ACL 2025 Papers — Page 15
Annual Meeting of the Association for Computational Linguistics · 1699 papers
Sliding Windows Are Not the End: Exploring Full Ranking with Long-Context Large Language Models
Wenhan Liu (Gaoling School of Artificial Intelligence, Renmin University of China), Zhicheng Dou (Gaoling School of Artificial Intelligence, Renmin University of China)
RetrievalComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Investigated achieving full ranking in retrieval tasks using large language models with long context, and compared it with traditional sliding-window strategies.
Small Changes, Big Impact: How Manipulating a Few Neurons Can Drastically Alter LLM Aggression
Jaewook Lee (Konkuk University), Harksoo Kim (Konkuk University)
Explainability and InterpretabilityAdversarial AttackTransformerLarge Language ModelText
🎯 What it does: This paper investigates the aggressiveness of models by identifying and manipulating 'attack neurons' within LLMs, demonstrating that activating or shielding a small number of neurons can significantly alter aggressive behaviors in outputs.
Smarter, Better, Faster, Longer: A Modern Bidirectional Encoder for Fast, Memory Efficient, and Long Context Finetuning and Inference
Benjamin Warner (Answer Ai), Iacopo Poli (Lighton)
ClassificationRetrievalAI Code AssistantTransformerLarge Language ModelText
🎯 What it does: This paper proposes ModernBERT, an efficient encoder-only Transformer architecture designed for long contexts (8192) and tailored for retrieval, classification, and code-related tasks.
SocialCC: Interactive Evaluation for Cultural Competence in Language Agents
Jincenzi Wu (Chinese University of Hong Kong), Helen M. Meng (Chinese University of Hong Kong)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes the SocialCC benchmark, which evaluates the cultural literacy of large language models (LLMs) through multi-round cross-cultural interaction scenarios.
SocialEval: Evaluating Social Intelligence of Large Language Models
Jinfeng Zhou (Tsinghua University), Minlie Huang (Tsinghua University)
TransformerTextBenchmarkChain-of-Thought
🎯 What it does: Built a bilingual script-based social intelligence evaluation benchmark named SOCIALEVAL, systematically assessing the goal achievement and interpersonal capability performance of large language models (LLMs) in social interactions using manually crafted 'World Tree' scripts;
SoftCoT: Soft Chain-of-Thought for Efficient Reasoning with LLMs
Yige Xu (Nanyang Technological University), Chunyan Miao (Nanyang Technological University)
Computational EfficiencyRepresentation LearningTransformerSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Designed and verified a SoftCoT framework that achieves efficient continuous-space chain-of-thought reasoning by generating soft thinking tokens through an auxiliary model and mapping them into the LLM representation space.
SongComposer: A Large Language Model for Lyric and Melody Generation in Song Composition
Shuangrui Ding (Chinese University of Hong Kong), Jiaqi Wang (Shanghai AI Laboratory)
GenerationTransformerLarge Language ModelTextSequential
🎯 What it does: Developed a large language model called SongComposer that can generate symbolic lyrics and corresponding melodies in one go based on prompts, forming complete and singable musical scores.
SoRFT: Issue Resolving with Subtask-oriented Reinforced Fine-Tuning
Zexiong Ma (Peking University), Bing Xie (Peking University)
AI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: Proposed a subtask-based reinforcement fine-tuning method called SoRFT, which enhances the performance of open-source LLMs in software problem-solving tasks by using rule-based rewards.
SOTOPIA-Ω: Dynamic Strategy Injection Learning and Social Instruction Following Evaluation for Social Agents
Wenyuan Zhang (Chinese Academy of Sciences), Tingwen Liu (Chinese Academy of Sciences)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextBenchmark
🎯 What it does: Propose the SOTOPIA- framework, which dynamically injects multiple strategies to generate high-quality social dialogue corpora, and trains a social agent that outperforms GPT-4;
Soundwave: Less is More for Speech-Text Alignment in LLMs
Yuhao Zhang (Chinese University of Hong Kong), Haizhou Li (Chinese University of Hong Kong)
Computational EfficiencyRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextMultimodalityChain-of-ThoughtAudio
🎯 What it does: Propose Soundwave, a three-stage efficient training framework, first aligning the audio and text representation space via CTC, then shortening the audio sequence length, and finally using LoRA for conversational instruction fine-tuning, constructing an end-to-end speech LLM.
