ACL 2026 Papers with AI Summaries
Annual Meeting of the Association for Computational Linguistics · 2296 papers
→ ACL 2026 papers with code (557)
Each paper below shows an AI-generated one-line summary. Get the full 6-part summary (innovation, method, data, results, limitations) and search all 2296 ACL 2026 papers by keyword, author or institution —
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)
Drug 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.
“I See What You Did There”: Can Large Vision-Language Models Understand Multimodal Puns?
Naen Xu (Zhejiang University), Shouling Ji (Zhejiang University)
RecognitionExplainability and InterpretabilityData-Centric LearningTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelGenerative Adversarial NetworkImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed the MULTIPUN dataset and a multi-task evaluation framework (identification, localization, and interpretation of multimodal puns), and designed Prompt-level method Pun-CoT and Model-level method Pun-Tuning to enhance the pun understanding capability of Vision-Language Models (VLMs).
"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)
ClassificationLarge 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 BERTology View of LLM Orchestrations: Token- and Layer-Selective Probes for Efficient Single-Pass Classification
Gonzalo Ariel Meyoyan (Universidad de Buenos Aires), Luciano Del Corro (Universidad de San Andres)
ClassificationSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: This paper proposes inserting a lightweight detection head into the forward computation of already deployed LLMs, using token and hierarchical hidden states for classification under single forward inference, avoiding the need for additional calls to safety or sentiment judgment models;
A Comprehensive Survey of Process Reward Models: Data Generation, Model Construction, and Usage
Congmin Zheng (Shanghai Jiao Tong University), Weinan Zhang (Shanghai Jiao Tong University)
Data-Centric LearningRobotic IntelligenceAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringMixture of ExpertsTextMultimodalitySequentialReview/Survey PaperBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Reviews the entire process of Process Reward Models (PRMs), including data generation, model construction, and practical applications, and systematically summarizes various techniques, datasets, and benchmarks.
A Data-Centric Approach to Generalizable Speech Deepfake Detection
Wen Huang (Shanghai Jiao Tong University), Yanmin Qian (Shanghai Jiao Tong University)
Anomaly DetectionData-Centric LearningTransformerSupervised Fine-TuningContrastive LearningAudio
🎯 What it does: Enhancing the Generalization Ability of Speech Deepfake Detection through Data Center Methods
A Data-Efficient Path to Multilingual LLMs: Language Expansion via Post-training PARAM\Delta Integration into Upcycled MoE
Hao Zhou (Nanjing University), Shujian Huang (Nanjing University)
Computational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsTextMultimodality
🎯 What it does: This paper proposes DeltaMoE, which efficiently scales multilingual LLMs while preserving original language capabilities by combining sparse MoE with delta merging.
A Dual-Task Paradigm to Investigate Sentence Comprehension Strategies in Language Models
Rei Emura (Tohoku University), Saku Sugawara (National Institute of Informatics)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Design a dual-task paradigm to evaluate the sentence comprehension strategies of language models under limited cognitive resources.
A Functionality-Grounded Benchmark for Evaluating Web Agents in E-commerce Domains
Xianren Zhang (Pennsylvania State University), Mat Hans (Amazon)
Recommendation 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.
A Game-Theoretic Negotiation Framework for Cross-Cultural Consensus
Guoxi Zhang (Peking University), Juntao Dai (Peking University)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes a cross-cultural consensus negotiation framework based on game theory (Nash Consensus Negotiation), which allows agents of different cultural large language models (LLMs) to iteratively generate, evaluate, and adjust natural language guidelines, eventually converging to a Nash equilibrium, thereby achieving a consensus result that is fair, interpretable, and balanced among all parties.
A Goal Without a Plan Is Just a Wish: Efficient and Effective Global Planner Training for Long-Horizon Agent Tasks
Shuzheng Si (University of Illinois Urbana-Champaign), Maosong Sun (Tsinghua University)
Autonomous DrivingOptimizationComputational EfficiencyRobotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIPrompt EngineeringWorld ModelTextBenchmark
🎯 What it does: Propose EAGLET, a global planner training framework for LLM agents in long-term sequential tasks. It first generates and filters high-quality plans using a senior LLM for SFT cold start, and then further improves the planner using rule-based RL (Executor Capability Gain Reward).
A Layer-wise Analysis of Supervised Fine-Tuning
Qinghua Zhao (Hefei University), Xinlu Li (Hefei University)
OptimizationExplainability 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 Lightweight Explainable Guardrail for Prompt Safety
Md Asiful Islam (University of Arizona), Mihai Surdeanu (University of Arizona)
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningText
🎯 What it does: Propose a lightweight and interpretable prompt safety guardian model called LEG, which can simultaneously determine whether a prompt is unsafe and provide keywords that lead to the judgment.
