EMNLP 2025 Papers — Page 7
Conference on Empirical Methods in Natural Language Processing · 1809 papers
FinRAGBench-V: A Benchmark for Multimodal RAG with Visual Citation in the Financial Domain
Suifeng Zhao (Peking University), Jun Gao (Peking University)
RetrievalTransformerVision Language ModelTextMultimodalityBenchmarkFinance RelatedRetrieval-Augmented Generation
🎯 What it does: Constructed FinRAGBench-V, which includes a multimodal retrieval corpus and a QA dataset, and implemented the RGenCite baseline to support visual citations.
FinTrust: A Comprehensive Benchmark of Trustworthiness Evaluation in Finance Domain
Tiansheng Hu (NYU Shanghai), Chen Zhao (NYU Shanghai)
Adversarial AttackTransformerLarge Language ModelTextMultimodalityTabularTime SeriesBenchmarkFinance Related
🎯 What it does: Proposed the FINTRUST benchmark to systematically evaluate the trustworthiness and alignment of large language models in the financial domain.
FIRE: Flexible Integration of Data Quality Ratings for Effective Pretraining
Xu Liangyu (Meituan), Xunliang Cai (Meituan)
Data-Centric LearningText
🎯 What it does: Propose the FIRE framework, which comprehensively evaluates pre-trained data using a multi-dimensional data quality assessor and achieves efficient filtering.
Firewall Routing: Blocking Leads to Better Hybrid Inference for LLMs
Runyu Peng (Fudan University), Xipeng Qiu (Fudan University)
Computational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposed the Firewall Routing dual-model hybrid inference framework, which uses a lightweight router to dynamically assign queries to strong or weak LLMs based on query difficulty, and improves inference efficiency and quality by blocking hard-to-solve long-tail queries through Hard Blocking and Soft Blocking mechanisms;
FISTAPruner: Layer-wise Post-training Pruning for Large Language Models
Pengxiang Zhao (University of Hong Kong), Xiaoming Yuan (University of Hong Kong)
OptimizationComputational EfficiencyLarge Language ModelText
🎯 What it does: Propose a hierarchical post-training pruning method called FISTAPruner based on FISTA for unstructured and 2:4 semi-structured sparsity in large language models;
FLARE: Faithful Logic-Aided Reasoning and Exploration
Erik Arakelyan (NVIDIA), Isabelle Augenstein (University of Copenhagen)
Explainability and InterpretabilityLarge Language ModelTextChain-of-Thought
🎯 What it does: Propose FLARE, a method that integrates large language models with logic programs (Prolog) by first generating a plan, then generating logical code, and finally simulating search within the LLM to complete a full workflow for question answering and reasoning; simultaneously providing a method to evaluate the credibility of reasoning.
FlashAdventure: A Benchmark for GUI Agents Solving Full Story Arcs in Diverse Adventure Games
Jaewoo Ahn (Seoul National University), Gunhee Kim (Seoul National University)
TransformerLarge Language ModelAgentic AIVision-Language-Action ModelTextBenchmark
🎯 What it does: Introduce the FlashAdventure benchmark, using 34 Flash adventure games to evaluate the ability of GUI agents to complete full storylines
Flaw or Artifact? Rethinking Prompt Sensitivity in Evaluating LLMs
Andong Hua (University Of Santa Barbara), Yao Qin (University Of Oxford)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Systematically evaluate the performance of large language models (LLMs) under different prompt templates, comparing traditional heuristic evaluation methods with the LLM-as-a-Judge approach, demonstrating that the latter significantly reduces prompt sensitivity.
Flexible-length Text Infilling for Discrete Diffusion Models
Andrew Zhang (Virginia Tech), Chris Thomas (Virginia Tech)
GenerationDiffusion modelText
🎯 What it does: Propose a discrete diffusion model called DDOT that can perform text filling at any length and position.
Flexibly Utilize Memory for Long-Term Conversation via a Fragment-then-Compose Framework
Cai Ke (Harbin Institute of Technology), Ruifeng Xu (Harbin Institute of Technology)
RetrievalTransformerLarge Language ModelTextSequentialRetrieval-Augmented Generation
🎯 What it does: Proposes the Fragmentation-Combination framework (FraCom), which achieves more flexible and efficient memory retrieval and response generation by splitting dialogue history into predicate-argument memory fragments and constructing proposition graphs.
FlightGPT: Towards Generalizable and Interpretable UAV Vision-and-Language Navigation with Vision-Language Models
Hengxing Cai (Sun Yat-Sen University), Renxin Zhong (Sun Yat-Sen University)
Autonomous DrivingExplainability and InterpretabilityTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: Proposed the FlightGPT framework, which utilizes vision-language models to achieve UAV visual and language navigation.
FLRC: Fine-grained Low-Rank Compressor for Efficient LLM Inference
Yu-Chen Lu (National Yang Ming Chiao Tung University), Kai-Chiang Wu (National Yang Ming Chiao Tung University)
CompressionComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes Fine-grained Low-Rank Compressor (FLRC), achieving refined low-rank compression through hierarchical Fisher importance assessment, and employs Progressive Low-Rank Decoding (PLRD) during inference to enhance text generation quality by progressively reducing ranks.
fLSA: Learning Semantic Structures in Document Collections Using Foundation Models
Weijia Xu (Microsoft Research), Nicolas Le Roux (Microsoft Research)
Representation LearningData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes a latent semantic analysis method based on foundational models, named fLSA, which utilizes LLMs to cluster and label document paragraphs during EM iterations, learns the latent semantic structure of documents, and employs labels to reconstruct original text and generate answers through hierarchical sampling.
