These 557 ACL 2026 papers come with a code repository. Each shows an AI one-line summary below — get the verified repo link + the full 6-part summary (innovation, method, data, results, limitations) and search every ACL 2026 paper, free trial on arXivSub.
"Excuse me, may I say something..." CoLabScience, A Proactive AI Assistant for Biomedical Discovery and LLM-Expert Collaborations
Yang Wu (Worcester Polytechnic Institute), Xiaozhong Liu (Worcester Polytechnic Institute)
CodeDrug DiscoveryReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIPrompt EngineeringTextBiomedical DataRetrieval-Augmented Generation
🎯 What it does: Proposes CoLabScience, an active large language model assistant that utilizes the PULI framework to automatically determine when and how to intervene in biomedical discussions, thereby improving the efficiency of scientific collaboration.
"Newspaper Eat" Means "Not Tasty": A Taxonomy and Benchmark for Coded Language in Real-World Chinese Online Reviews
Ruyuan Wan (Pennsylvania State University), Ting-Hao Kenneth Huang (Pennsylvania State University)
CodeClassificationLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed the CODEDLANG dataset, which contains span-level annotations of coded language in 7,744 Chinese Google Maps reviews, and proposed a seven-classification system, establishing a benchmark for detecting, classifying, and predicting review ratings of coded language.
A Functionality-Grounded Benchmark for Evaluating Web Agents in E-commerce Domains
Xianren Zhang (Pennsylvania State University), Mat Hans (Amazon)
CodeRecommendation SystemAutonomous DrivingReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringVision-Language-Action ModelImageTextTabularSequentialBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose Amazon-Bench, a functional-based e-commerce web agent evaluation benchmark, including diverse user query generation and security assessment.
Qinghua Zhao (Hefei University), Xinlu Li (Hefei University)
CodeOptimizationExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: This paper systematically analyzes the hierarchical evolution of large language models during the supervised fine-tuning (SFT) process using three types of metrics: information theory, geometry, and optimization. It is found that the middle layers (approximately 20%–80% depth) are relatively stable, while the end layers are highly sensitive to SFT. Based on this, the paper proposes Mid-Block Efficient Tuning, which updates only the middle layers and significantly improves alignment performance.
A Multi-Agent Framework for Feature-Constrained Difficulty Control in Reading Comprehension Item Generation
Seonjeong Hwang (POSTECH), Gary Lee
CodeGenerationData SynthesisTransformerLarge Language ModelTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the MAFIG framework, which utilizes multi-agent LLMs to collaboratively generate multiple-choice reading comprehension questions that comply with feature constraints, and achieves fine-grained difficulty control through iterative revision.
A Shared Geometry of Difficulty in Multilingual Language Models
Stefano Civelli (University of Queensland), Gianluca Demartini (University of Queensland)
CodeExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: Studied how large language models encode the difficulty of math problems in multilingual environments, and explored how the difficulty signal is shared and specialized in language-specific geometric structures at different levels.
🎯 What it does: Reviews the application and progress of deep learning in geometric problem solving (GPS), systematically organizing task types, method architectures, training and inference techniques, evaluation metrics, and performance;
A Survey of Inductive Reasoning for Large Language Models
Kedi Chen (East China Normal University), Wei Zhang (East China Normal University)
CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextReview/Survey PaperBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Systematically reviews the inductive reasoning capabilities of large language models (LLMs), proposes a classification of three improvement methods (post-training enhancement, exploration at test time, data augmentation), summarizes existing inductive reasoning benchmarks, and proposes a unified sandbox evaluation framework and a fine-grained observation coverage (OC) metric, along with theoretical analysis of the sources of inductive bias.
A Survey of Large Language Models for Text-Guided Molecular Discovery: From Molecule Generation to Optimization
Ziqing Wang (Northwestern University), Kaize Ding (Northwestern University)
CodeOptimizationDrug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextGraphBiomedical DataReview/Survey PaperBenchmarkRetrieval-Augmented Generation
🎯 What it does: A systematic review of the application of large language models in text-driven molecular generation and optimization, outlining four core challenges (effectiveness, synthesizability, property control, structural diversity), proposing a classification based on learning paradigms, and summarizing commonly used datasets, evaluation metrics, and benchmark frameworks.
