ACL 2025 Papers — Page 7
Annual Meeting of the Association for Computational Linguistics · 1699 papers
FactBench: A Dynamic Benchmark for In-the-Wild Language Model Factuality Evaluation
Farima Fatahi Bayat (University of Michigan), Lu Wang (University of Michigan)
Data-Centric LearningTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed the VERIFY evaluation pipeline and FACTBENCH dynamic benchmark for assessing the factualness of language models in real user interactions.
FairI Tales: Evaluation of Fairness in Indian Contexts with a Focus on Bias and Stereotypes
Janki Atul Nawale (Nilekani Centre at AI4Bharat), Mitesh M Khapra (Nilekani Centre at AI4Bharat)
Large Language ModelAgentic AIPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Proposed INDIC-BIAS, a benchmark for evaluating fairness in large language models (LLMs) within the Indian context, covering four identity categories: caste, religion, region, and tribe. It constructed over 20,000 real-world scenario templates reviewed by experts and assessed LLM bias and stereotypes across three tasks (feasibility, judgment, and generation).
Fairness Beyond Performance: Revealing Reliability Disparities Across Groups in Legal NLP
Santosh T.y.s.s, Irtiza Chowdhury (Technical University of Munich)
ClassificationTransformerTextBenchmark
🎯 What it does: This paper systematically evaluates the FairLex legal NLP benchmark to investigate differences in model accuracy and reliability (i.e., self-calibration and abstention behavior) across different groups, and proposes that fairness should not only focus on performance balance but also consider reliability equality.
Fairness through Difference Awareness: Measuring \textit{Desired} Group Discrimination in LLMs
Angelina Wang (Stanford University), Sanmi Koyejo (Stanford University)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes two metrics for measuring the fairness of large language models: Difference Awareness and Contextual Awareness, and constructs eight benchmark datasets containing 16,000 questions covering descriptive and prescriptive tasks;
Faithful and Robust LLM-Driven Theorem Proving for NLI Explanations
Xin Quan, Andre Freitas
Explainability and InterpretabilityLarge Language ModelText
🎯 What it does: Built the Faithful-Refiner framework, improving the interaction between large language models and theorem provers to verify and enhance NLI explanations through syntax parsing, quantifier consistency checks, logic expression-guided proof sketches, and detailed feedback.
FaithfulRAG: Fact-Level Conflict Modeling for Context-Faithful Retrieval-Augmented Generation
Qinggang Zhang (Xiamen University), Jinsong Su (Xiamen University)
GenerationTransformerPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the FaithfulRAG framework, which dynamically harmonizes through self-fact mining and self-reasoning to resolve knowledge conflicts in retrieval-augmented generation, thereby enhancing answer authenticity and consistency.
Fast or Slow? Integrating Fast Intuition and Deliberate Thinking for Enhancing Visual Question Answering
Songtao Jiang (Zhejiang University), Zuozhu Liu (Zhejiang University)
Computational EfficiencyTransformerPrompt EngineeringVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: To address the reasoning bottleneck of multimodal large language models in visual question answering (VQA), the FOCUS method is proposed, which can dynamically switch between two thinking modes: Fast Intuition and Deliberate Thinking, based on question difficulty. Additionally, it employs a 'Conceptualizing before Observation' strategy to extract key information and perform visual grounding when necessary, thereby improving the model's answer accuracy and reasoning efficiency.
Faster Speculative Decoding via Effective Draft Decoder with Pruned Candidate Tree
Huanran Zheng (East China Normal University), Xiaoling Wang (East China Normal University)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: Proposed two improvement methods: Effective Draft Decoder (EDD) utilizes large language models (LLMs) as encoders to generate soft prompts, enhancing the accuracy of the draft model; Pruned Candidate Tree (PCT) dynamically prunes the candidate tree based on confidence and expected time gain, reducing redundant computations and significantly improving the acceleration effect of speculative decoding.
FastMCTS: A Simple Sampling Strategy for Data Synthesis
Peiji Li (Fudan University), Qipeng Guo (Shanghai AI Laboratory)
Data SynthesisLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: Propose FastMCTS, an efficient multi-step reasoning data synthesis method based on MCTS;
FCMR: Robust Evaluation of Financial Cross-Modal Multi-Hop Reasoning
Seunghee Kim (Hanyang University), Taeuk Kim (Hanyang University)
Large Language ModelPrompt EngineeringImageTextMultimodalityTabularBenchmarkFinance RelatedChain-of-Thought
🎯 What it does: Proposed the financial cross-modal multi-hop reasoning benchmark FCMR, performing three-hop reasoning across text, tables, and charts, constructing easy/medium/hard three-level questions, and designing a multiple-choice answer format.
FEA-Bench: A Benchmark for Evaluating Repository-Level Code Generation for Feature Implementation
Wei Li (Peking University), Scarlett Li (Peking University)
AI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed FEA-Bench, a benchmark task set for evaluating large language models' ability to implement new features (i.e., adding new components) at the code repository level.
FEAT: A Preference Feedback Dataset through a Cost-Effective Auto-Generation and Labeling Framework for English AI Tutoring
Hyein Seo (Chungnam National University), Sangkeun Jung
Data-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelText
🎯 What it does: Propose the FEAT framework, which leverages large language models to automatically generate and annotate teacher feedback, constructing three preference datasets (DIRECT-Manual, DIRECT-Generated, DIRECT-Augmented) to support AI training based on rewards or rankings.
