ACL 2025 Papers — Page 5
Annual Meeting of the Association for Computational Linguistics · 1699 papers
Debiasing the Fine-Grained Classification Task in LLMs with Bias-Aware PEFT
Daiying Zhao (Xi'an Jiaotong University), Hang Chen (Xi'an Jiaotong University)
ClassificationTransformerSupervised Fine-TuningPrompt EngineeringContrastive LearningText
🎯 What it does: Propose a parameter-efficient fine-tuning (PEFT)-based intermediate layer bias balance optimization framework to eliminate prediction tendency bias and discriminative ability bias in large language models for fine-grained classification tasks.
Decoder-Only LLMs can be Masked Auto-Encoders
Dan Qiao (Soochow University), Min Zhang (Soochow University)
Representation LearningTransformerLarge Language ModelAuto EncoderText
🎯 What it does: This paper proposes the UniMAE method, which transforms the decoder-only LLM into a masked autoencoder capable of simultaneously performing generation and embedding tasks, using the EOS position as an implicit representation of sentence semantics;
Decoding by Contrasting Knowledge: Enhancing Large Language Model Confidence on Edited Facts
Baolong Bi (Key Laboratory of Network Data Science and Technology, ICT, CAS), Xueqi Cheng (Key Laboratory of Network Data Science and Technology, ICT, CAS)
TransformerLarge Language ModelContrastive LearningText
🎯 What it does: This paper proposes DeCK, a decoding strategy based on contrastive knowledge, to enhance the contextual editing capability of large language models (LLMs), thereby more effectively updating outdated facts.
Decoding Knowledge Attribution in Mixture-of-Experts: A Framework of Basic-Refinement Collaboration and Efficiency Analysis
Junzhuo Li (Hong Kong University of Science and Technology), Xuming Hu (Hong Kong University of Science and Technology)
Explainability and InterpretabilityComputational EfficiencyTransformerMixture of ExpertsTextBenchmark
🎯 What it does: This paper proposes a cross-level knowledge attribution algorithm to explain the dynamic expert collaboration in sparse MoE models. By comparing MoE architectures such as Qwen1.5-MoE, Mixtral, and OLMoE with dense models like Qwen1.5-7B, Llama-7B, and Mistral-7B, it reveals the basic-refinement collaboration pattern and efficiency advantages of MoE.
Decoding on Graphs: Faithful and Sound Reasoning on Knowledge Graphs through Generation of Well-Formed Chains
Kun Li (Chinese University of Hong Kong), Helen M. Meng (Chinese University of Hong Kong)
TransformerLarge Language ModelGraphChain-of-Thought
🎯 What it does: Proposed the DoG framework, which generates credible reasoning chains on knowledge graphs using LLMs;
Decoding Reading Goals from Eye Movements
Omer Shubi (Technion Israel Institute of Technology), Yevgeni Berzak (Technion Israel Institute of Technology)
ClassificationRecurrent Neural NetworkTransformerMultimodalityTime Series
🎯 What it does: This paper automatically distinguishes whether readers are engaged in general reading or information retrieval using eye-tracking data, and proposes the task of predicting the reading goal from the eye movement sequence of a single text segment.
Deep Temporal Reasoning in Video Language Models: A Cross-Linguistic Evaluation of Action Duration and Completion through Perfect Times
Olga Loginova (University of Trento), Sofía Ortega Loguinova (Maastricht University)
TransformerLarge Language ModelVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: This paper proposes a new cross-lingual video-language multimodal QA benchmark called Perfect Times, and uses it to evaluate the temporal reasoning capabilities of existing video language models.
DEEPER Insight into Your User: Directed Persona Refinement for Dynamic Persona Modeling
Aili Chen (Fudan University), Yanghua Xiao (Fudan University)
Recommendation SystemOptimizationTransformerLarge Language ModelReinforcement LearningTextTabular
🎯 What it does: Propose the DEEPER method to achieve continuous optimization of dynamic personas.
DeepReview: Improving LLM-based Paper Review with Human-like Deep Thinking Process
Minjun Zhu (Zhejiang University), Yue Zhang (University College London)
TransformerLarge Language ModelSupervised Fine-TuningTextReview/Survey PaperBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose DeepReviewer, a multi-stage LLM paper review framework that simulates the expert review process, including innovation verification, multi-dimensional evaluation, and reliability verification, and construct the DeepReview-13K dataset and DeepReviewBench benchmark, training a 14B parameter model.
DeepSolution: Boosting Complex Engineering Solution Design via Tree-based Exploration and Bi-point Thinking
Zhuoqun Li (Chinese Information Processing Laboratory, Institute of Software, Chinese Academy of Sciences), Le Sun (Chinese Information Processing Laboratory, Institute of Software, Chinese Academy of Sciences)
TextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes a new benchmark called SolutionBench and a RAG system named SolutionRAG based on tree exploration and dual-point thinking, for automatically generating complex engineering solutions that satisfy multiple constraints.
Defense Against Prompt Injection Attack by Leveraging Attack Techniques
Yulin Chen (National University of Singapore), Bryan Hooi (National University of Singapore)
Adversarial AttackTransformerPrompt EngineeringText
🎯 What it does: This paper proposes a novel defense framework that reverses known prompt injection attack techniques to construct defensive prompts (shield prompts), thereby preventing large language models from executing injected malicious instructions while maintaining correct responses to original instructions.