A Survey of Reinforcement Learning for Large Language Models under Data Scarcity: Challenges and Solutions
Zhiyin Yu (Peking University), Xiao Luo (University of Wisconsin-Madison)
CodeData-Centric LearningTransformerLarge Language ModelReinforcement LearningTextReview/Survey Paper
🎯 What it does: A systematic review of reinforcement learning for large language models under data-scarce conditions, and the proposal of a low-level hierarchical classification framework from three perspectives: data, training, and framework.
A Theoretically Grounded Approach to Summarizing Conversation Dynamics for Forecasting the Derailment of Online Conversations
Yingxue Fu (Universite Cote d'Azur), Anaïs Ollagnier (Universite Cote d'Azur)
CodeClassificationAnomaly DetectionTransformerLarge Language ModelPrompt EngineeringTextReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: Designed and implemented a dialogue pragmatics information annotation and summary generation method based on a theoretical framework, used for dynamically predicting toxic deviations in online conversations.
Accommodation and Epistemic Vigilance: A Pragmatic Account of Why LLMs Fail to Challenge Harmful Beliefs
Myra Cheng (Stanford University), Dan Jurafsky (Stanford University)
CodeSafty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Studied the pragmatic reasons why large language models do not challenge users' harmful beliefs in safety tasks, and verified the impact of human pragmatic factors (topic relevance, language encoding, information source reliability) on LLM behavior; meanwhile, proposed and tested two simple prompt interventions (EXPLICIT and WAIT) to improve the model's safety performance.
Achieving Multi-Hop Calculation and Safe Abstention in Financial Numerical Reasoning by Metric Graph Constrained LLMs
Aoyuan Jiang (Wuhan University), Rui Wang (Wuhan University of Technology)
CodeGraph Neural NetworkTransformerLarge Language ModelTextGraphTabularFinance RelatedRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes a graph-constrained neuro-symbolic framework called GBFR for safe and accurate multi-hop numerical reasoning in financial text.
Activation-Guided Local Editing for Jailbreaking Attacks
Jiecong Wang (Beihang University), Zhengtao Yu (Kunming University of Science and Technology)
CodeSafty and PrivacyExplainability and InterpretabilityAdversarial AttackTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Proposed a two-stage jailbreak method called AGILE, which first uses a generative LLM to inject malicious queries into multi-round safe dialogues and rephrase them, and then guides fine-grained editing through attention and hidden state, shifting the model's internal representation from malicious to safe;
AdabNER: Arabic Digital Archive Books with Nested Entity Recognition
Aya Mourad (Sorbonne Université), Mustafa Jarrar (Hamad Bin Khalifa University)
CodeRecognitionTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Constructed the AdabNer literary nested named entity recognition dataset, containing 876k tokens from 138 modern standard Arabic books and 21 entity types.
AdaFuse: Adaptive Ensemble Decoding for Large Language Models
Chengming Cui (University Of Illinois Urbana Champaign), Jingrui He (University Of Illinois Urbana Champaign)
CodeGenerationComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed a framework called AdaFuse for adaptive word-level fusion of large language models during inference;
🎯 What it does: The paper proposes the Text Frequency Law (TFL), which estimates sentence-level frequency using an online corpus and utilizes higher-frequency synonymous sentences for LLM prompting and fine-tuning.
Adaptive Prompt Structure Factorization: A Framework for Self-Discovering and Optimizing Compositional Prompt Programs
Haoyue Liu (Chinese University of Hong Kong), Xiaoying Tang (Chinese University of Hong Kong)
CodeOptimizationExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes an adaptive prompt structure decomposition (aPSF) framework that decomposes prompts into task-specific semantic factors using an API-only approach, and iteratively optimizes them factor by factor;
Jongho Kim (Seoul National University-LG AI Research Center), Moontae Lee (LG AI Research)
CodeRetrievalDomain AdaptationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes the REPAIR framework, which combines planning-based reordering with neighborhood adaptive retrieval, achieving mid-course correction by converting reasoning steps into dense rewards.
