ACL 2025 Papers — Page 8
Annual Meeting of the Association for Computational Linguistics · 1699 papers
GUICourse: From General Vision Language Model to Versatile GUI Agent
Wentong Chen (Renmin University of China), Maosong Sun (Tsinghua University)
Reinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningAgentic AIVision Language ModelVision-Language-Action ModelImageTextMultimodalitySequential
🎯 What it does: Developed a complete workflow for training a visual GUI agent from a general vision-language model (VLM) and created three dataset collections
GuideBench: Benchmarking Domain-Oriented Guideline Following for LLM Agents
Lingxiao Diao (Shanghai Jiao Tong University), Zhuosheng Zhang (Shanghai Jiao Tong University)
Large Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes the GUIDEBENCH benchmark to evaluate the ability of large language models to follow domain-specific knowledge rules in real business scenarios, covering seven categories with a total of 1272 instances;
Guiding not Forcing: Enhancing the Transferability of Jailbreaking Attacks on LLMs via Removing Superfluous Constraints
Junxiao Yang (Tsinghua University), Minlie Huang (Tsinghua University)
Adversarial AttackLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: By analyzing redundant constraints in gradient attacks, the Guided Jailbreaking Optimization method is proposed, significantly improving the transferability attack success rate of LLMs.
Gumbel Reranking: Differentiable End-to-End Reranker Optimization
Siyuan Huang (Shanghai Jiao Tong University), Zhouhan Lin (Shanghai Jiao Tong University)
RetrievalOptimizationLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes a differentiable attention mask framework, G-Rerank, based on Gumbel-Softmax and Relaxed Topk, achieving end-to-end re-ranker optimization in RAG systems.
HACo-Det: A Study Towards Fine-Grained Machine-Generated Text Detection under Human-AI Coauthoring
Zhixiong Su (Xi'an Jiaotong University), Minnan Luo (Xi'an Jiaotong University)
ClassificationGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: A fine-grained (word-level/sentence-level) machine-generated text detection task and its corresponding dataset HACO-Det were constructed for texts co-created by humans and large language models. Seven existing document-level detection methods were migrated to a word-level sequence labeling framework, and extensive evaluations were conducted both within and across domains.
HAF-RM: A Hybrid Alignment Framework for Reward Model Training
Shujun Liu (Fudan University), Zhongyu Wei (Fudan University)
Reinforcement Learning from Human FeedbackTransformerReinforcement LearningText
🎯 What it does: Proposes a hybrid alignment framework HAF-RM, incorporating token-level policy probability constraints during reward model training.
HAIC: Improving Human Action Understanding and Generation with Better Captions for Multi-modal Large Language Models
Xiao Wang (Harbin Institute of Technology), Liqiang Nie (Harbin Institute of Technology)
RecognitionGenerationPose EstimationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVision-Language-Action ModelVideoTextMultimodalityBenchmark
🎯 What it does: Propose a two-phase data annotation pipeline to construct high-quality human action video caption datasets HAICTrain and evaluation benchmark HAICBench, significantly enhancing the performance of multimodal large language models in human action understanding and text generation.
Hallucination Detox: Sensitivity Dropout (SenD) for Large Language Model Training
Shahrad Mohammadzadeh (McGill University), Golnoosh Farnadi (McGill University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: By analyzing the internal states during the training of large language models (LLMs), we propose the Sensitivity Dropout (SenD) training protocol, which selectively discards embedding indices that exhibit significant fluctuations during training, thereby reducing the variance of hallucinations and improving model confidence.
HalluLens: LLM Hallucination Benchmark
Yejin Bang (Hong Kong University of Science and Technology), Pascale Fung (Meta)
Large Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the HalluLens evaluation framework, defining the distinction between hallucination and factual accuracy, and designing three new external hallucination assessment tasks (PreciseWikiQA, LongWiki, NonExistentRefusal), reducing data leakage through dynamic test set generation.
HALoGEN: Fantastic LLM Hallucinations and Where to Find Them
Abhilasha Ravichander (University of Washington), Yejin Choi (Stanford University)
GenerationTransformerLarge Language ModelTextBenchmark
🎯 What it does: Created and released HALOGEN, a large-scale, cross-domain hallucination benchmark for evaluating the factuality and appropriate avoidance of generative large language models (LLMs); simultaneously collected and analyzed 150,000 generated results from 14 LLMs, constructing a rich set of hallucination case studies.
Hanging in the Balance: Pivotal Moments in Crisis Counseling Conversations
Vivian Nguyen (Cornell University), Cristian Danescu-Niculescu-Mizil (Cornell University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Detect pivotal moments in dialogues using an unsupervised method by generating multiple possible next replies and employing a conversation prediction model to estimate the impact of these replies on the final outcome, using the variance of these predictions as a measure of pivotal moments.
Has Machine Translation Evaluation Achieved Human Parity? The Human Reference and the Limits of Progress
Lorenzo Proietti (Sapienza University of Rome), Roberto Navigli (Sapienza University of Rome)
TextBenchmark
🎯 What it does: On multi-year WMT evaluation data, human evaluation baselines and automatic MT assessment metrics are integrated into the same meta-evaluation framework, comparing their performance across different annotation protocols and language directions.