FLUID QA: A Multilingual Benchmark for Figurative Language Usage in Dialogue across English, Chinese, and Korean
Seoyoon Park (Yonsei University), Hansaem Kim (Yonsei University)
Large Language ModelTextBenchmark
🎯 What it does: Researchers proposed the FLUID QA multilingual dialogic metaphor evaluation benchmark, based on FLUTE data with culturally adapted translations, assessing LLMs' ability to use metaphors in multi-turn dialogues.
Follow the Flow: Fine-grained Flowchart Attribution with Neurosymbolic Agents
Manan Suri (University of Maryland), Dinesh Manocha
Explainability and InterpretabilityAI Code AssistantGraph Neural NetworkTransformerLarge Language ModelAgentic AIVision Language ModelImageTextBenchmark
🎯 What it does: Studied the fine-grained flowchart attribution task, proposing FlowPathAgent which performs post-hoc attribution through graph reasoning after flowchart parsing, and constructed the FlowExplainBench benchmark dataset.
Following Length Constraints in Instructions
Weizhe Yuan (Meta FAIR), Jing Xu (Meta FAIR)
Reinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes the 'length instruction' evaluation method and constructs two length-constrained instruction-following benchmarks, AlpacaEval-LI and MT-Bench-LI; meanwhile, it designs a Length-Instruction Fine-Tuning (LIFT) training framework, enabling models to comply with length constraints during inference.
Following the Autoregressive Nature of LLM Embeddings via Compression and Alignment
Jingcheng Deng (Chinese Academy of Sciences), Xueqi Cheng (Chinese Academy of Sciences)
RetrievalRepresentation LearningTransformerLarge Language ModelAuto EncoderContrastive LearningText
🎯 What it does: This paper proposes AutoRegEmbed, a contrastive learning framework based on autoregressive language models, for generating high-quality text embeddings;
Fooling the LVLM Judges: Visual Biases in LVLM-Based Evaluation
Yerin Hwang (IPAI Seoul National University), Kyomin Jung (IPAI Seoul National University)
Explainability and InterpretabilityAdversarial AttackPrompt EngineeringVision Language ModelImageMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Investigate how visual biases affect the judgment of large vision-language models (LVLM) in text-image consistency evaluation, systematically assess and reveal their vulnerabilities for the first time.
Foot-In-The-Door: A Multi-turn Jailbreak for LLMs
Zixuan Weng (University Of Notre Dame), Xiangyu Zhang (Pennsylvania State University)
Safty and PrivacyAdversarial AttackLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposed a multi-round jailbreak method called FITD based on the psychological 'foot-in-the-door' principle, which induces LLMs to generate inappropriate outputs through progressively escalating prompts.
FoREST: Frame of Reference Evaluation in Spatial Reasoning Tasks
Tanawan Premsri (Michigan State University), Parisa Kordjamshidi (Michigan State University)
GenerationData SynthesisLarge Language ModelPrompt EngineeringDiffusion modelImageTextBenchmarkChain-of-Thought
🎯 What it does: Propose the FoREST benchmark to evaluate large language models' understanding of spatial reference frames (FoR) and extend it to layout generation in text-to-image generation;
Forget What You Know about LLMs Evaluations - LLMs are Like a Chameleon
Nurit Cohen Inger, Seffi Cohen (Harvard University)
Large Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Investigate the overfitting phenomenon of large language models on public benchmarks, and propose a meta-evaluation framework called C-BOD based on parameterized paraphrase transformation to detect whether models excessively rely on surface-level prompts;
Formalizing Style in Personal Narratives
Gustave Cortal (Université Paris-Saclay, CNRS, ENS Paris-Saclay), Alain Finkel (Université Paris-Saclay, CNRS, ENS Paris-Saclay)
Representation LearningLarge Language ModelText
🎯 What it does: This paper proposes a framework based on Systemic Functional Linguistics and sequence analysis to categorize style in personal narratives as patterns of language choices, validated through experiments on dream narrative corpora using automated tools.
Frame First, Then Extract: A Frame-Semantic Reasoning Pipeline for Zero-Shot Relation Triplet Extraction
Zehan Li (Northeastern University), Tianyue Peng (Northeastern University)
Explainability and InterpretabilityTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes the FrameRTE framework, which first constructs relation semantic frameworks (RSFs) and then guides a frozen LLM to complete zero-shot relation triplet extraction;
Frequency & Compositionality in Emergent Communication
Jean-Baptiste Sevestre (École Normale Supérieure Paris-Saclay), Emmanuel Dupoux (École Normale Supérieure Paris-Saclay)
Representation LearningData-Centric LearningRecurrent Neural NetworkReinforcement LearningTabular
🎯 What it does: In Lewis signaling games with large-scale neural networks, the authors systematically investigate the relationship between frequency and linguistic compositionality by introducing Zipfian frequency distribution, using the phenomenon of irregular verbs as an example to illustrate the impact of frequency on the evolution of compositionality.
Friend or Foe? A Computational Investigation of Semantic False Friends across Romance Languages
Ana Sabina Uban (University of Bucharest), Claudia Vlad (University of Bucharest)
ClassificationRepresentation LearningTransformerContrastive LearningText
🎯 What it does: Constructed a comprehensive semantic difference analysis of cognates and borrowings in Romance languages, automatically detected and corrected deceptive cognates/borrowings, and published the corresponding word pair distance table.