🎯 What it does: Propose an adversarial metric learning (AML) framework to improve representation learning for fine-grained emotion classification (FEC), replacing traditional fixed similarity metrics;
AgentGL: Towards Agentic Graph Learning with LLMs via Reinforcement Learning
Yuanfu Sun (New York University Shanghai), Qiaoyu Tan (New York University Shanghai)
CodeOptimizationRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelReinforcement LearningPrompt EngineeringContrastive LearningTextGraphRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the AgentGL framework, which transforms graph learning into an interactive process between graph structure exploration and LLM reasoning, enabling LLMs to autonomously navigate and reason through graph-local tools;
Aniket Rege (Meta Reality Labs Research), Hyo Jin Kim (University of Wisconsin-Madison)
CodeRetrievalRecommendation SystemExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelAgentic AIVideoTextMultimodalityRetrieval-Augmented GenerationChain-of-ThoughtAudio
🎯 What it does: Propose the EGAgent framework, which achieves question answering on extremely long temporal audio-visual videos by constructing time-annotated entity scene graphs and planning-based tool calls;
ALDEN: Reinforcement Learning for Active Navigation and Evidence Gathering in Long Documents
Tianyu Yang (University of Göttingen), Bela Gipp (University of Göttingen)
CodeTransformerReinforcement LearningVision Language ModelImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Train visual language models as active navigation agents, using multi-round reinforcement learning to actively retrieve pages and generate answers in long-form visual documents.
CodeRecommendation SystemFederated LearningExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed the cultural alignment dataset CULTURAX and proposed a two-stage ALIGNCULTURA process for systematically evaluating the Helpfulness, Harmlessness, and Honesty (HHH) performance of large language models across different cultural contexts.
Aligning Backchannel and Dialogue Context Representations via Contrastive LLM Fine-Tuning
Livia Qian (KTH Royal Institute of Technology), Gabriel Skantze (KTH Royal Institute of Technology)
CodeRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextMultimodalityAudio
🎯 What it does: This paper first fine-tunes a large language model on dialogue corpora to obtain rich contextual semantic representations, then combines them with WavLM's speech representations. By using contrastive learning, the paper maps contextual information and Backchannel audio features into a shared embedding space, achieving alignment between dialogue context and Backchannel forms.
CodeKnowledge DistillationRepresentation LearningData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextFinance RelatedRetrieval-Augmented Generation
🎯 What it does: Propose a sustainable updating Chinese real-time knowledge editing dataset CRAFT, and build a KEDAS framework based on LoRA, achieving a unified improvement in editing success, locality, and transferability in real-time knowledge editing.
All Changes May Have Invariant Principles: Improving Ever-Shifting Harmful Meme Detection via Design Concept Reproduction
Ziyou Jiang (State Key Laboratory of Complex System Modeling and Simulation Technology), Qing Wang (State Key Laboratory of Complex System Modeling and Simulation Technology)
CodeAnomaly DetectionExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringMultimodalityGraphRetrieval-Augmented Generation
🎯 What it does: This paper proposes a method called REPMD for detecting harmful Memes that are always changing, based on the reproduction of design concepts. It constructs a Design Concept Graph (DCG) and guides a multi-modal large language model to achieve detection.
Amadeus: Autoregressive Model with Bidirectional Attribute Modelling for Symbolic Music
Hongju Su (Beijing University of Posts and Telecommunications), Yi-Zhe Song (University of Surrey)
CodeGenerationData SynthesisTransformerLarge Language ModelDiffusion modelAuto EncoderTextSequentialAudio
🎯 What it does: Proposed the Amadeus framework, which employs a two-tier architecture: first autoregressively generating a sequence of notes, and then using a bidirectional discrete diffusion model to decode note attributes, achieving symbolic music generation.