A Linguistics-Aware LLM Watermarking via Syntactic Predictability
Shinwoo Park (Yonsei University), Yo-Sub Han (Yonsei University)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Designed and implemented a linguistically aware LLM watermarking framework called STELA, which dynamically adjusts watermark strength using syntactic predictability (POS n-gram conditional entropy) and realizes publicly verifiable detection without model logits.
A Mechanistic Account of Attention Sinks in GPT-2: One Circuit, Broader Implications for Mitigation
Yuval Ran-Milo (Tel Aviv University), Shahar Mendel (Tel Aviv University)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Mechanistically analyze the attention sink (i.e., excessive focus on the first position) in the GPT-2 architecture, and verify its formation path through structural analysis and causal intervention.
A Model of the Language Process
Brandon Duderstadt (Calcifer Computing), Hayden Helm (Helivan)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelContrastive LearningTextTime SeriesReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: This paper proposes Temporal Language Model 1 (TLM-1), an encoder that incorporates temporal information into the BERT structure, capable of simultaneously predicting document content and publication year; and introduces a unified kernel prior query method based on the Bayesian framework to separate temporal dynamics from era bias.
A Multi-Agent Framework for Feature-Constrained Difficulty Control in Reading Comprehension Item Generation
Seonjeong Hwang (POSTECH), Gary Lee
GenerationData 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 Multi-Agent Framework for High-Interaction Terminal Simulation
Kai Wei (University of South Florida), Guangjing Wang (University of South Florida)
AI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningAgentic AIPrompt EngineeringTextSequentialRetrieval-Augmented Generation
🎯 What it does: Proposed the MANTIS terminal simulation framework, utilizing a multi-agent system and agentic file system to achieve high-interaction, sustainable command-line simulation;
A Multilingual Social Bias Benchmark Incorporating Thinking Processes
Masahiro Kaneko (MBZUAI), Timothy Baldwin (MBZUAI)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Construct a multilingual social bias benchmark called MBTP, and evaluate the bias of LLMs using human-written pro and con bias reasoning texts;
A Scalable Multi-LLM Collaboration System with Retrieval-based Selection and Exploration-Exploitation-Driven Enhancement
Shengji Tang (Shanghai Artificial Intelligence Laboratory), Peng Ye (Shanghai Artificial Intelligence Laboratory)
Recommendation SystemOptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelAgentic AIPrompt EngineeringMixture of ExpertsTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes an scalable multi-LLM collaboration system, SMCS, which dynamically schedules and fuses answers from multiple open-source LLMs by utilizing retrieval-based prior selection and exploration-exploitation posterior enhancement mechanisms.
A Shared Geometry of Difficulty in Multilingual Language Models
Stefano Civelli (University of Queensland), Gianluca Demartini (University of Queensland)
Explainability 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.
A Structured Clustering Approach for Inducing Media Narratives
Rohan Das (University of Colorado Boulder), Maria Leonor Pacheco (University of Colorado Boulder)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: By constructing causal event chains, identifying roles and narrative roles, and performing structured clustering based on these, automatically induce interpretable media narrative patterns from large-scale news corpora.
A Survey of Deep Learning for Geometry Problem Solving
Jianzhe Ma (Renmin University of China), Qin Jin (Renmin University of China)
OptimizationRepresentation LearningData-Centric LearningGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringGenerative Adversarial NetworkTextMultimodalityGraphReview/Survey PaperBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 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)
TransformerLarge 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 Model-Based Search Agents
Yunjia Xi, Weinan Zhang (Shanghai Jiao Tong University)
Recommendation SystemAutonomous DrivingOptimizationExplainability and InterpretabilityComputational EfficiencyData-Centric LearningRobotic IntelligenceDrug DiscoveryAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIPrompt EngineeringTextMultimodalityTabularTime SeriesSequentialBiomedical DataReview/Survey PaperRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper provides a systematic review of search agents driven by large language models (LLMs), summarizing their workflow, architecture, optimization methods, application scenarios, and evaluation criteria, and proposes future research directions.