From A and B to A+B: Can Large Language Models Solve Compositional Math Problems?
Xisheng Xiao (South China Agricultural University), Hanlin Zhao (South China Agricultural University)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Studying how to combine two existing mathematical problems logically into a new 'composite question,' and using this method to evaluate the generalization ability of large language models (LLMs) in compositional reasoning.
From Automation to Autonomy: A Survey on Large Language Models in Scientific Discovery
Tianshi Zheng (HKUST), Yangqiu Song (HKUST)
TransformerLarge Language ModelAgentic AITextReview/Survey PaperBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper systematically reviews the application of large language models (LLMs) in scientific discovery, tracing their evolution by mapping three levels of autonomy—tool, analyst, and scientist—to the six stages of the scientific method, and identifying future development challenges;
From Capabilities to Performance: Evaluating Key Functional Properties of LLM Architectures in Penetration Testing
Lanxiao Huang (Virginia Tech), Ming Jin (Virginia Tech)
Large Language ModelAgentic AIText
🎯 What it does: Conducted a systematic evaluation of multiple LLM architectures in penetration testing, analyzing their success rates, failure modes, and critical functional improvements required during real attack phases;
From Charts to Fair Narratives: Uncovering and Mitigating Geo-Economic Biases in Chart-to-Text
Ridwan Mahbub (York University), Enamul Hoque (York University)
Explainability and InterpretabilityTransformerPrompt EngineeringVision Language ModelMultimodality
🎯 What it does: This paper systematically evaluates the geographic-economic bias of six vision-language models when generating chart summaries and attempts prompt-based debiasing methods.
From Chat Logs to Collective Insights: Aggregative Question Answering
Wentao Zhang (University of Waterloo), Yuntian Deng (University of Waterloo)
RetrievalTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: The study proposes the Aggregative Question Answering (AQA) task, requiring models to perform global reasoning across massive user-chatbot dialogue logs to answer aggregative questions.
From General Reward to Targeted Reward: Improving Open-ended Long-context Generation Models
Zhihan Guo (Chinese University of Hong Kong), Irwin King (Chinese University of Hong Kong)
GenerationData SynthesisLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Proposed a ProxyReward framework based on reinforcement learning to enhance open-ended long text generation (Open-LTG), utilizing automatically constructed meta-questions and Boolean proxy QA pairs to provide goal-oriented reward signals, eliminating reliance on human annotations and fixed answers.
From Generation to Judgment: Opportunities and Challenges of LLM-as-a-judge
Dawei Li, Huan Liu (Arizona State University)
RetrievalTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringMixture of ExpertsTextReview/Survey PaperBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper provides a systematic review of the LLM-as-a-judge paradigm, clarifying its input/output formats, evaluation attributes, methods, and benchmarks. Based on this, a three-dimensional classification framework (attributes, methods, benchmarks) is constructed, summarizing its applications and challenges across multiple domains such as evaluation, alignment, retrieval, and reasoning.
From Input Perception to Predictive Insight: Modeling Model Blind Spots Before They Become Errors
Maggie Mi (University of Sheffield), Nafise Sadat Moosavi (University of Sheffield)
Explainability and InterpretabilityTransformerText
🎯 What it does: By analyzing the token-level probability distribution of input sequences in language models, an error prediction method based solely on input-side likelihood is proposed.
From Language to Cognition: How LLMs Outgrow the Human Language Network
Badr AlKhamissi (EPFL), Martin Schrimpf (EPFL)
TransformerLarge Language ModelBiomedical DataMagnetic Resonance ImagingBenchmark
🎯 What it does: This paper conducts a large-scale comparison across 8 model sizes and 34 training checkpoints to evaluate the brain alignment of large language models (LLMs) with human language networks and investigate its relationship with language ability.
From Long to Lean: Performance-aware and Adaptive Chain-of-Thought Compression via Multi-round Refinement
JianZhi Yan, Buzhou Tang (Harbin Institute of Technology)
CompressionComputational EfficiencyTransformerPrompt EngineeringTextChain-of-Thought
🎯 What it does: Proposed the MACC multi-round adaptive chaining reasoning compression framework, which can significantly shorten the reasoning chain while maintaining accuracy;
From Parameters to Performance: A Data-Driven Study on LLM Structure and Development
Suqing Wang (Wuhan University), Qianren Wang (Shanghai Huawei Technologies)
OptimizationExplainability and InterpretabilityData-Centric LearningTransformerTextTabularBenchmark
🎯 What it does: Built a massive dataset containing open-source LLM structural configurations and their performance on six benchmarks, systematically analyzing the correlation between structure and performance through regression and mechanism explainability methods;
From perception to production: how acoustic invariance facilitates articulatory learning in a self-supervised vocal imitation model
Marvin Lavechin (Massachusetts Institute of Technology), Thomas Hueber (University Grenoble Alpes)
GenerationRepresentation LearningRecurrent Neural NetworkTransformerAuto EncoderContrastive LearningAudio
🎯 What it does: Built a self-supervised speech imitation model that maps auditory representations to articulator actions and investigated the impact of different acoustic representations on pronunciation learning.