Analytical FFN-to-MoE Restructuring via Activation Pattern Analysis
Zehua Pei (The Chinese University of Hong Kong), Bei Yu (The Chinese University of Hong Kong)
CodeComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringMixture of ExpertsText
🎯 What it does: During the inference phase of large-scale language models, dense feed-forward networks (FFN) are transformed into sparse Mixture-of-Experts (MoE) architectures. By utilizing a small amount of calibration data, activation pattern analysis is performed to quickly build shared experts and routing experts, and performance improvements can be achieved through minimal fine-tuning.
Anchoring the Cache: Mitigating Contextual Hallucination in KV-Compressed Long-Context Summarization
Yu Fu (University of California, Riverside), Yue Dong (Amazon)
CodeCompressionExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: A systematic study on the hallucination problem caused by KV cache compression techniques in long-text summarization is conducted, and a strategy is proposed to clear the KV cache of key retrieval heads during the decoding phase, called HalluKV. This strategy effectively anchors the retrieval heads' attention to the source text, thereby reducing the hallucination rate.
CodeSegmentationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose AnchorSeg, which achieves reasoning segmentation through a language-guided query bank, separating semantic reasoning from spatial localization;
🎯 What it does: Propose the ANDROiD COACH framework, which introduces a single-state multi-action (SSMA) paradigm in online reinforcement learning. By sampling multiple actions in bulk and using the Critic for evaluation, it improves sample efficiency and accelerates training without requiring additional simulation interactions.
CodeClassificationRecommendation SystemData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Constructed the first sentence-level social perception dataset, W&C-Sent, containing 1,633 English sentence-target pairs, with seven-point ratings for trust, likeability, and competence
Anti-Length Shift: Dynamic Outlier Truncation for Training Efficient Reasoning Models
Wei Wu (University of Science and Technology of China), Hui Xiong (Hong Kong University of Science and Technology)
CodeComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: During the reinforcement learning training phase, dynamic outlier truncation (DOT) is applied to fully correct answers to suppress redundant reasoning, thereby reducing the reasoning length; meanwhile, KL normalization and predictive dynamic sampling are introduced to maintain training stability.
Are LLMs Reliable Rankers? Rank Manipulation via Two-Stage Token Optimization
Tiancheng Xing (National University Of Singapore), Xiyang Hu (Arizona State University)
CodeRecommendation SystemOptimizationAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposed a two-stage token optimization method called RAF (Rank Anything First), which uses natural language text to induce target items to improve rankings in LLM rerankers;
Are they lovers or friends? Evaluating LLMs’ Social Reasoning in English and Korean Dialogues
Eunsu Kim (KAIST), Najoung Kim (Boston University)
CodeClassificationExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Evaluate the ability of LLMs to reason about speakers' social relationships (e.g., friends, lovers, etc.) in dialogues on the SCRIPTS dataset;
Are We Using the Right Benchmark: An Evaluation Framework for Visual Token Compression Methods
Chenfei Liao (Hong Kong University of Science and Technology), Xuming Hu (Hong Kong University of Science and Technology)
CodeCompressionVision Language ModelAuto EncoderMultimodalityBenchmark
🎯 What it does: This paper proposes an evaluation framework for visual token compression methods in multimodal large models and constructs a new benchmark called VTC-Bench;
arXiv2Table: Toward Realistic Benchmarking and Evaluation for LLM-Based Literature-Review Table Generation
Weiqi Wang (Johns Hopkins University), Daniel Khashabi (Johns Hopkins University)
CodeGenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringTextTabularBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Studied a more realistic scientific literature review table generation task, constructed the ARXIV2TABLE benchmark, and proposed an iterative batch generation method.