HateDay: Insights from a Global Hate Speech Dataset Representative of a Day on Twitter
Manuel Tonneau (University of Oxford), Paul Röttger
ClassificationTransformerLarge Language ModelTextBenchmark
🎯 What it does: Constructed and analyzed a global real-sample Twitter hate speech dataset named HATEDAY, covering 8 languages and 4 English-speaking countries, subsequently using it to evaluate the practical performance of public models and comparing differences with traditional academic datasets and functional tests.
Have We Designed Generalizable Structural Knowledge Promptings? Systematic Evaluation and Rethinking
Yichi Zhang (Zhejiang University), Huajun Chen (Zhejiang University)
Graph Neural NetworkTransformerLarge Language ModelPrompt EngineeringGraphBenchmark
🎯 What it does: Constructed a multi-granularity, multi-difficulty evaluation benchmark named SUBARU, systematically assessing the generalization capability of structured knowledge prompting (SKP) in large language models (LLMs) across four dimensions: granularity, transferability, scalability, and universality, and analyzed its performance through extensive experiments.
HD-NDEs: Neural Differential Equations for Hallucination Detection in LLMs
Qing Li (Mohamed bin Zayed University of Artificial Intelligence), Fakhri Karray (Alibaba International Digital Commerce)
Anomaly DetectionLarge Language ModelTextStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Propose the HD-NDEs method, which significantly enhances hallucination detection capability by mapping LLM internal activations to a latent space, modeling their dynamic trajectories using neural differential equations (Neural ODE/CDE/SDE), and then mapping them to a classification space for authenticity assessment.
HELIOS: Harmonizing Early Fusion, Late Fusion, and LLM Reasoning for Multi-Granular Table-Text Retrieval
Sungho Park (POSTECH), Wook-Shin Han (POSTECH)
RetrievalTransformerLarge Language ModelTextTabularBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed a three-stage retrieval framework named HELIOS, which first generates bilateral subgraphs at the edge level through early fusion, then enhances the retrieval scope by expanding query-related nodes, and finally performs aggregation and multi-hop reasoning on star subgraphs using LLMs;
Help Me Write a Story: Evaluating LLMs’ Ability to Generate Writing Feedback
Hannah Rashkin (Google DeepMind), Mirella Lapata (Google DeepMind)
GenerationTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Investigate the capability of LLMs in generating writing feedback, propose the StoryFeedback dataset, and analyze the effectiveness of model-generated feedback through automatic and human evaluation systems.
HelpSteer3: Human-Annotated Feedback and Edit Data to Empower Inference-Time Scaling in Open-Ended General-Domain Tasks
Zhilin Wang (NVIDIA), Oleksii Kuchaiev (NVIDIA)
Data-Centric LearningReinforcement Learning from Human FeedbackSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: This study constructs the HelpSteer3 dataset, collecting over 7,000 human-annotated open-ended task feedback and edits, and trains specialized feedback and editing models based on this dataset to achieve scalable improvements during inference;
HFT: Half Fine-Tuning for Large Language Models
Tingfeng Hui (Beijing University of Posts and Telecommunications), Hua Wu (Beijing University of Posts and Telecommunications)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose Half Fine-Tuning (HFT), which randomly freezes half of the parameters and only updates the other half in each round of fine-tuning, thereby alleviating catastrophic forgetting in LLMs and improving downstream performance without altering the model architecture.
HiAgent: Hierarchical Working Memory Management for Solving Long-Horizon Agent Tasks with Large Language Model
Mengkang Hu (University of Hong Kong), Ping Luo (University of Hong Kong)
Robotic IntelligenceTransformerLarge Language ModelReinforcement LearningAgentic AITextSequentialBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the HiAgent framework, which utilizes hierarchical management of subgoals to handle the working memory of LLM agents, automatically generates subgoals, and summarizes their trajectories upon completion while retaining essential information;
Hidden in Plain Sight: Evaluation of the Deception Detection Capabilities of LLMs in Multimodal Settings
Md Messal Monem Miah (Texas Aandm University), Ruihong Huang (Texas Aandm University)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelVideoTextMultimodalityRetrieval-Augmented GenerationAudio
🎯 What it does: Systematic evaluation of large language models (LLMs) and large multimodal models (LMMs) on three types of deception detection tasks (court interrogation, conversational deception, and online reviews), and exploration of the impact of different prompting methods, context example selection, auxiliary features, and fine-tuning on performance.
HiddenDetect: Detecting Jailbreak Attacks against Multimodal Large Language Models via Monitoring Hidden States
Yilei Jiang (Chinese University of Hong Kong), Xiangyu Yue (Chinese University of Hong Kong)
Anomaly DetectionSafty and PrivacyLarge Language ModelVision Language ModelTextMultimodality
🎯 What it does: Proposes HiddenDetect, a framework that detects and intercepts jailbreak attacks without fine-tuning, by merely monitoring the internal hidden states of large vision-language models (LVLMs).
HiDe-LLaVA: Hierarchical Decoupling for Continual Instruction Tuning of Multimodal Large Language Model
Haiyang Guo (University of Chinese Academy of Sciences), Cheng-Lin Liu (University of Chinese Academy of Sciences)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelMixture of ExpertsVision Language ModelContrastive LearningMultimodalityBenchmark
🎯 What it does: Studies how to enable multimodal large language models to retain memory of previous tasks during continuous instruction fine-tuning, and proposes the HiDe-LLaVA framework.