From Personas to Talks: Revisiting the Impact of Personas on LLM-Synthesized Emotional Support Conversations
Shenghan Wu (National University of Singapore), Yang Deng (Singapore Management University)
GenerationData SynthesisTransformerLarge Language ModelText
🎯 What it does: This paper constructs an LLM simulation framework based on psychological scales to systematically evaluate the stability, changes, and impact on strategy distribution of persona in emotional support dialogue generation.
From Problem-Solving to Teaching Problem-Solving: Aligning LLMs with Pedagogy using Reinforcement Learning
David Dinucu-Jianu (ETH Zurich), Mrinmaya Sachan (ETH Zurich)
Reinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose an online reinforcement learning framework that simulates student-teacher dialogues, enabling large language models to provide adaptive, step-by-step tutoring in educational scenarios, while balancing teaching quality and problem-solving accuracy through dialogue-level rewards.
From Reasoning to Answer: Empirical, Attention-Based and Mechanistic Insights into Distilled DeepSeek R1 Models
Jue Zhang (Microsoft), Dongmei Zhang (Microsoft)
Explainability and InterpretabilityKnowledge DistillationTransformerTextBenchmark
🎯 What it does: A three-stage study on three Distilled DeepSeek R1 models through experimental evaluation, attention analysis, and activation patching, investigating how reasoning trajectories influence the final answer generation;
From Schema to State: Zero-Shot Scheme-Only Dialogue State Tracking via Diverse Synthetic Dialogue and Step-by-Step Distillation
Huan Xu (Bournemouth University), Jian Jun Zhang (Bournemouth University)
Data SynthesisComputational EfficiencyKnowledge DistillationTransformerPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose a zero-shot, schema-only dialogue state tracking (DST) framework that generates diverse synthetic dialogues using dynamic complexity prompting and transfers large language models' reasoning capabilities to small models through a two-stage chain-of-thought (CoT) distillation process.
From Scores to Steps: Diagnosing and Improving LLM Performance in Evidence-Based Medical Calculations
Benlu Wang (Yale University), Zonghai Yao (VA Bedford Health Care)
TransformerLarge Language ModelPrompt EngineeringBiomedical DataElectronic Health RecordsBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This work reorganizes medical computation tasks, proposes fine-grained step evaluation and structured error attribution, and builds the MedRaC multi-agent to enhance LLM performance in medical calculations.
From Shortcuts to Balance: Attribution Analysis of Speech-Text Feature Utilization in Distinguishing Original from Machine-Translated Texts
Yongjian Chen (University of Groningen), Antonio Toral (Universitat d'Alacant)
ClassificationExplainability and InterpretabilityTransformerTextMultimodalityAudio
🎯 What it does: This paper integrates text and speech modalities to study the dependency on named entities when distinguishing machine-translated text from original text.
From Surveys to Narratives: Rethinking Cultural Value Adaptation in LLMs
M. Farid Adilazuarda (Mohamed bin Zayed University of Artificial Intelligence), Alham Fikri Aji (Technical University of Darmstadt)
Domain AdaptationRepresentation LearningData-Centric LearningLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper systematically evaluates methods for adapting large language models (LLMs) to cultural values using the World Values Survey (WVS), supplemented by encyclopedia (Wikipedia) and contextualized norms (NormAd) data, and explores the impact of different data sources on model cultural diversity, task performance, and factual knowledge retention.
From Tens of Hours to Tens of Thousands: Scaling Back-Translation for Speech Recognition
Tianduo Wang (Singapore University of Technology and Design), Shanbo Cheng (Nanjing University)
RecognitionGenerationData SynthesisTransformerTextAudio
🎯 What it does: Train TTS models using limited real transcribed speech, generate hundreds of thousands of hours of synthetic speech via text-to-speech inverse translation (Speech Back-Translation) to expand ASR training data for low-resource languages.
From Unaligned to Aligned: Scaling Multilingual LLMs with Multi-Way Parallel Corpora
Yingli Shen (Tsinghua University), Maosong Sun (Tsinghua University)
Representation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Constructed the TED2025 multilingual parallel corpus (113 languages, up to 50 languages in parallel) and used it for continued pre-training and instruction tuning of large multilingual LLMs, systematically evaluating the impact of factors such as parallelism, language combinations, and instruction targets on model performance.
From Understanding to Generation: An Efficient Shortcut for Evaluating Language Models
Viktor Hangya (Fraunhofer IIS), Darina Gold (Fraunhofer IIS)
Computational EfficiencyTextBenchmark
🎯 What it does: Propose rewriting natural language generation (NLG) evaluation into natural language understanding (NLU) forms (multiple-choice MC and log-likelihood LL) to significantly reduce evaluation computational cost, while verifying the correlation between the rewritten evaluation and original NLG; and conduct systematic experiments across multiple capabilities (mathematical reasoning, code generation, factual knowledge, reading comprehension).
From Word to World: Evaluate and Mitigate Culture Bias in LLMs via Word Association Test
Xunlian Dai, Haizhou Li (Shenzhen Research Institute Of Big Data)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Propose a cross-cultural evaluation framework based on the Word Association Test (WAT), design an LLM-adaptive free association task, and further introduce the CultureSteer model to explicitly guide culturally specific associations within the internal semantic space, thereby enhancing the cross-cultural cognitive capabilities of LLMs.
FuseChat: Knowledge Fusion of Chat Models
Fanqi Wan (Sun Yat-sen University), Xiaojun Quan (Sun Yat-sen University)
Knowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed the FUSECHAT framework, which combines various structures and scales of chat LLMs into a more powerful single model through two stages (pairwise lightweight knowledge fusion + parameter-level merging).