FedEx-LoRA: Exact Aggregation for Federated and Efficient Fine-Tuning of Large Language Models
Raghav Singhal (Mohamed bin Zayed University of Artificial Intelligence), Praneeth Vepakomma (Mohamed bin Zayed University of Artificial Intelligence)
Federated LearningTransformerLarge Language ModelText
🎯 What it does: Propose FedEx-LoRA, achieving precise aggregation of LoRA adapters in a federated learning environment, addressing the issue of inconsistent low-rank updates caused by traditional FedAvg.
FinanceReasoning: Benchmarking Financial Numerical Reasoning More Credible, Comprehensive and Challenging
Zichen Tang (Beijing University of Posts and Telecommunications), Qianhe Zheng (Beijing University of Posts and Telecommunications)
TextTabularBenchmarkFinance RelatedRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose FinanceReasoning, a high-quality, scalable, and challenging numerical reasoning benchmark in the financial domain, containing re-annotated public data, 908 newly generated and manually reviewed questions, and 3,133 Python function libraries;
Finding A Voice: Exploring the Potential of African American Dialect and Voice Generation for Chatbots
Sarah E. Finch (Emory University), Jinho D. Choi (Emory University)
GenerationTransformerLarge Language ModelPrompt EngineeringDiffusion modelTextAudio
🎯 What it does: This paper evaluates the affinity and functional performance of a multi-level African American English (AAE) text and spoken chatbot for AAE users by constructing it;
Finding Needles in Images: Can Multi-modal LLMs Locate Fine Details?
Parth Thakkar (Fujitsu Research India), Chaitanya Devaguptapu (Fujitsu Research India)
RetrievalLarge Language ModelVision Language ModelMultimodalityBenchmark
🎯 What it does: Proposed the NiM-Benchmark dataset for fine-grained information retrieval in document visual question answering, and developed the Spot-IT method
Finding the Sweet Spot: Preference Data Construction for Scaling Preference Optimization
Yao Xiao (Singapore University of Technology and Design), Roy Ka-Wei Lee (Singapore University of Technology and Design)
OptimizationReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningText
🎯 What it does: Proposed a preference data construction method based on reward distribution, and further introduced a scalable sampling strategy to enhance the alignment effectiveness of DPO (Direct Preference Optimization) in large-scale scenarios.
Fine-grained Video Dubbing Duration Alignment with Segment Supervised Preference Optimization
Chaoqun Cui (Alibaba Digital Media and Entertainment Group), Xiaofeng Liu (Huazhong University of Science and Technology)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningVideoTextAudio
🎯 What it does: To address the duration mismatch between source and target languages in video dubbing, a fine-grained paragraph preference optimization method is proposed to achieve duration alignment
Fine-Tuning on Diverse Reasoning Chains Drives Within-Inference CoT Refinement in LLMs
Haritz Puerto (TU Darmstadt), Iryna Gurevych (TU Darmstadt)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Investigated a fine-tuning method that generates multiple diverse reasoning chains (DCoT) in a single inference step, utilizing the generated chains to iteratively refine internal reasoning chains.
FineReason: Evaluating and Improving LLMs’ Deliberate Reasoning through Reflective Puzzle Solving
Guizhen Chen (Nanyang Technological University), Yu Rong (Alibaba Group)
TransformerReinforcement LearningTextTabularBenchmarkChain-of-Thought
🎯 What it does: This paper constructs the FINEREASON logic puzzle benchmark, using state checking and state transition tasks to finely evaluate the deep reasoning capabilities of LLMs.
Finite State Automata Inside Transformers with Chain-of-Thought: A Mechanistic Study on State Tracking
Yifan Zhang (Peking University), Zhi Jin (Peking University)
Explainability and InterpretabilityTransformerTextChain-of-Thought
🎯 What it does: Investigated and verified the learnability of Transformer+CoT in state tracking tasks, using mechanism interpretation techniques to reveal its internal formation of an implicit finite state machine (FSA) and assess its robustness in challenging scenarios such as skipping steps, noise, and length generalization.
FinMME: Benchmark Dataset for Financial Multi-Modal Reasoning Evaluation
Junyu Luo (Peking University), Yike Guo (Hong Kong University of Science and Technology)
TransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmarkFinance Related
🎯 What it does: Proposed and constructed the FINMME financial multimodal evaluation benchmark, covering 18 core financial fields, 6 asset categories, 10 types of charts, and 21 subcategories, containing over 11,000 high-quality samples, and designed the FinScore assessment framework;
Fixing Distribution Shifts of LLM Self-Critique via On-Policy Self-Play Training
Rong Bao (Fudan University), Minpeng Liao (Tongyi Lab)
Reinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningText
🎯 What it does: Propose an SCOP framework that synchronizes the training of reasoning, criticism, and correction processes through model self-dialogue, addressing distribution shift and reward hijacking issues.
FlashAudio: Rectified Flow for Fast and High-Fidelity Text-to-Audio Generation
Huadai Liu (Hong Kong University of Science and Technology), Wei Xue (Hong Kong University of Science and Technology)
GenerationFlow-based ModelRectified FlowAuto EncoderAudio
🎯 What it does: Designed and trained the FlashAudio model, utilizing rectified flow to achieve high-fidelity audio generation with one or very few steps under text input.
