EMNLP 2025 Papers — Page 17
Conference on Empirical Methods in Natural Language Processing · 1809 papers
ToM-SSI: Evaluating Theory of Mind in Situated Social Interactions
Matteo Bortoletto (University of Stuttgart), Andreas Bulling (University of Stuttgart)
Data SynthesisLarge Language ModelVision Language ModelMultimodalityBenchmark
🎯 What it does: Designed and released a multimodal benchmark called ToM-SSI to evaluate the theory of mind (ToM) capabilities of large models in multi-agent, space-constrained social interaction scenarios.
ToM: Leveraging Tree-oriented MapReduce for Long-Context Reasoning in Large Language Models
Jiani Guo (Wuhan University), Yujiu Yang (Tsinghua University)
Large Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Propose a tree-structured MapReduce framework called ToM to enhance the logical coherence and information integration capabilities of large language models in long-text reasoning.
TombRaider: Entering the Vault of History to Jailbreak Large Language Models
Junchen Ding (UNSW), Yuekang Li (UNSW)
Adversarial AttackLarge Language ModelAgentic AIPrompt EngineeringText
🎯 What it does: Proposed a multi-round Jailbreak framework TOMBRAIDER based on historical knowledge, utilizing two agents, Inspector-Attacker, to progressively guide LLMs to generate harmful outputs.
ToneCraft: Cantonese Lyrics Generation with Harmony of Tones and Pitches
Junyu Cheng (South China Normal University), Shuangyin Li (South China Normal University)
GenerationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Propose the ToneCraft framework to achieve Cantonese lyric generation based on large models, ensuring harmony between lyrics and melody in terms of pitch and tone;
Too Consistent to Detect: A Study of Self-Consistent Errors in LLMs
Hexiang Tan (Institute of Computing Technology, Chinese Academy of Sciences), Xueqi Cheng (Institute of Computing Technology, Chinese Academy of Sciences)
Anomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Studied the phenomenon of self-consistent errors generated by large language models (LLMs) during multi-sample generation, evaluated existing error detection methods, and proposed a cross-model probe to enhance the detection of self-consistent errors.
Too Helpful, Too Harmless, Too Honest or Just Right?
Gautam Siddharth Kashyap (Macquarie University), Usman Naseem (Macquarie University)
TransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsText
🎯 What it does: Propose TrinityX, a modular alignment framework that integrates low-rank task vector fine-tuning with Mixture of Calibrated Experts (MoCaE), achieving unified alignment for LLMs across three dimensions: Helpfulness, Harmlessness, and Honesty.
Tool Preferences in Agentic LLMs are Unreliable
Kazem Faghih (University of Maryland), Soheil Feizi (University of Maryland)
Explainability and InterpretabilityTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmark
🎯 What it does: By editing the text descriptions of tools, the study reveals that LLMs are highly susceptible to the influence of descriptions when selecting tools, leading to significant variations in tool usage rates;
ToolSafety: A Comprehensive Dataset for Enhancing Safety in LLM-Based Agent Tool Invocations
Yuejin Xie (Chinese University of Hong Kong), Pinjia He (Huawei)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningAgentic AITextBenchmark
🎯 What it does: This paper proposes the ToolSafety dataset, specifically designed for safety fine-tuning when LLM agents call tools, addressing insufficient safety in multi-step calls and indirect harm scenarios.
Topic Coverage-based Demonstration Retrieval for In-Context Learning
Wonbin Kweon (University of Illinois Urbana Champaign), Hwanjo Yu (Pohang University of Science and Technology)
RetrievalTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the TopicK framework, which leverages topic coverage to achieve targeted example retrieval, thereby enhancing the context learning effectiveness of large language models.
TopicAttack: An Indirect Prompt Injection Attack via Topic Transition
Yulin Chen (National University of Singapore), Bryan Hooi (National University of Singapore)
Adversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose TopicAttack, achieving indirect prompt injection attacks by constructing topic-transition dialogues and reminder prompts;
TORSO: Template-Oriented Reasoning Towards General Tasks
Minhyuk Kim (Korea University), Heuiseok Lim (Korea University)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposed a method called TORSO, which activates the model's inherent reasoning capabilities by injecting two template tokens, <reasoning> and <answer>, into the decoding stage of LLMs, avoiding the need for manually crafted few-shot examples.
TounsiBench: Benchmarking Large Language Models for Tunisian Arabic
Souha Ben Hassine (University of Michigan-Flint), Steven R. Wilson (University of Michigan-Flint)
Large Language ModelTextBenchmark
🎯 What it does: Created and released the TounsiBench benchmark, conducting a systematic evaluation of the instruction-following dialogue responses of 10 mainstream LLMs in Tunisian Arabic (Derja), and provided corresponding gold standard answers and evaluation framework.
Toward Efficient Sparse Autoencoder-Guided Steering for Improved In-Context Learning in Large Language Models
Ikhyun Cho (University of Illinois at Urbana-Champaign), Julia Hockenmaier (University of Illinois at Urbana-Champaign)
ClassificationLarge Language ModelPrompt EngineeringAuto EncoderText
🎯 What it does: Proposes a two-step method based on sparse autoencoders: Feature Detection through Prompt Variation (FDPV) for efficiently identifying features usable for steering; SISTER for selectively steering only on label words in ICL classification tasks, thereby enhancing model performance.
