EMNLP 2025 Papers — Page 5
Conference on Empirical Methods in Natural Language Processing · 1809 papers
Definition Generation for Word Meaning Modeling: Monolingual, Multilingual, and Cross-Lingual Perspectives
Francesco Periti (KU Leuven), Nina Tahmasebi (University of Gothenburg)
GenerationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Conduct research on the word sense definition generation task using a dataset based on Wiktionary/Dbnary across 22 languages.
DEL-ToM: Inference-Time Scaling for Theory-of-Mind Reasoning via Dynamic Epistemic Logic
Yuheng Wu (Stanford University), Zhaozhuo Xu (Stevens Institute of Technology)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposed the DEL-ToM framework, which structurally and verifiably reasons about belief updates in Theory-of-Mind (ToM) tasks during inference by scaling the reasoning time of large language models through Dynamic Epistemic Logic (DEL).
DELOC: Document Element Localizer
Hammad Ayyubi (Columbia University), Vlad I Morariu (Adobe Research)
Object DetectionData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelMultimodality
🎯 What it does: Propose the DELOC system, which uses a multimodal large language model (MLLM) to perform spatial localization for PDF editing requests, and achieves precise positioning of PDF elements through synthetic data.
Demystifying Domain-adaptive Post-training for Financial LLMs
Zixuan Ke (Salesforce AI Research), Shafiq Joty (Salesforce AI Research)
Domain AdaptationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmarkFinance Related
🎯 What it does: Proposed the FINDAP framework for systematic post-training of LLMs in the financial domain, comprising FinCap (core capabilities), FinRec (joint CPT+IT with preference alignment), FinTrain (fine-grained dataset), and FinEval (comprehensive evaluation).
Demystifying optimized prompts in language models
Rimon Melamed (George Washington University), H Howie Huang (George Washington University)
OptimizationExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Investigated the characteristics of 'optimized prompts' generated through gradient optimization in modern language models, focusing on their composition, lexical scarcity, and internal representations, and verified their recognizability using sparse detectors.
Demystifying Synthetic Data in LLM Pre-training: A Systematic Study of Scaling Laws, Benefits, and Pitfalls
Feiyang Kang (FAIR at Meta), Carole-Jean Wu (FAIR at Meta)
Data SynthesisTransformerLarge Language ModelText
🎯 What it does: This paper systematically evaluates the effectiveness of using synthetic data during the foundational pre-training phase by training over 1000 large language models (up to 3B parameters, 200B tokens).
Deriving Strategic Market Insights with Large Language Models: A Benchmark for Forward Counterfactual Generation
Keane Ong (National University of Singapore), Gianmarco Mengaldo (National University of Singapore)
TransformerLarge Language ModelTextBenchmarkFinance Related
🎯 What it does: Proposed the FIN-FORCE benchmark for forward counterfactual generation in the financial domain, supporting automated prediction of market opportunities and risks.
Detecting Corpus-Level Knowledge Inconsistencies in Wikipedia with Large Language Models
Sina Semnani, Monica Lam
Anomaly DetectionTransformerLarge Language ModelAgentic AITextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes the Corpus-Level Inconsistency Detection (CLID) task and constructs the WIKICOLLIDE dataset and CLAIRE system to identify internal contradictions in large corpora.
Detecting Knowledge Boundary of Vision Large Language Models by Sampling-Based Inference
Zhuo Chen (ShanghaiTech University), Kewei Tu (Alibaba Group)
RetrievalLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose two variants of Fine-Tuning methods, leveraging VLLM self-sampling and text LLM scoring to automatically construct a knowledge boundary dataset, training a model to determine whether VQA questions require external retrieval;
Detecting Legal Citations in United Kingdom Court Judgments
Holli Sargeant (University of Cambridge), Måns Magnusson (Uppsala University)
RecognitionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Construct a high-quality UK court judgment citation annotated corpus and systematically evaluate the performance of regular expressions, pre-trained encoders, and large language models in legal citation detection.
Detecting LLM Hallucination Through Layer-wise Information Deficiency: Analysis of Ambiguous Prompts and Unanswerable Questions
Hazel Kim (University of Oxford), Yarin Gal (University of Oxford)
Anomaly DetectionExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose an unsupervised layer-wise usable information (LI) method that detects hallucinations and unanswerable questions in large language models (LLMs) by analyzing information flow through each layer during inference.
Detoxifying Large Language Models via the Diversity of Toxic Samples
Ying Zhao (Jilin University), Yi Chang (Jilin University)
Safty and PrivacyComputational EfficiencyReinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningReinforcement LearningContrastive LearningTextBenchmark
🎯 What it does: This paper proposes a diversified detoxification framework called DivDetox, aiming to enhance the detoxification effectiveness of large language models by leveraging the diversity and specificity of toxic samples.
Diagnosing Memorization in Chain-of-Thought Reasoning, One Token at a Time
Huihan Li (University of Southern California), Xiang Ren (University of Southern California)
Explainability and InterpretabilityTransformerLarge Language ModelTextSequentialRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed the STIM (Source-aware Token-level Identification of Memorization) framework for fine-grained evaluation of memorization levels at the token level in chain-of-thought (CoT) reasoning, assessing each generated token's memorization degree from three sources (local, mid-range, long-range), and using these memorization scores to predict erroneous tokens.