Flexora: Flexible Low-Rank Adaptation for Large Language Models
Chenxing Wei (Shenzhen University), Fei Yu (Shenzhen University)
OptimizationComputational EfficiencyHyperparameter SearchTransformerLarge Language ModelText
🎯 What it does: This paper proposes Flexora, a method for automatically selecting key layers for LoRA fine-tuning;
Flipping Knowledge Distillation: Leveraging Small Models’ Expertise to Enhance LLMs in Text Matching
Mingzhe Li (ByteDance), Xiuying Chen (Mohamed bin Zayed University of Artificial Intelligence)
Knowledge DistillationRepresentation LearningTransformerLarge Language ModelContrastive LearningTextFinance Related
🎯 What it does: Proposes a reverse knowledge distillation method, enabling large language models (LLM) to learn text matching representations from small language models (SLM).
FloorPlan-LLaMa: Aligning Architects’ Feedback and Domain Knowledge in Architectural Floor Plan Generation
Jun Yin (Tsinghua University), Shuai Lu (Tsinghua University)
GenerationReinforcement Learning from Human FeedbackTransformerReinforcement LearningVision Language ModelAuto EncoderImageTextMultimodality
🎯 What it does: This paper proposes the autoregressive model FP-LLaMa and achieves RLHF alignment in generating architectural floor plans by introducing the ArchiMetricsNet dataset and the FloorPlan-MPS multi-dimensional preference scoring model, thereby enhancing the professional feasibility of the generated designs.
FocalPO: Enhancing Preference Optimizing by Focusing on Correct Preference Rankings
Tong Liu (LMU Munich), Volker Tresp (LMU Munich)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Propose a new preference optimization algorithm called FocalPO, which improves DPO's attention to misordered samples during training;
Focus on What Matters: Enhancing Medical Vision-Language Models with Automatic Attention Alignment Tuning
Aofei Chang (Pennsylvania State University), Fenglong Ma (Pennsylvania State University)
TransformerSupervised Fine-TuningPrompt EngineeringMixture of ExpertsVision Language ModelBiomedical Data
🎯 What it does: Propose the A3TUNE fine-tuning framework, which automatically generates prompt-aware weak labels using SAM + BioMedCLIP, combines LoRA with A3MOE, and only adjusts the visual key attention heads to achieve adaptive alignment of visual attention, thereby improving the visual anchoring and overall performance of Med-LVLM.
FOCUS: Evaluating Pre-trained Vision-Language Models on Underspecification Reasoning
Kankan Zhou (Singapore Management University), Jing Jiang (Singapore Management University)
Explainability and InterpretabilityVision Language ModelMultimodalityBenchmark
🎯 What it does: The study evaluates the reasoning ability of vision-language models in visual contexts for implicit sentences (with ambiguity), constructs the FOCUS dataset, and designs three probing question-answering scenarios.
FocusLLM: Precise Understanding of Long Context by Dynamic Condensing
Zhenyu Li (Tsinghua University), Jianyong Wang (Tsinghua University)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose the FocusLLM framework, which extends the context length of large models by dividing long texts into multiple blocks, dynamically injecting context and generating candidate tokens after each block, and then aggregating information through parallel decoding while preserving information integrity.
FoldMoE: Efficient Long Sequence MoE Training via Attention-MoE Pipelining
Guichao Zhu (University of Hong Kong), Heming Cui (University of Hong Kong)
Computational EfficiencyTransformerMixture of ExpertsText
🎯 What it does: Propose the FOLDMOE system, constructing an attention-MoE pipeline within the Transformer-MoE block to overlap All-to-All communication and computation, thereby accelerating long-sequence MoE training.
Follow-up Question Generation For Enhanced Patient-Provider Conversations
Joseph Gatto (Dartmouth), Sarah M. Preum (Dartmouth)
GenerationTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBiomedical DataElectronic Health Records
🎯 What it does: Propose a multi-agent framework FollowupQ that generates subsequent questions in asynchronous medical dialogues by leveraging three parallel approaches: EHR reasoning, differential diagnosis, and information clarification, significantly reducing the volume of follow-up communication required from physicians.
Forward Knows Efficient Backward Path: Saliency-Guided Memory-Efficient Fine-tuning of Large Language Models
Yeachan Kim (Korea University), SangKeun Lee (Korea University)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes a sparse gradient flow (SAGE) method based on activation saliency to significantly reduce the intermediate activation cache memory usage during fine-tuning of large language models (LLMs) while maintaining task performance.
FR-Spec: Accelerating Large-Vocabulary Language Models via Frequency-Ranked Speculative Sampling
Weilin Zhao (Tsinghua University), Maosong Sun (Tsinghua University)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: To accelerate inference for large-vocabulary LLMs, the FR-Spec framework is proposed, which compresses the LM Head computation by using only a subset of high-frequency words during the draft phase, significantly improving draft efficiency while maintaining the final output distribution unchanged.
FRACTAL: Fine-Grained Scoring from Aggregate Text Labels
Yukti Makhija (Google DeepMind), Aravindan Raghuveer (Google DeepMind)
Representation LearningData-Centric LearningReinforcement Learning from Human FeedbackContrastive LearningText
🎯 What it does: Propose a method called FRACTAL, which can split the limited overall response-level labels into sentence-level pseudo labels and further train the model using these pseudo labels;
Frictional Agent Alignment Framework: Slow Down and Don’t Break Things
Abhijnan Nath (Colorado State University), Nikhil Krishnaswamy (Colorado State University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: Propose a multi-player optimization framework named Frictional Agent Alignment Framework (FAAF) to generate 'friction'-type interventions that promote reflection and reasoning in dialogue collaboration tasks.
from Benign import Toxic: Jailbreaking the Language Model via Adversarial Metaphors
Yu Yan (Institute of Computing Technology, Chinese Academy of Sciences), Qi Li (Tsinghua University)
Adversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposed a black-box jailbreak attack framework AVATAR that gradually induces large language models (LLMs) to generate harmful content by utilizing safe and harmless metaphors.