Toward Machine Interpreting: Lessons from Human Interpreting Studies
Matthias Sperber (Apple), Matthias Paulik (Apple)
TransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextReview/Survey PaperBenchmark
🎯 What it does: This paper reviews research on human interpreting, distills its operational and quality objectives, and discusses how to transfer these principles to machine translation systems to enhance the interactive experience of interpreting.
Toward Machine Translation Literacy: How Lay Users Perceive and Rely on Imperfect Translations
Yimin Xiao (University of Maryland), Marine Carpuat (University of Maryland)
TextTabular
🎯 What it does: Conduct human experiments in museums to evaluate the impact of fluency and sufficiency errors on bilingual and non-bilingual users' use of machine translation.
Toward Multi-Session Personalized Conversation: A Large-Scale Dataset and Hierarchical Tree Framework for Implicit Reasoning
Xintong Li (University of California San Diego), Jingbo Shang (University of California San Diego)
RetrievalLarge Language ModelTextBenchmark
🎯 What it does: Proposed a large-scale multi-conversation dialogue dataset called IMPLEXCONV for evaluating implicit reasoning in long-term personalized dialogues; and designed a hierarchical tree retrieval framework named TACITREE to efficiently retrieve hidden implicit information buried in long conversation histories.
Towards a Holistic and Automated Evaluation Framework for Multi-Level Comprehension of LLMs in Book-Length Contexts
Yuho Lee (Korea Advanced Institute of Science and Technology), Hwanjun Song (Korea Advanced Institute of Science and Technology)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose the HAMLET framework for multi-level automatic evaluation of long-text understanding in large language models (LLMs) at the book level.
Towards a Unified Paradigm of Concept Editing in Large Language Models
Zhuowen Han (Tianjin University), Deyi Xiong (Tianjin University)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningAuto EncoderText
🎯 What it does: Propose a unified neuronal-level paradigm to uniformly describe the concept injection process and systematically evaluate four representative concept editing methods (Neuron Editing, Supervised Fine-tuning, Sparse Autoencoder, Steering Vector), while providing an efficient SAE feature interpretation and localization scheme.
Towards Advanced Mathematical Reasoning for LLMs via First-Order Logic Theorem Proving
Chuxue Cao (Hong Kong University of Science and Technology), Yike Guo (Hong Kong University of Science and Technology)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposes the DREAM framework for first-order logic theorem proving in large language models
Towards AI-Assisted Psychotherapy: Emotion-Guided Generative Interventions
Kilichbek Haydarov (King Abdullah University of Science and Technology), Mohamed Elhoseiny (King Abdullah University of Science and Technology)
GenerationTransformerLarge Language ModelPrompt EngineeringVideoTextMultimodalityAudio
🎯 What it does: Constructed a multimodal dataset containing 1,441 psychotherapy videos, and developed a prompting strategy based on emotional dissonance to enhance the performance of large language models (LLMs) in generating more empathetic psychological intervention texts.
Towards Author-informed NLP: Mind the Social Bias
Inbar Pendzel (University of Haifa), Einat Minkov (University of Haifa)
ClassificationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningTextGraph
🎯 What it does: Built a user embedding space (SocialVec) based on the large-scale Twitter social network, and combined it with text representations for author-driven social NLP tasks such as stance detection and toxicity detection, further analyzing the association between author stance and socio-demographic features.
Towards Automated Error Discovery: A Study in Conversational AI
Dominic Petrak (Technical University of Darmstadt), Iryna Gurevych (Technical University of Darmstadt)
Anomaly DetectionTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Proposed an automated error discovery framework and implemented the SEEED model to detect known and unknown errors in dialogue AI, and automatically generate definitions for new errors.
Towards Controllable Speech Synthesis in the Era of Large Language Models: A Systematic Survey
Tianxin Xie (Hong Kong University of Science and Technology (Guangzhou)), Li Liu (Hong Kong University of Science and Technology (Guangzhou))
GenerationTransformerLarge Language ModelDiffusion modelFlow-based ModelTextReview/Survey PaperAudio
🎯 What it does: This paper reviews the research progress in the field of controllable text-to-speech (TTS), covering model architectures, control strategies, feature representations, datasets, and evaluation methods, and outlines current challenges and future research directions.
Towards Event Extraction with Massive Types: LLM-based Collaborative Annotation and Partitioning Extraction
Wenxuan Liu (Chinese Academy of Sciences), Xueqi Cheng (University of Chinese Academy of Sciences)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose a collaborative annotation framework based on LLM and construct the largest-scale event extraction dataset EEMT, while introducing the LLM-PEE partitioned extraction method;
Towards Faithful Natural Language Explanations: A Study Using Activation Patching in Large Language Models
Wei Jie Yeo (Nanyang Technological University), Erik Cambria (Nanyang Technological University)
Explainability and InterpretabilityTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose a causal faithfulness metric based on activation patching to evaluate the causal consistency between natural language explanations (NLE) generated by large language models (LLMs) and the answers.
Towards General-Domain Word Sense Disambiguation: Distilling Large Language Model into Compact Disambiguator
Liqiang Ming (Shenzhen University), Yuncong Li (EnCopilot Inc.)
ClassificationKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose to utilize large language models (LLMs) for distilling or annotating silver-standard word sense disambiguation (WSD) data, training small models to achieve general-domain WSD.