A Survey of Large Language Models for Text-Guided Molecular Discovery: From Molecule Generation to Optimization
Ziqing Wang (Northwestern University), Kaize Ding (Northwestern University)
OptimizationDrug 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 Multimodal Mathematical Reasoning: From Perception, Alignment to Reasoning
Tianyu Yang (University of Notre Dame), Xiangliang Zhang (New York University)
ClassificationRecognitionImage TranslationRestorationObject DetectionSegmentationGenerationData SynthesisPose EstimationDepth EstimationSuper ResolutionCompressionDomain AdaptationRecommendation SystemAnomaly DetectionAutonomous DrivingOptimizationFederated LearningSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningAdversarial AttackHyperparameter SearchData-Centric LearningRobotic IntelligenceMeta LearningDrug DiscoveryAI Code AssistantReinforcement Learning from Human FeedbackNeural Architecture SearchProtein Structure PredictionTransformerLarge Language ModelReinforcement LearningPrompt EngineeringVision Language ModelDiffusion modelScore-based ModelFlow-based ModelRectified FlowNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkContrastive LearningGaussian SplattingSimultaneous Localization and MappingWorld ModelOptical FlowImageTextMultimodalityTabularReview/Survey PaperBenchmarkChain-of-ThoughtStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper provides a comprehensive review of the field of Multimodal Mathematical Reasoning (MMR), constructs the Perception-Alignment-Reasoning (PAR) framework and the Answer-Process-Executable (APE) evaluation hierarchy, and systematically organizes perception, alignment, and reasoning techniques and benchmarks from three major task families: geometry, charts/tables, and visual math problems.
A Survey of Reasoning-Intensive Retrieval: Progress and Challenges
Yiyang Wei (Zhejiang University), Yilun Zhao (Yale University)
RetrievalPrompt EngineeringTextMultimodalityReview/Survey PaperBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Reviews the research progress in reasoning-enhanced retrieval (RIR), systematizing benchmarks, methods, and challenges
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)
Data-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)
ClassificationAnomaly 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.
AAPO: Enhancing the Reasoning Capabilities of LLMs with Advantage Margin
Jian Xiong (Fudan University), Dejing Dou (Fudan University)
OptimizationComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes a new reinforcement learning algorithm called Advantage-Augmented Policy Optimization (AAPO), aimed at enhancing the reasoning capabilities of large language models in chain-of-thought (CoT) tasks.
ACBQ: Adaptive Cross-Block Quantization of Large Language Models
Hailing Wang (Northeastern University), Yun Fu (Northeastern University)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: Proposed a post-training quantization framework called ACBQ, which can perform joint quantization of weights and activations as well as extremely low-bit weight quantization for large language models.
Accommodation and Epistemic Vigilance: A Pragmatic Account of Why LLMs Fail to Challenge Harmful Beliefs
Myra Cheng (Stanford University), Dan Jurafsky (Stanford University)
Safty 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.
Accurate and Efficient Statistical Testing for Word Semantic Breadth
Yo Ehara (Tokyo Gakugei University)
Computational EfficiencyRepresentation LearningTransformerContrastive LearningText
🎯 What it does: This paper proposes a permutation test method based on Householder reflection by conducting directional dispersion measurement on the context embedding vector cloud of word meanings, to accurately compare the semantic width of two words.
ACE-Router: Generalizing History-Aware Routing from MCP Tools to the Agent Web
Zhiyuan Yao (Zhejiang University), Weiwen Liu (Shanghai Jiao Tong University)
Recommendation SystemAutonomous DrivingOptimizationFederated LearningComputational EfficiencyMeta LearningReinforcement Learning from Human FeedbackGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningDiffusion modelAuto EncoderGenerative Adversarial NetworkTextGraphSequentialBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Trained and deployed ACE-Router, a history-aware router for Agent Web, which supports dynamically selecting the most suitable execution node in a massive space of tools/agents.
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)
Graph 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.
ACIArena: Toward Unified Evaluation for Agent Cascading Injection
Hengyu An (Zhejiang University), Shouling Ji (Zhejiang University)
Safty and PrivacyAdversarial AttackReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmark
🎯 What it does: Propose the ACIARENA unified framework, systematically evaluate the robustness of multi-agent systems under Agent Cascading Injection attacks, and construct 1,356 test cases;
Act as you think: Reinforcing Consistent Reasoning in Medical Visual Question Answering
Songtao Jiang (Zhejiang University), Zuozhu Liu (Zhejiang University)
Representation LearningData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelDiffusion modelContrastive LearningMultimodalityBiomedical DataBenchmark
🎯 What it does: Propose the Med-Zero-17K large-scale multimodal medical dataset and design the Multi-Consistency Reward (MCR) framework, using reinforcement learning (RLVR) to enhance trustworthy reasoning in medical visual question answering (Med-VQA).