From Citations to Criticality: Predicting Legal Decision Influence in the Multilingual Swiss Jurisprudence
Ronja Stern (University of Bern), Joel Niklaus (University of Bern)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: This paper proposes a new multilingual legal case priority prediction task called Criticality Prediction, which uses algorithm-generated Leading Decision (LD) and Citation tags to measure case impact.
From English to Second Language Mastery: Enhancing LLMs with Cross-Lingual Continued Instruction Tuning
Linjuan Wu (Zhejiang University), Weiming Lu (Zhejiang University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a cross-lingual continual instruction tuning (X-CIT) method, which first fine-tunes large language models on English instruction data, and then further fine-tunes them using parallel translation instructions and two-round dialogues cross-lingual instruction data to enhance the model's capabilities in non-English languages.
From Human Reading to NLM Understanding: Evaluating the Role of Eye-Tracking Data in Encoder-Based Models
Luca Dini (Istituto di Linguistica Computazionale 'Antonio Zampolli' (CNR-ILC), ItaliaNLP Lab, Pisa), Felice Dell’Orletta
Explainability and InterpretabilityTransformerSupervised Fine-TuningTextBiomedical Data
🎯 What it does: This study investigates injecting eye-tracking data into the RoBERTa model and conducts systematic evaluation across three dimensions: downstream task performance, attention alignment, and embedding space compression.
From Informal to Formal – Incorporating and Evaluating LLMs on Natural Language Requirements to Verifiable Formal Proofs
Jialun Cao (Hong Kong University of Science and Technology), Cong Tian (Xidian University)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Constructed 18k instruction-response pairs between natural language and formal verification code, forming the FM-ALPACA and FM-BENCH datasets, and evaluated LLM performance on six subtasks of formal verification.
From Information to Insight: Leveraging LLMs for Open Aspect-Based Educational Summarization
Yang Zhong (University of Pittsburgh), Diane Litman (University of Pittsburgh)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Construct an open-ended aspect-based summary (OABS) dataset named ReflectASP in the education field to evaluate the capability of large language models (LLMs) in generating aspect-based summaries under zero-shot scenarios, and propose two improved frameworks, E2A and E2A w/ MC-Refine, based on extraction-abstraction and fact-checking, with a fine-grained analysis of editing intentions during the improvement process.
From Isolates to Families: Using Neural Networks for Automated Language Affiliation
Frederic Blum (University of Passau), Johann-Mattis List (University of Passau)
ClassificationText
🎯 What it does: Train a neural network on lexical and grammatical data of over 1200 classified languages to automatically assign languages to their language families, predict family affiliations for unclassified or isolated languages, and test cross-generational relationships.
From Lists to Emojis: How Format Bias Affects Model Alignment
Xuanchang Zhang (University of Illinois Urbana-Champaign), Tong Zhang (University of Illinois Urbana-Champaign)
Reinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Study the bias of models in RLHF regarding formatting elements (such as bold, lists, emojis, etc.), evaluate its impact on reward models and downstream alignment, and propose a debiasing method combining two-headed reward models with relevance constraints.
From Neurons to Semantics: Evaluating Cross-Linguistic Alignment Capabilities of Large Language Models via Neurons Alignment
Chongxuan Huang (Xiamen University), Xiaodong Shi (Xiamen University)
RetrievalExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Evaluate the cross-lingual alignment capability of large language models and propose NeuronXA, an assessment method based on neuron activation.
From Objectives to Questions: A Planning-based Framework for Educational Mathematical Question Generation
Cheng Cheng (University Of Science And Technology Of China), Shijin Wang (State Key Laboratory Of Cognitive Intelligence)
GenerationOptimizationTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: This paper proposes an educational math problem generation framework EQPR based on the plan-evaluation-optimization cycle, which automatically generates math problems that meet multidimensional educational objectives.
From Outcomes to Processes: Guiding PRM Learning from ORM for Inference-Time Alignment
Bin Xie (Chinese Academy of Sciences), Huawei Shen (Chinese Academy of Sciences)
Reinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes an SP-PRM framework to address the granularity mismatch between partial and complete sequence evaluations in reward model training by incorporating process reward modeling during reward-guided search.
From Perceptions to Decisions: Wildfire Evacuation Decision Prediction with Behavioral Theory-informed LLMs
Ruxiao Chen (Johns Hopkins University), Susu Xu (University of Florida)
TransformerLarge Language ModelReinforcement LearningTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Developed FLARE, an LLM framework integrating behavioral theory for predicting wildfire evacuation decisions
From Real to Synthetic: Synthesizing Millions of Diversified and Complicated User Instructions with Attributed Grounding
Chiwei Zhu (University of Science and Technology of China), Zhendong Mao (Metastone Technology)
Data SynthesisTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes a two-step framework for synthesizing instruction data, first assigning documents, users, and motivations to real instructions, and then generating contexts and instructions from web documents, constructing a high-quality instruction dataset named SYNTHQUESTIONS with 1 million entries.