Diagram-Driven Course Questions Generation
Xinyu Zhang (Xi'an Jiaotong University), Jun Liu (Xi'an Jiaotong University)
GenerationTransformerLarge Language ModelVision Language ModelMultimodalityBenchmark
🎯 What it does: Propose the Diagram-Driven Course Questions Generation (DDCQG) task, construct the DiagramQG dataset, and design a hierarchical knowledge integration framework (HKI-DDCQG) to generate course-related diagram questions.
Dial-In LLM: Human-Aligned LLM-in-the-loop Intent Clustering for Customer Service Dialogues
Mengze Hong (Hong Kong Polytechnic University), Di Jiang (Hong Kong Polytechnic University)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextBenchmark
🎯 What it does: This paper proposes an iterative intent clustering framework based on LLM, combining refined semantic consistency evaluation, clustering naming, and post-merging to achieve high-quality intent clustering in customer service dialogues.
Dialect-SQL: An Adaptive Framework for Bridging the Dialect Gap in Text-to-SQL
Jie Shi (Fudan University), Wei Wang (Fudan University)
GenerationAI Code AssistantTransformerLarge Language ModelTextTabularRetrieval-Augmented Generation
🎯 What it does: Proposes an adaptive framework called Dialect-SQL that uses ORM code as an intermediate language to bridge text-to-SQL tasks across different SQL dialects.
DICE: Structured Reasoning in LLMs through SLM-Guided Chain-of-Thought Correction
Yiqi Li (Shanghai Jiao Tong University), Yu Wang (Shanghai Jiao Tong University)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: Propose a lightweight framework called DICE, which first lets a large model (LLM) generate natural language answers, then uses a small model (SLM) to analyze and correct the answers through chain-of-thought (CoT), ultimately outputting answers in a structured format.
DiCoRe: Enhancing Zero-shot Event Detection via Divergent-Convergent LLM Reasoning
Tanmay Parekh (University of California), Nanyun Peng (University of California)
RecognitionLarge Language ModelTextChain-of-Thought
🎯 What it does: Developed the DICORE framework for zero-shot event detection without training data, achieving zero-shot event detection through phased reasoning (Dreamer-Open Discovery, Grounder-Alignment, Judge-Validation);
DIDS: Domain Impact-aware Data Sampling for Large Language Model Training
Weijie Shi (Hong Kong University of Science and Technology), Xiaofang Zhou (Hong Kong University of Science and Technology)
Data-Centric LearningLarge Language ModelText
🎯 What it does: Propose a dynamic data sampling framework called DIDS based on domain influence to optimize data allocation for large language models in multi-domain training.
Diffusion vs. Autoregressive Language Models: A Text Embedding Perspective
Siyue Zhang (Nanyang Technological University), Chen Zhao (NYU Shanghai)
RetrievalTransformerDiffusion modelText
🎯 What it does: This paper proposes a new text embedding model called DIFFEMBED, which leverages diffusion language models to improve text embedding tasks, particularly excelling in long document retrieval and reasoning-intensive retrieval.
DiffusionAttacker: Diffusion-Driven Prompt Manipulation for LLM Jailbreak
Hao Wang (Beihang University), Lei Sha (Beihang University)
Adversarial AttackLarge Language ModelPrompt EngineeringDiffusion modelText
🎯 What it does: Propose an end-to-end text rewriting framework called DiffusionAttacker based on sequence-to-sequence diffusion models, designed to generate 'escape' prompts capable of bypassing LLM security mechanisms.
DiMo-GUI: Advancing Test-time Scaling in GUI Grounding via Modality-Aware Visual Reasoning
Hang Wu, Yiwei Wang
Object DetectionSegmentationImageMultimodality
🎯 What it does: Proposed DiMo-GUI, a zero-training, plug-and-play GUI localization framework for precisely locating text and icon elements in high-resolution interfaces.
DiNaM: Disinformation Narrative Mining with Large Language Models
Witold Sosnowski (Polish-Japanese Academy of Information Technology), Adam Wierzbicki (Polish-Japanese Academy of Information Technology)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposed the DiNaM algorithm, which detects, verifies, and refines misinformation in fact-checking articles using LLMs, then generates misinformation narratives via an embedding + clustering approach.
DINT Transformer
Yueyang Cang, Li Shi
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposed the DINT Transformer model, combining differential attention and integral mechanisms to reduce attention noise and enhance global information capture capabilities.
DiplomacyAgent: Do LLMs Balance Interests and Ethical Principles in International Events?
Jianxiang Peng (Tianjin University), Deyi Xiong (Macau University of Science and Technology)
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelAgentic AIPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Constructed a multi-agent framework named DiplomacyAgent to evaluate the decision-making tendencies of large language models (LLMs) when facing conflicts between interests and ethics, revealing potential security risks by simulating diplomatic decision-making processes.
Dipper: Diversity in Prompts for Producing Large Language Model Ensembles in Reasoning Tasks
Wenyang Hu (National University of Singapore), Bryan Kian Hsiang Low (National University of Singapore)
OptimizationComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Construct a training-free, parallelizable 'Prompt Diversification Integration' framework (DIPPER) that inputs diverse prompts in parallel during inference for a single large language model (LLM), thereby enhancing performance on complex reasoning tasks.