Defining and Evaluating Visual Language Models’ Basic Spatial Abilities: A Perspective from Psychometrics
Wenrui Xu (Tsinghua University), Yong Li (Tsinghua University)
Vision Language ModelImageBenchmarkChain-of-Thought
🎯 What it does: Proposed and implemented a psychometric-based framework for assessing five fundamental spatial abilities (BSAs), systematically evaluating 13 mainstream vision-language models (VLMs) using nine classical psychological tests, and comparing them with human benchmarks.
Dehumanizing Machines: Mitigating Anthropomorphic Behaviors in Text Generation Systems
Myra Cheng (Stanford University), Alexandra Olteanu (Microsoft Research)
GenerationTextReview/Survey Paper
🎯 What it does: Through literature review and crowdsourcing experiments, systematically organized and constructed a checklist of 28 interventions to reduce anthropomorphic behaviors in text generation systems, and proposed a four-dimensional conceptual framework.
Déjà Vu? Decoding Repeated Reading from Eye Movements
Yoav Meiri (Technion Israel Institute of Technology), Yevgeni Berzak (Technion Israel Institute of Technology)
ClassificationTransformerTextMultimodalityTime Series
🎯 What it does: This paper studies how to determine whether a reader has previously read the same text by analyzing their eye movement trajectories, and proposes two prediction tasks: single-sample and paired-sample.
Deliberate Reasoning in Language Models as Structure-Aware Planning with an Accurate World Model
Siheng Xiong, Faramarz Fekri (Georgia Institute Of Technology)
TransformerLarge Language ModelWorld ModelTextBenchmark
🎯 What it does: Propose the SWAP framework, which views language model reasoning as structured inference graph construction, and gradually updates the graph structure through the interaction between a policy network and a world model;
Delta-KNN: Improving Demonstration Selection in In-Context Learning for Alzheimer’s Disease Detection
Chuyuan Li (University of British Columbia), Giuseppe Carenini (University of British Columbia)
ClassificationTransformerLarge Language ModelPrompt EngineeringTextAlzheimer's DiseaseChain-of-Thought
🎯 What it does: Propose a demonstration selection method based on Delta-KNN, which evaluates example gains using a Delta matrix and dynamically selects optimal examples by combining KNN retrieval, to enhance the ICL performance of large language models in Alzheimer's text detection.
Delving into Multilingual Ethical Bias: The MSQAD with Statistical Hypothesis Tests for Large Language Models
Seunguk Yu (Chung-Ang University), YoungBin Kim
Safty and PrivacyExplainability and InterpretabilityLarge Language ModelText
🎯 What it does: This paper constructs a multilingual sensitive question-and-answer dataset called MSQAD and evaluates the ethical biases of large language models in different languages regarding sensitive questions through statistical hypothesis testing.
Demons in the Detail: On Implementing Load Balancing Loss for Training Specialized Mixture-of-Expert Models
Zihan Qiu (Alibaba Group), Junyang Lin (Alibaba Group)
Computational EfficiencyMixture of ExpertsText
🎯 What it does: Studied the implementation of Load-Balancing Loss (LBL) during training in Mixture-of-Experts (MoE) models, and proposed computing LBL at the global batch level to relax microbatch constraints, improving expert specialization and model performance.
Demystifying Small Language Models for Edge Deployment
Zhenyan Lu (Beijing University of Posts and Telecommunications), Mengwei Xu (Beijing University of Posts and Telecommunications)
Computational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Evaluated and benchmarked 68 small language models (SLMs) with 100M-5B parameters, constructing an end-to-end mobile evaluation suite and leaderboard, systematically measuring their inference capabilities and runtime costs on edge devices.
DenseLoRA: Dense Low-Rank Adaptation of Large Language Models
Lin Mu (Anhui University), Yiwen Zhang (Anhui University)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: DenseLoRA introduces a shared Encoder-Decoder structure in large language models (e.g., LLaMA2-7B and LLaMA3-8B). It first compresses and refines the hidden representations, then adapts them using a dense low-rank matrix, significantly reducing the number of trainable parameters while maintaining the model's expressive power.
Deontological Keyword Bias: The Impact of Modal Expressions on Normative Judgments of Language Models
Bumjin Park (KAIST AI), Jaesik Choi (INEEJI)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Study how large language models (LLMs) produce judgment biases in deontological scenarios due to modal words (e.g., 'must,' 'ought to') and propose a debiasing method based on few examples and reasoning prompts.
Design Choices for Extending the Context Length of Visual Language Models
Mukai Li (University of Hong Kong), Qi Liu (University of Hong Kong)
Representation LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoTextMultimodality
🎯 What it does: Built the GIRAFFE system, leveraging the ETVLM dataset, M-RoPE++ position embeddings, and hybrid-resolution training to successfully extend the context length of the Qwen-VL series of vision-language models to 128K while maintaining short-context performance.
Detecting Sockpuppetry on Wikipedia Using Meta-Learning
Luc Raszewski (University of Melbourne), Christine de Kock (University of Melbourne)
ClassificationAnomaly DetectionMeta LearningTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Detect malicious sockpuppetry on Wikipedia using meta-learning methods, treating each investigation as an independent task and performing binary classification based on written text;
Detection of Human and Machine-Authored Fake News in Urdu
Muhammad Zain Ali (University of Waikato), Tony C Smith (University of Waikato)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Developed a four-class fake news detection framework for Urdu, covering human real, human-generated real, human-generated fake, and machine-generated fake categories; generated machine-generated texts using GPT-4o and proposed a joint approach that splits the four-class task into two subtasks: machine-generated text detection and authenticity detection.