From Selection to Generation: A Survey of LLM-based Active Learning
Yu Xia (University of California San Diego), Julian McAuley (University of Oregon)
Data-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextReview/Survey PaperChain-of-Thought
🎯 What it does: This paper reviews the application of large language models (LLMs) in active learning, systematically explaining how LLMs play a role in two core aspects: query (selection/generation) and annotation, and providing a unified classification framework, case studies, and future challenges.
From Sub-Ability Diagnosis to Human-Aligned Generation: Bridging the Gap for Text Length Control via MarkerGen
Peiwen Yuan (Beijing Institute of Technology), Kan Li (Beijing Institute of Technology)
GenerationTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose the MARKERGEN method, which first decomposes the sub-capabilities of Length-Controllable Text Generation (LCTG) and conducts error analysis at the foundational level, and based on this, designs external tool assistance, dynamic length marker insertion, and a three-stage decoupled generation strategy to significantly enhance the performance of large language models in length-controllable generation.
From Tools to Teammates: Evaluating LLMs in Multi-Session Coding Interactions
Nathanaël Carraz Rakotonirina (Universitat Pompeu Fabra), Marco Del Tredici (Cohere)
Data SynthesisAI Code AssistantTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposed the MEMORYCODE dataset to evaluate the ability of large language models to track and execute coding instructions in multi-turn conversations.
From Trade-off to Synergy: A Versatile Symbiotic Watermarking Framework for Large Language Models
Yidan Wang (Institute of Information Engineering, Chinese Academy of Sciences), Binxing Fang (Hainan Province Fang Binxing Academician Workstation)
Safty and PrivacyTransformerLarge Language ModelText
🎯 What it does: Propose SymMark, a compositional watermarking framework for large language models, integrating logits-based and sampling-based watermarks, and designing serial, parallel, and hybrid strategies.
Fusing Highly Specialized Language Models for Comprehensive Expertise
Ning Ding (Tsinghua University), Maosong Sun (Tsinghua University)
Data SynthesisAI Code AssistantLarge Language ModelSupervised Fine-TuningMixture of ExpertsTextBenchmark
🎯 What it does: Proposed and implemented the ULTRAFUSER framework, which directly integrates high-precision specialized models in the domains of text, programming code, and mathematical reasoning, and constructed a 300k multi-domain instruction-response dataset named ULTRACHAT 2 for training and evaluation.
G-Safeguard: A Topology-Guided Security Lens and Treatment on LLM-based Multi-agent Systems
Shilong Wang (University of Science and Technology of China), Yang Wang (University of Science and Technology of China)
Anomaly DetectionSafty and PrivacyGraph Neural NetworkLarge Language ModelText
🎯 What it does: Designed and implemented a framework named G-Safeguard to construct multi-agent speech graphs in large language model (LLM)-driven multi-agent systems (MAS), detecting attacks and pruning transmission edges through graph neural networks (GNNs) to identify and neutralize attackers.
GainRAG: Preference Alignment in Retrieval-Augmented Generation through Gain Signal Synthesis
Yi Jiang (Harbin Institute of Technology), Bing Qin (Harbin Institute of Technology)
GenerationRetrievalTransformerContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Propose the GainRAG framework, which first calculates the 'gain' metric through contrastive decoding and perplexity, then trains a selector using a small number of synthetic samples, and combines pseudo-document strategies to perform refined filtering of retrieval results, thereby achieving preference alignment between the retrieval module and large language models in knowledge utilization;
GALLa: Graph Aligned Large Language Models for Improved Source Code Understanding
Ziyin Zhang (Shanghai Jiao Tong University), Rui Wang (Shanghai Jiao Tong University)
Representation LearningAI Code AssistantGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningTextGraph
🎯 What it does: By using graph neural networks and alignment tasks during the pre-training phase, the structured graphs of code (AST, DFG) are injected into large language models to enhance source code understanding.
Game Development as Human-LLM Interaction
Jiale Hong (Shanghai Jiao Tong University), Hai Zhao (Shanghai Jiao Tong University)
GenerationData SynthesisAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose Chat Game Engine (ChatGE), enabling large language models (LLM) to complete game development through natural language interaction, with implementation and evaluation on poker (Texas Hold'em).
GAMEBoT: Transparent Assessment of LLM Reasoning in Games
Wenye Lin (University of Hong Kong), Kai Han (University of Hong Kong)
Explainability and InterpretabilityTransformerPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Introduce the GAMEBOT benchmark, which uses a game environment with chain-of-thought (CoT) prompts and rule-based subproblems to transparently evaluate the reasoning process of large language models (LLMs) in strategy games.
GAPO: Learning Preferential Prompt through Generative Adversarial Policy Optimization
Zhouhong Gu (Fudan University), Yanghua Xiao (Fudan University)
Reinforcement Learning from Human FeedbackReinforcement LearningPrompt EngineeringGenerative Adversarial NetworkTextBenchmark
🎯 What it does: Proposes the GAPO framework, combining GAN and PPO, and using an encoder-only reward model to learn preference prompts, enhancing the constraint-following ability of large language models.
GeLLM³O: Generalizing Large Language Models for Multi-property Molecule Optimization
Vishal Dey (Ohio State University), Xia Ning (Ohio State University)
OptimizationDrug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningTextGraph
🎯 What it does: Proposed the MuMOInstruct dataset and trained the GeLLM O 3 series large language models (LLMs) based on it for multi-attribute molecular optimization tasks.