Direct Judgement Preference Optimization
PeiFeng Wang, Shafiq Joty (Salesforce AI Research)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: Trained large-scale base evaluation models (8B, 12B, 70B) through direct preference optimization (DPO) to learn three types of evaluation tasks: chain-of-thought (CoT) criticism, standard judgment, and response reasoning.
Direct Value Optimization: Improving Chain-of-Thought Reasoning in LLMs with Refined Values
Hongbo Zhang (Zhejiang University), Yue Zhang (Southern University of Science and Technology)
OptimizationTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: Propose Direct Value Optimization (DVO), which directly optimizes the reasoning path of large language models by using step-level value signals during the Chain-of-Thought (CoT) process;
Disambiguation in Conversational Question Answering in the Era of LLMs and Agents: A Survey
Mehrab Tanjim, Chanyoung Park (Adobe Inc.)
TransformerLarge Language ModelSupervised Fine-TuningAgentic AIPrompt EngineeringTextReview/Survey PaperRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper reviews the current state of large language models in detecting and disambiguating in conversational question answering, and proposes a unified classification of three categories of ambiguity and three mainstream disambiguation strategies;
DischargeSim: A Simulation Benchmark for Educational Doctor–Patient Communication at Discharge
Zonghai Yao (VA Bedford Health Care), Hong Yu (VA Bedford Health Care)
TransformerLarge Language ModelAgentic AITextBiomedical DataElectronic Health RecordsBenchmark
🎯 What it does: Propose the DischargeSim benchmark, simulating multi-round dialogues between doctors and patients after discharge, personalized summaries, and comprehension tests to evaluate the performance of large language models in discharge education.
DiscoSG: Towards Discourse-Level Text Scene Graph Parsing through Iterative Graph Refinement
Shaoqing Lin (Wuhan University), Zhuang Li (Monash University)
GenerationData SynthesisTransformerLarge Language ModelText
🎯 What it does: Designed and implemented the text scene graph parsing task for multi-sentence visual descriptions, DiscoSG, and proposed a lightweight iterative refinement framework, DiscoSG-Refiner.
Discourse-Driven Code-Switching: Analyzing the Role of Content and Communicative Function in Spanish-English Bilingual Speech
Debasmita Bhattacharya (Columbia University), Julia Hirschberg (Columbia University)
ClassificationData-Centric LearningText
🎯 What it does: On a Spanish-English bilingual corpus, the authors systematically analyzed how discourse content (named entities) and discourse function (dialogue acts) influence the occurrence and structure of code-switching, and built two predictive models (logistic regression and hidden Markov models) based on statistical findings to predict whether code-switching would appear in subsequent sentences.
Discrepancy Detection at the Data Level: Toward Consistent Multilingual Question Answering
Lorena Calvo-Bartolomé (Universidad Carlos III de Madrid), Jordan Lee Boyd-Graber
Anomaly DetectionData-Centric LearningTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Designed and implemented a four-stage LLM-assisted pipeline named MIND for detecting factual and cultural differences in multilingual knowledge bases, first aligning documents via a multilingual topic model, then generating questions, retrieving evidence, generating answers, and determining differences.
Discriminating Form and Meaning in Multilingual Models with Minimal-Pair ABX Tasks
Maureen de Seyssel (Apple), Natalie Schluter
Representation LearningTransformerContrastive LearningText
🎯 What it does: Proposes a training-agnostic ABX-style contrastive task to measure the internal representation differences in multilingual models for language identity (form) and semantic content (meaning), evaluated in a zero-shot, unsupervised manner;
Discursive Circuits: How Do Language Models Understand Discourse Relations?
Yisong Miao (National University of Singapore), Min-Yen Kan (National University of Singapore)
Explainability and InterpretabilityComputational EfficiencyTransformerSupervised Fine-TuningText
🎯 What it does: Conduct an interpretability study on the mechanisms in Transformer language models for handling discourse relations, introducing the concept of 'Discursive Circuits.'
Disentangled Information Bottleneck for Adversarial Text Defense
Yidan Xu (China University of Petroleum (East China)), Weifeng Liu (University of Technology Sydney)
ClassificationRepresentation LearningAdversarial AttackAuto EncoderGenerative Adversarial NetworkText
🎯 What it does: Proposes the Disentangled Information Bottleneck (DisIB) method, which uses a two-branch structure to decouple robust and non-robust features, thereby enhancing the robustness of text models against adversarial attacks.
Disentangling Subjectivity and Uncertainty for Hate Speech Annotation and Modeling using Gaze
Özge Alacam (Bielefeld University), Barbara Plank (LMU Munich)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelTextMultimodality
🎯 What it does: Study how to separate the subjectivity and uncertainty in hate speech annotations through confidence scores and eye-tracking data, and evaluate their impact on model behavior.
DisLoRA: Task-specific Low-Rank Adaptation via Orthogonal Basis from Singular Value Decomposition
She Yifei (Beijing University of Posts and Telecommunications), Yulong Wang (Beijing University of Posts and Telecommunications)
OptimizationComputational EfficiencySupervised Fine-TuningTextBenchmark
🎯 What it does: Proposes a low-rank adaptation framework called DisLoRA based on SVD, decomposing pre-trained weights into trunk and task-specific subspaces to achieve efficient and interpretable parameterized fine-tuning.