Hierarchical Attention Generates Better Proofs
Jianlong Chen (Chinese University of Hong Kong), Andrew C Yao (Shanghai Qi Zhi Institute)
AI Code AssistantTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposed a new regularization method called Hierarchical Attention, aiming to improve the performance of large language models (LLMs) in mathematical proofs by establishing a five-level hierarchy to guide the model's attention mechanism.
Hierarchical Bracketing Encodings for Dependency Parsing as Tagging
Ana Ezquerro (Universidade da Coruña), Carlos Gómez-Rodríguez (Universidade da Coruña)
Representation LearningTransformerLarge Language ModelText
🎯 What it does: Studied a dependency syntactic parsing label encoding based on hierarchical parentheses, proposed an optimal hierarchical parenthesis encoding and extended it to non-projective trees.
Hierarchical Document Refinement for Long-context Retrieval-augmented Generation
Jiajie Jin (Renmin University of China), Zhicheng Dou (Huawei)
GenerationRetrievalCompressionTransformerSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: Propose LongRefiner, a long document refinement framework for retrieval-augmented generation (RAG), which improves the quality of long-text inputs through three steps: dual-layer query analysis, hierarchical document structuring, and adaptive refinement;
Hierarchical Level-Wise News Article Clustering via Multilingual Matryoshka Embeddings
Hans William Alexander Hanley (Stanford University), Zakir Durumeric (Stanford University)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Proposed and trained a multilingual Matryoshka embedding model, and designed a level-based hierarchical recursive clustering algorithm based on its hierarchical structure for clustering news articles into stories, topics, and subtopics.
Hierarchical Memory Organization for Wikipedia Generation
Eugene J. Yu (Peking University), Sujian Li (Peking University)
GenerationTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Proposed a Wikipedia automatic generation framework MOG based on hierarchical memory structure, which can refine massive web information into factual units and organize them according to Wiki structure, ultimately generating complete and traceable entries.
Hierarchical-Task-Aware Multi-modal Mixture of Incremental LoRA Experts for Embodied Continual Learning
Ziqi Jia (Ping An Technology Co., Ltd.), Jianzong Wang (Tsinghua University)
Robotic IntelligenceTransformerMixture of ExpertsTextMultimodality
🎯 What it does: This paper proposes a hierarchical embodied continual learning framework (Hierarchical Embodied Continual Learning Setups, HEC) and designs the Task-aware MoILE method, addressing the problem of catastrophic forgetting in embodied agents during high-level instruction and low-level action learning.
HintsOfTruth: A Multimodal Checkworthiness Detection Dataset with Real and Synthetic Claims
Michiel Van Der Meer, Lonneke Van Der Plas
ClassificationData SynthesisTransformerLarge Language ModelVision Language ModelDiffusion modelImageTextMultimodalityBenchmark
🎯 What it does: This paper constructs the HINTSOFTRUTH dataset, focusing on multimodal (image + text) verification detection, incorporating both real and synthetic samples; through experiments, the performance of various text, image, and multimodal models is evaluated;
HoH: A Dynamic Benchmark for Evaluating the Impact of Outdated Information on Retrieval-Augmented Generation
Jie Ouyang (University of Science and Technology of China), Qi Liu (University of Science and Technology of China)
RetrievalLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Designed and implemented the HOH benchmark for systematically evaluating the performance of Retrieval-Augmented Generation (RAG) when facing outdated information, and automatically generated large-scale dynamic QA datasets and simulated search engines through token-level diff and LLM pipelines;
HomeBench: Evaluating LLMs in Smart Homes with Valid and Invalid Instructions Across Single and Multiple Devices
Silin Li (Beijing Institute of Technology), Haifeng Wang (Baidu Inc)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the HomeBench dataset to evaluate the ability of LLMs to handle valid/invalid single-device and multi-device instructions in smart home environments, constructed a customizable virtual home environment and generated approximately 170,000 instructions;
HoPE: A Novel Positional Encoding Without Long-Term Decay for Enhanced Context Awareness and Extrapolation
Yuhan Chen (Xiaomi Inc), Wei Liu (Xiaomi Inc)
Representation LearningTransformerLarge Language ModelText
🎯 What it does: Proposed a new relative position encoding called HoPE, breaking the traditional long-term decay principle and enhancing context awareness and extrapolation capabilities.
HotelMatch-LLM: Joint Multi-Task Training of Small and Large Language Models for Efficient Multimodal Hotel Retrieval
Arian Askari (Leiden University), Moran Beladev (Booking.com)
RetrievalTransformerLarge Language ModelVision Language ModelContrastive LearningMultimodality
🎯 What it does: This paper proposes HotelMatch-LLM, a multimodal dense retrieval model for hotel search that supports natural language queries;
How does Misinformation Affect Large Language Model Behaviors and Preferences?
Miao Peng (Hong Kong University of Science and Technology (Guangzhou)), Jia Li (Hong Kong University of Science and Technology (Guangzhou))
Anomaly DetectionPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed MISBENCH, a large-scale and diverse misinformation benchmark, to evaluate the performance of large language models (LLMs) under different types of knowledge conflicts and text styles; simultaneously proposed the RtD method based on retrieval and reconstruction to enhance LLMs' ability to identify misinformation.
How Humans and LLMs Organize Conceptual Knowledge: Exploring Subordinate Categories in Italian
Andrea Pedrotti (Istituto di Scienza e Tecnologie dell'Informazione 'A. Faedo'), Marianna Bolognesi (Università di Bologna)
ClassificationTransformerLarge Language ModelVision Language ModelText
🎯 What it does: Collected and analyzed 187 examples of Italian basic-level concrete vocabulary, and evaluated the performance of large language models in tasks such as example generation, category induction, and typicality judgment.