Act-Adaptive Margin: Dynamically Calibrating Reward Models for Subjective Ambiguity
Feiteng Fang (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences), Yongbin Li (Tongyi Lab, Alibaba Group)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: Propose an adaptive margin method (AAM) based on the dynamic calibration of internal knowledge of a reward model, aimed at improving the training of reward models in subjective tasks (e.g., role-playing), eliminating the reliance on manual margin annotations;
Action Boundary Blindness: When LLM Agents Cannot Tell Where One Action Ends and Another Begins
Zhangyi Wang (Nanyang Technological University), Zongze Li (Nanyang Technological University)
Federated LearningExplainability and InterpretabilityComputational EfficiencyRepresentation LearningAdversarial AttackRobotic IntelligenceMeta LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringWorld ModelTextSequentialBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Studied the phenomenon where large language model (LLM) agents fail to correctly determine action boundaries (granularity, scope, completeness) when performing multi-step tasks, and proposed automated evaluation metrics and diagnostic methods.
Activation Decomposition and Steering for LLM Backdoor Remediation
Lingfeng Zhong (Macquarie University), Usman Naseem (Macquarie University)
Safty and PrivacyExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: A lightweight activation decomposition and guidance method named CS-ADS, which does not rely on auxiliary data or additional models, is proposed. It utilizes contrastive weak samples to decouple clean and malicious features, and suppresses backdoors during inference by incorporating clean vectors.
Activation-Guided Local Editing for Jailbreaking Attacks
Jiecong Wang (Beihang University), Zhengtao Yu (Kunming University of Science and Technology)
Safty 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)
RecognitionTransformerLarge 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.
AdaDPI: Document-level Translation Adaptive Agent via Dynamic Parametric Internalization
Hong Ren (Tianjin University), Deyi Xiong (Tianjin University)
Domain AdaptationComputational EfficiencyRepresentation LearningTransformerLarge Language ModelAgentic AIPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose AdaDPI, an adaptive agent framework that achieves context-continual learning at the document level through dynamic parameter internalization.
AdaFuse: Adaptive Ensemble Decoding for Large Language Models
Chengming Cui (University Of Illinois Urbana Champaign), Jingrui He (University Of Illinois Urbana Champaign)
GenerationComputational 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;
AdaJudge: Adaptive Multi-Perspective Judging for Reward Modeling
Yongliang Miao (Emory University), Mengnan Du (Chinese University of Hong Kong Shenzhen)
Computational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextBenchmark
🎯 What it does: Designed and implemented the AdaJudge framework, which combines adaptive representation refinement with multi-perspective aggregation to enhance the reward model's ability to judge human preferences.
Adam’s Law: Textual Frequency Law on Large Language Models
Hongyuan Lu (Chinese University of Hong Kong), Wai Lam (Chinese University of Hong Kong)
Computational EfficiencyKnowledge DistillationRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningTextRetrieval-Augmented Generation
🎯 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.
AdaMix: Adaptive Mixing for Short and Long Reasoning Adapters
Hao Luo (Wuhan University), Jiawei Jiang (Wuhan University)
Computational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningMixture of ExpertsTextBenchmarkChain-of-Thought
🎯 What it does: To address the overthinking problem in large reasoning models (LRM), the AdaMix framework is proposed. It trains two specialized adapters (a short adapter focusing on efficiency and a long adapter focusing on accuracy) and designs a difficulty-aware router. During inference, the two adapters are dynamically interpolated and mixed based on problem difficulty, achieving adaptive balance between reasoning depth and efficiency.
AdapShot: Adaptive Many-Shot In-Context Learning with Semantic-Aware KV Cache Reuse
Jie Ou (University Of Electronic Science And Technology Of China), Wenhong Tian (University Of Electronic Science And Technology Of China)
Computational EfficiencyMeta LearningTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose AdapShot, which determines the scale of diverse context through exploratory dynamic assessment for adaptability, and achieves efficient Many-Shot ICL by utilizing semantic-aware KV cache reuse and position-decoupled re-encoding.
ADAPT: Benchmarking Commonsense Planning under Unspecified Affordance Constraints
Pei-An Chen (National Taiwan University), Winston H. Hsu (National Taiwan University)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVision-Language-Action ModelImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the DynAfford benchmark and the ADAPT adaptation module to verify an agent's common-sense reasoning and task execution capabilities in dynamic, incomplete object availability environments.
Adaptive Constraint Propagation: Scaling Structured Inference for Large Language Models via Meta-Reinforcement Learning
Ibne Farabi Shihab (Iowa State University), Anuj Sharma (Iowa State University)
Meta LearningGraph Neural NetworkTransformerLarge Language ModelReinforcement LearningPrompt EngineeringMixture of ExpertsTextGraph
🎯 What it does: Propose MetaJuLS, an adaptive constraint propagation scheduler based on meta-reinforcement learning, for structured reasoning and constraint decoding in large language models.