SpaRE: Enhancing Spatial Reasoning in Vision-Language Models with Synthetic Data
Michael Ogezi (University of Waterloo Vector Institute), Freda Shi (University of Waterloo Vector Institute)
Data SynthesisLarge Language ModelVision Language ModelMultimodality
🎯 What it does: Enhancing the spatial reasoning ability of vision-language models by generating synthetic question-answer pairs from ultra-detailed image descriptions.
Sparse Latents Steer Retrieval-Augmented Generation
Chunlei Xin (Chinese Information Processing Laboratory, Institute of Software, Chinese Academy of Sciences), Le Sun (Chinese Academy of Sciences)
Explainability and InterpretabilityTransformerAuto EncoderTextRetrieval-Augmented Generation
🎯 What it does: Leverage sparse autoencoders (SAE) within the LLaMA Scope framework to uncover and manipulate internal sparse latent variables in Retrieval-Augmented Generation (RAG) systems, enabling precise control over key behaviors such as context/memory prioritization, generation, and rejection.
Sparse Logit Sampling: Accelerating Knowledge Distillation in LLMs
Anshumann (Samsung Research), Joohyung Lee (Samsung Research)
Computational EfficiencyKnowledge DistillationTransformerText
🎯 What it does: Use sparse teacher logits for knowledge distillation during the pre-training phase, and propose a random sampling KD method based on importance sampling to accelerate the distillation training of large language models.
Sparse-to-Dense: A Free Lunch for Lossless Acceleration of Video Understanding in LLMs
Xuan Zhang (Singapore Management University), Qian Liu (Sea AI Lab)
Computational EfficiencyTransformerLarge Language ModelVision Language ModelVideoTextBenchmark
🎯 What it does: Utilize sparse topK attention as a draft model, parallel with full attention validation, achieving lossless acceleration in video LLM inference.
Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues
Youngmin Kim (Yonsei University), Youngjae Yu (Yonsei University)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelAuto EncoderVideoTextMultimodality
🎯 What it does: Proposed a large-scale multimodal dialogue dataset VENUS and trained a multimodal language model MARS capable of understanding and generating text along with non-linguistic cues
SPECTRA: Faster Large Language Model Inference with Optimized Internal and External Speculation
Nguyen-Khang Le (Japan Advanced Institute of Science and Technology), Le-Minh Nguyen (Japan Advanced Institute of Science and Technology)
RetrievalComputational EfficiencyTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Proposes SPECTRA, a plug-and-play internal and external speculation mechanism that accelerates LLM inference without requiring additional training.
SpeechFake: A Large-Scale Multilingual Speech Deepfake Dataset Incorporating Cutting-Edge Generation Methods
Wen Huang (Shanghai Jiao Tong University), Yanmin Qian (Shanghai Jiao Tong University)
GenerationData SynthesisBenchmarkAudio
🎯 What it does: Constructed a cross-platform deepfake speech dataset called SpeechFake containing over 3 million synthetic speech samples, spanning more than 3000 hours and covering 46 languages.
SpeechIQ: Speech-Agentic Intelligence Quotient Across Cognitive Levels in Voice Understanding by Large Language Models
Zhen Wan (Kyoto University), Sadao Kurohashi (Kyoto University)
Large Language ModelBenchmarkAudio
🎯 What it does: Propose SpeechIQ, a multidimensional evaluation framework based on Bloom's cognitive hierarchy, to assess the memory, understanding, and application capabilities of large speech understanding models (LLMVoice);
Speed Up Your Code: Progressive Code Acceleration Through Bidirectional Tree Editing
Longhui Zhang (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
OptimizationComputational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose the BITE method, enhancing LLM code acceleration capabilities through bidirectional tree editing and progressive learning.
SPHERE: Unveiling Spatial Blind Spots in Vision-Language Models Through Hierarchical Evaluation
Wenyu Zhang (Agency for Science, Technology and Research), Lu Wang (Agency for Science, Technology and Research)
Explainability and InterpretabilityVision Language ModelImageMultimodalityBenchmark
🎯 What it does: Proposes the SPHERE framework, which systematically evaluates the spatial perception and reasoning capabilities of vision-language models.
SpindleKV: A Novel KV Cache Reduction Method Balancing Both Shallow and Deep Layers
Zicong Tang (Wuhan University), Ping Wang (Wuhan University)
CompressionComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose a KV cache compression method called SpindleKV, which addresses redundancy elimination in both shallow and deep layers, employing attention-weight-based pruning and codebook-based replacement techniques.