Distribution Prompting: Understanding the Expressivity of Language Models Through the Next-Token Distributions They Can Produce
Haojin Wang (University of Waterloo), Freda Shi (University of Waterloo)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Systematically evaluate the expressive capability of large language models (LLMs) in approximating arbitrary given next-word probability distributions, without altering model parameters, by using a soft-hard hybrid prompt tuning method. The study first constructs target distributions with different entropy values, then attempts to make the model output as close as possible to the target distribution, and measures the approximation difficulty using KL divergence.
Diverse, not Short: A Length-Controlled Data Selection Strategy for Improving Response Diversity of Language Models
Vijeta Deshpande (University of Massachusetts Lowell), Anna Rumshisky (University of Massachusetts Lowell)
GenerationData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposed and validated a length-controlled data selection strategy (Diverse-NS), enhancing the output diversity of aligned language models through self-supervised self-learning and preference optimization.
DivScore: Zero-Shot Detection of LLM-Generated Text in Specialized Domains
Zhihui Chen (National University of Singapore), Mengling Feng (National University of Singapore)
Domain AdaptationAnomaly DetectionKnowledge DistillationTransformerLarge Language ModelTextBiomedical Data
🎯 What it does: Proposes a zero-shot detection framework called DivScore, leveraging normalized entropy and cross-entropy scores, and constructing domain-adapted language models through unsupervised domain knowledge distillation, specifically for detecting LLM-generated text in medical and legal domains.
DIWALI - Diversity and Inclusivity aWare cuLture specific Items for India: Dataset and Assessment of LLMs for Cultural Text Adaptation in Indian Context
Pramit Sahoo (Indian Institute of Technology Hyderabad), Maunendra Sankar Desarkar (Indian Institute of Technology Hyderabad)
Domain AdaptationData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed the DIWALI project, which is specific to multi-regional Indian culture, and used it to evaluate the performance of LLMs in cultural text adaptation tasks.
DMDTEval: An Evaluation and Analysis of LLMs on Disambiguation in Multi-domain Translation
Zhibo Man (Beijing Jiaotong University), Jinan Xu (Beijing Jiaotong University)
Domain AdaptationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Constructed a test set for ambiguous words in multi-domain translation and designed multiple prompt-based strategies to systematically evaluate the ambiguity resolution capabilities of large language models (LLMs) in multi-domain translation; subsequently evaluated five open-source LLMs across four language pairs and thirteen domains, investigating the impact of prompt strategies and domain knowledge on translation performance and ambiguity resolution.
DnDScore: Decontextualization and Decomposition for Factuality Verification in Long-Form Text Generation
Miriam Wanner (Johns Hopkins University), Mark Dredze (Johns Hopkins University)
GenerationLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose a verification framework that integrates decontextualization and claim decomposition in long-text generation, design the DNDSCORE verification method, and implement a joint decomposition-decontextualization prompt (DnD).
Do All Autoregressive Transformers Remember Facts the Same Way? A Cross-Architecture Analysis of Recall Mechanisms
Minyeong Choe (Chosun University), Hyunil Kim (Kongju National University)
Explainability and InterpretabilityTransformerText
🎯 What it does: Conduct causal analysis on hierarchical and module-level fact-related memory mechanisms of various autoregressive Transformers (GPT, LLaMA, Qwen, DeepSeek).
Do Large Language Models excel in Complex Logical Reasoning with Formal Language?
Jin Jiang (Peking University), Liangcai Gao (Peking University)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkChain-of-Thought
🎯 What it does: This paper constructs a multi-dimensional evaluation framework to systematically evaluate the performance of large language models across various logical reasoning tasks and formal language trajectories.
Do Large Language Models Truly Grasp Addition? A Rule-Focused Diagnostic Using Two-Integer Arithmetic
Yang Yan (Zhejiang University), Zhenzhong Lan (Westlake University)
Explainability and InterpretabilityLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: Evaluate the rule understanding of 12 mainstream LLMs for two-integer addition (0~2^64-1), propose three diagnostic metrics: digit length consistency, reversible symbol mapping, and commutativity, and assess their performance in zero-shot, symbolic input, prompting, and fine-tuning experiments.
Do Large Language Models Understand Word Senses?
Domenico Meconi (Babelscape), Roberto Navigli (Babelscape)
ClassificationGenerationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Evaluate the performance of large language models in word sense disambiguation (WSD) and generative word sense explanation tasks
Do LLMs Adhere to Label Definitions? Examining Their Receptivity to External Label Definitions
Seyedali Mohammadi (UMBC), Manas Gaur (UMBC)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Systematically investigated the dependence of LLMs on label definitions, explored two scenarios of definition conflicts and integration, and evaluated model performance across four domains (natural language reasoning, mental health, hate speech detection, and fact-checking).
Do LLMs Behave as Claimed? Investigating How LLMs Follow Their Own Claims using Counterfactual Questions
Haochen Shi (Harbin Institute of Technology), Zhenzhou Ji (Harbin Institute of Technology)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Proposes the Behave as Claimed (BaC) framework, which evaluates whether large language models (LLMs) act as claimed by dynamically generating and validating 'what-if' dialogues;
Do LLMs Encode Frame Semantics? Evidence from Frame Identification
Jayanth Krishna Chundru (University of Cincinnati), Tianyu Jiang (University of Cincinnati)
RecognitionExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Investigate whether large language models (LLMs) have internalized frame semantic knowledge, and perform frame identification tasks on FrameNet through prompting and fine-tuning.