How LLMs Comprehend Temporal Meaning in Narratives: A Case Study in Cognitive Evaluation of LLMs
Karin De Langis, Dongyeop Kang (University of Minnesota)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This study constructs an 'expert-in-the-loop' detection pipeline to systematically evaluate large language models (LLMs) in their understanding and reasoning of aspect (tense meaning) in narrative texts.
How Much Do Encoder Models Know About Word Senses?
Simone Teglia, Roberto Navigli (Sapienza University of Rome)
ClassificationComputational EfficiencyRepresentation LearningTransformerSupervised Fine-TuningText
🎯 What it does: Without task fine-tuning, this study probes four mainstream encoder pre-trained language models (BERT, RoBERTa, ELECTRA, DeBERTa-v3). It calculates cluster centers for each sense using training set samples and assigns test samples to the nearest cluster center via cosine similarity, evaluating the models' sense discrimination ability on two semantic lists: WordNet (SemCor) and the Oxford English Dictionary (ODE).
How to Compare Things Properly? A Study of Argument Relevance in Comparative Question Answering
Irina Nikishina (University of Hamburg), Chris Biemann (University of Hamburg)
GenerationRetrievalTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Constructed a manually annotated comparative argument relevance dataset, and generated answers for comparison questions based on this dataset (the silver dataset was automatically generated by ChatGPT, while the gold dataset was manually refined by human editors).
How to Enable Effective Cooperation Between Humans and NLP Models: A Survey of Principles, Formalizations, and Beyond
Chen Huang (Sichuan University), Jimmy Xiangji Huang (York University)
Explainability and InterpretabilityReinforcement LearningTextReview/Survey Paper
🎯 What it does: Conduct a systematic review of human-machine collaboration, explaining its definition, principles, formalization methods, and proposing a unified three-class classification (sequential collaboration, divergent collaboration, joint collaboration), finally discussing frontier challenges at both technical and social levels.
How to Mitigate Overfitting in Weak-to-strong Generalization?
Junhao Shi (Fudan University), Xipeng Qiu (Fudan University)
OptimizationData-Centric LearningTransformerSupervised Fine-TuningTextBenchmarkChain-of-Thought
🎯 What it does: Proposes a two-stage weak-to-strong generalization framework, first improving supervision quality through model uncertainty filtering, then enhancing problem quality by re-annotating filtered questions with a fine-tuned strong model.
How to Train Long-Context Language Models (Effectively)
Tianyu Gao (Princeton University), Danqi Chen (Princeton University)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a complete training process for long-context language models (LLMs) and trains the ProLong 8B model based on this, significantly improving performance for context lengths of 128K and 512K.
HSCR: Hierarchical Self-Contrastive Rewarding for Aligning Medical Vision Language Models
Songtao Jiang (Zhejiang University), Zuozhu Liu (Zhejiang University)
Reinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningVision Language ModelContrastive LearningBiomedical Data
🎯 What it does: This paper proposes a Hierarchical Self-Contrastive Rewarding (HSCR) method, which generates high-quality preference data through self-contrastive rewarding and optimizes preferences at multiple levels, thereby enhancing modality alignment and reliability in medical vision-language models (Med-VLM).
Human Alignment: How Much Do We Adapt to LLMs?
Cazalets Tanguy (Ghent University), Joni Dambre (Ghent University)
TransformerLarge Language ModelTextSequential
🎯 What it does: Investigating how humans adapt their communication strategies when collaborating with large language models (LLMs) in lexical synchronization games
HumT DumT: Measuring and controlling human-like language in LLMs
Myra Cheng (Stanford University), Dan Jurafsky (Stanford University)
GenerationOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes two metrics based on LLM probabilities, HUMT and SOCIOT, to quantify the human-like tone in generated text, and employs the DUMT method to reduce the level of human-likeness while maintaining model performance.
HybGRAG: Hybrid Retrieval-Augmented Generation on Textual and Relational Knowledge Bases
Meng-Chieh Lee (Carnegie Mellon University), Christos Faloutsos (Carnegie Mellon University)
RetrievalExplainability and InterpretabilityAgentic AITextGraphBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes HYBGRAG, a hybrid retrieval-augmented generation framework for semi-structured knowledge bases, combining text and graph retrieval with a self-reflective evaluation module to address hybrid question answering (HQA);
Hybrid Preferences: Learning to Route Instances for Human vs. AI Feedback
Lester James Validad Miranda, Pradeep Dasigi (Allen Institute for AI)
Data-Centric LearningReinforcement Learning from Human FeedbackLarge Language ModelText
🎯 What it does: Propose the HYPER framework, which uses a routing mechanism to assign preference instances to humans or LLMs to construct a hybrid labeled set, and trains a reward model to improve performance.
HyKGE: A Hypothesis Knowledge Graph Enhanced RAG Framework for Accurate and Reliable Medical LLMs Responses
Xinke Jiang (Peking University), Yasha Wang (Peking University)
TransformerLarge Language ModelPrompt EngineeringTextGraphRetrieval-Augmented Generation
🎯 What it does: A retrieval-augmented generation framework called HyKGE based on a Hypothesis Knowledge Graph (HKG) was constructed to enhance the accuracy and reliability of large language models in the medical domain.