Developmentally-plausible Working Memory Shapes a Critical Period for Language Acquisition
Masato Mita (University of Tokyo), Yohei Oseki (University of Tokyo)
TransformerLarge Language ModelText
🎯 What it does: By introducing an exponential growth limit on working memory in Transformer language model training, simulating the cognitive development during the critical period of human language acquisition, thus enhancing the model's syntactic learning efficiency.
Dialectal Coverage And Generalization in Arabic Speech Recognition
Amirbek Djanibekov (Mohamed bin Zayed University of Artificial Intelligence), Hanan Aldarmaki (Mohamed bin Zayed University of Artificial Intelligence)
RecognitionDomain AdaptationTransformerSupervised Fine-TuningContrastive LearningAudio
🎯 What it does: Proposed an automatic speech recognition (ASR) model for multivariate Arabic (including Modern Standard Arabic, regional dialects, and code-switching spoken languages with English, French, etc.), and released pre-trained and fine-tuned model weights;
Dialogue Systems for Emotional Support via Value Reinforcement
Juhee Kim (Seoul National University), Yohan Jo (Seoul National University)
Reinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This study proposes a value reinforcement-based emotional support dialogue system that enhances support quality by identifying and reinforcing the core values of the seeker;
Dialogue-RAG: Enhancing Retrieval for LLMs via Node-Linking Utterance Rewriting
Qiwei Li (Wuhan University), Hai Zhao (Shanghai Jiao Tong University)
GenerationRetrievalRecurrent Neural NetworkTransformerTextSequentialBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed a retrieval-augmented generation framework called Dialogue-RAG based on Incomplete Utterance Rewriting (IUR), which utilizes a lightweight node-link model to complete omissions and anaphora in dialogues, thereby improving retrieval accuracy and generation quality.
DialUp! Modeling the Language Continuum by Adapting Models to Dialects and Dialects to Models
Niyati Bafna (Johns Hopkins University), Hale Sirin (Johns Hopkins University)
Domain AdaptationData-Centric LearningTransformerSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper proposes the DialUp method, which improves machine translation performance for low-resource dialects by manually dialectalizing high-resource languages (HRL) during training (M→D) and mixing dialect inputs with HRL codes during inference (D→M).
Did Translation Models Get More Robust Without Anyone Even Noticing?
Ben Peters (Instituto de Telecomunicações), Andre Martins
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper systematically evaluates the performance of various translation models (traditional NMT, cross-lingual models, LLM) under synthetic noise and social media noise, demonstrating that modern large models exhibit higher noise robustness without targeted robustness training; it also explores the feasibility of improving small model robustness through fine-tuning and source text error correction pipelines.
Different Speech Translation Models Encode and Translate Speaker Gender Differently
Dennis Fucci (University of Trento), Giuseppe Attanasio (Instituto de Telecomunicações)
Explainability and InterpretabilityRepresentation LearningTransformerTextAudio
🎯 What it does: Investigate how different speech translation (ST) models encode speaker gender in their internal representations and evaluate the impact on gender translation.
DiffPO: Diffusion-styled Preference Optimization for Inference Time Alignment of Large Language Models
Ruizhe Chen (Zhejiang University), Zuozhu Liu (Zhejiang University)
OptimizationLarge Language ModelDiffusion modelTextBenchmark
🎯 What it does: Align large language models at the sentence level during the inference phase, proposing the DIFFPO framework to achieve efficient, plug-and-play alignment;
DiffuseDef: Improved Robustness to Adversarial Attacks via Iterative Denoising
Zhenhao Li (Imperial College London), Lucia Specia (Imperial College London)
ClassificationAdversarial AttackTransformerDiffusion modelText
🎯 What it does: Insert a diffusion layer into the classifier of a pre-trained language model, iteratively denoise the hidden representations, thereby constructing an adversarial defense framework named DiffuseDef.
Diffusion Directed Acyclic Transformer for Non-Autoregressive Machine Translation
Quan Nguyen-Tri (FPT Software AI Center), Hoang Thanh-Tung (University of Engineering and Technology Vietnam National University Hanoi)
GenerationTransformerDiffusion modelText
🎯 What it does: Propose Diffusion Directed Acyclic Transformer (Diff-DAT), integrating diffusion models into Directed Acyclic Transformer (DAT) to replace GLAT, addressing multimodal problems and enhancing decoding quality;
Diffusion Models Through a Global Lens: Are They Culturally Inclusive?
Zahra Bayramli (KAIST), Alice Oh (KAIST)
GenerationData SynthesisTransformerDiffusion modelContrastive LearningImageTextBenchmark
🎯 What it does: Investigated the generation capability of text-to-image diffusion models for cultural elements across ten countries, and constructed and evaluated the CULTDIFF benchmark.
Digest the Knowledge: Large Language Models empowered Message Passing for Knowledge Graph Question Answering
Junhong Wan (Hikvision Research Institute), Jiang Zhu (Hikvision Research Institute)
Explainability and InterpretabilityGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextGraphRetrieval-Augmented Generation
🎯 What it does: Propose a Language Message Passing (LMP) framework based on large language models, which first iteratively aggregates neighboring entities in a knowledge graph and converts them into semantic facts to form a fact graph, then converts it into a multi-level list through topological reading to enhance the question-answering capabilities of LLMs.