G2: Guided Generation for Enhanced Output Diversity in LLMs
Zhiwen Ruan (Southern University of Science and Technology), Guanhua Chen (Southern University of Science and Technology)
GenerationTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: Propose a training-agnostic, plug-and-play decoding strategy G2, which significantly enhances the diversity of LLM-generated text while maintaining quality. This is achieved by constructing a base generator, Diversity Guide, and Dedupe Guide using different prompt templates on the same LLM, combined with entropy-gated selective intervention and center selection sampling.
Gamma-Guard: Lightweight Residual Adapters for Robust Guardrails in Large Language Models
Lijia Lv (Chinese Academy of Sciences), Songlin Hu (Chinese Academy of Sciences)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelText
🎯 What it does: Propose Gamma-Guard, which enhances the robustness of the LLM Guard model against adversarial attacks by inserting a lightweight learnable residual adapter into it
GAP: a Global Adaptive Pruning Method for Large Language Models
Zhihua Ban (CVTE Research), Ming Yang (CVTE Research)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposed a global adaptive pruning method called GAP, achieving LLM structural pruning by assigning different pruning rates to each layer.
GATEAU: Selecting Influential Samples for Long Context Alignment
Shuzheng Si (Tsinghua University), Maosong Sun (Tsinghua University)
Data-Centric LearningLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose the GATEAU framework to select samples rich in long-range dependencies from synthetic long instruction training data, aiming to enhance the long-context alignment and instruction-following capabilities of large language models (LLMs).
GCML: Gradient Coherence Guided Meta-Learning for Cross-Domain Emerging Topic Rumor Detection
Zejiang He (National University of Defense Technology), Dongsheng Li (National University of Defense Technology)
Domain AdaptationMeta LearningTransformerText
🎯 What it does: Propose the GCML method, utilizing gradient consistency and meta-learning for cross-domain emerging topic rumor detection, addressing the challenges of domain transfer between source and target domains and sample scarcity.
Generative Annotation for ASR Named Entity Correction
Yuanchang Luo (Huawei Translation Service Center), Hao Yang (Huawei Translation Service Center)
RecognitionGenerationRetrievalTransformerAuto EncoderTextRetrieval-Augmented GenerationAudio
🎯 What it does: Proposed a post-editing named entity correction (NEC) method based on generative annotation, which retrieves candidate entities using phonetic features of speech, annotates errors in ASR transcriptions with a generative model, and then replaces them with correct entities.
Generative or Discriminative? Revisiting Text Classification in the Era of Transformers
Siva Rajesh Kasa (Amazon Inc.), Vijay Huddar (Amazon Inc.)
ClassificationTransformerDiffusion modelText
🎯 What it does: What was done: Systematically evaluated the performance of generative (AR, MLM, DIFF, pseudo-generated AR) and discriminative (Encoder) text classification models based on Transformer in different model scales and sample sizes, including accuracy, robustness, calibration, and ordinality.
Generator-Assistant Stepwise Rollback Framework for Large Language Model Agent
Xingzuo Li, Min Zhang (Harbin Institute Of Technology)
TransformerLarge Language ModelAgentic AITextChain-of-Thought
🎯 What it does: Proposes a framework called GA-Rollback, where LLM agents delegate action execution to an independent assistant for error checking and rollback, thereby reducing one-time error propagation;
GenLink: Generation-Driven Schema-Linking via Multi-Model Learning for Text-to-SQL
Zhifeng Hao (Guangdong University of Technology), Boyan Xu (Guangdong University of Technology)
GenerationAI Code AssistantTransformerLarge Language ModelText
🎯 What it does: Propose a generation-driven Schema Linking framework called GenLink based on multi-model generation for cross-domain reasoning in text-to-SQL tasks;
Genre Matters: How Text Types Interact with Decoding Strategies and Lexical Predictors in Shaping Reading Behavior
Lena Sophia Bolliger (University of Zurich), Lena Ann Jäger (University of Zurich)
TransformerLarge Language ModelTime Series
🎯 What it does: This study investigates how different text types and large language model (LLM) generation strategies influence human reading behavior, using eye-tracking data to examine interactions between text types, decoding strategies, and word-level predictors (surprisal, word frequency, word length).
GeoEdit: Geometric Knowledge Editing for Large Language Models
Yujie Feng (Hong Kong Polytechnic University), Xiao-Ming Wu (Hong Kong Polytechnic University)
Representation LearningTransformerLarge Language ModelAuto EncoderText
🎯 What it does: Proposes the GeoEdit framework, which identifies and edits knowledge in LLMs using geometric relationships.
GER-LLM: Efficient and Effective Geospatial Entity Resolution with Large Language Model
Haojia Zhu (Southeast University), Jiahui Jin (Southeast University)
TransformerLarge Language ModelTabular
🎯 What it does: Proposed a framework called GER-LLM for geospatial entity resolution using LLM.
Glider: Global and Local Instruction-Driven Expert Router
Pingzhi Li (University of North Carolina at Chapel Hill), Tianlong Chen (University of North Carolina at Chapel Hill)
Large Language ModelMixture of ExpertsTextBenchmark
🎯 What it does: Designed a multi-scale expert routing mechanism called GLIDER, which combines global semantics with local token-level routing, uses LLM-generated task descriptions to guide expert selection, and conducts experiments on the T5 model.