Gender Inclusivity Fairness Index (GIFI): A Multilevel Framework for Evaluating Gender Diversity in Large Language Models
Zhengyang Shan, Jiawei Zhou
TransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes the Gender Inclusivity Fairness Index (GIFI) framework to systematically evaluate the performance of large language models (LLMs) in terms of gender diversity (including non-binary gender).
GenderAlign: An Alignment Dataset for Mitigating Gender Bias in Large Language Models
Tao Zhang (South China University of Technology), Shimei Pan (University of Maryland, Baltimore County)
Data-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Constructed a public gender bias alignment dataset called GenderAlign, containing 8,000 single-round human-machine dialogues, for training and evaluating the gender fairness of large language models.
Generalized Attention Flow: Feature Attribution for Transformer Models via Maximum Flow
Behrooz Azarkhalili (Simon Fraser University), Maxwell W. Libbrecht (Simon Fraser University)
ClassificationExplainability and InterpretabilityTransformerFlow-based ModelText
🎯 What it does: Proposes a Transformer feature attribution method called Generalized Attention Flow (GAF), which constructs an information tensor using attention weights and their gradients, and generates feature importance scores by solving a max-flow problem with log-barrier regularization.
Generate First, Then Sample: Enhancing Fake News Detection with LLM-Augmented Reinforced Sampling
Zhao Tong (Institute of Information Engineering Chinese Academy of Sciences), Xiao-Yu Zhang (Institute of Information Engineering Chinese Academy of Sciences)
ClassificationData SynthesisTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Leverage large language models (LLM) to generate multi-perspective fake news and dynamically adjust the ratio of real/fake news in the training set through reinforcement learning to enhance fake news detection performance.
Generating Diverse Training Samples for Relation Extraction with Large Language Models
Zexuan Li (Nanjing University of Aeronautics and Astronautics), Piji Li (Nanjing University of Aeronautics and Astronautics)
Data SynthesisTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: By incorporating a 'diversity' instruction in the prompts and adopting a per-sample generation approach, combined with Direct Preference Optimization (DPO) fine-tuning, large language models generate relation extraction training samples that are both diverse and accurate, which are then used to train few-shot relation extraction models.
Generating Plausible Distractors for Multiple-Choice Questions via Student Choice Prediction
Yooseop Lee (Seoul National University), Yohan Jo (Seoul National University)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Designed a contrastive sorter trained on student selection data and a suspicious distractor generator based on DPO to automatically generate more suspicious multiple-choice distractors.
Generative Psycho-Lexical Approach for Constructing Value Systems in Large Language Models
Haoran Ye (Peking University), Guojie Song (Peking University)
Data-Centric LearningReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningAgentic AITextBenchmark
🎯 What it does: The study proposes a Generative Psychological Lexical Approach (GPLA) to automatically construct a value system for large language models (LLMs), building a five-factor psychological value system (social responsibility, risk-taking, compliance, personal competitiveness, rationality). It evaluates the system's structure, predictive safety, and alignment capabilities through three benchmark tasks.
Generative Reward Modeling via Synthetic Criteria Preference Learning
Xiaobo Liang (Soochow University), Min Zhang (Soochow University)
Data SynthesisReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: Built a tree-based preference learning framework called SyncPL based on synthetic evaluation criteria, used for fine-grained process supervision and long CoT training in generative reward models (GenRM), significantly improving the model's consistency with human preferences.
Genius: A Generalizable and Purely Unsupervised Self-Training Framework For Advanced Reasoning
Fangzhi Xu (Xi'an Jiaotong University), Zhiyong Wu (Shanghai AI Lab)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: Propose a fully unsupervised self-training framework called Genius, which generates, evaluates, and optimizes reasoning sequences by leveraging the LLM's own reasoning steps to enhance the reasoning capabilities of large language models.
GenKnowSub: Improving Modularity and Reusability of LLMs through General Knowledge Subtraction
Mohammadtaha Bagherifard (Tehran Institute for Advanced Studies Khatam University), Yadollah Yaghoobzadeh (University of Tehran)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed a method called GenKnowSub that subtracts general knowledge LoRA from task-specific LoRA, combined with the Arrow routing algorithm to achieve zero-shot transfer;
Geometric Signatures of Compositionality Across a Language Model’s Lifetime
Jin Hwa Lee (UCL), Emily Cheng (Universitat Pompeu Fabra)
Data SynthesisRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Investigated how language models reflect input composability through geometric features of internal representations during the training lifecycle, computing nonlinear intrinsic dimensions and linear dimensions using controlled datasets and natural corpora.
GETReason: Enhancing Image Context Extraction through Hierarchical Multi-Agent Reasoning
Shikhhar Siingh (Arizona State University), Vivek Gupta (Arizona State University)
RecognitionAgentic AIVision Language ModelImage
🎯 What it does: Constructed a multi-layered multi-agent framework called GETREASON, specifically designed for extracting events, time, and geographic information from public event images.
GIFT-SW: Gaussian noise Injected Fine-Tuning of Salient Weights for LLMs
Maxim Zhelnin (Skolkovo Institute of Science and Technology), Evgeny Burnaev (Skolkovo Institute of Science and Technology)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose a parameter-efficient fine-tuning method called GIFT-SW, which only fine-tunes the weights of a small number of significant columns and injects Gaussian noise into non-significant columns, helping the model maintain performance in full precision and after quantization.