Digital Gatekeepers: Google’s Role in Curating Hashtags and Subreddits
Amrit Poudel (University of Notre Dame), Jürgen Pfeffer (Technical University of Munich)
Recommendation SystemTransformerText
🎯 What it does: By comparing non-sampled Reddit and Twitter/X data with Google Search Engine Results (SERP), the study investigates how Google filters, suppresses, or promotes specific subreddits and hashtags, and explores its impact on information visibility and public discourse.
DioR: Adaptive Cognitive Detection and Contextual Retrieval Optimization for Dynamic Retrieval-Augmented Generation
Hanghui Guo (Zhejiang Normal University), Jiajie Xu (Zhejiang Normal University)
GenerationRetrievalOptimizationRecurrent Neural NetworkTransformerTextRetrieval-Augmented Generation
🎯 What it does: Propose a Dynamic Retrieval-Augmented Generation (Dynamic RAG) framework named DioR, which can actively determine when and what to retrieve during the generation process, thereby significantly reducing the hallucination problem in LLMs.
Direct Prompt Optimization with Continuous Representations
Yangkun Wang (University of California San Diego), Jingbo Shang (University of California San Diego)
ClassificationOptimizationAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes a direct prompt optimization framework called SGCG (Soft Greedy Coordinate Gradient), which updates the prompt distribution first and then samples Gumbel variables at each step, using greedy strategies and sliding window techniques to improve search efficiency.
Disambiguating Reference in Visually Grounded Dialogues through Joint Modeling of Textual and Multimodal Semantic Structures
Shun Inadumi (Nara Institute of Science and Technology), Koichiro Yoshino (Institute of Science Tokyo)
TransformerSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose a unified framework that integrates text reference resolution (core anaphora, predicate-argument structure, bridge anaphora) with multimodal reference resolution (correspondence between text and image objects), resolving referential ambiguity in visual dialogues by extracting mention and object embeddings and calculating similarity.
DISC: Plug-and-Play Decoding Intervention with Similarity of Characters for Chinese Spelling Check
Ziheng Qiao (Soochow University), Fei Huang (Alibaba Group)
Anomaly DetectionText
🎯 What it does: Proposes a lightweight, pluggable decoding intervention module called DISC (Decoding Intervention based on Character Phonetic-Shape Similarity), which can enhance the performance of Chinese spell-check models without additional training.
Discourse Relation-Enhanced Neural Coherence Modeling
Wei Liu (Heidelberg Institute for Theoretical Studies gGmbH), Michael Strube (Heidelberg Institute for Theoretical Studies gGmbH)
Representation LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper investigates and verifies the impact of discourse relations on text coherence, and proposes a neural model that integrates text and relation features to assess coherence.
Disentangling Biased Knowledge from Reasoning in Large Language Models via Machine Unlearning
Zheyuan Liu (University of Notre Dame), Meng Jiang (University of Notre Dame)
Safty and PrivacyData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: Propose a three-stage machine unlearning framework SDU to remove biased knowledge from large language models while preserving inference capabilities.
Disentangling Language and Culture for Evaluating Multilingual Large Language Models
Jiahao Ying (Singapore Management University), Wenxuan Zhang (Singapore University of Technology and Design)
Explainability and InterpretabilityLarge Language ModelText
🎯 What it does: Proposed a dual-dimensional evaluation framework (language and culture) to assess the capabilities of multilingual large models, and under this framework, discovered the cultural-linguistic synergy phenomenon;
Disentangling Memory and Reasoning Ability in Large Language Models
Mingyu Jin (Rutgers University), Yongfeng Zhang (Rutgers University)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Propose a new LLM reasoning paradigm that splits the reasoning process into two steps: memory recall and reasoning, using special markers 〈memory〉 and 〈reason〉 to guide the model in distinguishing between these two types of operations;
Disentangling the Roles of Representation and Selection in Data Pruning
Yupei Du (Utrecht University), Dong Nguyen (Utrecht University)
Representation LearningData-Centric LearningTransformerSupervised Fine-TuningText
🎯 What it does: Systematically decomposes data pruning into two components: data representation and selection algorithm, and conducts theoretical and experimental studies on their impacts on instance selection and model performance.
Distilling an End-to-End Voice Assistant Without Instruction Training Data
William Held (Stanford University), Diyi Yang (Stanford University)
Knowledge DistillationRepresentation LearningData-Centric LearningTransformerLarge Language ModelTextMultimodalityAudio
🎯 What it does: Trained an end-to-end voice assistant DiVA by converting audio to text and using the response of a text LLM as a self-supervised signal, without the need for labeled instruction data.
DIVE into MoE: Diversity-Enhanced Reconstruction of Large Language Models from Dense into Mixture-of-Experts
Yuchen Feng (Chinese Academy of Sciences), Weiping Wang (Chinese Academy of Sciences)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsText
🎯 What it does: Proposes the DIVE method, converting dense LLMs into Mixture-of-Experts (MoE) architecture through pruning and restructuring while maintaining or improving model performance.
