter segmentation, comparing text-driven models, audio-only models, and multimodal models, and proposes an audio-only model called AudioSeg as well as a temporal-spatial evaluation protocol.
Beyond Value Benchmarks: Measuring Value-Structure Alignment in Large Language Models via Symmetric Q-Sorts
Jingting Zheng (Tianjin University), Deyi Xiong (Tianjin University)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmark
🎯 What it does: Designed and implemented a symmetric Q-sort evaluation framework, forcing both humans and large language models to perform a forced distribution ranking on the same 140 moral value statements, generating comparable geometric value structures;
Beyond Variance: Knowledge-Aware LLM Compression via Fisher-Aligned Subspace Diagnostics
Ibne Farabi Shihab (Iowa State University), Anuj Sharma (Iowa State University)
CompressionComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: Studied post-training activation compression, and proposed a knowledge-aware subspace compression framework called FASC based on the Fisher information matrix.
Beyond Word Boundaries: A Hebrew Coreference Benchmark and an Evaluation Protocol for Morphologically Complex Text
Refael Shaked Greenfeld (Bar-Ilan University), Reut Tsarfaty (Bar-Ilan University)
Data-Centric LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmarkChain-of-Thought
🎯 What it does: Constructed the Hebrew coreference benchmark KibutzR, annotated multiple mentions from the word level to the subword level, and proposed a segmentation-aware evaluation protocol addressing differences in tokenization boundaries.
BhashaSutra: A Task-Centric Unified Survey of Indian NLP Datasets, Corpora, and Resources
Raghvendra Kumar (Indian Institute of Technology Patna), Sriparna Saha (Indian Institute of Technology Patna)
Data-Centric LearningLarge Language ModelContrastive LearningTextMultimodalityReview/Survey PaperBenchmark
🎯 What it does: This work systematically organizes and unifies NLP resources for Indian languages, constructing a cross-task, cross-modal resource directory;
Bias Fitting to Mitigate Length Bias of Reward Model in RLHF
Kangwen Zhao (University of Science and Technology of China), Houqiang Li (University of Science and Technology of China)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: A FiMi-RM framework is constructed to eliminate length bias in RLHF by training a reward model and fitting its length bias, thereby improving alignment with human preferences.
Bidirectional LMs are Better Knowledge Memorizers? A Benchmark for Real-world Knowledge Injection
Yuwei Zhang (University Of California San Diego), Jingbo Shang (University Of California San Diego)
Knowledge DistillationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed a real-time, scalable knowledge injection benchmark called WIKIDYK based on Wikipedia's 'Did You Know...' pages, and evaluated the knowledge memory and reasoning capabilities of models through multi-dimensional question-answering assessments;
Big AI is Accelerating the Metacrisis: What Can We Do?
Steven Bird (Charles Darwin University)
Federated LearningSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningAdversarial AttackHyperparameter SearchData-Centric LearningReinforcement Learning from Human FeedbackNeural Architecture SearchLarge Language ModelTextReview/Survey Paper
🎯 What it does: The paper discusses the 'meta-crisis' of the large-scale artificial intelligence acceleration ecosystem, the significance crisis, and the language crisis, and calls on the computational linguistics community to take action to limit corporate influence, focus on public interest, and reconstruct research paradigms.
BiMind: A Dual-Head Reasoning Model with Attention-Geometry Adapter for Incorrect Information Detection
Zhongxing Zhang (University of Minnesota, Twin Cities), Jaideep Srivastava (University of Minnesota, Twin Cities)
Anomaly DetectionExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerPrompt EngineeringContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Propose the BiMind dual-head model, combining the attention geometry adapter, self-retrieval knowledge module, and uncertainty fusion to detect misinformation in social media.
BiMol-Diff: A Unified Diffusion Framework for Molecular Generation and Captioning
Aditya Hemant Shahane (Indian Institute of Technology Delhi), Sandeep Kumar (Indian Institute of Technology Delhi)
Drug DiscoveryTransformerDiffusion modelAuto EncoderGenerative Adversarial NetworkTextBiomedical Data
🎯 What it does: Propose BiMol‑Diff, a unified diffusion framework for molecular generation (text→SMILES) and molecular description (SMILES→text), and introduce a token‑aware noise schedule.