Spurious Correlations and Beyond: Understanding and Mitigating Shortcut Learning in SDOH Extraction with Large Language Models
Fardin Ahsan Sakib (George Mason University), Ozlem Uzuner
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextTabularBiomedical DataElectronic Health Records
🎯 What it does: Investigated the 'shortcut learning' phenomenon in large language models (LLMs) during social determinants of health (SDOH) extraction tasks, focusing on misclassification in drug use temporal classification;
Squeezed Attention: Accelerating Long Context Length LLM Inference
Coleman Richard Charles Hooper (UC Berkeley), Amir Gholami (UC Berkeley)
Computational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Perform K-means semantic clustering on keys in a fixed context, using centroids as representatives. During inference, dynamically retrieve important keys by matching queries with centroids, and apply precise attention only to the retrieved keys, significantly reducing KV cache loading and computational costs.
SR-LLM: Rethinking the Structured Representation in Large Language Model
Jiahuan Zhang (Autolab, Westlake University), Kaicheng Yu (Autolab, Westlake University)
Representation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Propose the SR-LLM framework, enhancing LLM reasoning capabilities by converting abstract structural representations (AMR, PST, FOL) into natural language descriptions or incorporating structural data during training.
STaR-SQL: Self-Taught Reasoner for Text-to-SQL
Mingqian He (Zhejiang University), Weiming Lu (Zhejiang University)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextTabularChain-of-Thought
🎯 What it does: Redefines the text-to-SQL task as an inference-driven process, leveraging LLMs to generate step-by-step reasoning (chain-of-thought) and iteratively self-finetune based on the generated reasonable reasoning, while verifying and selecting multiple candidate SQLs using a result-supervised reward model (ORM) at the end of inference.
State-offset Tuning: State-based Parameter-Efficient Fine-Tuning for State Space Models
Wonjun Kang (Seoul National University), Nam Ik Cho (Seoul National University)
Supervised Fine-TuningTextBenchmark
🎯 What it does: Investigated parameter-efficient fine-tuning (PEFT) on state-space models (SSMs) and proposed a novel State-offset Tuning method.
Statistical Deficiency for Task Inclusion Estimation
Loïc Fosse (Orange Research), Pablo Piantanida (International Laboratory on Learning Systems)
Representation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a task inclusion relationship framework based on the theory of statistical deficiency and information sufficiency, aiming to quantify the 'inclusion' degree between natural language processing (NLP) tasks;
Stealing Training Data from Large Language Models in Decentralized Training through Activation Inversion Attack
Chenxi Dai (Huazhong University of Science and Technology), Pan Zhou (Huazhong University of Science and Technology)
Federated LearningSafty and PrivacyAdversarial AttackLarge Language ModelText
🎯 What it does: This paper proposes Activation Inversion Attack (AIA), which steals private training data by exploiting intermediate activation values during decentralized training, demonstrating its feasibility and efficiency in a multi-layer pipeline parallel training framework.
Steering into New Embedding Spaces: Analyzing Cross-Lingual Alignment Induced by Model Interventions in Multilingual Language Models
Anirudh Sundar (Georgia Institute of Technology), Masha Fedzechkina (Georgia Institute of Technology)
RetrievalRepresentation LearningTransformerLarge Language ModelMixture of ExpertsContrastive LearningText
🎯 What it does: Apply the 'finding experts' intervention to multilingual large models to investigate its impact on cross-lingual representation alignment and retrieval performance.
Steering off Course: Reliability Challenges in Steering Language Models
Patrick Queiroz Da Silva (Ohio State University), Sachin Kumar (Ohio State University)
Explainability and InterpretabilityComputational EfficiencyHyperparameter SearchTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper systematically evaluates the robustness of three mainstream steering methods (DoLa, Function Vector (FV), and Task Vector (TV)) across 36 different scales (1.5B–70B) and 14 model families, finding limited improvements or even performance degradation on most models.