Towards Holistic Evaluation of Large Audio-Language Models: A Comprehensive Survey
Chih-Kai Yang, Hung-yi Lee (National Taiwan University)
MultimodalityReview/Survey PaperBenchmarkAudio
🎯 What it does: Review the evaluation framework for large audio-language models (LALMs) and propose a systematic classification based on four dimensions (general auditory perception and processing, knowledge and reasoning, conversational ability, and fairness, safety, and trustworthiness)
Towards Infinite-Long Prefix in Transformer
Yingyu Liang (University of Hong Kong), Chiwun Yang (Sun Yat-sen University)
Computational EfficiencyRepresentation LearningTransformerPrompt EngineeringImageText
🎯 What it does: Investigate the feasibility of infinite-length prefixes in prefix learning, prove their convergence using NTK, and propose NTK-Attention to achieve efficient prefix approximation.
Towards Language-Agnostic STIPA: Universal Phonetic Transcription to Support Language Documentation at Scale
Jacob Lee Suchardt (University of Gothenburg), Pierluigi Cassotti (University of Gothenburg)
RecognitionTransformerSupervised Fine-TuningAudio
🎯 What it does: Achieve speech-to-International Phonetic Alphabet (IPA) transcription (STIPA) by fine-tuning the Whisper model, and construct an IPA dataset for Southern Levantine Arabic;
Towards Optimal Evaluation Efficiency for Large Language Models
Guohong Li (Tianjin University), Deyi Xiong (Tianjin University)
OptimizationComputational EfficiencyLarge Language ModelAuto EncoderText
🎯 What it does: This paper studies the method of efficiently selecting a subset of questions in large language model evaluation by using pre-test results, and models subset selection as an optimizable objective function, solved using optimal random sampling and simulated annealing algorithms.
Towards Robust Mathematical Reasoning
Thang Luong (Google DeepMind), Junehyuk Jung (Brown University)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Designed and released IMO-Bench, comprising AnswerBench, ProofBench, and GradingBench, to evaluate the mathematical reasoning and proof capabilities of large language models at the International Mathematical Olympiad (IMO) level.
Towards Statistical Factuality Guarantee for Large Vision-Language Models
Zhuohang Li (Vanderbilt University), Bradley A. Malin (Vanderbilt University)
Explainability and InterpretabilityTransformerLarge Language ModelVision Language ModelImageMultimodalityBiomedical Data
🎯 What it does: This paper proposes the CONFLVLM framework, which decomposes generated text into individual propositions and achieves statistical guarantees on the factual outputs of large vision-language models (LVLMs) through distribution-free synthetic prediction;
Towards Transferable Personality Representation Learning based on Triplet Comparisons and Its Applications
Kai Tang (Zhejiang University), Haobo Wang (Zhejiang University)
Representation LearningTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Constructed a triplet-based personality tendency comparison dataset (PTCD) and obtained measurable sentence-level personality embeddings through triplet learning;
Toxicity Red-Teaming: Benchmarking LLM Safety in Singapore’s Low-Resource Languages
Yujia Hu (Singapore University of Technology and Design), Roy Ka-Wei Lee (Singapore University of Technology and Design)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: In a multilingual environment in Singapore, the SGToxicGuard dataset and evaluation framework were proposed to systematically assess the toxicity of LLMs in three real-world scenarios: dialogue, question-answering, and content creation.
TP-RAG: Benchmarking Retrieval-Augmented Large Language Model Agents for Spatiotemporal-Aware Travel Planning
Hang Ni (Hong Kong University of Science and Technology (Guangzhou)), Hao Liu (Hong Kong University of Science and Technology (Guangzhou))
Recommendation SystemOptimizationLarge Language ModelAgentic AITextTime SeriesSequentialBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper constructs the first trajectory-level retrieval-enhanced LLM travel planning benchmark, TP-RAG, and proposes the EvoRAG evolutionary optimization framework, systematically evaluating the performance of LLM agents in spatiotemporal-aware travel planning.
Tracing L1 Interference in English Learner Writing: A Longitudinal Corpus with Error Annotations
Poorvi Acharya (George Mason University), Antonios Anastasopoulos (George Mason University)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper created and made publicly available the LENS (Longitudinal English Nonnative Speaker) corpus, which includes 15 academic writing samples from graduate students with different first languages (L1), along with fine-grained error annotations, particularly L1 interference tags.
TracSum: A New Benchmark for Aspect-Based Summarization with Sentence-Level Traceability in Medical Domain
Bohao Chu (University of Duisburg-Essen), Norbert Fuhr (University of Duisburg-Essen)
TransformerLarge Language ModelBiomedical DataBenchmarkChain-of-Thought
🎯 What it does: Constructed the TRACSUM benchmark in the medical field, providing 500 clinical abstracts with 3.5K summary-citation pairs across seven key medical dimensions, and proposed a fine-grained evaluation framework and the TRACK-THEN-SUM two-stage baseline model; evaluated multiple LLMs and conducted human evaluations.
Train It and Forget It: Merge Lists are Unnecessary for BPE Inference in Language Models
Tomohiro Sawada, Kartik Goyal (Georgia Institute Of Technology)
Safty and PrivacyComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Explores inference algorithms for large language models that do not rely on BPE merge lists, and systematically evaluates their impact on model performance and safety.