Diversity Explains Inference Scaling Laws: Through a Case Study of Minimum Bayes Risk Decoding
Hidetaka Kamigaito (Nara Institute of Science and Technology), Taro Watanabe (Nara Institute of Science and Technology)
GenerationExplainability and InterpretabilityImageText
🎯 What it does: Theoretically explain the mechanism of performance improvement through bias-diversity decomposition of Minimum Bayes Risk (MBR) decoding, and propose the pseudo bias and diversity-enhanced MBR (MAMBR) method based on approximate bias measurement, followed by experimental verification on three NLP tasks: machine translation, text summarization, and image captioning.
Diversity-oriented Data Augmentation with Large Language Models
Zaitian Wang (Chinese Academy of Sciences), Yuanchun Zhou (Chinese Academy of Sciences)
Data SynthesisTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Proposed and implemented a diversity-oriented text data augmentation framework called DoAug.
Divide-Then-Aggregate: An Efficient Tool Learning Method via Parallel Tool Invocation
Dongsheng Zhu (Baidu Inc), Dawei Yin (Baidu Inc)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningAgentic AITextBenchmark
🎯 What it does: Proposes DTA-Llama, a framework that enhances LLM task completion capabilities through parallel tool calls
Divide-Then-Align: Honest Alignment based on the Knowledge Boundary of RAG
Xin Sun (University of Science and Technology of China), Liang Wang (National Laboratory of Pattern Recognition)
Explainability and InterpretabilityTransformerSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: Propose the Divide-Then-Align (DTA) post-training framework to improve the robustness of retrieval-augmented generation models under retrieval noise and enable the model to actively respond with 'I don't know' when the retrieval exceeds the parameter knowledge boundary.
DNASpeech: A Contextualized and Situated Text-to-Speech Dataset with Dialogues, Narratives and Actions
Chuanqi Cheng (Wuhan Textile University), Rui Yan (Wuhan University)
GenerationPrompt EngineeringTextBenchmarkAudio
🎯 What it does: Proposed the Contextual and Scenario-based Text-to-Speech task (CS-TTS) and constructed a new dataset called DNASpeech, which includes dialogue, narrative, and action descriptions.
DNCASR: End-to-End Training for Speaker-Attributed ASR
Xianrui Zheng (University of Cambridge), Phil Woodland
RecognitionTransformerSupervised Fine-TuningContrastive LearningAudio
🎯 What it does: Developed an end-to-end jointly trainable speaker-attributed ASR system named DNCASR, which achieves speaker-attributed transcription for multi-speaker meetings by utilizing two independent encoders (global speaker features and local waveform information) and a linked decoder.
Do Language Models Have Semantics? On the Five Standard Positions
Anders Søgaard (University of Copenhagen)
Explainability and InterpretabilityReview/Survey Paper
🎯 What it does: This paper systematically reviews and compares five academic positions regarding whether large language models (LLMs) possess semantic understanding, proposes a new sixth comprehensive stance, and evaluates the logical consistency and explainability of each position.
Do Language Models Understand Honorific Systems in Javanese?
Mohammad Rifqi Farhansyah (Institut Teknologi Bandung), Derry Tanti Wijaya (Monash Indonesia)
ClassificationGenerationRecurrent Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: This work proposes and makes publicly available the first Javanese honorific hierarchy annotated corpus, UNGGAH-UNG... containing 4,024 sentences and 16 dialogue cases. Based on this corpus, four NLP tasks were designed (honorific hierarchy classification, honorific style transfer, cross-lingual honorific translation, and dialogue generation), systematically evaluating the capabilities of various language models in processing Javanese honorifics.
Do Large Language Models have an English Accent? Evaluating and Improving the Naturalness of Multilingual LLMs
Yanzhu Guo (Inria Paris), Henry Xiao (Apple)
GenerationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Evaluate and enhance the naturalness of multilingual LLMs, propose automatic metrics at lexical and syntactic levels, and improve model naturalness based on preference learning
Do LLMs Understand Dialogues? A Case Study on Dialogue Acts
Ayesha Qamar (Texas A&M University), Ruihong Huang (Texas A&M University)
ClassificationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper systematically evaluates the performance of large language models (LLMs) in fine-grained dialogue act (DA) classification within multi-party conversations. It proposes and constructs three fundamental prerequisite tasks for dialogue understanding (turn management, interactive function, and dialogue structure), quantifies the correlation between LLM errors on these tasks and DA classification errors, and compares them with human annotations.
Do Multimodal Large Language Models Truly See What We Point At? Investigating Indexical, Iconic, and Symbolic Gesture Comprehension
Noriki Nishida (RIKEN), Katsuya Takanashi (University of Shiga Prefecture)
TransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoTextMultimodality
🎯 What it does: This paper constructs a video and text dataset based on interviews from the Miraikan Science Museum to evaluate the ability of Multimodal Large Language Models (MLLM) to understand indexical, symbolic, and iconic gestures.
Do not Abstain! Identify and Solve the Uncertainty
Jingyu Liu (Renmin University of China), Yong Liu (Renmin University of China)
RetrievalOptimizationExplainability and InterpretabilityTransformerTextBenchmarkChain-of-Thought
🎯 What it does: Propose the ConfuseBench benchmark and methodology to investigate overconfidence of large language models (LLMs) in uncertain scenarios, aiming to identify and resolve uncertainty sources;
Doc-React: Multi-page Heterogeneous Document Question-answering
Junda Wu (University of California San Diego), Ritwik Sinha (Adobe Inc)
RetrievalTransformerLarge Language ModelContrastive LearningMultimodalityRetrieval-Augmented Generation
🎯 What it does: This paper proposes Doc-React, an adaptive iterative retrieval-generation framework suitable for multi-page heterogeneous document question answering.