Stepwise Reasoning Disruption Attack of LLMs
Jingyu Peng (University of Science and Technology of China), Qi Liu (University of Science and Technology of China)
Adversarial AttackLarge Language ModelTextChain-of-Thought
🎯 What it does: Proposes the SEED attack, which disrupts LLM's reasoning process by injecting subtle errors during multi-step reasoning;
Sticking to the Mean: Detecting Sticky Tokens in Text Embedding Models
Kexin Chen (Zhejiang University), Wenhai Wang (Zhejiang University)
RetrievalAnomaly DetectionRepresentation LearningTransformerText
🎯 What it does: Systematically detect and define 'sticky tokens' in text embedding models, propose an efficient detection method called STD, and identify 868 sticky tokens across 40 checkpoints and 14 model families.
StitchLLM: Serving LLMs, One Block at a Time
Bodun Hu (University of Texas at Austin), Aditya Akella (University of Texas at Austin)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose a dynamic routing framework named StitchLLM, which achieves fine-grained control over inference resources and performance by stitching (stitch) layer blocks between pre-trained LLMs of different sizes.
Stochastic Chameleons: Irrelevant Context Hallucinations Reveal Class-Based (Mis)Generalization in LLMs
Ziling Cheng (Mila), Jackie CK Cheung (Mila)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Investigate the hallucination behavior of large language models when receiving irrelevant context, and propose the 'category-based (mis)generalization' mechanism, conducting a joint analysis of model behavior and internal computations.
STRICTA: Structured Reasoning in Critical Text Assessment for Peer Review and Beyond
Nils Dycke, Iryna Gurevych (Synthetic Rna Biology Technical University Of Darmstadt)
Large Language ModelBiomedical DataReview/Survey PaperChain-of-Thought
🎯 What it does: Proposed the STRICTA framework, decomposing text evaluation into causal structured reasoning steps, and constructed the first expert reasoning steps dataset for biomedical paper evaluation;
StrucText-Eval: Evaluating Large Language Model’s Reasoning Ability in Structure-Rich Text
Zhouhong Gu (Fudan University), Yanghua Xiao (Fudan University)
Data SynthesisTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Proposes an automatically generated evaluation dataset called StrucText-Eval for structured text, designed to assess the ability of large language models in structured text reasoning.
Structural Reasoning Improves Molecular Understanding of LLM
Yunhui Jang (KAIST), Sungsoo Ahn (KAIST)
Drug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningTextGraphChain-of-Thought
🎯 What it does: Propose a molecular structure reasoning (MSR) framework that enhances performance on molecular tasks by reasoning about molecular structure details before large language models (LLMs) generate answers.
Structure-aware Domain Knowledge Injection for Large Language Models
Kai Liu (Zhejiang University), Jieping Ye (Alibaba Cloud)
Domain AdaptationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: In the paper, the authors propose a two-stage structured knowledge injection method (SCPT+SSFT), embedding the hierarchical structure of domain knowledge into large language models to achieve efficient domain adaptation.
STUN: Structured-Then-Unstructured Pruning for Scalable MoE Pruning
Jaeseong Lee (Snowflake AI Research), Yuxiong He (Snowflake AI Research)
Computational EfficiencyLarge Language ModelMixture of ExpertsText
🎯 What it does: Propose a structured-then-unstructured pruning method (STUN), first removing redundant experts via expert-level structured pruning, then performing fine-grained unstructured pruning within the remaining experts;
SubLIME: Subset Selection via Rank Correlation Prediction for Data-Efficient LLM Evaluation
Gayathri Saranathan (Hewlett Packard Labs), Suparna Bhattacharya (Hewlett Packard Labs)
Computational EfficiencyData-Centric LearningLarge Language ModelTextBenchmark
🎯 What it does: Propose the SubLIME framework, which trains a Rank Correlation Prediction (RCP) model to select subsets that maintain the complete ranking list by leveraging a small number of evaluation results from anchor LLMs and intrinsic benchmark metrics (difficulty, quality, distribution divergence), significantly reducing LLM evaluation costs.
Substance over Style: Evaluating Proactive Conversational Coaching Agents
Vidya Srinivas (University of Washington), Tim Althoff (University of Washington)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIText
🎯 What it does: Designed and evaluated five multi-round health coaching dialogue agents, exploring the impact of directive, inquisitive, and guiding interaction styles on user experience.