Do RAG Systems Really Suffer From Positional Bias?
Florin Cuconasu (Sapienza University of Rome), Fabrizio Silvestri (Technology Innovation Institute)
GenerationRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Investigate the impact of position bias (differences in LLM's attention to the location of information in context) on answer accuracy in retrieval-augmented generation (RAG) systems, quantify the high-interference paragraphs generated by the retrieval pipeline, and evaluate the effects of different paragraph arrangement strategies on actual question-answering performance.
Do Slides Help? Multi-modal Context for Automatic Transcription of Conference Talks
Supriti Sinhamahapatra (Karlsruhe Institute of Technology), Jan Niehues (Karlsruhe Institute of Technology)
RecognitionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkAudio
🎯 What it does: This paper investigates how to leverage the visual information from academic lecture slides to enhance Automatic Speech Recognition (ASR) performance, constructing a multimodal evaluation benchmark incorporating slide text/images, and implementing context integration methods based on slide content on existing large pre-trained models (e.g., SALMONN, Phi-4-multimodal, Whisper). Additionally, an unsupervised data augmentation scheme is proposed to automatically generate slide images for fine-tuning training.
Do You Know About My Nation? Investigating Multilingual Language Models’ Cultural Literacy Through Factual Knowledge
Eshaan Tanwar (Indian Institute of Technology Delhi), Tanmoy Chakraborty (Indian Institute of Technology Delhi)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Constructed a cross-lingual, cross-cultural literacy evaluation dataset called XNationQA, and evaluated multiple multilingual large language models (LLMs) on it
Doc2Chart: Intent-Driven Zero-Shot Chart Generation from Documents
Akriti Jain (Adobe Research), Apoorv Saxena (Adobe Research)
GenerationExplainability and InterpretabilityTransformerLarge Language ModelImageTextTabularFinance RelatedRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes a task and method for automatically generating statistical charts from long documents based on user intent
DocAgent: An Agentic Framework for Multi-Modal Long-Context Document Understanding
Li Sun (Boston University), Chenyu You (Boston University)
RetrievalTransformerLarge Language ModelAgentic AIMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Built a multi-modal long document understanding framework called DocAgent, which uses hierarchical outlines to guide LLM agents for efficient retrieval, followed by a review agent to verify answers and a memory module to enable cross-task knowledge transfer.
DocReRank: Single-Page Hard Negative Query Generation for Training Multi-Modal RAG Rerankers
Navve Wasserman (Weizmann Institute of Science), Michal Irani (IBM Research Israel)
RetrievalTransformerLarge Language ModelSupervised Fine-TuningMultimodalityTabularFinance RelatedRetrieval-Augmented Generation
🎯 What it does: This study proposes a single-page hard negative query generation method to improve the training of multi-modal RAG re-rankers;
Does Acceleration Cause Hidden Instability in Vision Language Models? Uncovering Instance-Level Divergence Through a Large-Scale Empirical Study
Yizheng Sun (University of Manchester), Jingyuan Sun (University of Manchester)
Computational EfficiencyVision Language ModelMultimodalityBiomedical DataBenchmark
🎯 What it does: Systematically evaluated the instance-level instability of multiple post-training acceleration methods (such as token reduction and quantization) in vision-language models, and proposed DR and NDR metrics.
Does Context Matter? A Prosodic Comparison of English and Spanish in Monolingual and Multilingual Discourse Settings
Debasmita Bhattacharya (Columbia University), Julia Hirschberg (Columbia University)
RecognitionTransformerAudio
🎯 What it does: Compare the prosodic feature differences between English and Spanish in monolingual and multilingual contexts, and verify that they can be captured by LID models.
Does Localization Inform Unlearning? A Rigorous Examination of Local Parameter Attribution for Knowledge Unlearning in Language Models
Hwiyeong Lee (Hanyang University), Taeuk Kim (Hanyang University)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: This paper designs a controlled experiment to test whether local parameter localization in large language models truly benefits knowledge deletion, assessing the causal impact of localization.
Does quantization affect models’ performance on long-context tasks?
Anmol Mekala (UMass Amherst), Mohit Iyyer (University of Maryland)
Computational EfficiencyTransformerTextBenchmark
🎯 What it does: Systematic evaluation of the performance of multiple open-source large language models after quantization on long context tasks (≥64K token input or long output), covering 9.7K test samples, 5 quantization methods (FP8, GPTQ int8/int4, AWQ int4, BNB nf4), and 5 models (Llama-3.1 8B/70B, Qwen-2.5 7B/32B/72B).
Don’t Sweat the Small Stuff: Segment-Level Meta-Evaluation Based on Pairwise Difference Correlation
Colten DiIanni (Google), Daniel Deutsch (Google)
TextBenchmark
🎯 What it does: Propose a Pairwise Difference Pearson (PDP) based on pairwise differences for segment-level meta-evaluation in machine translation.