GLIMPSE: Do Large Vision-Language Models Truly Think With Videos or Just Glimpse at Them?
Yiyang Zhou (UNC Chapel Hill), Huaxiu Yao (UNC Chapel Hill)
TransformerLarge Language ModelVision Language ModelVideoMultimodalityBenchmark
🎯 What it does: This paper introduces the GLIMPSE benchmark to evaluate the real-world capabilities of large-scale vision-language models in video understanding and reasoning.
Good Intentions Beyond ACL: Who Does NLP for Social Good, and Where?
Grace LeFevre (Northwestern University), Rob Voigt (Northwestern University)
TransformerLarge Language ModelText
🎯 What it does: This paper conducts a large-scale scientometric analysis of authors and publication contexts across 309,208 NLP papers, systematically mapping the global research ecosystem of NLP-for-Social-Good (NLP4SG);
Governance in Motion: Co-evolution of Constitutions and AI models for Scalable Safety
Chenhao Huang (Fudan University), Xuanjing Huang (Fudan University)
Safty and PrivacyReinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Proposes the COCOA framework to achieve co-evolution between models and constitutional principles for dynamic alignment of LLM safety.
Graceful Forgetting in Generative Language Models
Chunyang Jiang (Hong Kong University of Science and Technology), Yike Guo
GenerationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposed an elegant forgetting framework for generative language models (Learning With Forgetting, LWF), which mines pre-trained knowledge through self-generated text, calculates forgetting confidence, and performs periodic gradient ascent-based forgetting during fine-tuning to enhance learning plasticity for downstream tasks.
GRADA: Graph-based Reranking against Adversarial Documents Attack
Jingjie Zheng (University of Melbourne), Qiongkai Xu (Mohamed bin Zayed University of Artificial Intelligence)
RetrievalAdversarial AttackGraph Neural NetworkTextRetrieval-Augmented Generation
🎯 What it does: Propose the GRADA framework, which utilizes a document similarity graph for re-ranking in RAG systems to defend against adversarial document attacks.
GraDaSE: Graph-Based Dataset Search with Examples
Jing He (Nanjing University), Gong Cheng (Nanjing University)
RetrievalGraph Neural NetworkGraphBenchmark
🎯 What it does: Propose GraDaSE, a graph-based 'dataset retrieval and example' (DSE) method, which constructs a domain-agnostic dataset graph by leveraging the provenance and topic relationships of datasets, generates target-biased queries and candidate representations on this graph, and finally ranks results by jointly utilizing graph structure and textual information.
Graders Should Cheat: Privileged Information Enables Expert-Level Automated Evaluations
Jin Peng Zhou (Cornell University), Fei Sha (Google Research)
Large Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Propose using Privileged Information (PI) to enhance language model evaluators, enabling expert-level automatic evaluation on cutting-edge tasks.
Gradient-Attention Guided Dual-Masking Synergetic Framework for Robust Text-based Person Retrieval
Tianlu Zheng (Northeastern University), Qichuan Ding (Northeastern University)
RetrievalTransformerVision Language ModelImageTextMultimodality
🎯 What it does: Construct the WebPerson dataset and propose the GA-DMS framework to enhance person retrieval performance in text retrieval
GRAID: Synthetic Data Generation with Geometric Constraints and Multi-Agentic Reflection for Harmful Content Detection
Melissa Kazemi Rad (Capital One), Mohammad Shahed Sorower (Capital One)
Data SynthesisTransformerLarge Language ModelSupervised Fine-TuningAgentic AITextChain-of-Thought
🎯 What it does: GRAID automatically synthesizes harmful text and augments training data through a two-phase process involving geometric constraint generation and multi-agent reflective evaluation.
Grammar Pruning: Enabling Low-Latency Zero-Shot Task-Oriented Language Models for Edge AI
Octavian Alexandru Trifan (University of California, Irvine), Alexander Veidenbaum (University of California, Irvine)
OptimizationComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Research on methods for achieving low-latency zero-shot task-oriented semantic parsing on edge devices.
Graph-Based Multi-Trait Essay Scoring
Shengjie Li (University of Texas at Dallas), Vincent Ng (University of Texas at Dallas)
ClassificationGraph Neural NetworkTransformerLarge Language ModelTextGraph
🎯 What it does: Modeling essays as graphs and using Graph Attention Networks (GAT) to model and predict multi-trait scores
Graph-Guided Textual Explanation Generation Framework
Shuzhou Yuan (TU Dresden), Isabelle Augenstein (University of Copenhagen)
Explainability and InterpretabilityGraph Neural NetworkTransformerLarge Language ModelTextGraph
🎯 What it does: Propose the G-TEx framework, which enhances the faithfulness of natural language explanations by injecting high-confidence highlight explanations from the model's internal mechanisms into the generative model as a graph structure.
Graph-R1: Incentivizing the Zero-Shot Graph Learning Capability in LLMs via Explicit Reasoning
Yicong Wu (Beihang University), Junjie Wu (Beihang University)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextGraphChain-of-Thought
🎯 What it does: This paper proposes a graph learning framework called GRAPH-R1 that completely does not rely on graph neural networks, transforming graph tasks such as node classification, link prediction, and graph classification into text reasoning problems, and generating interpretable chain-of-thought reasoning through large reasoning models (LRM).