Subtle Errors in Reasoning: Preference Learning via Error-injected Self-editing
Kaishuai Xu (Hong Kong Polytechnic University), Wenjie Li (Huawei Noah's Ark Lab)
Reinforcement Learning from Human FeedbackPrompt EngineeringTextChain-of-Thought
🎯 What it does: Proposed the RISE framework, which leverages the LLM itself to inject subtle errors into key tokens of correct answers, thereby generating hard negative samples for preference learning;
Subword models struggle with word learning, but surprisal hides it
Bastian Bunzeck (Bielefeld University), Sina Zarrieß (Bielefeld University)
TransformerLarge Language ModelText
🎯 What it does: Investigated the performance of subword and character-level language models in lexical decision tasks, comparing their accuracy in distinguishing words from non-words in contexts of no context, surprisal, and anti-surprisal, and tracking the temporal relationship between word learning and syntactic learning using BLiMP.
SudoLM: Learning Access Control of Parametric Knowledge with Authorization Alignment
Qin Liu (University of California Davis), Muhao Chen (University of California Davis)
Safty and PrivacyLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBiomedical Data
🎯 What it does: Propose the SUDOLM framework, utilizing the SUDO key to achieve authorized access control over LLM parameters;
SURVEYFORGE : On the Outline Heuristics, Memory-Driven Generation, and Multi-dimensional Evaluation for Automated Survey Writing
Xiangchao Yan (Shanghai Artificial Intelligence Laboratory), Bo Zhang (Shanghai Artificial Intelligence Laboratory)
Large Language ModelAgentic AITextReview/Survey PaperBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes the SURVEYFORGE framework to automate the generation of academic review papers, incorporating heuristic outline generation and memory-based content generation, and constructs the SurveyBench evaluation benchmark.
SurveyPilot: an Agentic Framework for Automated Human Opinion Collection from Social Media
Viet Thanh Pham, Gholamreza Haffari (Monash University)
TransformerLarge Language ModelAgentic AITextRetrieval-Augmented Generation
🎯 What it does: Built SURVEYPILOT, an LLM-based finite state machine agent framework for automatically collecting and analyzing real human opinions from social media, replacing traditional surveys and synthetic methods.
SwiLTra-Bench: The Swiss Legal Translation Benchmark
Joel Niklaus, Niko Grupen
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposed the Swiss multilingual legal translation benchmark SwiLTra-Bench, containing over 180,000 parallel translations of legal texts, case law, and press releases, and developed an LLM-based evaluation tool called SwiLTra-Judge.
Symmetrical Visual Contrastive Optimization: Aligning Vision-Language Models with Minimal Contrastive Images
Shengguang Wu (Stanford University), Nick Haber (Stanford University)
OptimizationRepresentation LearningTransformerLarge Language ModelReinforcement LearningVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Symmetric Visual Contrast Optimization (S-VCO) for large-scale vision-language models (VLM) and construction of a Minimum Visual Contrast Dataset (MVC) to enhance alignment and utilization of detailed visual information.
Synergistic Weak-Strong Collaboration by Aligning Preferences
Yizhu Jiao (Microsoft Research), Huaxiu Yao (Microsoft Research)
Computational EfficiencyKnowledge DistillationReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Propose a weak-strong model collaboration framework called COWEST, where the small model first generates a draft and provides domain knowledge, followed by the large model performing reasoning refinement.
Synergizing LLMs with Global Label Propagation for Multimodal Fake News Detection
Shuguo Hu (Inner Mongolia University), Huaiwen Zhang (Inner Mongolia University)
ClassificationGraph Neural NetworkLarge Language ModelVision Language ModelContrastive LearningMultimodality
🎯 What it does: This paper proposes a multi-modal fake news detection framework named GLPN-LLM, which combines a global label propagation network with pseudo labels generated by large language models (LLMs). It leverages high-confidence pseudo labels from LLMs and enhances detection accuracy by fusing multi-modal features through global label propagation and a mask mechanism.
Synergizing Unsupervised Episode Detection with LLMs for Large-Scale News Events
Priyanka Kargupta (University of Illinois at Urbana-Champaign), Jiawei Han (University of Illinois at Urbana-Champaign)
SegmentationTransformerLarge Language ModelText
🎯 What it does: This paper proposes an unsupervised episode detection method that can automatically identify and segment different episode paragraphs from news corpora based on key events.
SynthesizeMe! Inducing Persona-Guided Prompts for Personalized Reward Models in LLMs
Michael J. Ryan (Stanford University), Diyi Yang (Stanford University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This study proposes the SynthesizeMe method, which generates reasoning, synthesizes user personas, and selects the most informative examples from users' sparse binary preference interactions to construct interpretable and transferable personalized reward model prompts; subsequently, it is compared with various existing personalized reward models and LLM-as-a-Judge baselines on the newly constructed PersonalRewardBench.