Doubling Your Data in Minutes: Ultra-fast Tabular Data Generation via LLM-Induced Dependency Graphs
Shuo Yang (Technical University of Munich), Gjergji Kasneci (Technical University of Munich)
Data SynthesisTransformerLarge Language ModelFlow-based ModelTabular
🎯 What it does: Automatically extract sparse feature dependency graphs using LLM, then synthesize tabular data on this graph in topological order with KDE or conditional normalizing flows.
Dovetail: A CPU/GPU Heterogeneous Speculative Decoding for LLM inference
Libo Zhang (National University of Defense Technology), Dongsheng Li (National University of Defense Technology)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Implementing the lossless inference acceleration method Dovetail in a CPU/GPU heterogeneous environment, combined with a draft-verification mechanism for LLM inference acceleration.
DPED: Multi-Layer Noise Distillation for Privacy-Preserving Text Embeddings
Shuya Feng (University of Alabama at Birmingham), Yuan Hong (University of Connecticut)
Safty and PrivacyKnowledge DistillationRepresentation LearningText
🎯 What it does: Proposed the DPED framework, which achieves differential privacy text embedding training by utilizing teacher-student distillation and multi-layer noise injection
Draft Model Knows When to Stop: Self-Verification Speculative Decoding for Long-Form Generation
Ziyin Zhang (Shanghai Jiao Tong University), Zhaopeng Tu (Tencent)
GenerationTransformerLarge Language ModelText
🎯 What it does: Proposed a training-agnostic, dynamic-length self-verification sampling strategy (SVIP) that determines when to stop generating drafts and hand them over to the target model for verification by checking the entropy of the draft model.
DrDiff: Dynamic Routing Diffusion with Hierarchical Attention for Breaking the Efficiency-Quality Trade-off
Jusheng Zhang (Sun Yat-sen University), Keze Wang (Alibaba Group)
GenerationComputational EfficiencyTransformerMixture of ExpertsDiffusion modelTextBenchmark
🎯 What it does: Developed a long-text generation framework called DrDiff, enhancing generation efficiency and quality through dynamic expert scheduling, hierarchical sparse attention, and semantic anchor states.
DRES: Fake news detection by dynamic representation and ensemble selection
Faramarz Farhangian (École de Technologie Supérieure), Rafael M. O. Cruz (École de Technologie Supérieure)
ClassificationRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Propose a framework named DRES that can dynamically select the most suitable text representations and corresponding subsets of classifiers based on the instance hardness of test samples during fake news detection, ultimately obtaining prediction results through majority voting.
DrFrattn: Directly Learn Adaptive Policy from Attention for Simultaneous Machine Translation
Libo Zhao (Hong Kong Polytechnic University), Ziqian Zeng (South China University of Technology)
Computational EfficiencyTransformerReinforcement LearningText
🎯 What it does: Proposes a DrFrattn method that directly learns adaptive read/write strategies from the cross-attention mechanism of Transformers;
Drift-Adapter: A Practical Approach to Near Zero-Downtime Embedding Model Upgrades in Vector Databases
Harshil Vejendla (Rutgers University)
RetrievalDomain AdaptationRepresentation LearningTransformerSupervised Fine-TuningImageText
🎯 What it does: Proposed a lightweight embedding space mapping layer (Drift-Adapter), which enables retrieval using a new model without re-encoding the database by transforming the query embeddings of the new model, thereby achieving near-zero downtime upgrades.
DRISHTIKON: A Multimodal Multilingual Benchmark for Testing Language Models’ Understanding on Indian Culture
Arijit Maji (Indian Institute of Technology Patna), Sriparna Saha (Indian Institute of Technology Patna)
Large Language ModelVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposes DRISHTIKON, a multimodal multilingual evaluation benchmark covering 15 Indian languages, spanning 28 states + 8 federal territories, with 64,288 image-text aligned pairs, designed to assess visual-language models' performance in understanding and reasoning about Indian culture.
Drivel-ology: Challenging LLMs with Interpreting Nonsense with Depth
Yang Wang (University of Manchester), Chenghua Lin (University of Manchester)
Explainability and InterpretabilityTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose the concept of 'Drivelology' (nonsensical language containing deep implicit meaning), construct a cross-lingual DRIVELHUB dataset, and design four evaluation tasks: detection, annotation, generation, and multiple-choice selection.
DSCD: Large Language Model Detoxification with Self-Constrained Decoding
Ming Dong (Central China Normal University), Tingting He (Central China Normal University)
Safty and PrivacyTransformerLarge Language ModelText
🎯 What it does: Proposed a LLM detoxification method called DSCD based on self-constrained decoding, which utilizes token-level toxicity layers to locate and dynamically adjust the distribution of the next token.
DSG-MCTS: A Dynamic Strategy-Guided Monte Carlo Tree Search for Diversified Reasoning in Large Language Models
Rui Ha (Beijing University of Posts and Telecommunications), Sen Su (Beihang University)
OptimizationReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes a Dynamic Strategy-Guided Monte Carlo Tree Search (DSG-MCTS) framework to enhance the thinking diversity and accuracy of large language models in multi-step reasoning tasks.