Adaptive Instruction Composition for Automated LLM Red-Teaming
Jesse Zymet (Capital One), Emily Chen (Capital One)
Adversarial AttackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper proposes an adaptive instruction composition framework for automated LLM red team testing;
Adaptive Planning for Multi-Attribute Controllable Summarization with Monte Carlo Tree Search
Sangwon Ryu, Jungseul Ok (POSTECH)
GenerationOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningMixture of ExpertsText
🎯 What it does: Propose PACO, an untrained adaptive planning framework that gradually adjusts multi-attribute control at the summary level using a custom Monte Carlo Tree Search (MCTS), ultimately generating summaries that comply with users' multi-attribute constraints.
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)
OptimizationExplainability 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;
Adaptive Retrieval for Reasoning
Jongho Kim (Seoul National University-LG AI Research Center), Moontae Lee (LG AI Research)
RetrievalDomain 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.
Adaptive Spatial and Temporal Redundancy Optimization for Efficient Reasoning in Large Language Models
Tianle Chen (Xi'an Jiaotong University), Jiacheng Liu (Hong Kong University of Science and Technology)
OptimizationComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed the ASTRO framework, which trains freely to eliminate both temporal and spatial redundancies simultaneously in the chain-of-thought reasoning of large language models.
Adaptive Text2GQL: Integrating Structural Twig Linking and Evolutionary In-Context Learning
Fang Niu (Tsinghua University), Songyao Wang (Tsinghua University)
GenerationAdversarial AttackData-Centric LearningAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextGraphTabularBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes a training-free unified framework called Adaptive Text2GQL, which utilizes structured Twig Linking, evolutionary unsupervised In-Context learning, and adversarial execution correction to achieve automatic generation of natural language to graph queries across multiple graph query languages (Cypher, Gremlin, SPARQL).
AdvancedIF: Rubric-Based Benchmarking and Reinforcement Learning for Advancing LLM Instruction Following
Yun He (Meta Superintelligence Labs), Manaal Faruqui (Meta Superintelligence Labs)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: Proposed a new advanced instruction following benchmark, AdvancedIF, and built a rule-based reinforcement learning post-training framework, RIFL, to enhance the complex, multi-turn, and system-level instruction following capabilities of large language models.
Advancing Reasoning in Diffusion Language Models with Denoising Process Rewards
Shaoan Xie (Carnegie Mellon University), Kun Zhang (Carnegie Mellon University)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelReinforcement LearningPrompt EngineeringDiffusion modelTextChain-of-Thought
🎯 What it does: This paper proposes Denoising Process Reward (DPR), based on diffusion language models, which provides process-level rewards by evaluating the contribution of intermediate denoising steps to the final answer. It combines DPR with GRPO reinforcement learning, significantly improving multi-step reasoning performance.
Adversarial Metric Learning for Fine-Grained Emotion Classification
Junfan Chen (Beihang University), Chunming Hu (Beihang University)
ClassificationRepresentation LearningAdversarial AttackTransformerSupervised Fine-TuningGenerative Adversarial NetworkContrastive LearningText
🎯 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;
ADVICE: Answer-Dependent Verbalized Confidence Estimation
KiJung Seo, Taeuk Kim (Hanyang University)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningText
🎯 What it does: Analyze the overconfidence problem that LLMs encounter in speech-based confidence estimation, and propose the ADVICE framework, which enhances the model's reliance on answers during confidence generation through fine-tuning.
AEA: Adaptive Expert Allocation Improves Sentence Embeddings from Mixture-of-Experts LLM
Shufan Yang (Nanjing University), Qing Gu (Nanjing University)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Propose the Adaptive Expert Allocation (AEA) framework for training-free extraction of high-quality sentence embeddings in Mixture-of-Experts (MoE) language models.
AEGIS: A Holistic Benchmark for Evaluating Forensic Analysis of AI-Generated Academic Images
Bo Zhang, Haihong E (Beijing University of Posts and Telecommunications)
ClassificationRecognitionAnomaly DetectionTransformerPrompt EngineeringVision Language ModelDiffusion modelGenerative Adversarial NetworkImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed the AEIGS benchmark, constructed a multidimensional forensic evaluation system containing 8,210 high-quality academic images, and simulated four typical academic image forgery strategies.