Train One Sparse Autoencoder Across Multiple Sparsity Budgets to Preserve Interpretability and Accuracy
Nikita Balagansky (T Tech), Daniil Gavrilov (T Tech)
Explainability and InterpretabilityRepresentation LearningAuto EncoderText
🎯 What it does: Proposed the HierarchicalTopK training objective, enabling a single sparse autoencoder to maintain high interpretability and reconstruction quality under various sparsity budgets.
Training a Utility-based Retriever Through Shared Context Attribution for Retrieval-Augmented Language Models
Yilong Xu (Chinese Academy of Sciences), Xueqi Cheng (Chinese Academy of Sciences)
Data SynthesisRetrievalExplainability and InterpretabilityContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Propose the SCARLet framework, which trains the retriever through shared context synthesis and perturbation attribution, enabling the retriever to focus on the practical value of downstream tasks.
Training compute-optimal transformer encoder models
Megi Dervishi (FAIR at Meta), Yann LeCun (FAIR at Meta)
OptimizationComputational EfficiencyHyperparameter SearchData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: Systematically studied the computational optimal pre-training of Transformer encoders, investigated the scaling laws of learning rate, batch size, and data-model ratio through large-scale experiments, and proposed the OptiBERT series models based on these findings.
Training LLMs to be Better Text Embedders through Bidirectional Reconstruction
Chang Su (Shanghai Jiao Tong University), Zhouhan Lin (Shanghai Jiao Tong University)
RetrievalRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextBenchmark
🎯 What it does: Propose a two-stage training framework that introduces bidirectional reconstruction tasks (EBQ2D and EBD2Q) before contrastive learning, leveraging EOS token vectors as anchors to enable LLM text embeddings to better capture semantic relationships between queries and documents.
Transferable Direct Prompt Injection via Activation-Guided MCMC Sampling
Minghui Li (Huazhong University of Science and Technology), Jing Wang (Huazhong University of Science and Technology)
Adversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose a transferable direct prompt injection attack framework based on activation guidance, generating efficient and natural attack prompts in black-box environments using energy models and MCMC sampling.
Transformer-Based Temporal Information Extraction and Application: A Review
Xin Su (Intel), Steven Bethard (University of Arizona)
TransformerLarge Language ModelPrompt EngineeringTextReview/Survey Paper
🎯 What it does: This paper systematically reviews the application of Transformer models in temporal information extraction, covering datasets, methods, and downstream tasks.
Transitive self-consistency evaluation of NLI models without gold labels
Wei Wu (Ben Gurion University of the Negev), Mark Last (Ben Gurion University of the Negev)
ClassificationTransformerLarge Language ModelText
🎯 What it does: This paper proposes an automatic evaluation method without manual annotation, using antonym replacement to generate adversarial samples to detect self-consistency of NLI models in transmission consistency.
Translate Smart, not Hard: Cascaded Translation Systems with Quality-Aware Deferral
António Farinhas (Sword Health), Andre Martins
GenerationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Designed and implemented a machine translation cascade system based on quality estimation (QE) metrics: first, a small model generates translations, then a lightweight QE model evaluates the quality and decides whether to re-generate the translation using a large model; automatic and human evaluations were conducted under various configurations.
Translating Domain-Specific Terminology in Typologically-Diverse Languages: A Study in Tax and Financial Education
Arturo Oncevay (JPMorgan Chase), Charese Smiley
Large Language ModelPrompt EngineeringTextFinance Related
🎯 What it does: Constructed a gold-standard terminology resource covering seven languages (English, Spanish, Russian, Vietnamese, Korean, Traditional/Simplified Chinese, and Haitian Creole) in the tax and financial education domain, with approximately 3,000 terms annotated for domain specificity; evaluated the performance of multiple MT systems and LLMs in overall translation quality and terminology accuracy using this terminology library and parallel corpora extracted from IRS public articles; further experimented with term-assisted translation prompts and term extraction capabilities in parallel texts.
Translation in the Hands of Many: Centering Lay Users in Machine Translation Interactions
Beatrice Savoldi (Fondazione Bruno Kessler), Luisa Bentivogli (Fondazione Bruno Kessler)
Large Language ModelTextReview/Survey Paper
🎯 What it does: This paper reviews the evolution of machine translation (MT) technology from specialized scenarios to mass-market adoption, focusing on the experiences and needs of non-expert users in MT interactions, and proposes a user-centered research framework centered on three factors: usability, trust, and literacy.
Translationese-index: Using Likelihood Ratios for Graded and Generalizable Measurement of Translationese
Yikang Liu (Shanghai Jiao Tong University), Hai Hu (City University of Hong Kong)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: Investigated the 'translation flavor' of translated text, proposing a continuous and quantifiable translation flavor metric called T-index.
Transparent and Coherent Procedural Mistake Detection
Shane Storks (University of Michigan), Joyce Chai (University of Michigan)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelVideoMultimodalityChain-of-Thought
🎯 What it does: In this work, the authors extend Programmed Mistake Detection (PMD) to require the generation of explainable and coherent visual self-dialog reasoning. They construct a single-frame version of the PMD dataset, Ego4D-PMD, based on Ego4D. Additionally, they propose two coherence metrics using NLI models to quantify question relevance and answer informativeness, and introduce interventions in Vision-Language Models (VLMs), such as re-ranking, context learning, and preference optimization, based on these metrics.