DSMoE: Matrix-Partitioned Experts with Dynamic Routing for Computation-Efficient Dense LLMs
Minxuan Lv (Institute of Information Engineering Chinese Academy of Sciences), Songlin Hu (Institute of Information Engineering Chinese Academy of Sciences)
Computational EfficiencyTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Propose DSMoE, which achieves sparsity by matrix partitioning of pre-trained FFN layers and introducing dynamic routing, allowing each token to adaptively activate different experts based on input complexity;
DSVD: Dynamic Self-Verify Decoding for Faithful Generation in Large Language Models
YiQiu Guo (Fudan University), Yu Wang (Shanghai AI Laboratory)
GenerationTransformerLarge Language ModelText
🎯 What it does: Propose a decoding framework named DSVD (Dynamic Self-Verify Decoding), which realizes real-time self-verification and dynamic rollback during the generation process of large language models, aiming to reduce hallucinations and factual errors.
Dual-Path Counterfactual Integration for Multimodal Aspect-Based Sentiment Classification
Rui Liu (Chinese Academy of Sciences), Yanan Cao (Chinese Academy of Sciences)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningMultimodality
🎯 What it does: Propose a dual-path counterfactual integration framework (DPCI), which explicitly eliminates pseudo correlations in multi-modal sentiment classification by generating and selecting aspect and context counterfactual samples;
Dual-Path Dynamic Fusion with Learnable Query for Multimodal Sentiment Analysis
Miao Zhou (Guangxi University), Xinru Zhang (Guangxi University)
ClassificationTransformerMultimodality
🎯 What it does: Propose the Dual-Path Dynamic Fusion with Learnable Query (DPDF-LQ) framework, which combines global and local paths to simultaneously capture cross-modal global dependencies and fine-grained sentiment information;
DyePack: Provably Flagging Test Set Contamination in LLMs Using Backdoors
Yize Cheng (University of Maryland), Soheil Feizi (University of Maryland)
Anomaly DetectionAdversarial AttackTransformerTextBenchmark
🎯 What it does: Propose the DyePack framework, which embeds multiple backdoor samples into public benchmark test sets to detect whether LLMs used the test set during training without requiring access to internal model information.
Dynamic Collaboration of Multi-Language Models based on Minimal Complete Semantic Units
Chao Hao (Great Bay University), Zitong Yu (Great Bay University)
Computational EfficiencyTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: This paper studies a token-level collaborative inference method for multiple models based on the dynamic selection strategy (DDS) using minimum complete semantic units (MCSU) and distribution distance.
Dynamic Energy-Based Contrastive Learning with Multi-Stage Knowledge Verification for Event Causality Identification
Ya Su (Shanxi University), Hongye Tan (Shanxi University)
ClassificationRepresentation LearningTransformerLarge Language ModelContrastive LearningTextBenchmarkStochastic Differential Equation
🎯 What it does: Propose a dynamic energy-based contrastive learning method that enhances the quality of training samples for event causality recognition through multi-stage knowledge verification
Dynamic Expert Specialization: Towards Catastrophic Forgetting-Free Multi-Domain MoE Adaptation
Junzhuo Li (Hong Kong University of Science and Technology), Xuming Hu (Hong Kong University of Science and Technology)
Domain AdaptationKnowledge DistillationMixture of ExpertsText
🎯 What it does: Propose the DES-MoE framework, dynamic expert allocation to achieve multi-domain Mixture-of-Experts fine-tuning without catastrophic forgetting
Dynamic Jointly Batch Selection for Data Efficient Machine Translation Fine-Tuning
Mohammad Amin Ghanizadeh (University of Tehran), Mohammad Javad Dousti (University of Tehran)
Computational EfficiencyData-Centric LearningTransformerSupervised Fine-TuningText
🎯 What it does: For fine-tuning machine translation models, this paper proposes a dynamic batch selection method based on the learnability metric, which can automatically select data samples that are both easy to learn and unlabeled during training.
Dynamic Model-Bank Test-Time Adaptation for Automatic Speech Recognition
Yanshuo Wang (Hong Kong Polytechnic University), Xuesong Li (Australian National University)
RecognitionDomain AdaptationTransformerAudio
🎯 What it does: Proposes a dynamic model bank-based single-sentence test-time adaptation framework, DMSUTA, enabling ASR to maintain robustness under continuous domain drift.
Dynamic Retriever for In-Context Knowledge Editing via Policy Optimization
Mahmud Wasif Nafee (Rensselaer Polytechnic Institute), Yanfu Zhang (William & Mary)
RetrievalOptimizationTransformerReinforcement LearningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose a dynamic retriever based on policy gradient training to dynamically select and rank examples for knowledge editing while keeping LLM weights unchanged.
DynamicNER: A Dynamic, Multilingual, and Fine-Grained Dataset for LLM-based Named Entity Recognition
Hanjun Luo (New York University Abu Dhabi), Zuozhu Liu (Zhejiang University)
RecognitionComputational EfficiencyData-Centric LearningLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes a multilingual, multi-granularity named entity recognition (NER) dataset called DynamicNER and its dynamic classification method, and designs a two-stage lightweight LLM framework named CascadeNER for efficient and deployable NER tasks.
Dyve: Thinking Fast and Slow for Dynamic Process Verification
Jianyuan Zhong (Chinese University of Hong Kong), Qiang Xu (Chinese University of Hong Kong)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkChain-of-Thought
🎯 What it does: Propose Dyve—a process validator capable of dynamically switching between fast (System 1) and deep (System 2) verification during large language model reasoning.