AttnPO: Attention-Guided Process Supervision for Efficient Reasoning
Shuaiyi Nie (Institute of Information Engineering Chinese Academy of Sciences), Tingwen Liu (Baidu Inc)
CodeComputational EfficiencyAI Code AssistantTransformerLarge Language ModelReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: Propose a low-cost process supervision reinforcement learning framework called ATTNP, which utilizes key attention heads in model attention to allocate credit to reasoning steps, reducing redundant thinking.
Automatic Correction of Writing Anomalies in Hausa Texts
Ahmad Mustapha Wali (University of Bucharest), Sergiu Nisioi (University of Bucharest)
CodeData SynthesisAnomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Constructed approximately 400k pairs of noisy-clean Hausa parallel corpora, and fine-tuned multiple transformers on this data for writing anomaly correction.
Autonomous Knowledge Graph Exploration with Adaptive Breadth-Depth Retrieval
Joaquin Polonuer, Marinka Zitnik (Harvard Medical School)
CodeRetrievalFederated LearningExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningReinforcement Learning from Human FeedbackGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningAgentic AIPrompt EngineeringTextGraphBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Designed and implemented a tool-use knowledge graph retrieval framework called ARK, which utilizes large language models to achieve adaptive control over the breadth and depth of retrieval through two tools: global retrieval and neighborhood exploration;
Awakening Dormant Experts:Counterfactual Routing to Mitigate MoE Hallucinations
Wentao Hu (Xi'an Jiaotong University), Xuelong Li (China Telecom)
CodeExplainability and InterpretabilityComputational EfficiencyLarge Language ModelPrompt EngineeringMixture of ExpertsTextRetrieval-Augmented Generation
🎯 What it does: Proposed a training-agnostic inference framework called Counterfactual Routing (CoR), which wakes up dormant experts through causal analysis, thereby improving the factual accuracy of MoE models on long-tail knowledge.
BaseCal: Unsupervised Confidence Calibration via Base Model Signals
Hexiang Tan (State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences), Xueqi Cheng (University of Chinese Academy of Sciences)
CodeExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: Proposes an unsupervised framework called BaseCal based on the signal of a base model to calibrate the confidence of post-training large language models (LLMs), restoring the overconfidence problem of post-training LLMs without modifying model parameters.
BatonVoice: An Operationalist Framework for Enhancing Controllable Speech Synthesis with Linguistic Intelligence from LLMs
Yue Wang (Soochow University), Min Zhang (Soochow University)
CodeGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented GenerationAudio
🎯 What it does: Propose the BATONVOICE framework, which decouples user instruction understanding from speech generation. It first uses an LLM to generate quantifiable text-based speech feature plans, and then a specialized BATONTTS model synthesizes audio based on these plans.
Bayesian Social Deduction with Graph-Informed Language Models
Shahab Rahimirad (Purdue University), Joseph Campbell (Purdue University)
CodeExplainability and InterpretabilityReinforcement Learning from Human FeedbackGraph Neural NetworkTransformerLarge Language ModelTextGraph
🎯 What it does: Proposes GRAIL, a social reasoning framework that combines LLMs with graph-structured Bayesian inference, specifically designed for the Avalon game.
Benchmarking and Enabling Efficient Chinese Medical Retrieval via Asymmetric Encoders
Angqing Jiang (University of Science and Technology of China), Defu Lian (University of Science and Technology of China)
CodeRetrievalComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningTextBiomedical DataBenchmark
🎯 What it does: This paper proposes the Chinese medical text embedding benchmark CMedTEB and implements efficient retrieval based on the heterogeneous retrieval architecture CARE.
CodeComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelVision Language ModelContrastive LearningTextMultimodalityBenchmark
🎯 What it does: This paper systematically evaluates post-training quantization (PTQ) of large-scale language models and multimodal language models in the micro-scale floating-point (MXFP) format, covering 7 PTQ algorithms, 15 evaluation benchmarks, and 3 model families.