BioHiCL: Hierarchical Multi-Label Contrastive Learning for Biomedical Retrieval with MeSH Labels
Mengfei Lan (University of Illinois Urbana-Champaign), Halil Kilicoglu (University of Illinois Urbana-Champaign)
RetrievalRepresentation LearningTransformerSupervised Fine-TuningContrastive LearningTextBiomedical Data
🎯 What it does: Use MeSH hierarchical labels to perform multi-label contrastive learning on document pairs, training a dense retrieval model
Biomedical Question Answering via Multi-Level Summarization on a Local Knowledge Graph
Lingxiao Guan (University of Michigan), Jie Liu (University of Michigan)
Drug DiscoveryGraph Neural NetworkTransformerLarge Language ModelTextBiomedical DataRetrieval-Augmented Generation
🎯 What it does: Through the Retrieval-Augmented Generation (RAG) framework, propositional claims are extracted from retrieved documents to construct a local knowledge graph, followed by hierarchical graph summarization, enabling a small LLM to perform biomedical question answering with high-quality context.
BioProAgent: Neuro-Symbolic Grounding for Constrained Scientific Planning
Yuyang Liu (Peking University), Yonghong Tian (Peking University)
Autonomous DrivingOptimizationRobotic IntelligenceDrug DiscoveryRecurrent Neural NetworkTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextSequentialBiomedical DataBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose BioProAgent, a neuro-symbolic framework that combines probabilistic LLM reasoning with deterministic finite state machines (FSM) for safely performing experimental operations in irreversible laboratory environments, incorporating a design-validate-correct (DVR) workflow and semantic symbolic induction.
BioTool: A Comprehensive Tool-Calling Dataset for Enhancing Biomedical Capabilities of Large Language Models
Xin Gao (UC San Diego), Pengtao Xie (UC San Diego)
Drug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBiomedical DataBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed and publicly released the BIOTOOL dataset to train large language models (LLMs) for precise biomedical tool calling and to verify their effectiveness in answering medical-related questions.
Bit-by-Bit: Progressive QAT Strategy with Outlier Channel Splitting for Stable Low-Bit LLMs
Binxing Xu (Zhejiang University), Yike Guo (Hong Kong University of Science and Technology)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose a progressive quantization-aware training (QAT) framework called BIT-BY-BIT, which combines progressive bit-width reduction, one-time multi-precision training, aligned channel splitting, and micro-scale quantization, and implements custom high-performance GPU kernels for 2-bit operations;
Black-Box Membership Inference Attacks for Video Training Data in Multimodal Large Language Models
Jinrui Wang (Beijing University of Posts and Telecommunications), Tao Qi (Beijing University of Posts and Telecommunications)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelPrompt EngineeringVideoTextMultimodalityBenchmark
🎯 What it does: Proposed a black-box video member inference attack framework called VideoMIA, which utilizes the temporal dependencies between video frames to determine whether a video was used as training data
BlindGuard: Safeguarding LLM-based Multi-Agent Systems under Unknown Attacks
Rui Miao (Jilin University), Xin Wang (Jilin University)
Anomaly DetectionSafty and PrivacyGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextGraph
🎯 What it does: Designed and implemented an unsupervised defense framework called BlindGuard to detect and isolate malicious agents in LLM-driven multi-agent systems without attack labels.
Bloom-Eval: A Hierarchical Evaluation Benchmark for Automatic Survey Generation Based on Bloom’s Taxonomy
Fei Zhang (City University of Hong Kong), Ye Wei (City University of Hong Kong)
TransformerLarge Language ModelPrompt EngineeringTextReview/Survey PaperBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes a hierarchical evaluation benchmark called Bloom-Eval based on Bloom's Taxonomy for systematic assessment of Automatic Survey Generation (ASG) systems.