Document-Level Event-Argument Data Augmentation for Challenging Role Types
Joseph Gatto (Dartmouth College), Sarah M. Preum (Dartmouth College)
Data SynthesisTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes two document-level event argument extraction (DocEAE) data generation methods based on large language models (LLMs) - Mad Lib Generation (MLG) and Struct2Text (S2T) - for data augmentation in low-resource cross-domain scenarios;
Document-Level Text Generation with Minimum Bayes Risk Decoding using Optimal Transport
Yuu Jinnai (CyberAgent)
GenerationText
🎯 What it does: Propose an MBR decoding method (MBR-OT) that utilizes optimal transport (Wasserstein distance) as a document-level utility function, and verify its effectiveness in document-level text generation tasks.
Does Context Matter? ContextualJudgeBench for Evaluating LLM-based Judges in Contextual Settings
Austin Xu (Salesforce AI Research), Shafiq Joty (Salesforce AI Research)
Adversarial AttackTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes ContextualJudgeBench to evaluate context-based LLM discriminators, constructing 2,000 contrast samples covering four levels: rejection, factualness, completeness, and conciseness.
Does the Emotional Understanding of LVLMs Vary Under High-Stress Environments and Across Different Demographic Attributes?
Jaewook Lee (Konkuk University), Harksoo Kim (Konkuk University)
Data SynthesisPrompt EngineeringVision Language ModelImageBenchmark
🎯 What it does: Systematically evaluate the emotional understanding capabilities of large vision-language models (LVLMs) under high-pressure environments and across diverse combinations of race, gender, and age.
Does Time Have Its Place? Temporal Heads: Where Language Models Recall Time-specific Information
Yein Park (Korea University), Jaewoo Kang (Korea University)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: This paper discovers through circuit analysis that large language models contain specialized attention heads (Temporal Heads) for processing time-related knowledge, and proves that they are key subcomponents for temporal knowledge recall.
Dolphin: Moving Towards Closed-loop Auto-research through Thinking, Practice, and Feedback
Jiakang Yuan (Fudan University), Bowen Zhou (Shanghai Artificial Intelligence Laboratory)
AI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringImageTextMultimodalityPoint CloudRetrieval-Augmented Generation
🎯 What it does: Built a closed-loop self-driven research framework called DOLPHIN, covering the complete research cycle from paper retrieval, idea generation, experimental validation to result feedback;
DoMIX: An Efficient Framework for Exploiting Domain Knowledge in Fine-Tuning
Dohoon Kim (Seoul National University), Taesup Moon (Seoul National University)
Domain AdaptationTransformerSupervised Fine-TuningText
🎯 What it does: This paper proposes the DoMIX framework, which achieves efficient domain adaptive pre-training and knowledge utilization through LoRA modules.
Don’t Erase, Inform! Detecting and Contextualizing Harmful Language in Cultural Heritage Collections
Orfeas Menis Mastromichalakis (National Technical University of Athens), Giorgos Stamou (National Technical University of Athens)
ClassificationRetrievalTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Built a multilingual, community-created list of controversial terms and developed an AI-driven tool to detect harmful language in cultural heritage metadata, providing historical context and alternative terms, and integrated it into the Europeana and MINT platforms as well as a public web application.
Don’t Get Lost in the Trees: Streamlining LLM Reasoning by Overcoming Tree Search Exploration Pitfalls
Ante Wang (Xiamen University), Dong Yu (Tencent AI Lab)
Computational EfficiencyLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Propose the FETCH framework, which alleviates over-exploration and under-exploration issues in LLM tree search by clustering semantically similar states, and reduces verifier variance through TD(λ) and multi-model integration, thereby improving inference accuracy and computational efficiency.
Don’t Half-listen: Capturing Key-part Information in Continual Instruction Tuning
Yongquan He (Meituan), Xunliang Cai (Meituan)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a continuous instruction fine-tuning method based on Key Part Information Gain (KPIG) to address catastrophic forgetting and partial forgetting phenomena that occur during the continuous instruction fine-tuning of large language models;
Don’t Reinvent the Wheel: Efficient Instruction-Following Text Embedding based on Guided Space Transformation
Yingchaojie Feng (Zhejiang University), Wei Chen (Zhejiang University)
Computational EfficiencyRepresentation LearningLarge Language ModelAuto EncoderContrastive LearningText
🎯 What it does: Propose an efficient instruction-following text embedding framework GSTransform, leveraging guided space transformation to achieve dynamic embeddings without re-encoding the entire corpus.
Donate or Create? Comparing Data Collection Strategies for Emotion-labeled Multimodal Social Media Posts
Christopher Bagdon (University of Bamberg), Roman Klinger (University of Bamberg)
ClassificationData SynthesisData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: This paper compares three methods for collecting emotion-annotated multimodal social media data: asking participants to create posts that align with specific emotion labels (CREATION), donating real posts (DONATION), and submitting recent posts along with self-assessed emotions (RECENT).