Beyond "I Don’t Know": Evaluating LLM Self-Awareness in Discriminating Data and Model Uncertainty
Jingyi Ren (Tsinghua University), Yang Liu (Tsinghua University)
CodeExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the UA-Bench benchmark to evaluate the self-awareness of LLMs in identifying data uncertainty and model uncertainty, and enhance this ability through lightweight reinforcement learning.
Beyond Atomic Characters: Glyph-Aware Sub-character Alignment for Low-Resource Multilingual OCR
Mengxiao Zhu (North China University of Technology), Ge Shi (Beijing Institute of Technology)
CodeRecognitionData SynthesisConvolutional Neural NetworkTransformerSupervised Fine-TuningVision Language ModelContrastive LearningImageTextBenchmark
🎯 What it does: Proposed the BASA OCR framework, integrating high-resolution visual encoders, Glyph-Aware Fine-grained Adapter (GAFA) sub-character alignment module, two-stage curriculum learning, and Glyph-Aware Reverse Synthesis data generation technique with zero-cost sub-character labels, and constructed the BASA-Bench benchmark containing 11 low-resource languages and 23 real-world scenarios.
Beyond Examples: Towards Automated Thought-level In-Context Reasoning for Large Language Models
Jinyang Wu (Tsinghua University), Jianhua Tao (Tsinghua University)
CodeExplainability 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 Literal Mapping: Benchmarking and Improving Non-Literal Translation Evaluation
Yanzhi Tian (Beijing Institute of Technology), Yuhang Guo (Beijing Institute of Technology)
CodeTransformerLarge 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.
🎯 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)
CodeRetrievalTransformerLarge 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 Query Memorization: Large Language Model Routing with Query Decomposition and Historical Matching
Bo Lv (Tencent Hunyuan), Yue Yu (Peng Cheng Laboratory)
CodeComputational 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;
CodeTransformerLarge 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-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)
CodeClassificationExplainability 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 the Last Frame: Process-aware Evaluation for Generative Video Reasoning
Yifan Li (Renmin University of China), Minghui Qiu (ByteDance)
CodeGenerationData 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.
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)
CodeAnomaly 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.
🎯 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)
CodeDrug 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.
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)
CodeTransformerLarge 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.
Boundary-Guided Policy Optimization for Memory-efficient RL of Diffusion Large Language Models
Nianyi Lin, Juanzi Li (Zhipu)
CodeComputational 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.
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)
CodeAutonomous 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)
CodeComputational 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 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)
CodeDomain 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.
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)
CodeRetrievalRecommendation 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;
CodeData 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.
Building LLMs Like LEGO: Two-dimensional Architecture Reassembly of Large Language Models
Xingyu Wu (Hong Kong Polytechnic University), KC Tan (Hong Kong Polytechnic University)
CodeKnowledge 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.
CAKE: Causal-Guided Adaptive Knowledge Editing for LLMs
Shuxin Liu (University of Chinese Academy of Sciences), Jianhao Zhang (University of Chinese Academy of Sciences)
CodeOptimizationExplainability 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.
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)
CodeClassificationTransformerLarge 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 Large Language Models Keep Up? Benchmarking Online Adaptation to Continual Knowledge Streams
Jiyeon Kim (KAIST AI), Minjoon Seo (KAIST AI)
CodeFederated 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 Safety Be Ensured by Constraining Parameter Regions?
Zongmin Li (Nanyang Technological University), Aixin Sun (Nanyang Technological University)
CodeSafty 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.
CodeTransformerLarge 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)
CodeClassificationExplainability 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 Reasoning Path still be Effective as Input? Bridging Post-Reasoning to Chain-of-Thought Compression
Chengzhengxu Li (Xi'an Jiaotong University), Chao Shen (Xi'an Jiaotong University)
CodeComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Proposed post-reasoning and the UCoT framework, which reduces the length of the reasoning output by adding compressed CoT to the input of the LLM.