Transplant Then Regenerate: A New Paradigm for Text Data Augmentation
Guangzhan Wang (Shanghai Jiao Tong University), Xiaodong Gu (Shanghai Jiao Tong University)
ClassificationRecognitionData SynthesisTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose LMTransplant, a novel text data augmentation paradigm that embeds seed text into a bidirectional context generated by large language models (LLMs) and regenerates text within this context.
Tree-of-Quote Prompting Improves Factuality and Attribution in Multi-Hop and Medical Reasoning
Justin Xu (University of Oxford), David W Eyre (University of Oxford)
Large Language ModelPrompt EngineeringTextBiomedical DataChain-of-Thought
🎯 What it does: Propose the Tree-of-Quote (ToQ) prompting framework, which generates and evaluates citations at each step in multi-hop reasoning and medical question answering to enhance factual accuracy and traceability.
TreeRare: Syntax Tree-Guided Retrieval and Reasoning for Knowledge-Intensive Question Answering
Boyi Zhang (University of Rochester), Hangfeng He (University of Rochester)
RetrievalLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the TreeRare framework, which utilizes syntactic trees for semantic decomposition and performs sub-component retrieval and answer generation at each node, eventually aggregating to produce the final answer.
TreeReview: A Dynamic Tree of Questions Framework for Deep and Efficient LLM-based Scientific Peer Review
Yuan Chang (Chinese Academy of Sciences), Ngai Wong (University of Hong Kong)
Computational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose a dynamic tree-based question-answering framework called TreeReview, based on large language models (LLMs), to enable automatic peer review of scientific papers.
TRIAL: Token Relations and Importance Aware Late-interaction for Accurate Text Retrieval
Hyukkyu Kang (POSTECH), Wook-Shin Han (POSTECH)
RetrievalTransformerContrastive LearningText
🎯 What it does: Propose TRIAL, a late-interaction text retrieval model based on token relationships and importance-awareness;
TrInk: Ink Generation with Transformer Network
Zezhong Jin (The Hong Kong Polytechnic University), Shujie Liu (Microsoft Corporation)
GenerationData SynthesisTransformerImageText
🎯 What it does: TrInk proposes a Transformer-based online handwriting synthesis model capable of generating coherent pen trajectories from text input.
TrojanStego: Your Language Model Can Secretly Be A Steganographic Privacy Leaking Agent
Dominik Meier (University of Göttingen), Bela Gipp (Bocconi University)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose the TrojanStego threat model, which leverages LLMs to embed sensitive contextual information into natural outputs through linguistic steganography, achieving data leakage without prompts or explicit control;
TrojanWave: Exploiting Prompt Learning for Stealthy Backdoor Attacks on Large Audio-Language Models
Asif Hanif (Mohamed Bin Zayed University of Artificial Intelligence), Karthik Nandakumar (Mohamed Bin Zayed University of Artificial Intelligence)
Adversarial AttackTransformerPrompt EngineeringAudio
🎯 What it does: This paper proposes TrojanWave, a method for implanting covert backdoors in frozen large audio language models by leveraging learnable prompts, and provides a corresponding lightweight defense scheme called TrojanWave-Defense;
Trojsten Benchmark: Evaluating LLM Problem-Solving in Slovak STEM Competition Problems
Adam Zahradník (Comenius University in Bratislava), Marek Suppa
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Constructed the Trojsten Benchmark—a collection of 1,108 Slovak high school STEM competition problems and their solutions—and designed an automatic grading framework based on LLMs.
TRUST-VL: An Explainable News Assistant for General Multimodal Misinformation Detection
Zehong Yan (National University of Singapore), Mong-Li Lee
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: Propose a unified interpretable vision-language model TRUST-VL for detecting text, visual, and cross-modal multimodal fake news.
Trustworthy Medical Question Answering: An Evaluation-Centric Survey
Yinuo Wang (Shandong University), Xindi Wang (Shandong University)
Safty and PrivacyExplainability and InterpretabilityReinforcement Learning from Human FeedbackBiomedical DataReview/Survey PaperRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Reviews the dimensions of trustworthiness evaluation in medical question-answering systems (factuality, robustness, fairness, safety, explainability, calibration) and systematically organizes corresponding assessment methods, benchmarks, and evaluation-driven improvement pathways.
TS-CLIP: Time Series Understanding by CLIP
Ziwen Chen (Huazhong University of Science and Technology), Ming Zhu (Huazhong University of Science and Technology)
ClassificationPrompt EngineeringMixture of ExpertsContrastive LearningTime Series
🎯 What it does: Built a TS-CLIP model based on the CLIP contrastive learning framework to align time series data with natural language, enabling zero-shot and few-shot time series classification.
TSVer: A Benchmark for Fact Verification Against Time-Series Evidence
Marek Strong (University of Cambridge), Andreas Vlachos (University of Cambridge)
Data-Centric LearningTransformerLarge Language ModelTextTime SeriesBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the TSVer benchmark dataset, which focuses on fact-checking using time series evidence.
TTT-Bench: A Benchmark for Evaluating Reasoning Ability with Simple and Novel Tic-Tac-Toe-style Games
Prakamya Mishra (Advanced Micro Devices, Inc.), Emad Barsoum (Advanced Micro Devices, Inc.)
TransformerTextBenchmarkChain-of-Thought
🎯 What it does: This paper designs and releases the TTT-Bench benchmark to evaluate the reasoning and strategic capabilities of large reasoning models in four simple two-player board games.