Afri-MCQA: Multimodal Cultural Question Answering for African Languages
Atnafu Lambebo Tonja (Ucl), Thamar Solorio (Mbzuai)
TransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkAudio
🎯 What it does: This study constructs Afri-MCQA, the first multilingual and multimodal cultural visual question answering benchmark covering 15 African languages and including both text and speech modalities;
AfriqueLLM: How Data Mixing and Model Architecture Impact Continued Pre-training for African Languages
Hao Yu (McGill University), David Ifeoluwa Adelani (McGill University)
Computational EfficiencyRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringMixture of ExpertsContrastive LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Launched the AfriqueLLM series covering 20 African languages through continued pre-training on 26B tokens across multiple base models.
AFT-Tab: Adversarial Fine-Tuning for Tabular Data Synthesis with Long Text Columns
Yuhao Zhang (Harbin Institute of Technology), Chuanyi Liu (Harbin Institute of Technology)
Data SynthesisTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningGenerative Adversarial NetworkTextMultimodalityTabularBenchmark
🎯 What it does: Proposes the AFT-Tab framework for synthesizing cross-modal tabular data containing long text columns.
AG-GRPO: Answer-Guided GRPO for Masked Diffusion Language Models
Juhyeong Kim (Gachon University), Sangwoo Kang (Gachon University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningDiffusion modelText
🎯 What it does: For masked diffusion language models (dLLMs), this paper proposes Answer‑Guided Group Relative Policy Optimization (AG‑GRPO), which uses the answer as a suffix anchor during inference and re-predicts the answer after generation to build answer-guided rollouts. It also utilizes group relative advantages to jointly train answer-free and answer-guided rollouts.
AgencyBench: Benchmarking the Frontiers of Autonomous Agents in 1M-Token Real-World Contexts
Keyu Li (Shanghai Jiao Tong University), Pengfei Liu (Shanghai Jiao Tong University)
Autonomous DrivingRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringWorld ModelTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes AGENTBENCH, which includes 32 real-world scenarios and 138 tasks, evaluating six core agent capabilities, requiring an average of 90 tool calls and 1 million tokens for long-term interactions.
Agent Newsroom: Efficient Chronological Report Generation via Dynamic Multi-Agent Collaboration
Zhenhua Wang (Huazhong University of Science and Technology), Bang Wang (Huazhong University of Science and Technology)
GenerationOptimizationComputational EfficiencyTransformerLarge Language ModelAgentic AIPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Proposes MAS-TLS, a multi-agent news studio-style timeline summarization framework that can efficiently generate event reports in chronological order from document streams.
Agent-based Substructure Counting under Local Differential Privacy
Yuting Zhang (University of New South Wales), Wenjie Zhang (University of New South Wales)
Federated LearningSafty and PrivacyGraph Neural NetworkTransformerLarge Language ModelAgentic AIGraph
🎯 What it does: This paper proposes a technique for counting substructures in graphs under the condition of edge local differential privacy (ε-edge LDP), utilizing a multi-agent framework called PSC.
AgentAsk: Multi-Agent Systems Need to Ask
Bohan Lin (University of Science and Technology of China), Yang Wang (University of Science and Technology of China)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed a lightweight clarification module called AgentAsk, used for edge-level error intervention at information handover points between agents in multi-agent systems (MAS), to prevent error chain propagation.
AgentCoMa: A Compositional Benchmark Mixing Commonsense and Mathematical Reasoning in Real-World Scenarios
Lisa Alazraki (Imperial College London), Marek Rei (Imperial College London)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Researchers create the AgentCoMa benchmark to evaluate the performance of LLMs on composite tasks that combine common sense and mathematical reasoning.
AgentGL: Towards Agentic Graph Learning with LLMs via Reinforcement Learning
Yuanfu Sun (New York University Shanghai), Qiaoyu Tan (New York University Shanghai)
OptimizationRepresentation 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;
AgentGym2: Benchmarking Large Language Model Agents in De-Idealized Real-World Environments
Zhiheng Xi (Fudan University), Xuanjing Huang (Fudan University)
Autonomous DrivingOptimizationFederated LearningSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningAdversarial AttackHyperparameter SearchData-Centric LearningRobotic IntelligenceMeta LearningDrug DiscoveryAI Code AssistantReinforcement Learning from Human FeedbackNeural Architecture SearchProtein Structure PredictionTransformerLarge Language ModelAgentic AIPrompt EngineeringWorld ModelTextTabularSequentialBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Designed and implemented the AgentGym2 evaluation framework and benchmark for assessing the complete task execution capability of LLM agents in de-idealized real-world environments.