GraphAgent: Agentic Graph Language Assistant
Yuhao Yang (University of Hong Kong), Chao Huang (University of Hong Kong)
Representation LearningGraph Neural NetworkTransformerLarge Language ModelAgentic AITextGraph
🎯 What it does: Propose a multi-agent framework called GraphAgent that automatically generates semantic knowledge graphs (SKG), understands user queries, and performs prediction and generation tasks;
GraphKV: Breaking the Static Selection Paradigm with Graph-Based KV Cache Eviction
Xuelin Li (Shanghai Jiao Tong University), Linfeng Zhang (Shanghai Jiao Tong University)
Computational EfficiencyLarge Language ModelTextBenchmark
🎯 What it does: Propose the GraphKV framework, leveraging graph structures to dynamically update the importance of tokens in the KV cache, thereby achieving efficient KV cache compression without additional training.
GRASP: Replace Redundant Layers with Adaptive Singular Parameters for Efficient Model Compression
Kainan Liu (Ping An Technology Co., Ltd.), Jing Xiao (Ping An Technology Co., Ltd.)
CompressionComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Replace redundant layers in large language models with gradient-driven singular value selection to achieve efficient compression.
GRIT: Guided Relational Integration for Efficient Multi-Table Understanding
Yujin Kang (Chung-Ang University), Yoon-Sik Cho (Chung-Ang University)
Computational EfficiencyData-Centric LearningLarge Language ModelPrompt EngineeringTabular
🎯 What it does: Propose a lightweight multi-table relation extraction and prompt generation method called GRIT, which uses hashing techniques to pre-discover primary/foreign key relationships and encodes these relationships as natural language prompts for LLMs to use in table column retrieval and Text-to-SQL tasks.
Grounded Semantic Role Labelling from Synthetic Multimodal Data for Situated Robot Commands
Claudiu Daniel Hromei (University of Rome Tor Vergata), Roberto Basili (University of Rome Tor Vergata)
GenerationData SynthesisRobotic IntelligenceLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: Propose a multimodal semantic role labeling (G-SRL) framework that aligns robot commands with perceptual context and designs an automated synthetic image generation and validation pipeline, producing approximately 11,000 image-command pairs with structured annotations.
Grounding Multilingual Multimodal LLMs With Cultural Knowledge
Jean De Dieu Nyandwi, Graham Neubig (Carnegie Mellon University)
Data SynthesisRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelMultimodalityRetrieval-Augmented Generation
🎯 What it does: This paper proposes a multilingual cultural multimodal data generation method based on knowledge graphs and image retrieval, constructing a CulturalGround dataset with a scale of 22M, and training a culture-rich multimodal large language model called CulturalPangea on this dataset.
Group-Aware Reinforcement Learning for Output Diversity in Large Language Models
Oron Anschel (Amazon), Gerard Medioni (Amazon)
GenerationTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: This paper proposes a new reinforcement learning framework called Group-Aware Policy Optimization (GAPO), which promotes higher diversity in large language models (LLMs) during generation tasks by calculating rewards for entire output groups (rather than individual samples) in a single update, thereby mitigating the mode collapse problem.
Group-SAE: Efficient Training of Sparse Autoencoders for Large Language Models via Layer Groups
Davide Ghilardi (University of Milan-Bicocca), Matteo Palmonari (London School of Economics)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelAuto EncoderText
🎯 What it does: Significantly reduces the computational overhead of SAE training by clustering adjacent layers into groups and training a single sparse autoencoder (SAE) per group to reconstruct multi-layer activations.
Grouping Entities with Shared Properties using Multi-Facet Prompting and Property Embeddings
Amit Gajbhiye, Steven Schockaert (Cardiff University)
ClassificationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Leverage large language models (LLM) to generate multi-faceted (facet) attribute pairs (facet-property) for each entity, then use pre-trained text embeddings (LLM2Vec) to map these attribute pairs into vectors and cluster them. Finally, group entities under the same category if they appear in the same attribute cluster, thereby achieving general-purpose grouping of large-scale entity sets.
GRPO-LEAD: A Difficulty-Aware Reinforcement Learning Approach for Concise Mathematical Reasoning in Language Models
Jixiao Zhang (Johns Hopkins University), Chunsheng Zuo (Johns Hopkins University)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Improve GRPO (Group Relative Policy Optimization) to achieve more concise and accurate reasoning processes in mathematical reasoning tasks through reinforcement learning.
GuessingGame: Measuring the Informativeness of Open-Ended Questions in Large Language Models
Dylan Hutson (University of Cincinnati), Tianyu Jiang (University of Cincinnati)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Designed and implemented GuessingGame, an open-source protocol for evaluating the strategic performance of large language models in open-domain, open-ended questioning scenarios.
GUI-Bee: Align GUI Action Grounding to Novel Environments via Autonomous Exploration
Yue Fan (University Of California Santa Cruz), Gang Wu (Adobe Research)
Data-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIPrompt EngineeringVision-Language-Action ModelMultimodalityBenchmark
🎯 What it does: Align GUI action induction models in new environments by using self-driven GUI-Bee agents to explore the environment, generate high-quality environment-specific data, and improve model performance through continuous fine-tuning in these environments.
Hallucination Detection in LLMs Using Spectral Features of Attention Maps
Jakub Binkowski (Wroclaw University of Science and Technology), Tomasz Jan Kajdanowicz
Anomaly DetectionTransformerLarge Language ModelText
🎯 What it does: Researchers propose using spectral features from attention maps to detect hallucinations in large language models (LLMs), treating attention maps as adjacency matrices of graphs, constructing Laplacian matrices, and taking their first k eigenvalues as features, then training a logistic regression probe to detect hallucinations.