Tuning Less, Prompting More: In-Context Preference Learning Pipeline for Natural Language Transformation
Shuyun Yang (Sichuan University), Mingjie Tang (Sichuan University)
TransformerPrompt EngineeringContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Proposed a framework that combines prompt learning and small model fine-tuning, enhancing data by retrieving similar context ICE (In-Context Examples) to achieve better context awareness and preference alignment in machine translation (MT) and text style transfer (TST) tasks.
TurBLiMP: A Turkish Benchmark of Linguistic Minimal Pairs
Ezgi Başar (University of Groningen), Arianna Bisazza (University of Groningen)
TextBenchmark
🎯 What it does: This paper proposes TurBLiMP, a Turkish language model evaluation benchmark covering 16 grammatical phenomena, with 1,000 minimal sentence pairs for each phenomenon, and expands lexical diversity through manual and semi-automated methods, further incorporating experimental paradigms for word order and subclauses.
TurboRAG: Accelerating Retrieval-Augmented Generation with Precomputed KV Caches for Chunked Text
Songshuo Lu (Moore Threads AI), Yaohua Tang (Moore Threads AI)
RetrievalComputational EfficiencyTransformerTextRetrieval-Augmented Generation
🎯 What it does: Propose TurboRAG by precomputing KV caches for each retrieved segment and concatenating them during inference, reducing prefill computation and latency in RAG systems.
TurnaboutLLM: A Deductive Reasoning Benchmark from Detective Games
Yuan Yuan (Drexel University), Li Zhang (University of Pennsylvania)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes the TURNABOUTLLM framework and dataset to evaluate the deductive reasoning ability of large language models in推理 based on Ace Attorney and Danganronpa.
TurnBack: A Geospatial Route Cognition Benchmark for Large Language Models through Reverse Route
Hongyi Luo (Huawei Riemann Lab), Liqiu Meng (TU Munchen)
TransformerLarge Language ModelGraphBenchmark
🎯 What it does: This paper introduces the TurnBack benchmark, which systematically evaluates the spatial cognition ability of large language models (LLMs) in route reversal tasks, using PathBuilder to convert natural language navigation instructions and geometric paths into each other;
Turning Logic Against Itself: Probing Model Defenses Through Contrastive Questions
Rachneet Singh Sachdeva (Technical University of Darmstadt), Iryna Gurevych (Technical University of Darmstadt)
Adversarial AttackLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose the POATE two-phase adversarial generation method, which utilizes opposite-polarity questions and adversarial templates to induce LLMs to generate harmful responses, and designs two chain-of-thought (CoT) defenses: Intent-Aware CoT and Reverse Thinking CoT.
TVQACML: Benchmarking Text-Centric Visual Question Answering in Multilingual Chinese Minority Languages
Sha Jiu (Minzu University of China), Jialedongzhu (Minzu University of China)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Constructed a text-centric visual question answering benchmark TVQACML for multilingual Chinese minority languages, containing 8,000 real images and 32,000 high-quality QA pairs, covering 8 languages and 30 scenarios.
Two Heads Are Better Than One: Dual-Model Verbal Reflection at Inference-Time
Jiazheng Li (King's College London), Yulan He (King's College London)
ClassificationSupervised Fine-TuningContrastive LearningTextChain-of-Thought
🎯 What it does: Propose a dual-model reflective scoring framework (Reasoner+Critic), generating precise language feedback through comparative analysis of thought trees to enhance automated scoring of student answers
Type-Less yet Type-Aware Inductive Link Prediction with Pretrained Language Models
Alessandro De Bellis, Eugenio Di Sciascio (Politecnico di Bari)
Representation LearningGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringGraph
🎯 What it does: Propose an inductive link prediction model called TyleR that captures implicit types without explicit type information, enhancing node semantic representations through pre-trained language models.
UI-Hawk: Unleashing the Screen Stream Understanding for Mobile GUI Agents
Jiwen Zhang (Fudan University), Zhongyu Wei (Huawei Technologies Co., Ltd.)
TransformerLarge Language ModelAgentic AIVision Language ModelVision-Language-Action ModelImageMultimodalityBenchmark
🎯 What it does: Propose UI-Hawk, a multimodal GUI agent with screen flow understanding capability.
UICOMPASS: UI Map Guided Mobile Task Automation via Adaptive Action Generation
Yuanzhang Lin (Beihang University), Hailong Sun (Beihang University)
TransformerLarge Language ModelAgentic AIGraph
🎯 What it does: Proposes UICOMPASS, a framework that automatically generates UI maps using static analysis and LLM, and achieves mobile task automation through adaptive path planning
UltraIF: Advancing Instruction Following from the Wild
Kaikai An (Peking University), Baobao Chang (Peking University)
Data-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Designed the ULTRAIF method, which first uses LLM to decompose real user instructions into simplified instructions, constraints, and corresponding evaluation questions; then trains UltraComposer to generate complex instructions with constraints in one go; subsequently uses a generate-evaluate process combined with preference learning (SFT+DPO/NC) to build high-quality instruction-following data; finally achieves performance comparable to the Instruct version using an 8B base model.
UNCERTAINTY-LINE: Length-Invariant Estimation of Uncertainty for Large Language Models
Roman Vashurin (Mohamed bin Zayed University of Artificial Intelligence), Maxim Panov (Mohamed bin Zayed University of Artificial Intelligence)
GenerationTransformerLarge Language ModelText
🎯 What it does: Propose a post-hoc debiasing method called UNCERTAINTY-LINE to eliminate the impact of length bias in the generated text of large language models on uncertainty estimation.