Bootstrapping Code Translation with Weighted Multilanguage Exploration
Yuhan Wu (Nanjing University), Wei Hu (Nanjing University)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextSequential
🎯 What it does: Propose BootTrans, a reinforcement learning training method for multilingual code translation that utilizes transferable unit tests as verification oracles, achieving this through a dual-pool mechanism (seed pool and exploration pool);
BOSCH: Black-Box Binary Optimization for Short-Context Attention-Head Selection in LLMs
Abbas Ghaddar (Huawei Noah's Ark Lab), Yufei Cui (Huawei Noah's Ark Lab)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose a no-training required black-box binary optimization method called BOSCH, used for selecting sliding window attention (SWA) heads in large language models according to short context requirements;
Boundary-Guided Policy Optimization for Memory-efficient RL of Diffusion Large Language Models
Nianyi Lin, Juanzi Li (Zhipu)
Computational EfficiencyTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposes Boundary‑Guided Policy Optimization (BGPO), an algorithm that can use large-scale Monte Carlo samples to approximate log-likelihood and optimize policies on diffusion large language models (dLLM), while maintaining low memory consumption.
BoYaEval: Evaluating Multimodal Large Language Models on Understanding Ancient Chinese Musical Scores
Jiajia Li (Wuhan University), Hai Zhao (Shanghai Jiao Tong University)
TransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed the BoYaEval benchmark to evaluate the understanding of ancient Chinese musical notation by multimodal large language models (MLLMs), covering five notation systems and setting up three-tier tasks (structure parsing, instruction translation, and music reasoning).
BracketRank: Large Language Model Document Ranking via Reasoning-based Competitive Elimination
Abdelrahman Abdallah (University of Innsbruck), Adam Jatowt (University of Innsbruck)
RetrievalTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes a document ranking framework called BracketRank based on large language models, utilizing competitive elimination tournaments and explicit reasoning to improve retrieval effectiveness for complex reasoning tasks.
Branch-and-Browse: Efficient and Controllable Web Exploration with Tree-Structured Reasoning and Action Memory
Shiqi He (University of Michigan), Mosharaf Chowdhury (University of Michigan)
Autonomous DrivingOptimizationComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AITextMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the Branch-and-Browse framework, which realizes an LLM-driven efficient Web agent by utilizing techniques such as subtask management, tree structure exploration, and page action memory.
Breaking Block Boundaries: Anchor-based History-stable Decoding for Diffusion Large Language Models
Shun Zou (University of Science and Technology of China), Xiangxiang Chu (AMAP, Alibaba Group)
Computational EfficiencyAI Code AssistantTransformerLarge Language ModelDiffusion modelImageTextMultimodalityChain-of-ThoughtAudio
🎯 What it does: Propose Anchor-based History-stable Decoding (AHD), which dynamically unlocks cross-block stable Tokens through dynamic anchors and historical consistency evaluation, breaking the block boundary limitations of traditional Semi-AR decoding;
Breaking Down and Building Up: Mixture of Skill-Based Vision-and-Language Navigation Agents
Tianyi Ma (Michigan State University), Parisa Kordjamshidi (KU Leuven)
Autonomous DrivingRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningMixture of ExpertsVision Language ModelDiffusion modelGenerative Adversarial NetworkContrastive LearningImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose SkillNav, a modular framework that decomposes the visual language navigation task into reusable atomic skills and dynamically schedules them via a VLM router; train each specialized skill using synthetic data and evaluate them on multiple VLN benchmarks.
Breaking the Generator Barrier: Disentangled Representation for Generalizable AI-Text Detection
Xiao Pu (Chongqing University of Posts and Telecommunications), Xiuli Bi (Chongqing University of Posts and Telecommunications)
Domain AdaptationAnomaly DetectionRepresentation LearningTransformerLarge Language ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningTextBenchmark
🎯 What it does: Construct a progressively structured separation framework in AI text detection that separates semantic content from generator-related features, thereby enhancing generalization ability for unknown generators.