E2LLM: Encoder Elongated Large Language Models for Long-Context Understanding and Reasoning
Zihan Liao (East China Normal University), Wei Zhang (East China Normal University)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Chunk long texts, compress each chunk into soft prompts using a pre-trained text encoder, align them to the decoder's input space via an adapter, and complete understanding and reasoning tasks through the decoder.
Easy as PIE? Identifying Multi-Word Expressions with LLMs
Kai Golan Hashiloni (Reichman University), Kfir Bar (Reichman University)
RecognitionTransformerPrompt EngineeringTextChain-of-Thought
🎯 What it does: Utilizing large language models without fine-tuning to identify idioms (multi-word expressions) in text through prompt engineering
EasyRec: Simple yet Effective Language Models for Recommendation
Xubin Ren (University of Hong Kong), Chao Huang (University of Hong Kong)
Recommendation SystemTransformerLarge Language ModelContrastive LearningText
🎯 What it does: This paper proposes EasyRec, a recommendation framework that combines language models with collaborative filtering, achieving excellent performance in both text zero-shot recommendation and text-enhanced collaborative filtering scenarios.
ECC: An Emotion-Cause Conversation Dataset for Empathy Response
Yuanyuan He (University of Electronic Science and Technology of China), Fuji Ren (University of Electronic Science and Technology of China)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes an end-to-end automated framework, EC-DD, which generates dialogues with emotion and causal labels using large language models combined with common-sense knowledge, and builds the ECC dataset based on this; subsequently, the CAER model is trained to achieve causality-aware empathetic responses.
ECO Decoding: Entropy-Based Control for Controllability and Fluency in Controllable Dialogue Generation
Seungmin Shin (SungKyunKwan University), Youngjoong Ko (SungKyunKwan University)
GenerationText
🎯 What it does: Propose an entropy-based dynamic control strength decoding method (ECO Decoding) to balance attribute control and text fluency in controllable dialogue generation.
EcoLoRA: Communication-Efficient Federated Fine-Tuning of Large Language Models
Han Liu (Washington University in St. Louis), Ning Zhang (Washington University in St. Louis)
Federated LearningComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose the ECOLORA framework to achieve communication-efficient federated fine-tuning for large language models, primarily by reducing upload volume through polling segment sharing, adaptive sparsification, and lossless encoding.
EcoTune: Token-Efficient Multi-Fidelity Hyperparameter Optimization for Large Language Model Inference
Yuebin Xu (Hong Kong University of Science and Technology), Zeyi Wen (Hong Kong University of Science and Technology)
Computational EfficiencyHyperparameter SearchLarge Language ModelTextBenchmark
🎯 What it does: Propose the EcoTune method for token-efficient multi-precision optimization of decoding hyperparameters during large language model inference;
Editing Across Languages: A Survey of Multilingual Knowledge Editing
Nadir Durrani (Qatar Computing Research Institute, Hamad Bin Khalifa University), Fahim Dalvi (Qatar Computing Research Institute, Hamad Bin Khalifa University)
Supervised Fine-TuningPrompt EngineeringTextReview/Survey PaperBenchmark
🎯 What it does: Systematically reviews existing methods, evaluation benchmarks, and research trends in multilingual knowledge editing (MKE), proposes a unified classification framework for four major method families, outlines key challenges such as cross-lingual transfer, model size, and sustainability, and provides future research directions.
EduAdapt: A Question Answer Benchmark Dataset for Evaluating Grade-Level Adaptability in LLMs
Numaan Naeem (Mohamed bin Zayed University of Artificial Intelligence), Muhammad Abdul-Mageed (University of British Columbia)
Large Language ModelTextBenchmark
🎯 What it does: Proposed the EDUADAPT benchmark, covering all K-12 education stages, nine subjects, and nearly 50,000 QA pairs, to evaluate the grade-level adaptability of large language models.
EduVidQA: Generating and Evaluating Long-form Answers to Student Questions based on Lecture Videos
Sourjyadip Ray (Indian Institute of Technology Kharagpur), Pawan Goyal (Indian Institute of Technology Kharagpur)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: Constructed the EduVidQA dataset and evaluated the performance of multimodal large language models in generating long-form responses
Effective Red-Teaming of Policy-Adherent Agents
Itay Nakash (IBM Research AI), Ateret Anaby Tavor
Safty and PrivacyTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmark
🎯 What it does: Proposed a red team framework called CRAFT for policy-compliant LLM agents, and transformed the original task completion benchmark τ-bench into a security assessment benchmark τ-break;
Efficient Beam Search for Large Language Models Using Trie-Based Decoding
Brian J Chan (National Chengchi University), Hen-Hsen Huang (Academia Sinica)
GenerationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose a parallel beam search decoding method based on a prefix tree, significantly reducing memory usage while maintaining generation quality through shared KV cache.
Efficient Compositional Multi-tasking for On-device Large Language Models
Ondrej Bohdal (Samsung R&D Institute UK), Umberto Michieli (Samsung R&D Institute UK)
Computational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes a compositional multi-tasking method for implementing large language models (LLMs) on mobile devices and creates a benchmark containing four practical compositional tasks for this scenario; simultaneously, it designs a Learnable Calibration method that only requires minimal additional parameters to calibrate existing task adapters (LoRA), enabling multi-tasking with a single inference.