DRAE: Dynamic Retrieval-Augmented Expert Networks for Lifelong Learning and Task Adaptation in Robotics
Yayu Long (Chongqing Institute of Green and Intelligent Technology, Chinese Academy of Sciences), Mingsheng Shang (Chongqing Institute of Green and Intelligent Technology, Chinese Academy of Sciences)
Robotic IntelligenceMeta LearningReinforcement LearningMixture of ExpertsImageVideoPoint CloudSequentialRetrieval-Augmented Generation
🎯 What it does: Proposed Dynamic Retrieval-Augmented Expert Network (DRAE) to enable lifelong learning and task adaptation in robots.
DRAG: Distilling RAG for SLMs from LLMs to Transfer Knowledge and Mitigate Hallucination via Evidence and Graph-based Distillation
Jennifer Chen (Mohamed bin Zayed University of AI), Zhiqiang Shen (Mohamed bin Zayed University of AI)
RetrievalKnowledge DistillationTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Generate evidence and knowledge graphs using a large language model teacher, then compress and distill them to small language models, enabling them to have retrieval-enhanced generation capabilities.
DRAMA: Diverse Augmentation from Large Language Models to Smaller Dense Retrievers
Xueguang Ma (University of Waterloo), Xilun Chen (FAIR at Meta)
RetrievalRepresentation LearningData-Centric LearningTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Propose the DRAMA framework: Utilize pruned decoder-only LLMs as the backbone of the retriever, and construct rich training data through multiple LLM-based generation and re-ranking methods. A single-stage contrastive learning approach trains a general-purpose retrieval model with less than 1B parameters.
DREsS: Dataset for Rubric-based Essay Scoring on EFL Writing
Haneul Yoo (Korea Advanced Institute of Science and Technology), Alice Oh (Korea Advanced Institute of Science and Technology)
ClassificationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Released a large-scale, standardized English writing automatic scoring dataset DREsS based on scoring rubrics, and provided a synthetic data generation method called CASE
Drift: Enhancing LLM Faithfulness in Rationale Generation via Dual-Reward Probabilistic Inference
Jiazheng Li (King's College London), Yulan He (King's College London)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposed a dual-reward probabilistic reasoning framework called Drift, which uses task rewards and prospective reasoning rewards to guide LLMs in generating more accurate and faithful answers and explanations during the reasoning phase.
DRPruning: Efficient Large Language Model Pruning through Distributionally Robust Optimization
Hexuan Deng (Harbin Institute of Technology), Zhaopeng Tu (Harbin Institute of Technology)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: During structured pruning and continued pre-training stages, distributed robust optimization (DRO) dynamically adjusts the training data distribution to achieve efficient pruning and performance recovery for large language models;
DS^2-ABSA: Dual-Stream Data Synthesis with Label Refinement for Few-Shot Aspect-Based Sentiment Analysis
Hongling Xu (Peng Cheng Laboratory), Ruifeng Xu (Peng Cheng Laboratory)
ClassificationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Propose the DS-ABSA dual-stream data synthesis framework, which combines keypoint-driven and instance-driven synthesis and incorporates a label refinement module to generate diverse, high-quality end-to-end ABSA data with a very limited number of training samples.
DTCRS: Dynamic Tree Construction for Recursive Summarization
Guanran Luo (Xiamen University), Qingqiang Wu (Xiamen University)
ClassificationGenerationTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Proposes a recursive summarization method called DTCRS that dynamically generates hierarchical summary trees based on document structure and query semantics.
DualGuard: A Parameter Space Transformation Approach for Bidirectional Defense in Split-Based LLM Fine-Tuning
Zihan Liu (Zhejiang University), Sai Wu (Zhejiang Key Laboratory of Big Data Intelligent Computing)
Federated LearningSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Propose DualGuard, a bidirectional defense mechanism for split LLM fine-tuning, which utilizes parameter space transformation to prevent data reconstruction attacks;
Dually Self-Improved Counterfactual Data Augmentation Using Large Language Model
Luhao Zhang (Beijing Institute of Technology), Liqiang Nie (Harbin Institute of Technology)
ClassificationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Studied a counterfactual data augmentation method (DICT) using large language models for bidirectional self-improvement to enhance model robustness in natural language inference tasks.
DualRAG: A Dual-Process Approach to Integrate Reasoning and Retrieval for Multi-Hop Question Answering
Rong Cheng (Tianjin University), Jianye Hao (Tianjin University)
RetrievalTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: DualRAG proposes a dual-process iterative retrieval-generation framework, achieving active retrieval and knowledge integration for multi-hop QA through Reasoning-augmented Querying (RaQ) and Progressive Knowledge Aggregation (pKA).
Dynamic and Generalizable Process Reward Modeling
Zhangyue Yin (Fudan University), Xuanjing Huang (Fudan University)
Reinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Proposed a dynamic general process reward model DG-PRM, which constructs multi-dimensional fine-grained rewards using a reward tree and dynamically allocates rewards at each step, combined with Pareto dominance estimation to enhance reward quality.
Dynamic Chunking and Selection for Reading Comprehension of Ultra-Long Context in Large Language Models
Boheng Sheng (East China Normal University), Guoxiu He (East China Normal University)
Knowledge DistillationTransformerLarge Language ModelContrastive LearningTextBenchmark
🎯 What it does: To address the understanding bottleneck of LLMs in reading and reasoning over ultra-long texts, a dynamic chunking and selection mechanism is proposed, which dynamically splits long texts based on semantic similarity and selects the most relevant chunks for LLM processing using a question-aware classifier.