UNCLE: Benchmarking Uncertainty Expressions in Long-Form Generation
Ruihan Yang (Fudan University), Deqing Yang (Fudan University)
GenerationSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Developed the ULCLE benchmark to evaluate the ability of large language models to express uncertainty in short and long text question-answering;
UnCo: Uncertainty-Driven Collaborative Framework of Large and Small Models for Grounded Multimodal NER
Jielong Tang (Sun Yat-sen University), Jian Yin (Sun Yat-sen University)
RecognitionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringImageTextMultimodalityChain-of-Thought
🎯 What it does: This paper addresses the image-based named entity recognition (GMNER) task in visual-text pairs by proposing the UnCo two-stage collaborative framework: first, a small fine-tuned model generates entity triplets, and suspicious predictions are filtered through a unified uncertainty estimation; then, a multi-modal large language model (MLLM) performs refined corrections via an uncertainty-driven hierarchical error correction mechanism.
UNComp: Can Matrix Entropy Uncover Sparsity? — A Compressor Design from an Uncertainty-Aware Perspective
Jing Xiong (University of Hong Kong), Ngai Wong (University of Hong Kong)
CompressionComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose the UNCOMP framework, which employs an uncertainty-aware method based on truncated matrix entropy to dynamically perform two-stage compression on the KV cache and hidden states of LLMs, achieving significant acceleration and memory savings for long-context reasoning.
Unconditional Truthfulness: Learning Unconditional Uncertainty of Large Language Models
Artem Vazhentsev (Center for Artificial Intelligence), Artem Shelmanov (MBZUAI)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose an unconditional uncertainty quantification method (TAD) based on attention weight learning to evaluate the reliability of text generated by large language models (LLMs), thereby enabling selective generation.
Uncovering Argumentative Flow: A Question-Focus Discourse Structuring Framework
Yini Wang (Shanghai Jiao Tong University), Xiaoying Bai (PLA Academy of Military Science)
Explainability and InterpretabilityTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes a Question-Focus structured framework to reveal the internal logical flow of analytical argument texts and guide the argument reconstruction task.
Uncovering the Bigger Picture: Comprehensive Event Understanding Via Diverse News Retrieval
Yixuan Tang (National University Of Singapore), Anthony Kum Hoe Tung
RetrievalTextBenchmark
🎯 What it does: Proposed a two-stage news retrieval framework called NEWSCOPE, aiming to maximize the diversity of results while ensuring relevance to events, thereby helping users gain a more comprehensive understanding of events.
Understanding and Leveraging the Expert Specialization of Context Faithfulness in Mixture-of-Experts LLMs
Jun Bai (State Key Laboratory of General Artificial Intelligence), Zilong Zheng (State Key Laboratory of General Artificial Intelligence)
Large Language ModelSupervised Fine-TuningMixture of ExpertsText
🎯 What it does: This paper identifies experts in Mixture-of-Experts LLMs specialized in context fidelity using Router Lens, and proposes the CEFT method that fine-tunes only these experts.
Understanding and Mitigating Overrefusal in LLMs from an Unveiling Perspective of Safety Decision Boundary
Licheng Pan (Zhejiang University), Zhixuan Chu (Zhejiang University)
Safty and PrivacyExplainability and InterpretabilityRepresentation LearningLarge Language ModelReinforcement LearningPrompt EngineeringMixture of ExpertsTextBenchmark
🎯 What it does: This paper investigates the overrefusal phenomenon of large language models (LLMs) on legitimate queries, proposing the RASS framework based on representation learning. By analyzing the safety decision boundary, it generates and filters overrefusal prompts closest to the boundary, further constructing a multilingual overrefusal benchmark, MORBENCH, to evaluate models' rejection behaviors across different languages and safety categories.
Understanding LLMs’ Cross-Lingual Context Retrieval: How Good It Is And Where It Comes From
Changjiang Gao (National Key Laboratory for Novel Software Technology, Nanjing University), Shujian Huang (National Key Laboratory for Novel Software Technology, Nanjing University)
RetrievalExplainability and InterpretabilityTransformerLarge Language ModelTextBenchmark
🎯 What it does: Investigated the performance and mechanisms of large language models (LLMs) in cross-lingual context retrieval (xMRC).
Understanding Subword Compositionality of Large Language Models
Qiwei Peng (University of Copenhagen), Anders Søgaard (University of Copenhagen)
ClassificationExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Conduct experimental studies on subword composition methods in large language models, analyzing structural similarity, semantic decomposability, and form preservation.
Understanding the Information Propagation Effects of Communication Topologies in LLM-based Multi-Agent Systems
Xu Shen (Jilin University), Xin Wang (Jilin University)
OptimizationExplainability and InterpretabilityGraph Neural NetworkLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: A causal analysis-based LLM multi-agent system communication topology optimization framework named EIB-LEARNER is proposed and studied.
Understanding the Modality Gap: An Empirical Study on the Speech-Text Alignment Mechanism of Large Speech Language Models
Bajian Xiang (Beike Inc), Wei Zou (Beike Inc)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelContrastive LearningTextMultimodalityAudio
🎯 What it does: Systematically study modality gaps in large speech language models, analyze alignment mechanisms between speech and text at both coarse and fine-grained levels, propose alignment path score (APS), and verify the causal relationship between alignment quality and speech understanding performance through interventions such as angle projection and length normalization.