HyperFM: Fact-Centric Multimodal Fusion for Link Prediction over Hyper-Relational Knowledge Graphs
Yuhuan Lu (University of Macau), Dingqi Yang (University of Macau)
Representation LearningGraph Neural NetworkTransformerVision Language ModelMultimodalityGraph
🎯 What it does: Proposed HyperFM, a fact-centric multimodal fusion technique for link prediction in hyper-relational knowledge graphs.
I0T: Embedding Standardization Method Towards Zero Modality Gap
Na Min An (KAIST), Hyunjung Shim (KAIST)
RetrievalRepresentation LearningTransformerVision Language ModelContrastive LearningMultimodalityBenchmark
🎯 What it does: Propose and study a framework named I0T, aiming to significantly reduce the modality gap between image and text embeddings in CLIP models through methods such as embedding normalization or batch normalization.
IAM: Efficient Inference through Attention Mapping between Different-scale LLMs
Yi Zhao (Shanghai Jiao Tong University), Hai Zhao (Shanghai Jiao Tong University)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Leverage the attention matrix of a small LLM to map to a large LLM, accelerating inference and reducing KV cache usage while maintaining model performance.
ICR Probe: Tracking Hidden State Dynamics for Reliable Hallucination Detection in LLMs
Zhenliang Zhang (Peking University), Xiaojun Wan (Peking University)
Anomaly DetectionTransformerLarge Language ModelText
🎯 What it does: Introduce the ICR Score to quantify the residual flow update process of LLMs, and construct the ICR Probe by aggregating ICR Scores across all layers for unsupervised detection of hallucinations in text generation.
Identifying Cellular Niches in Spatial Transcriptomics: An Investigation into the Capabilities of Large Language Models
Huanhuan Wei (Yale School of Medicine), Xiting Yan (Yale School of Medicine)
RecognitionExplainability and InterpretabilityData-Centric LearningGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringBiomedical Data
🎯 What it does: Leverage large language models (LLMs) for spatial environment identification in spatial transcriptomics data, proposing the LLMiniST method with zero-shot and two-stage fine-tuning to automatically generate interpretable spatial niche labels.
Identifying Open Challenges in Language Identification
Rob Van Der Goot
RecognitionRecurrent Neural NetworkTransformerTextBenchmark
🎯 What it does: Constructed a multi-dimensional language identification benchmark, systematically evaluated the impact of input length, training samples, number of languages, domains, scripts, and language families on model performance, and compared five mainstream models.
Identifying Reliable Evaluation Metrics for Scientific Text Revision
Leane Jourdan, Richard Dufour (Nantes Université)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Study how to reliably evaluate the quality of scientific text revisions, systematically assess the effectiveness of traditional similarity metrics, cross-domain metrics, and LLM-as-a-judge methods.
If Attention Serves as a Cognitive Model of Human Memory Retrieval, What is the Plausible Memory Representation?
Ryo Yoshida (University of Tokyo), Yohei Oseki (University of Tokyo)
Explainability and InterpretabilityRepresentation LearningTransformerText
🎯 What it does: This study investigates whether the attention mechanism in Transformer Grammar (TG) can serve as a cognitive model for human memory retrieval, and evaluates its predictive power on reading time through Normalized Attention Entropy (NAE).
If Eleanor Rigby Had Met ChatGPT: A Study on Loneliness in a Post-LLM World
Adrian de Wynter (Microsoft)
Safty and PrivacyTransformerLarge Language ModelText
🎯 What it does: This study analyzed 79,951 conversations between ChatGPT and users in non-task-oriented scenarios, focusing on interactions deemed 'lonely,' and assessed their impact on user emotions and potential risks.
IMOL: Incomplete-Modality-Tolerant Learning for Multi-Domain Fake News Video Detection
Zhi Zeng (Xi'an Jiaotong University), Qinghua Zheng (Xi'an Jiaotong University)
Domain AdaptationAnomaly DetectionConvolutional Neural NetworkTransformerMixture of ExpertsContrastive LearningVideoMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the IMOL framework and two multi-domain incomplete-modal benchmarks for fake news video detection (FakeSVIM, FakeTTIM), achieving detection in multi-domain and modal-incomplete scenarios.
ImPart: Importance-Aware Delta-Sparsification for Improved Model Compression and Merging in LLMs
Yan Yang (Shanghai University of Finance and Economics), Guanhua Chen (Southern University of Science and Technology)
CompressionTransformerText
🎯 What it does: Propose the IMPART algorithm, which performs importance-aware sparsification on the incremental parameters of a fine-tuned LLM to maintain performance under high compression rates.
Impartial Multi-task Representation Learning via Variance-invariant Probabilistic Decoding
Dou Hu (Chinese Academy of Sciences), Songlin Hu (Chinese Academy of Sciences)
Representation LearningTransformerTextBenchmark
🎯 What it does: In multi-task learning, to address the biased learning problem caused by differences in task representation spaces, the VIP-MTL framework is proposed. It utilizes probabilistic decoding to map shared representations to task-specific probability distributions and achieves unbiased learning by normalizing variances to unify the distribution scales across tasks.
ImpliHateVid: A Benchmark Dataset and Two-stage Contrastive Learning Framework for Implicit Hate Speech Detection in Videos
Mohammad Zia Ur Rehman (Indian Institute of Technology Indore), Dr. Nagendra Kumar
ClassificationContrastive LearningVideoTextMultimodalityBenchmarkAudio
🎯 What it does: Constructed a video dataset named ImpliHateVid for implicit hate speech detection, and proposed a two-stage contrastive learning framework to identify implicit hate content in videos.