GiFT: Gibbs Fine-Tuning for Code Generation
Haochen Li (Nanyang Technological University), Zhiqi Shen (Nanyang Technological University)
GenerationAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose the GiFT method, which samples code from the marginal distribution in the description-code joint space during self-training of code generation via Gibbs sampling, and employs perplexity-guided sampling strategies to select training samples.
GigaSpeech 2: An Evolving, Large-Scale and Multi-domain ASR Corpus for Low-Resource Languages with Automated Crawling, Transcription and Refinement
Yifan Yang (Shanghai Jiao Tong University), Xie Chen (Shanghai Jiao Tong University)
RecognitionTransformerBenchmarkAudio
🎯 What it does: Constructed a large-scale, multi-domain, multi-lingual low-resource language ASR corpus named GigaSpeech 2, and completed the process from YouTube audio scraping, transcription, alignment, filtering to label refinement through an automated pipeline, ultimately obtaining approximately 30k hours of automatically transcribed data (raw) and refined data of approximately 10k hours of Thai, 6k hours of Indonesian, and 6k hours of Vietnamese.
Global Eye: Breaking the “Fixed Thinking Pattern” during the Instruction Expansion Process
Wenxuan Lu (Institute of Information Engineering, Chinese Academy of Sciences), Tianning Zang (Institute of Information Engineering, Chinese Academy of Sciences)
GenerationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposed a dynamic prompt updating instruction expansion method called Global Eye to break the 'fixed mindset' of large language models in instruction generation, thereby enhancing the diversity of generated instructions.
Global MMLU: Understanding and Addressing Cultural and Linguistic Biases in Multilingual Evaluation
Shivalika Singh (Cohere Labs), Sara Hooker (Cohere Labs)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper constructs the Global-MMLU benchmark covering 42 languages through large-scale human and community translation with post-editing, and adds cultural sensitivity annotations to the original MMLU. Subsequently, the performance of 14 large language models on the cultural neutral (CA) and culturally sensitive (CS) subsets is evaluated, revealing significant differences in model rankings and performance across subsets.
GODBench: A Benchmark for Multimodal Large Language Models in Video Comment Art
Yiming Lei (Beihang University), Yunhong Wang (Beihang University)
GenerationTransformerLarge Language ModelVideoTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Constructed a large-scale multimodal benchmark GODBench covering 67,000 videos and GOD-level comments, and proposed the Ripple of Thought (RoT) framework to enhance the video comment creativity of Multimodal Large Language Models (MLLMs).
Gödel Agent: A Self-Referential Agent Framework for Recursively Self-Improvement
Xunjian Yin (Peking University), William Yang Wang (University of California, Santa Barbara)
Large Language ModelAgentic AIText
🎯 What it does: Proposes the Godel Agent framework, enabling agents to dynamically read and modify their own code during reasoning to achieve recursive self-improvement.
GPT-4 as a Homework Tutor Can Improve Student Engagement and Learning Outcomes
Alessandro Vanzo (University of Zurich), Mrinmaya Sachan (ETH Zurich)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A randomized controlled trial (RCT) was conducted over 8 weeks in four high school classes at a technical college, using GPT-4 to generate interactive homework assignments in place of traditional homework to evaluate its impact on students' learning experiences and academic performance.
GRACE: A Granular Benchmark for Evaluating Model Calibration against Human Calibration
Yoo Yeon Sung (University of Maryland College Park), Jordan Lee Boyd-Graber
Large Language ModelTextBenchmark
🎯 What it does: Developed the GRACE benchmark, designing incremental clue-based question answering to evaluate differences in calibration between language models and humans.
Gradient-Adaptive Policy Optimization: Towards Multi-Objective Alignment of Large Language Models
Chengao Li (Key Lab of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences), Qing He (Key Lab of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences)
OptimizationReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningText
🎯 What it does: Propose Gradient-Adaptive Policy Optimization (GAPO) and its user-preference extension P-GAPO, utilizing gradient rescaling in multi-gradient descent to achieve Pareto optimal alignment for large language models across multiple objectives (e.g., usefulness and safety).
GradOT: Training-free Gradient-preserving Offsite-tuning for Large Language Models
Kai Yao (Zhejiang University), Wei Wang (Ant Group)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Introduces GradOT, a training-free gradient preservation offline tuning framework for efficiently adapting large language models while preserving privacy.
GRAM: Generative Recommendation via Semantic-aware Multi-granular Late Fusion
Sunkyung Lee (Sungkyunkwan University), Jongwuk Lee (Sungkyunkwan University)
Recommendation SystemTransformerLarge Language ModelText
🎯 What it does: Propose the GRAM model, which applies large language models to generative recommendation by leveraging semantic-to-lexical translation and multi-grained late fusion.
GrammaMT: Improving Machine Translation with Grammar-Informed In-Context Learning
Rita Ramos (Apple), Natalie Schluter (Apple)
TransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Proposes a zero-training, few-shot, grammar-based prompting method called GRAMMAMT, which enhances the context of large language models for machine translation using Interlinear Glossed Text (IGT).
GRaMPa: Subword Regularisation by Skewing Uniform Segmentation Distributions with an Efficient Path-counting Markov Model
Thomas Bauwens (KU Leuven), Miryam de Lhoneux (KU Leuven)
Representation LearningTransformerText
🎯 What it does: Propose the GRaMPa algorithm, which implements single-pass, uniform sampling for subword segmentation using a path-counting Markov model, and the bias can be adjusted via temperature and minimum token length parameters.