Breaking the Impasse: Dual-Scale Evolutionary Policy Training for Social Language Agents
Minzheng Wang (University of Chinese Academy of Sciences), Wenji Mao (Chinese Academy of Sciences)
Large Language ModelReinforcement LearningText
🎯 What it does: In self-play reinforcement learning for social language games, we propose Dual-Scale Evolutionary Policy Training (DEPT), which detects evolutionary stagnation and dynamically adjusts the optimization landscape to prevent policy degradation.
Bridging Distance and Spectral Positional Encodings via Anchor-Based Diffusion Geometry Approximation
Yan Zimo (National University of Defense Technology), Wumei Du (National University of Defense Technology)
Drug DiscoveryGraph Neural NetworkDiffusion modelContrastive LearningGraphBiomedical Data
🎯 What it does: Explored the mathematical bridge between anchor distance encoding (DE) and Laplacian spectral encoding (LapPE), and verified their effectiveness in drug-drug interaction prediction on molecular graphs.
Bridging Internal Consistency and External Alignment: A Causal and Dynamic Interpretability Framework for LLM Generation
Shuyao Xiao (Beijing Normal University), Ke Chao (Beijing Normal University)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Proposes a causal and dynamic interpretable framework for autoregressive text generation in large language models, and defines two reference-free causal influence metrics based on this framework: Contextual Causal Influence (CCI) and Question-Answer Causal Influence (QACI).
Bridging Language and Items for Retrieval and Recommendation: Benchmarking LLMs as Semantic Encoders
Yupeng Hou (University of California San Diego), Julian McAuley (University of California San Diego)
RetrievalRecommendation SystemTransformerLarge Language ModelContrastive LearningTextSequentialBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the BLAIR benchmark to evaluate the effectiveness of large language models (LLMs) as semantic encoders in recommendation systems, and released the new Amazon Reviews 2023 large-scale dataset with a complex query product search subtask;
Bridging the Memorization-Utilization Gap: Near-Lossless Context Compression via Reinforcement Learning
Yujan Ting (United Imaging Intelligence), Weijing Huang (United Imaging Intelligence)
CompressionReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: By compressing the context of LLMs through average pooling and combining it with reinforcement learning, the gap between compressed memory and downstream task utilization is addressed.
Bridging the Sensory Gap: Visual Injection for Taxonomy Completion
Yuhang Niu (Nankai University), Xiaojie Yuan (Nankai University)
Data SynthesisRepresentation LearningTransformerLarge Language ModelVision Language ModelDiffusion modelGenerative Adversarial NetworkContrastive LearningImageTextMultimodality
🎯 What it does: By mapping synthetic images to pseudo-words and injecting them into a text encoder, the method utilizes visual information to achieve insertion prediction of new concepts within an existing hierarchical structure.
Bringing Real-World Relations into Video Generation with Graph-Structured Knowledge
Joonhyung Park (KAIST), Eunho Yang (KAIST)
GenerationData SynthesisGraph Neural NetworkTransformerVision Language ModelDiffusion modelVideoTextGraph
🎯 What it does: Introduce graph-structured knowledge into video generation models, using video scene graphs to guide diffusion models in generating short videos with higher physical realism and compositional coherence.
BrowseComp-Plus: A Fair and Disentangled Evaluation Benchmark for Deep Search Agents
Zijian Chen (University Of Waterloo), Jimmy Lin (University Of Waterloo)
RetrievalTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This work creates the BrowseComp-Plus benchmark, utilizing a fixed and human-verified document library to provide reproducible and separable deep search evaluation;
BTC-LLM: Efficient Sub-1-Bit LLM Quantization via Learnable Transformation and Binary Codebook
Hao Gu (Hong Kong University of Science and Technology), Yike Guo (Hong Kong University of Science and Technology)
CompressionComputational EfficiencyRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Proposes the BTC-LLM framework, which can compress large language models into binary weights with sub-1-bit precision, achieving extreme compression while maintaining high accuracy.