Hanfu-Bench: A Multimodal Benchmark on Cross-Temporal Cultural Understanding and Transcreation
Li Zhou (Chinese University of Hong Kong Shenzhen), Haizhou Li (University of Copenhagen)
Vision Language ModelDiffusion modelImageMultimodalityBenchmark
🎯 What it does: Proposed Hanfu-Bench, a cross-temporal multimodal benchmark dataset, to evaluate the ability of vision-language models in understanding Hanfu culture and cross-temporal creative transformation.
Harmful Prompt Laundering: Jailbreaking LLMs with Abductive Styles and Symbolic Encoding
Seongho Joo (Seoul National University), Kyomin Jung (Seoul National University)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposed a black-box jailbreaking method called HaPLa, which bypasses LLM security mechanisms by rephrasing harmful requests using reductio ad absurdum and hiding sensitive words through symbolic encoding.
HD-PiSSA: High-Rank Distributed Orthogonal Adaptation
Yiding Wang (Peking University), Muhan Zhang (Peking University)
OptimizationComputational EfficiencyTransformerSupervised Fine-TuningText
🎯 What it does: Propose a High-Rank Distributed Orthogonal Adapter (HD-PiSSA) for parameter-efficient fine-tuning of large-scale language models.
HealthCards: Exploring Text-to-Image Generation as Visual Aids for Healthcare Knowledge Democratizing and Education
Qian Wu (Chinese University of Hong Kong), Qi Dou (Chinese University of Hong Kong)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelFlow-based ModelImageTextBiomedical Data
🎯 What it does: Explored how to use text-to-image models to generate health education flashcards, constructed a high-quality medical knowledge flashcard dataset, and fine-tuned and validated the model.
HELENE: Hessian Layer-wise Clipping and Gradient Annealing for Accelerating Fine-tuning LLM with Zeroth-order Optimization
Huaqin Zhao (University of Georgia), Jin Lu (University of Georgia)
OptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: For the fine-tuning tasks of large language models (LLMs), we propose an efficient optimizer called HELENE based on zeroth-order optimization (ZO), which can estimate gradients and second-order information using only forward propagation, thereby achieving low memory consumption and accelerated convergence.
HESEIA: A community-based dataset for evaluating social biases in large language models, co-designed in real school settings in Latin America
Guido Ivetta (Universidad Nacional de Córdoba), Luciana Benotti (Universidad Nacional de Córdoba)
Data SynthesisData-Centric LearningLarge Language ModelTextBenchmark
🎯 What it does: Through professional development courses, teachers and students co-created a language dataset named HESEIA containing 45,416 sentences with multiple intersecting forms of discrimination;
HICode: Hierarchical Inductive Coding with LLMs
Mian Zhong, Anjalie Field (Johns Hopkins University)
Representation LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose the HICode pipeline, which first uses LLMs to generate fine-grained labels for each text segment, then merges them into themes through hierarchical clustering, achieving automatic inductive coding for large-scale corpora.
Hidden in Plain Sight: Reasoning in Underspecified and Misspecified Scenarios for Multimodal LLMs
Qianqi Yan, Xin Eric Wang (University Of California Santa Barbara)
Explainability and InterpretabilityTransformerPrompt EngineeringImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Construct the RUMS diagnostic dataset to evaluate the performance of nine multimodal large models on four categories of implicit reasoning tasks (missing objects, anaphora ambiguity, factual conflicts, and goal infeasibility), and explore the models' latent reasoning capabilities through explicit prompting, Chain-of-Thought, and intervention during reasoning (system roles and clarifying questions).
Hierarchical Bracketing Encodings Work for Dependency Graphs
Ana Ezquerro (Universidade da Coruña), David Vilares (Universidade da Coruña)
TransformerLarge Language ModelTextGraph
🎯 What it does: Apply hierarchical bracket encoding to dependency graph parsing and achieve linear-time inference within a sequence labeling framework
HMoE: Heterogeneous Mixture of Experts for Language Modeling
An Wang (Tencent Hunyuan), Cheng-zhong Xu (University of Macau)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Propose a heterogeneous hybrid expert (HMoE) framework that uses experts of different scales to enhance the performance and efficiency of language models, and introduces a parameter penalty loss during training to balance expert activation.
HookMoE: A learnable performance compensation strategy of Mixture-of-Experts for LLM inference acceleration
Cheng Longkai (Nankai University), Tao Li (Nankai University)
Computational EfficiencyTransformerSupervised Fine-TuningMixture of ExpertsText
🎯 What it does: Proposed a lightweight Hook compensation module combined with the TLNLE scheme to reduce the number of activated experts in MoE models for faster inference, while restoring performance through fine-tuning.
Hope vs. Hate: Understanding User Interactions with LGBTQ+ News Content in Mainstream US News Media through the Lens of Hope Speech
Jonathan Pofcher (Rochester Institute of Technology), Ashiqur R. KhudaBukhsh (Rochester Institute of Technology)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningVideoText
🎯 What it does: Collect and analyze YouTube comments from the three major U.S. cable news networks (Fox News, CNN, MSNBC), construct a 'hope speech' detector focused on LGBTQ+ themes, and explore the impact of commenters' political inclinations on labeling.