Improve Safety Training of Large Language Models with Safety-Critical Singular Vectors Localization
Peijian Gu (University Of Science And Technology Of China), Zhendong Mao (University Of Science And Technology Of China)
Safty and PrivacyComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: The study proposes a pluggable method that initializes LoRA by locating safety-critical singular vectors in model parameters, enabling safe training to update only safety-related parameters and minimizing impact on the model's general functionality.
Improve Vision Language Model Chain-of-thought Reasoning
Ruohong Zhang (CMULTI), Yiming Yang (Apple)
TransformerSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: This paper enhances the chain-of-thought (CoT) reasoning capability of vision-language models through a two-stage post-training strategy.
Improved Unbiased Watermark for Large Language Models
Ruibo Chen (University of Maryland, College Park), Heng Huang (University of Maryland, College Park)
GenerationTransformerLarge Language ModelText
🎯 What it does: Propose a distortion-free multi-channel watermark method called MCMARK for text generation in large language models;
Improving Automatic Evaluation of Large Language Models (LLMs) in Biomedical Relation Extraction via LLMs-as-the-Judge
Md Tahmid Rahman Laskar (York University), Jimmy Xiangji Huang (York University)
Drug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringBiomedical Data
🎯 What it does: This paper explores the feasibility of using large language models (LLMs) as judges (LLM-Judge) to automatically evaluate biomedical relation extraction models, and proposes structured output formats and domain adaptation techniques to improve judgment accuracy.
Improving Chain-of-Thought Reasoning via Quasi-Symbolic Abstractions
Leonardo Ranaldi (Idiap Research Institute), Andre Freitas
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Propose a new Chain-of-Thought variant called QuaSAR, which utilizes 'quasi-symbolic abstraction' to guide LLMs in performing abstraction, formalization, interpretation, and answer generation between natural language and symbolic expressions;
Improving Contextual Faithfulness of Large Language Models via Retrieval Heads-Induced Optimization
Lei Huang (Harbin Institute of Technology), Bing Qin (Harbin Institute of Technology)
OptimizationTransformerSupervised Fine-TuningReinforcement LearningContrastive LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the RHIO framework, which generates realistic untrustworthy samples through masked retrieval heads, trains models to distinguish between trustworthy and untrustworthy answers using control codes, and further enhances contextual credibility during inference via self-contrast decoding.
Improving Dialogue Discourse Parsing through Discourse-aware Utterance Clarification
Yaxin Fan (Soochow University), Qiaoming Zhu (Soochow University)
Data SynthesisTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Proposed and implemented a discourse-aware clarification module (DCM) to eliminate ambiguity in dialogues, combined with contribution-aware preference optimization (CPO) to adaptively improve the module, thereby enhancing the accuracy of discourse parsing in dialogues.
Improving Dialogue State Tracking through Combinatorial Search for In-Context Examples
Haesung Pyun (Seoul National University), Yohan Jo (Seoul National University)
RetrievalOptimizationPrompt EngineeringContrastive LearningText
🎯 What it does: Proposed CombiSearch, a retrieval example method based on combinatorial search for dialogue state tracking tasks, which selects examples that synergistically improve model performance during retrieval.
Improving Factuality with Explicit Working Memory
Mingda Chen (Meta FAIR), Wen-tau Yih (Meta FAIR)
GenerationRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Designed and implemented the Ewe system, introducing explicit working memory in long-text generation, which receives real-time retrieval and fact-checking feedback, periodically pauses generation, refreshes memory, and corrects errors;
Improving Fairness of Large Language Models in Multi-document Summarization
Haoyuan Li (University of North Carolina at Chapel Hill), Snigdha Chaturvedi (University of North Carolina at Chapel Hill)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes FairPO, a preference optimization method for multi-document summarization, aiming to simultaneously improve fairness at both the summary level and the corpus level.
Improving Language and Modality Transfer in Translation by Character-level Modeling
Ioannis Tsiamas (FAIR at Meta), Marta R. Costa-jussà (FAIR at Meta)
Domain AdaptationKnowledge DistillationRepresentation LearningTransformerTextMultimodalityBenchmarkAudio
🎯 What it does: Studied a multilingual and speech translation method based on character-level encoding, achieving cross-lingual and cross-modal knowledge transfer through the SONAR and MMS models.
Improving Low-Resource Morphological Inflection via Self-Supervised Objectives
Adam Wiemerslage (University of Colorado Boulder), Katharina Von Der Wense (University of Colorado Boulder)
GenerationTransformerAuto EncoderText
🎯 What it does: Explored and compared the effectiveness of various self-supervised auxiliary tasks (such as autoencoding, character masked language models, etc.) on morphological inflection under extremely low-resource conditions.
Improving Medical Large Vision-Language Models with Abnormal-Aware Feedback
Yucheng Zhou (University of Macau), Jianbing Shen (University of Macau)
Anomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningPrompt EngineeringVision Language ModelBiomedical Data
🎯 What it does: Proposed UMed-LVLM, a vision-language model specifically designed for medical image diagnosis, emphasizing visualization localization and interpretation of abnormal regions.