Building LLMs Like LEGO: Two-dimensional Architecture Reassembly of Large Language Models
Xingyu Wu (Hong Kong Polytechnic University), KC Tan (Hong Kong Polytechnic University)
Knowledge DistillationNeural Architecture SearchTransformerLarge Language ModelMixture of ExpertsTextMultimodality
🎯 What it does: This paper proposes the LEGO-LLM framework, which recombines the Transformer blocks of pre-trained large language models in two dimensions: vertical depth and horizontal layer-wise, to construct a new feasible architecture.
Business as Rulesual: A Benchmark and Framework for Business Rule Flow Modeling with LLMs
Chen Yang (Zhejiang University of Technology), Jing Fan (Zhejiang University of Technology)
AI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmarkFinance RelatedChain-of-Thought
🎯 What it does: Constructed a cross-domain benchmark dataset BREX oriented towards business rule flows, and proposed a structured extraction framework ExIde based on executable pseudocode;
BWLA: Breaking the Barrier of W1AX Post-Training Quantization for LLMs
Zhixiong Zhao (Houmo AI), Dawei Yang (Houmo AI)
CompressionComputational EfficiencyRepresentation LearningTransformerLarge Language ModelAuto EncoderContrastive LearningText
🎯 What it does: Proposed a post-training quantization framework called BWLA, which can compress large language models into 1-bit weights and low-bit activations (e.g., 6-bit) without fine-tuning, achieving end-to-end acceleration.
C-World: A Computer Use Agent Environment Creator
Ziqiao Xi (University Of California San Diego), Kun Zhou (University Of California San Diego)
Autonomous DrivingRobotic IntelligenceAI Code AssistantTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringWorld ModelTextTabularBenchmarkRetrieval-Augmented Generation
🎯 What it does: Built a system called C-World that can generate multi-tool environments on demand, integrating a large-scale tool library, task synthesis, state transition control, and reward signals, supporting both real-time and synthetic modes.
C2: Scalable Rubric-Augmented Reward Modeling from Binary Preferences
Akira Kawabata (Graduate University for Advanced Studies), Saku Sugawara (Graduate University for Advanced Studies)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: Propose the C2 framework, achieving scalable Rubric-augmented reward models using only binary preference data
CaBSALLM: Efficient Context-Aware Batch Annotation of Conversational Streams with Large Language Models
Mohammadsadegh Abolhasani (University of Utah), Paul Jen-Hwa Hu (University of Utah)
Computational EfficiencyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Proposed an efficient annotation pipeline called CaBSALLM, combining lightweight dialogue context with dynamic batch control, using large models to automatically annotate PSI clues in live chat.
CachePrune: Teaching LLMs What Not to Follow via KV-Cache Editing
Rui Wang (Adobe Research), Julian McAuley (University of California San Diego)
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Proposes a novel defense method called CachePrune, which prunes instruction-following-related neurons during KV cache encoding to prevent LLMs from responding to injected instructions.
CADMate: Generating CAD Assembly Plan with Geometric Chain-of-Thought and Spatial Physical Rewards
Jiali Chen (South China University of Technology), Yi Cai (Key Laboratory of Big Data and Intelligent Robot Ministry of Education)
GenerationPose EstimationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelImageTextMultimodalityChain-of-Thought
🎯 What it does: Achieve CAD assembly based on reference images through multi-modal large language models (MLLM), generating step-by-step 6-DoF transformation actions;
CAKE: Causal-Guided Adaptive Knowledge Editing for LLMs
Shuxin Liu (University of Chinese Academy of Sciences), Jianhao Zhang (University of Chinese Academy of Sciences)
OptimizationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposes a causal-guided adaptive knowledge editing framework called CAKE, which can collaboratively update the knowledge of large language models across multiple layers.
Calibrated Speculative Decoding: Frequency-Guided Candidate Selection for Efficient Inference
Zhouxuwen, Haibing Guan (Shanghai Jiao Tong University)
Computational EfficiencyAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: A no-training, lightweight calibrated inference framework called CSD was developed to recover valid tokens from the model's false rejections, thereby improving inference speed and efficiency.