CAPruner: Conceptual-Adjacent Scene Graph Pruner for Enhancing 3D Spatial Reasoning of Large Language Models
Shengli Zhou (Southern University of Science and Technology), Feng Zheng (Southern University of Science and Technology)
CodeComputational EfficiencyRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningTextPoint CloudGraph
🎯 What it does: A lightweight concept adjacency scene graph pruning model called CAPruner is proposed to retain the most important spatial relationships in 3D vision-language tasks under limited budget, thereby improving the efficiency and accuracy of large language models in 3D spatial reasoning.
CAR-bench: Evaluating the Consistency and Limit-Awareness of LLM Agents under Real-World Uncertainty
Johannes Kirmayr (BMWGroup Research and Technology), Elisabeth Andre
CodeAutonomous DrivingExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelAgentic AIPrompt EngineeringTextTabularTime SeriesBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Construct and evaluate CAR-bench, a multi-turn interaction benchmark for in-vehicle assistants, used to test the consistency, self-awareness of uncertainty and capabilities of large language models in dynamic environments.
CaseFacts: A Benchmark for Legal Fact-Checking and Precedent Retrieval
Akshith Reddy Putta (University of Texas at Arlington), Chengkai Li (University of Texas at Arlington)
CodeRetrievalTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed a legal fact-checking benchmark called CaseFacts, targeting U.S. Supreme Court precedents, covering three types of spoken legal claims: supporting, refuting, and being overturned.
Chain-of-Thought as a Lens: Evaluating Structured Reasoning Alignment between Human Preferences and Large Language Models
Boxuan Wang (University of Liverpool), Yi Dong (University of Liverpool)
CodeExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose Alignment Score, which quantifies the alignment between the chain-of-thought reasoning generated by large language models and human preference reference chains using a semantic entropy matrix, and design alignment-based chain sampling and selection methods (ACSS, SC-Align) based on this.
Challenging the Boundaries of Reasoning: An Olympiad-Level Math Benchmark for Large Language Models
Haoxiang Sun (Renmin University of China), Ji-Rong Wen (Renmin University of China)
CodeLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper constructs a new bilingual (English-Chinese) Olympiad-level mathematics benchmark called OlymMATH, which includes 200 arithmetic problems (EASY/HARD) verifiable by Sympy and 150 formal proof problems using Lean4;
CodeRecommendation SystemExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Constructed the first multi-turn role-playing dataset for emotional support, named ChatAnime, and proposed the Emotionally Supportive Role-Playing (ESRP) framework based on users' emotional needs;
CheckRLM: Effective Knowledge–Thought Coherence Checking in Retrieval-Augmented Reasoning
Dingling Xu (Beijing Normal University), Maosong Sun (Chinese Academy Of Sciences)
CodeExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Proposes the CheckRLM framework, which reduces error accumulation during RLM inference by promptly checking and correcting factual errors through retrieval-augmented generation (RAG).
ChipSeek: Optimizing Verilog Generation via EDA-Integrated Reinforcement Learning
Zhirong Chen (Chinese Academy of Sciences), Ying Wang (Chinese Academy of Sciences)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Proposes the ChipSeek framework, which utilizes reinforcement learning combined with complete EDA toolchain feedback to directly optimize the functional correctness of RTL (Verilog) code and PPA (Power, Performance, Area) metrics.
CIA: Inferring the Communication Topology from LLM-based Multi-Agent Systems
Yongxuan Wu (Chinese Academy of Sciences), Yanan Cao (Chinese Academy of Sciences)
CodeExplainability and InterpretabilityAdversarial AttackReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringTextSequentialRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose an attack method called Communication Inference Attack (CIA) to infer the communication topology of multi-agent systems (MAS) in large language models under a black-box setting;
Chenghao Zhang (Chinese Academy of Sciences), Yi Du (Chinese Academy of Sciences)
CodeClassificationRecommendation SystemGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextGraphBenchmark
🎯 What it does: This paper constructs the first large-scale heterogeneous text-attribute graph dataset, CITE, containing 438K nodes, 1.22 million edges, four node types (papers, authors, journals, keywords), and four relationship types. Based on this dataset, various learning paradigms (homogeneous GNN, heterogeneous GNN, LLM, and LLM+Graph) are systematically evaluated for node classification and link prediction tasks.