Improving Model Factuality with Fine-grained Critique-based Evaluator
Yiqing Xie (Carnegie Mellon University), Hejia Zhang (Meta GenAI)
GenerationExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: This paper first trains a fine-grained critical factuality evaluator called FENCE, which provides criticism and scores during the generator's training. Subsequently, the evaluator revises the generator's output to generate high-quality training pairs, significantly improving the generator's factuality.
Improving Parallel Sentence Mining for Low-Resource and Endangered Languages
Shu Okabe (Technische Universität München), Alexander Fraser (Technische Universität München)
Data SynthesisRetrievalRepresentation LearningData-Centric LearningTransformerContrastive LearningTextBenchmark
🎯 What it does: Developed the BELOPSEM benchmark and improved and evaluated unsupervised sentence alignment and clustering-based isometric enhancement for sentence mining on three pairs of low-resource/endangered language pairs.
Improving Preference Extraction In LLMs By Identifying Latent Knowledge Through Classifying Probes
Sharan Maiya (University of Cambridge), Anna Korhonen (University of Cambridge)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningLarge Language ModelContrastive LearningText
🎯 What it does: Extract preference judgments from LLM hidden layers using a linear classification probe
Improving the Calibration of Confidence Scores in Text Generation Using the Output Distribution’s Characteristics
Lorenzo Jaime Yu Flores (Mila Quebec AI Institute), Jackie CK Cheung (Mila Quebec AI Institute)
GenerationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes two task-agnostic confidence evaluation methods based on the characteristics of the generated output distribution (ratio and tail thinness), and verifies their effectiveness in various text generation tasks.
In Prospect and Retrospect: Reflective Memory Management for Long-term Personalized Dialogue Agents
Zhen Tan (Arizona State University), Tomas Pfister (Google Cloud AI Research)
Recommendation SystemSafty and PrivacyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the Reflective Memory Management (RMM) framework, which decomposes dialogues into memory libraries by topics through forward reflection, and performs online reinforcement learning on retrievers using LLM attribution signals via backward reflection to achieve long-term personalized dialogue;
In-the-wild Audio Spatialization with Flexible Text-guided Localization
Tianrui Pan (Nanjing University), Gangshan Wu (Nanjing University)
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelAuto EncoderTextAudio
🎯 What it does: Propose a text-guided audio spatialization framework (TAS) that can convert monaural audio into stereo audio with a sense of space;
Incongruity-aware Tension Field Network for Multi-modal Sarcasm Detection
Jiecheng Zhang (South China University of Technology), Tong Zhang (South China University of Technology)
ClassificationTransformerVision Language ModelContrastive LearningMultimodality
🎯 What it does: Propose an inconsistency-priority learning framework called ITFNet for multimodal sarcasm detection, utilizing a tension field to capture inconsistent information in text-image pairs;
Inconsistent Tokenizations Cause Language Models to be Perplexed by Japanese Grammar
Andrew Gambardella (University of Tokyo), Yutaka Matsuo (University of Tokyo)
TransformerLarge Language ModelText
🎯 What it does: This paper constructs minimal contrastive sentence pairs and evaluates the performance of several open-source large language models with 7–10B parameters on the subtle Japanese grammatical rule of 'first-person psychological predicate restrictions' using perplexity and machine translation experiments, while analyzing the impact of tokenization on model outputs.
Incorporating Domain Knowledge into Materials Tokenization
Yerim Oh (Korea University), SangKeun Lee (Korea University)
ClassificationTransformerTextPhysics Related
🎯 What it does: Propose MATTER, a framework that integrates material knowledge into subword tokenization, enhancing the performance of language models on materials science texts.
IndicSynth: A Large-Scale Multilingual Synthetic Speech Dataset for Low-Resource Indian Languages
Divya V Sharma (Indian Institute of Information Technology Delhi), Anubha Gupta (Indian Institute of Information Technology Delhi)
GenerationData SynthesisTransformerSupervised Fine-TuningAuto EncoderAudio
🎯 What it does: This paper created the IndicSynth dataset, containing approximately 4000 hours of multilingual synthetic speech (12 low-resource Indian languages) from 989 speakers, divided into two subsets: realistic mimicry and diversity.
Inducing lexicons of in-group language with socio-temporal context
Christine de Kock (University of Melbourne)
GenerationRepresentation LearningText
🎯 What it does: Propose the LISTN method to automatically generate and evaluate in-group lexicons for online communities by leveraging dynamic word and user embeddings, combined with social and temporal information.
InductionBench: LLMs Fail in the Simplest Complexity Class
Wenyue Hua (University of California Santa Barbara), William Yang Wang (University of California Santa Barbara)
TransformerTextBenchmarkChain-of-Thought
🎯 What it does: Propose InductionBench to evaluate the ability of large language models (LLMs) in inductive reasoning (inferring string-to-string transformation rules from limited samples) and systematically test the model's performance under different context window sizes, alphabet sizes, number of rules, and sample scales.
iNews: A Multimodal Dataset for Modeling Personalized Affective Responses to News
Tiancheng Hu (University of Cambridge), Nigel Collier (University of Cambridge)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelPrompt EngineeringImageTextMultimodalityBenchmark
🎯 What it does: Construct the iNews dataset by collecting multimodal (text-image) emotional response labels from 291 UK participants to 2899 Facebook news posts, along with detailed personal profile information.