Understanding the Thinking Process of Reasoning Models: A Perspective from Schoenfeld’s Episode Theory
Ming Li (University of Maryland), Tianyi Zhou (University of Maryland)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Analyze the chain-of-thought processes of large-scale reasoning models, construct a fine-grained annotated dataset based on Schoenfeld's theory of mathematical problem-solving, and publicly release the annotation protocol.
UniDebugger: Hierarchical Multi-Agent Framework for Unified Software Debugging
Cheryl Lee (Chinese University of Hong Kong), Michael R. Lyu (Chinese University of Hong Kong)
TransformerLarge Language ModelAgentic AITextRetrieval-Augmented Generation
🎯 What it does: This paper proposes UniDebugger, a multi-agent based end-to-end software debugging framework that can automatically complete the full debugging process from localization to repair within CI/CD pipelines.
UnifiedVisual: A Framework for Constructing Unified Vision-Language Datasets
Pengyu Wang (Fudan University), Xipeng Qiu (Fudan University)
Data SynthesisTransformerLarge Language ModelVision Language ModelDiffusion modelContrastive LearningImageTextMultimodalityChain-of-Thought
🎯 What it does: Proposed the UnifiedVisual framework for constructing datasets that simultaneously support multi-modal understanding and generation, and built the UnifiedVisual-240K dataset based on this, helping to mutually reinforce VLLM in visual understanding and generation tasks;
Uniform Information Density and Syntactic Reduction: Revisiting *that*-Mentioning in English Complement Clauses
Hailin Hao (University of Southern California), Elsi Kaiser (University of Southern California)
TransformerLarge Language ModelText
🎯 What it does: Test and extend Jaeger 2010's 'Uniform Information Density' hypothesis in large-scale natural dialogue corpora, investigating the relationship between the optional omission of the complementizer 'that' and information density.
Unilaw-R1: A Large Language Model for Legal Reasoning with Reinforcement Learning and Iterative Inference
Hua Cai (UniDT), Tianke Ban (Fudan University)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextBenchmarkChain-of-Thought
🎯 What it does: Built a 7B-parameter legal reasoning large language model called Unilaw-R1, enhancing legal reasoning accuracy and interpretability through two-stage training (supervised fine-tuning + reinforcement learning) and multi-agent iterative reasoning.
UnitCoder: Scalable Code Synthesis from Pre-training Corpora
Yichuan Ma (Fudan University), Kai Chen (Shanghai AI Laboratory)
Data SynthesisAI Code AssistantTransformerLarge Language ModelAgentic AIText
🎯 What it does: Propose the UnitCoder framework, which uses automatically generated unit tests to supervise the quality of pre-trained code libraries, and generates over 500K executable high-quality code data through an iterative repair and refinement process.
UniversalCEFR: Enabling Open Multilingual Research on Language Proficiency Assessment
Joseph Marvin Imperial (University of Bath), Harish Tayyar Madabushi (University of Bath)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Constructed a large-scale, multilingual, cross-level, and cross-text-type CEFR annotated dataset named UNIVERSALCEFR, making it readable and reusable through a unified standardization process;
Unlearning vs. Obfuscation: Are We Truly Removing Knowledge?
Guangzhi Sun (University of Cambridge), Mark Gales (University of Cambridge)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a method to distinguish LLM's 'true unlearning' from 'obfuscation' from an uncertainty perspective, and designs probing questions (open-ended, Yes/No, MCQ) automatically generated to evaluate the differences between the two. Subsequently, a new unlearning technique called DF-MCQ is proposed, which achieves the effect of knowledge 'erasure' by normalizing the answer distribution of multiple-choice questions (MCQ) to a uniform distribution using KL-divergence.
Unleashing the Reasoning Potential of LLMs by Critique Fine-Tuning on One Problem
Yubo Wang (University of Waterloo), Wenhu Chen (Independent)
Computational EfficiencyData-Centric LearningTransformerSupervised Fine-TuningText
🎯 What it does: Improve the mathematical and logical reasoning performance of Qwen and Llama series models through Critique Fine-Tuning (CFT), using only diverse answers to a single question and criticism data generated by a teacher LLM.
Unmasking Deceptive Visuals: Benchmarking Multimodal Large Language Models on Misleading Chart Question Answering
Zixin Chen (Hong Kong University of Science and Technology), Huamin Qu (Hong Kong University of Science and Technology)
Large Language ModelMultimodalityTabularBenchmarkChain-of-Thought
🎯 What it does: Created the Misleading ChartQA benchmark dataset (3,026 multimodal QA instances) and evaluated 24 mainstream multimodal large language models on this dataset; proposed the Region-Aware Misleader Reasoning (RAMR) pipeline to enhance models' reasoning capabilities for misleading charts.
Unmasking Fake Careers: Detecting Machine-Generated Career Trajectories via Multi-layer Heterogeneous Graphs
Michiharu Yamashita (Pennsylvania State University), Dongwon Lee (Pennsylvania State University)
Anomaly DetectionGraph Neural NetworkLarge Language ModelTextGraph
🎯 What it does: This paper constructs a structured dataset for machine-generated resumes and proposes the CareerScape model based on a global heterogeneous multilayer graph to detect fake career trajectories generated by LLMs.