Graph Counselor: Adaptive Graph Exploration via Multi-Agent Synergy to Enhance LLM Reasoning
Junqi Gao (Shanghai Artificial Intelligence Laboratory), Jianxing Liu (Harbin Institute of Technology)
RetrievalTransformerGraphBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes a multi-agent collaborative graph retrieval augmented generation framework called Graph Counselor, which can adaptively extract graph structural information in multi-round interactions and improve the reasoning quality of LLMs through self-reflection.
Graph of Records: Boosting Retrieval Augmented Generation for Long-context Summarization with Graphs
Haozhen Zhang (University of Illinois at Urbana-Champaign), Jiaxuan You (University of Illinois at Urbana-Champaign)
GenerationRetrievalGraph Neural NetworkLarge Language ModelContrastive LearningTextGraphRetrieval-Augmented Generation
🎯 What it does: Proposed the Graph of Records (GoR) method, which constructs a graph structure to associate historical responses generated by LLMs with retrieved text blocks, thereby enhancing the effectiveness of global summarization for long texts.
Graph-Structured Trajectory Extraction from Travelogues
Aitaro Yamamoto (NAIST), Taro Watanabe (NAIST)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose a graph-structure-based trajectory extraction method and construct a corresponding evaluation dataset;
GraphCheck: Breaking Long-Term Text Barriers with Extracted Knowledge Graph-Powered Fact-Checking
Yingjian Chen (University of Tokyo), Irene Li (University of Tokyo)
Graph Neural NetworkTransformerLarge Language ModelText
🎯 What it does: Propose the GraphCheck framework, which enhances large language models (LLMs) for fact-checking long texts using knowledge graphs.
Graphically Speaking: Unmasking Abuse in Social Media with Conversation Insights
Célia Nouri (INRIA), Chloé Clavel (INRIA)
ClassificationGraph Neural NetworkTransformerLarge Language ModelTextGraph
🎯 What it does: In social media conversations, modeling the structure of comments using Graph Neural Networks (GAT) and combining contextual information to determine whether a comment is abusive.
GraphInsight: Unlocking Insights in Large Language Models for Graph Structure Understanding
Yukun Cao (University of Science and Technology of China), S Kevin Zhou
TransformerLarge Language ModelGraphBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the GraphInsight framework, enhancing LLMs' understanding of graph structures through rearranging graph description sequences and lightweight retrieval augmentation.
GraphNarrator: Generating Textual Explanations for Graph Neural Networks
Bo Pan (Emory University), Liang Zhao (Emory University)
Explainability and InterpretabilityKnowledge DistillationTransformerLarge Language ModelGraph
🎯 What it does: Propose GraphNarrator, a model-agnostic post-hoc interpreter that maps the decision-making process of graph neural networks to natural language explanations using generative language models.
GRAT: Guiding Retrieval-Augmented Reasoning through Process Rewards Tree Search
Xianshu Peng (Huazhong University of Science and Technology), Wei Wei (Huazhong University of Science and Technology)
RetrievalSupervised Fine-TuningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed and implemented a retrieval-augmented generation framework named GRAT based on Monte Carlo Tree Search (MCTS) for multi-hop question answering, enabling multi-path exploration, strategy prediction, step-by-step evaluation, and global selection, while supporting self-training to enhance reasoning capabilities.
Grounded, or a Good Guesser? A Per-Question Balanced Dataset to Separate Blind from Grounded Models for Embodied Question Answering
Miles Shelton (University of Richmond), Catherine Finegan-Dollak
Data SynthesisRobotic IntelligenceVision-Language-Action ModelMultimodalityBenchmark
🎯 What it does: Proposed and constructed the first question-balanced immersive question answering (EQA) dataset, PQB-EQA, to test whether models truly rely on environmental perception rather than merely linguistic priors;
Growing Through Experience: Scaling Episodic Grounding in Language Models
Chunhui Zhang (Dartmouth College), Soroush Vosoughi (Dartmouth College)
Computational EfficiencyKnowledge DistillationReinforcement Learning from Human FeedbackTransformerReinforcement LearningText
🎯 What it does: Propose a weak-to-strong short-term experience learning framework that transfers experience from small and medium models to large models, achieving efficient event memory integration.
GuessArena: Guess Who I Am? A Self-Adaptive Framework for Evaluating LLMs in Domain-Specific Knowledge and Reasoning
Qingchen Yu (MemTensor (Shanghai) Technology Co., Ltd.), Zhiyu Li (MemTensor (Shanghai) Technology Co., Ltd.)
Large Language ModelPrompt EngineeringTextBenchmarkFinance RelatedRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed the GUESSARENA framework, which utilizes a gamified interactive approach called 'Guess Who I Am' to dynamically generate domain knowledge cards and assess LLMs' knowledge mastery and reasoning capabilities in vertical domains through multi-round Q&A.
GUI-explorer: Autonomous Exploration and Mining of Transition-aware Knowledge for GUI Agent
Bin Xie, Liqiang Nie (Harbin Institute of Technology)
TransformerLarge Language ModelAgentic AIVision Language ModelMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose a GUI automation agent called GUI-explorer that integrates no-training, automatic exploration, and unsupervised knowledge mining.