Closing the Modality Reasoning Gap for Speech Large Language Models
Chaoren Wang (Chinese University of Hong Kong), Zhizheng Wu (Chinese University of Hong Kong)
CodeRepresentation LearningData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningContrastive LearningTextMultimodalityAudio
🎯 What it does: Propose a trajectory alignment framework based on reinforcement learning called TARS, aimed at closing the performance gap between speech and text in reasoning.
ClusterRAG: Cluster-Based Collaborative Filtering for Personalized Retrieval-Augmented Generation
Gibson Nkhata (University of Arkansas), Susan Gauch (University of Arkansas)
CodeRetrievalRecommendation SystemTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Propose ClusterRAG, a clustering-based collaborative filtering method aimed at enhancing the effectiveness of personalized retrieval-augmented generation (RAG);
CoCoNUTS: Concentrating on Content while Neglecting Uninformative Textual Styles for AI-Generated Peer Review Detection
Yihan Chen (University of Chinese Academy of Sciences), Le Sun (Chinese Information Processing Laboratory, Institute of Software, Chinese Academy of Sciences)
CodeClassificationAnomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Propose the CoCoNUTS benchmark and the CoCoDet detector, focusing on content rather than text style, achieving precise identification of AI-generated peer reviews.
CodeEvo: Interaction-Driven Synthesis of Code-centric Data through Hybrid and Iterative Feedback
Qiushi Sun (University of Hong Kong), Fei Yuan (University of Hong Kong)
CodeData SynthesisAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringText
🎯 What it does: Propose CodeEvo, an interactive framework based on dual agents (Coder and Reviewer), for automatically generating high-quality, executable, and logically complex instruction-code pairs, and build the CodeEvo-100K dataset.
CodeJudgeBench: Benchmarking LLM-as-a-Judge for Coding Tasks
Hongchao Jiang (ASUS Intelligent Cloud Services), Robby T. Tan (ASUS Intelligent Cloud Services)
CodeAI Code AssistantTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmark
🎯 What it does: Construct the CodeJudgeBench benchmark to evaluate the judgment capabilities of LLMs in code generation, code repair, and unit test generation tasks.
CODESTRUCT: Code Agents over Structured Action Spaces
Myeongsoo Kim (AWS AI Labs), Murali Krishna Ramanathan (AWS AI Labs)
CodeAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the CODESTRUCT framework, allowing code agents to read and write structured named entities (such as files, classes, functions, methods) through AST, rather than traditional text strings;
CoG: Controllable Graph Reasoning via Relational Blueprints and Failure-Aware Refinement over Knowledge Graphs
Yuanxiang Liu (Zhejiang University), Wen Zhang (Zhejiang University)
CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextGraphBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed the CoG framework, which employs a dual-process (intuition + reasoning) mechanism to achieve controllable, multi-hop reasoning on knowledge graphs.
CogEvolve: A Multimodal Benchmark for Evaluating Relational Reasoning in Semantic Extension
Jingjie Zeng (Dalian University of Technology), Hongfei Lin (Dalian University of Technology)
CodeExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringVision Language ModelTextMultimodalityGraphBenchmarkChain-of-Thought
🎯 What it does: Propose the CogEvolve benchmark, specifically designed to evaluate models' generative reasoning capabilities in semantic evolution (analogy, metaphor, metonymy);
Collaboration of Fusion and Independence: Hypercomplex-driven Robust Multi-Modal Knowledge Graph Completion
Zhiqiang Liu (Zhejiang University), Wen Zhang (Ant Group)
CodeKnowledge DistillationRepresentation LearningGraph Neural NetworkMixture of ExpertsContrastive LearningImageTextMultimodalityGraph
🎯 What it does: Proposes a multi-modal knowledge graph completion model called M-Hyper based on biquaternion space, combining fusion and independent modes to achieve collaborative representation of multi-modal information.