Calibrated? Not for Everyone: How Sexual Orientation and Religious Markers Distort LLM Accuracy and Confidence in Medical QA
Alberto Testoni (Amsterdam University Medical Center), Iacer Calixto (Amsterdam Public Health)
Federated LearningExplainability and InterpretabilityComputational EfficiencyDrug DiscoveryAI Code AssistantReinforcement Learning from Human FeedbackNeural Architecture SearchProtein Structure PredictionTransformerLarge Language ModelPrompt EngineeringTextBiomedical DataElectronic Health RecordsBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Assessing the impact of sexual orientation and religious identity markers on the accuracy and confidence calibration of large language models (LLMs) in medical question-answering tasks.
Calibrating Inference Time Alignment with Sequence-level Risk Accumulation
Shanwen Tan (Sichuan University), Ziyue Qiao (Great Bay University)
Safty and PrivacyTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Propose a reasoning-time safe decoding framework called SEAT, which dynamically calibrates LLM generation using a multi-branch reward model and sequence-level risk accumulation.
Calibration-Aware Policy Optimization for Reasoning LLMs
Ziqi Wang (National Key Laboratory of Cognition and Decision Intelligence for Complex Systems, Institution of Automation, Chinese Academy of Sciences), Junge Zhang (National Key Laboratory of Cognition and Decision Intelligence for Complex Systems, Institution of Automation, Chinese Academy of Sciences)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: This paper provides a theoretical analysis of overconfidence and calibration degradation caused by reinforcement learning methods such as GRPO, and proposes the CAPO method to achieve simultaneous improvements in accuracy and calibration.
CAMEC: Complexity-Aware Multi-Expert Collaboration for Reliable Chinese Medical Question Answering
Yukang Wu (Hangzhou Dianzi University), Wu Guohua
Domain AdaptationExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsTextBiomedical DataElectronic Health RecordsBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the CAMEC framework, combining three-stage LoRA self-supervised fine-tuning, complexity-aware expert routing, and LLM review, to achieve multi-expert collaboration and iterative optimization for Chinese medical question answering.
CAML: A Conflict-Aware Molecular Language Model Merging Framework for Multi-Constraint Molecular Generation
Xuanbai Ren (Hunan University), Xiangxiang Zeng (Hunan University)
Drug DiscoveryTransformerLarge Language ModelReinforcement LearningMixture of ExpertsContrastive LearningTextGraphBiomedical Data
🎯 What it does: Propose the CAML framework, which treats multi-attribute molecular generation as a cooperative game among expert models, achieving attribute conflict minimization through dynamic parameter fusion, and generating molecules that meet multiple constraints without retraining the models.
CamoQuery: Language-Guided Reasoning Camouflaged Object Segmentation
Tianxin Han (Northeastern University), Fu Zhang (Northeastern University)
SegmentationExplainability and InterpretabilityRepresentation LearningConvolutional Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: Propose the Language-Guided Concealed Object Segmentation task (LRCOS), construct the CamoQuery dataset and the multi-instance dataset MCD, and develop the COSA model to achieve reasoning-based segmentation and explanation generation for concealed objects.
Can AI Be a Good Peer Reviewer? A Survey of Peer Review Process, Evaluation, and the Future
Sihong Wu (Yale University), Arman Cohan (Yale University)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: This paper reviews the entire process of using large language models (LLMs) and related technologies to support academic peer review, including tasks such as review generation, response, meta-review, and paper revision.
Can AI-Generated Persuasion Be Detected? Persuaficial Benchmark and AI vs. Human Linguistic Differences
Arkadiusz Modzelewski (NASK National Research Institute), Giovanni Da San Martino (NASK National Research Institute)
ClassificationTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Built and utilized a multilingual benchmark of human-generated persuasive texts (Persuaficial) to evaluate the differences between LLM-generated persuasive texts and human-written persuasive texts in terms of automatic detection and linguistic features.