Inference Compute-Optimal Video Vision Language Models
Peiqi Wang (MIT), Qifan Wang (Meta)
Computational EfficiencyLarge Language ModelSupervised Fine-TuningVision Language ModelVideoText
🎯 What it does: The study investigates how to optimally allocate three key scale factors (language model size, number of frames, and per-frame visual token count) of video vision-language models under a fixed inference computational budget, and solves the optimal frontier of inference computation through large-scale training sweeps and parameterized performance models.
Inferring from Logits: Exploring Best Practices for Decoding-Free Generative Candidate Selection
Mingyu Derek Ma (University of California, Los Angeles), Wei Wang (University of California, Los Angeles)
Computational EfficiencyTransformerLarge Language ModelTextBiomedical Data
🎯 What it does: This paper proposes and systematically evaluates the 'decoding-free generative candidate selection' method, which estimates the probability of candidate answers directly using the logits from the initial output step of a language model, without requiring complete autoregressive decoding.
Inferring Functionality of Attention Heads from their Parameters
Amit Elhelo (Tel Aviv University), Mor Geva (Tel Aviv University)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose the MAPS framework, which directly infers the functionality of attention heads using their parameters without requiring training or inference.
Influences on LLM Calibration: A Study of Response Agreement, Loss Functions, and Prompt Styles
Yuxi Xia (University of Vienna), Benjamin Roth (University of Vienna)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose the Calib-n framework, which aggregates responses from multiple LLMs and trains an auxiliary Transformer for confidence estimation, while exploring the impact of response consistency, loss functions, and prompt styles on LLM calibration.
Infogen: Generating Complex Statistical Infographics from Documents
Akash Ghosh (Indian Institute of Technology Patna), Sriparna Saha (Indian Institute of Technology Patna)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningAgentic AIImageTextBenchmark
🎯 What it does: This paper proposes the task of generating complex statistical infographics (containing multiple subgraphs) from text documents, creates the first corresponding benchmark dataset Infodat, and designs a two-stage framework Infogen. The framework first generates structured metadata via an LLM, then converts it into executable code to produce the final infographic.
Information Extraction from Visually Rich Documents using LLM-based Organization of Documents into Independent Textual Segments
Aniket Bhattacharyya (Amazon), Maneesh Gupta (Amazon)
TransformerLarge Language ModelPrompt EngineeringMultimodalityBenchmark
🎯 What it does: Propose the BLOCKIE method, which leverages large language models to decompose visually rich documents (VRD) into self-contained semantic blocks, performs reasoning on individual blocks, and merges results to achieve final information extraction.
Information Locality as an Inductive Bias for Neural Language Models
Taiga Someya (University of Tokyo), Ryan Cotterell (ETH Zürich)
Representation LearningRecurrent Neural NetworkTransformerText
🎯 What it does: This paper defines m-local entropy to measure the local uncertainty of language, and investigates the learning difficulty of neural language models under different levels of local entropy.
INJONGO: A Multicultural Intent Detection and Slot-filling Dataset for 16 African Languages
Hao Yu, David Ifeoluwa Adelani (Saarland University)
ClassificationRecognitionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Developed the INJONGO dataset, providing cross-cultural intent detection and slot filling annotations for 16 African languages.
Inner Thinking Transformer: Leveraging Dynamic Depth Scaling to Foster Adaptive Internal Thinking
Yilong Chen (Chinese Academy of Sciences), Haifeng Wang (Baidu)
Computational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose Inner Thinking Transformer (ITT), which enables the model to perform deeper internal reasoning on key tokens during inference through dynamic depth scheduling and residual thinking connections.
Innovative Image Fraud Detection with Cross-Sample Anomaly Analysis: The Power of LLMs
QiWen Wang, Chen Lin (Xiamen University)
Anomaly DetectionTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmarkFinance RelatedRetrieval-Augmented Generation
🎯 What it does: Propose the CSIAD framework for image fraud detection based on cross-sample logical reasoning, combining retrieval, rule generation, and fact verification to accurately locate fine-grained forged regions and provide explanations.
InSerter: Speech Instruction Following with Unsupervised Interleaved Pre-training
Dingdong Wang (Chinese University of Hong Kong), Junyang Lin (Chinese University of Hong Kong)
RecognitionRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningTextMultimodalityBenchmarkAudio
🎯 What it does: Propose an scalable unsupervised interleaved pre-training method (InSerter) for speech LLMs, which interleaves speech segments generated by text-to-speech synthesis with original text, enabling the model to learn the task of 'predicting the next text word given a speech segment' during pre-training; simultaneously, construct a specialized benchmark called SpeechInstructBench for evaluating speech instruction following ability; and validate its advantages on multiple speech tasks including VoiceBench;
Insight Over Sight: Exploring the Vision-Knowledge Conflicts in Multimodal LLMs
Xiaoyuan Liu (Chinese University of Hong Kong), Zhaopeng Tu (Tencent)
Data SynthesisExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodalityBenchmark
🎯 What it does: This paper designs and implements an automated framework for generating images and question-answer pairs that conflict with common sense, constructing the CONFLICTVIS benchmark containing 374 images and 1122 high-quality QA pairs, and evaluates nine multimodal large language models (MLLMs).
InspireDebate: Multi-Dimensional Subjective-Objective Evaluation-Guided Reasoning and Optimization for Debating
Fuyu Wang (Tongji University), Changjun Jiang (Tongji University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the InspireScore evaluation framework and the InspireDebate optimization framework for multi-dimensional debate assessment and improvement.