Agentic Memory: Learning Unified Long-Term and Short-Term Memory Management for Large Language Model Agents
Yi Yu (Wuhan University), Libing Wu (Wuhan University)
OptimizationTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose the Agentic Memory (AgeMem) unified framework, enabling large language model agents to autonomously manage long-term and short-term memory through tool-based actions, and achieve end-to-end optimization via a three-stage reinforcement learning approach.
Agentic Oversight via Dialectic Reasoning
Leonardo Ranaldi (University of Edinburgh), Federico Ranaldi
Federated LearningExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningAgentic AITextMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Introduces a supervised framework based on dialectical reasoning (Dialectic Argumentation), combining it with Debate/Consultancy protocols in multi-modal and multi-lingual tasks to achieve interpretable and auditable supervision;
Agentic Rubrics as Contextual Verifiers for SWE Agents
Mohit Raghavendra (Scale AI), Yunzhong He (Scale AI)
AI Code AssistantTransformerLarge Language ModelAgentic AIPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes an Agentic Rubrics validator based on context generation, which can perform fine-grained scoring of software engineering (SWE) patches without executing the code;
Agentic Very Long Video Understanding
Aniket Rege (Meta Reality Labs Research), Hyo Jin Kim (University of Wisconsin-Madison)
RetrievalRecommendation 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;
AgentMark: Utility-Preserving Behavioral Watermarking for Agents
Kaibo Huang (Beijing University of Posts and Telecommunications), Linna Zhou (Beijing University of Posts and Telecommunications)
Computational EfficiencyData-Centric LearningTransformerLarge Language ModelReinforcement LearningAgentic AIDiffusion modelScore-based ModelTextSequentialRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose AgentMark, a multi-bit watermarking framework that preserves the distribution of planning behaviors in LLM agents;
AgentOCR: Reimagining Agent History via Optical Self-Compression
Lang Feng (Nanyang Technological University), Bo An (Nanyang Technological University)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelReinforcement LearningAgentic AIVision Language ModelAuto EncoderImageTextMultimodality
🎯 What it does: Propose the AgentOCR framework, which converts the long-text history of LLM agents into compressible image representations, significantly reducing token consumption through visual tokens while maintaining task performance.
AgentRouter: A Knowledge-Graph-Guided LLM Router for Collaborative Multi-Agent Question Answering
Zheyuan Zhang (University Of Notre Dame), Yanfang Ye (University Of Notre Dame)
Recommendation SystemExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningReinforcement Learning from Human FeedbackGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextGraphBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose AgentRouter, a multi-agent question-answering routing framework based on knowledge graphs;
AgentSlimming: Towards Efficient and Cost-Aware Multi-Agent Systems
Yulang Chen (Shanghai Jiao Tong University), Linfeng Zhang (Shanghai Jiao Tong University)
OptimizationComputational EfficiencyReinforcement Learning from Human FeedbackGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextGraphBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose AgentSlimming, a compression framework for graph-structured multi-agent workflows, which first identifies redundant nodes using a hybrid importance evaluation, and then prunes nodes, quantizes the model, and optionally performs MCTS fine-tuning in order of importance, significantly reducing token costs while maintaining performance.
AGSC: Adaptive Granularity and Semantic Clustering for Uncertainty Quantification in Long-text Generation
Guanran Luo (Xiamen University), Qingqiang Wu (Xiamen University)
GenerationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Proposes the AGSC framework, which uses adaptive granularity decomposition and GMM semantic clustering to achieve confidence quantification in long text generation.
AI use in American newspapers is widespread, uneven, and rarely disclosed
Jenna Russell (University of Maryland College Park), Mohit Iyyer (University of Maryland College Park)
ClassificationRecognitionTransformerLarge Language ModelPrompt EngineeringTextReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: This paper conducts a large-scale audit of 186,000 newspaper articles from the United States in the summer of 2025, using the Pangram AI detector to identify and label AI-generated or hybrid text;
AIM-CoT: Active Information-driven Multimodal Chain-of-Thought for Vision-Language Reasoning
Xiping Li (Chinese University of Hong Kong), Jianghong Ma (Chinese University of Hong Kong)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose a multi-modal chain-of-thought framework named AIM-CoT based on active information foraging and dynamic triggering, significantly enhancing the reasoning performance of VLMs in visual question answering tasks.
AIPO: Adaptive Information Guided Token-Level Reinforcement Learning for Large Language Model Reasoning
Bin Chen (University of Chinese Academy of Sciences), Furui Liu (Zhejiang Lab)
OptimizationComputational EfficiencyTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Proposes the AIPO (Adaptive Information Policy Optimization) method, which uses real-time mutual information to evaluate the influence of each token on the final answer, dynamically selects key tokens, and performs policy gradient updates only on these tokens, thereby improving the multi-step reasoning ability of large language models.