Dynamic Evaluation with Cognitive Reasoning for Multi-turn Safety of Large Language Models
Lanxue Zhang (Chinese Academy of Sciences), Yangxi Li (National Computer Network Emergency Response Technical Team)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed a dynamic multi-round safety evaluation framework called CogSafe based on cognitive theory, which can automatically generate multi-round dialogues and assess the safety of LLMs, significantly reducing the risk of evaluation leakage.
Dynamic Head Selection for Neural Lexicalized Constituency Parsing
Yang Hou (Soochow University), Zhenghua Li (Soochow University)
Representation LearningTransformerLarge Language ModelText
🎯 What it does: Re-derive and implement an implicit lexicalization syntactic parsing framework that can dynamically infer word heads during training, avoiding traditional fixed head lookup rules;
Dynamic Label Name Refinement for Few-Shot Dialogue Intent Classification
Gyutae Park (Chung-Ang University), Hwanhee Lee (Chung-Ang University)
ClassificationTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed a dynamic label name refinement method, combining retrieval-enhanced context learning to address confusion caused by semantic overlap of labels in dialog intent classification.
Dynamic Order Template Prediction for Generative Aspect-Based Sentiment Analysis
Yonghyun Jun (Chung-Ang University), Hwanhee Lee (Chung-Ang University)
GenerationTransformerLarge Language ModelText
🎯 What it does: Designed and implemented the Dynamic Order Template (DOT) method, which uses a two-phase process to separately predict the number of sentiment quadruples to be generated in a sentence and the content of each quadruple, ultimately outputting aspect-based sentiment quadruples.
Dynamic Parallel Tree Search for Efficient LLM Reasoning
Yifu Ding (Beihang University), Dacheng Tao (Nanyang Technological University)
Computational EfficiencyAI Code AssistantLarge Language ModelTextBenchmark
🎯 What it does: Propose the Dynamic Parallel Tree Search (DPTS) framework, combining parallel tree search with dynamic search/transition mechanisms to significantly enhance the computational efficiency of LLM logical reasoning.
Dynamic Scaling of Unit Tests for Code Reward Modeling
Zeyao Ma (Renmin University of China), Jie Tang (Tsinghua University)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: The study employs dynamic expansion of unit test quantity for code reward modeling to enhance the selection of optimal answers by code generation models.
EAC-MoE: Expert-Selection Aware Compressor for Mixture-of-Experts Large Language Models
Yuanteng Chen (Chinese Academy of Sciences), Jian Cheng (Chinese Academy of Sciences)
CompressionComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Researched and implemented EAC-MoE, a compression method for Mixture-of-Experts (MoE) large language models, incorporating low-bit quantization and dynamic pruning.
EAGLE: Expert-Guided Self-Enhancement for Preference Alignment in Pathology Large Vision-Language Model
Meidan Ding (Shenzhen University), Linlin Shen (Shenzhen University)
Supervised Fine-TuningVision Language ModelMultimodalityBiomedical Data
🎯 What it does: This paper proposes the EAGLE framework, achieving preference alignment for pathological vision-language models through expert-guided self-enhancement to reduce multimodal hallucinations and biases.
ECERC: Evidence-Cause Attention Network for Multi-Modal Emotion Recognition in Conversation
Tao Zhang (Northeastern University), Zhenhua Tan (Northeastern University)
RecognitionRecurrent Neural NetworkTransformerMultimodality
🎯 What it does: Proposed a model called ECERC that interacts emotional evidence with emotional causality in dialogue context to achieve multimodal dialogue emotion recognition.
ECLM: Entity Level Language Model for Spoken Language Understanding with Chain of Intent
Shangjian Yin (South China Agricultural University), Yuhong Xu (Tsinghua University)
RecognitionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose the ECLM framework, transforming traditional token-level slot-filling into entity-level identification, and introduce Chain of Intent to progressively decompose multi-intent, addressing alignment and length issues.
EcomScriptBench: A Multi-task Benchmark for E-commerce Script Planning via Step-wise Intention-Driven Product Association
Weiqi Wang (HKUST), Yangqiu Song (HKUST)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes and implements the e-commerce script planning task, and constructs the first large-scale multi-task benchmark, ECOMSCRIPTBENCH, covering script generation, product association, and overall feasibility verification.
EdiText: Controllable Coarse-to-Fine Text Editing with Diffusion Language Models
Che Hyun Lee (Seoul National University), Sungroh Yoon (Seoul National University)
GenerationDiffusion modelText
🎯 What it does: Propose EdiText, a text editing framework based on embedded diffusion models, capable of achieving both coarse-grained and fine-grained editing;
EditInspector: A Benchmark for Evaluation of Text-Guided Image Edits
Ron Yosef (Hebrew University of Jerusalem), Moran Yanuka (Tel Aviv University)
Supervised Fine-TuningVision Language ModelImageTextBenchmark
🎯 What it does: Proposed the EditInspector benchmark for systematic evaluation of text-driven image editing across multiple dimensions of quality, including accuracy, seamlessness, visual quality, scene integration, common-sense compliance, and difference description.
EducationQ: Evaluating LLMs’ Teaching Capabilities Through Multi-Agent Dialogue Framework
Yao Shi (South China University of Technology), Yong Xu (South China University of Technology)
Large Language ModelAgentic AITextBenchmark
🎯 What it does: This paper proposes and implements the EducationQ multi-agent dialogue framework, which quantitatively and qualitatively evaluates the interactive effectiveness of large language models in educational contexts through a three-round process: prediction-dialogue-post-test.