Can Compact Language Models Search Like Agents? Distillation-Guided Policy Optimization for Preserving Agentic RAG Capabilities
Rikuto Kotoge (University of Osaka), Jiaxin Ma (OMRON SINIC X Corporation)
RetrievalOptimizationKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextRetrieval-Augmented Generation
🎯 What it does: Train agentic RAG on language models with a scale of 0.5-1B, and propose the DGPO method, evaluating its capabilities using ARCap.
Can Continual Pretraining Bridge the Performance Gap between General-purpose and Specialized Language Models in the Medical Domain?
Niclas Doll (Fraunhofer IAIS), Katrin Klug (Fraunhofer IAIS)
Domain AdaptationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsContrastive LearningTextBiomedical DataElectronic Health RecordsBenchmark
🎯 What it does: A specialized large language model (DeFineMed) with a scale of 7B-24B, applicable to the German medical field, was constructed using continuous pre-training and model fusion, and domain adaptation was performed on it using high-quality medical corpus FineMedde.
Can Factual Opinions Be Edited (Manipulated) in Large Language Models?
Yuanpu Cao (Pennsylvania State University), Jinghui Chen (Pennsylvania State University)
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes an editing evaluation framework for factual opinions (such as the stance of public figures on social and political issues) in large language models, and constructs the FOE benchmark dataset.
Can Large Language Models Infer Causal Relationships from Real-World Text?
Ryan Saklad (Worcester Polytechnic Institute), Raha Moraffah (Worcester Polytechnic Institute)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextGraphTabularBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This study constructs a causal reasoning benchmark called ReCITE based on real academic papers, and evaluates the performance of ten advanced large language models on this benchmark.
Can Large Language Models Keep Up? Benchmarking Online Adaptation to Continual Knowledge Streams
Jiyeon Kim (KAIST AI), Minjoon Seo (KAIST AI)
Federated LearningExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed the OAKS benchmark to evaluate the online adaptation and tracking capabilities of large language models in a continuous streaming knowledge environment.
Can LLM Agents Simulate Multi-Turn Human Behavior? Evidence from Real Online Customer Behavior Data
Yuxuan Lu (Northeastern University), Dakuo Wang (Northeastern University)
Recommendation SystemTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextSequentialBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper constructs the ShopCART 2 dataset based on real user session data from an online shopping scenario. It conducts the first large-scale quantitative evaluation of the accuracy of large language models (LLMs) in generating the next user action in multi-turn interactions. Furthermore, the paper improves the behavioral simulation accuracy of the model by fine-tuning LLMs on real clickstream data and incorporating synthetic reasoning trajectories.
Can LLM Safety Be Ensured by Constraining Parameter Regions?
Zongmin Li (Nanyang Technological University), Aixin Sun (Nanyang Technological University)
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelContrastive LearningTextBenchmark
🎯 What it does: Systematically evaluated four safe region identification methods (SNIP & Wanda, SafeNeuron, SafeLayer, NLSR) by extracting safe regions on multiple safety datasets and calculating IoU and utility-isolated IoU, investigating the consistency and reproducibility of safe regions across different datasets, models, and semantic distributions.
Can LLMs Act as Historians? Evaluating Historical Research Capabilities of LLMs via the Chinese Imperial Examination
Lirong Gao (Zhejiang University), Junbo Zhao (Zhejiang University)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed ProHist-Bench, a professional historical research benchmark centered on the Chinese imperial examination system, and conducted systematic evaluations on 18 LLMs;
Can LLMs Estimate Cognitive Complexity of Reading Comprehension Items?
Seonjeong Hwang (POSTECH), Gary Lee (POSTECH)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Constructed the RECO dataset and evaluated the ability of large language models to estimate the cognitive complexity of reading comprehension questions (evidence scope and transformation level) on this dataset.
Can LLMs Learn to Map the World from Local Descriptions?
Sirui Xia (Fudan University), Yanghua Xiao (Fudan University)
Representation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningWorld ModelTextSequential
🎯 What it does: In a simulated urban environment, a large language model (LLM) learns to construct a global spatial cognition from local relative position descriptions and shortest path trajectories through two-stage continual pre-training; further evaluations assess its generalization and robustness in spatial perception and navigation tasks.