AIRCoder: Adaptive Integration of Multi-dimensional Retrieval for Repository-level Code Completion
Chuanqi Shi (Southeast University), Zhiqiang Gao (Southeast University)
AI Code AssistantTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextSequentialBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed a multi-dimensional retrieval framework called AIRCoder to achieve adaptive retrieval and fusion in code completion at the code repository level;
ALDEN: Reinforcement Learning for Active Navigation and Evidence Gathering in Long Documents
Tianyu Yang (University of Göttingen), Bela Gipp (University of Göttingen)
TransformerReinforcement 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.
Alexandria: A Multi-Domain Dialectal Arabic Machine Translation Dataset for Culturally Inclusive and Linguistically Diverse LLMs
Abdellah EL Mekki, Muhammad Abdul-Mageed (University of British Columbia)
Data-Centric LearningLarge Language ModelSupervised Fine-TuningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Built and made publicly available a large-scale, community-driven Arabic dialect machine translation dataset called Alexandria, covering 13 countries and 11 high-impact domains, containing 107,000 multi-turn dialogues with city-level dialect annotations and speaker and addressee gender configurations.
ALIGN: Word Association Learning for Cultural Alignment in Large Language Models
Chunhua Liu (University of Melbourne), Sukai Huang (Monash University)
Domain AdaptationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Using native speakers' free association data, the large language model is fine-tuned for cultural alignment, and its cultural consistency is evaluated at both the lexical and value levels.
AlignCultura: Towards Culturally Aligned Large Language Models?
Gautam Siddharth Kashyap (Macquarie University), Usman Naseem (Macquarie University)
Recommendation 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.
Aligned Multi-View Scripts for Universal Chart-to-Code Generation
Zhihan Zhang (Singapore Management University), Lizi Liao (Singapore Management University)
AI Code AssistantTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Studied multilingual chart-to-code generation, proposed the Chart2NCode dataset and the CharLuMA model, addressing the limitations of focusing only on Python;
Aligning Agents via Planning: A Benchmark for Trajectory-Level Reward Modeling
Jiaxuan Wang (Nanjing University), Lan-Zhe Guo (Nanjing University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes a benchmark called Plan-RewardBench specifically for reward modeling at the trajectory level in tool-integrated environments, and provides a complete construction and evaluation process.
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)
Representation 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.
Aligning Language Models with Real-time Knowledge Editing
Chenming Tang (Peking University), Yunfang Wu (Peking University)
Knowledge 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.
Aligning Large Language Models via Fully Self-Synthetic Data
Shangjian Yin (University of California Riverside), Yu Meng (University of Virginia)
Data SynthesisReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper proposes a fully self-synthesized LLM alignment method called SAO, which enables the model to generate diverse prompts, responses, and self-evaluate them without relying on human annotations or external reward models, thus obtaining preference data for fine-tuning.
AlignUSER: Human-Aligned LLM Agents via World Models for Recommender System Evaluation
Nicolas Bougie (Woven by Toyota), Narimawa Watanabe
Recommendation SystemReinforcement Learning from Human FeedbackTransformerLarge Language ModelGenerative Adversarial NetworkWorld ModelImageTextSequentialChain-of-Thought
🎯 What it does: Construct a synthetic user framework called ALIGNUSER based on large language models, which enables synthetic users to more realistically simulate human behavior in recommendation systems through world model pre-training and adversarial reflection.
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)
Anomaly 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.
All Languages Matter: Understanding and Mitigating Language Bias in Multilingual RAG
Dan Wang (Chinese Academy of Sciences), Le Sun (Chinese Academy of Sciences)
GenerationData SynthesisRetrievalRecommendation SystemTransformerSupervised Fine-TuningPrompt EngineeringTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Investigate language bias in multilingual retrieval-augmented generation (mRAG), quantify its impact on generation performance, and propose a two-stage training framework called LAURA based on generated Utility, significantly reducing language bias and improving generation quality.
All That Glisters Is Not Gold: A Benchmark for Reference-Free Counterfactual Financial Misinformation Detection
Yuechen Jiang, Sophia Ananiadou (University Of Manchester)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkFinance RelatedRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Constructed RFC-BENCH, a benchmark for detecting misleading financial text, providing paragraph-level 'reference-free' misleading judgment and 'comparative diagnosis' tasks, and evaluated multiple large language models on this benchmark.