Synthesizing Post-Training Data for LLMs through Multi-Agent Simulation
Shuo Tang (Shanghai Jiao Tong University), Siheng Chen (Shanghai Jiao Tong University)
Data SynthesisData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextBenchmarkChain-of-Thought
🎯 What it does: Leverage large-scale multi-agent simulations (MATRIX) to generate realistic, controllable social scenarios, and use MATRIX-Gen to generate high-quality instruction-response pairs based on these scenarios for LLM post-training (SFT, DPO, reasoning, and domain-specific tasks).
SYNTHIA: Novel Concept Design with Affordance Composition
Hyeonjeong Ha (University of Illinois Urbana-Champaign), Heng Ji (University of California Los Angeles)
GenerationData SynthesisPrompt EngineeringDiffusion modelContrastive LearningImageText
🎯 What it does: Proposed a text-to-image model framework called SYNTHIA based on text prompts, for generating visually novel and functionally consistent innovative design concepts.
SynWorld: Virtual Scenario Synthesis for Agentic Action Knowledge Refinement
Runnan Fang (Zhejiang University), Huajun Chen (Zhejiang University)
Data SynthesisOptimizationTransformerLarge Language ModelAgentic AIText
🎯 What it does: Built the SynWorld framework, which synthesizes multi-step tool call tasks in virtual scenarios using LLMs, and iteratively optimizes action knowledge through Monte Carlo Tree Search (MCTS);
T-REG: Preference Optimization with Token-Level Reward Regularization
Wenxuan Zhou (Zoom Communications), Tao Meng
Reinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: Proposes a self-generated token-level reward regularization (T-REG) method, incorporating token-level rewards as regularization within preference optimization (e.g., DPO) to achieve finer-grained credit assignment.
T2A-Feedback: Improving Basic Capabilities of Text-to-Audio Generation via Fine-grained AI Feedback
Zehan Wang (Zhejiang University), Zhou Zhao (University of Hong Kong)
GenerationLarge Language ModelReinforcement LearningContrastive LearningBenchmarkAudio
🎯 What it does: Proposes enhancing the fundamental capabilities of text-to-audio (T2A) models through AI feedback learning, and constructs a large-scale AI feedback dataset and a long-story-level evaluation benchmark.
T2I-FactualBench: Benchmarking the Factuality of Text-to-Image Models with Knowledge-Intensive Concepts
Ziwei Huang (Zhejiang University), Leilei Gan (Zhejiang University)
GenerationLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodalityBenchmark
🎯 What it does: Constructed the T2I-FactualBench benchmark, using three-tier knowledge-intensive text-to-image (T2I) tasks to evaluate model factual accuracy, and proposed a multi-round VQA assessment framework for fine-grained evaluation of generated images.
Table-Critic: A Multi-Agent Framework for Collaborative Criticism and Refinement in Table Reasoning
Peiying Yu, Jingjing Wang (Soochow University)
Large Language ModelAgentic AITabularChain-of-Thought
🎯 What it does: Proposed the Table-Critic multi-agent framework, which utilizes four specialized agents—Judge, Critic, Refiner, and Curator—to collaboratively critique and progressively refine the table reasoning process through a self-evolving template tree, addressing error propagation in multi-step reasoning.
TableLoRA: Low-rank Adaptation on Table Structure Understanding for Large Language Models
Xinyi He (Xi'an Jiaotong University), Dongmei Zhang (Microsoft)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTabular
🎯 What it does: Proposed and implemented the TableLoRA module, enabling large language models to better understand and process tabular data under parameter-efficient fine-tuning (PEFT).
TACLR: A Scalable and Efficient Retrieval-based Method for Industrial Product Attribute Value Identification
Yindu Su (Xianyu of Alibaba), Jufeng Chen (Xianyu of Alibaba)
RetrievalTransformerContrastive LearningText
🎯 What it does: Proposed a retrieval-based product attribute value identification method called TACLR
Takin-VC: Expressive Zero-Shot Voice Conversion via Adaptive Hybrid Content Encoding and Enhanced Timbre Modeling
Yang Yuguang (Ximalaya Inc), Jianjun Zhao (Kyushu University)
GenerationTransformerFlow-based ModelAudio
🎯 What it does: This paper proposes a zero-shot emotional voice conversion framework named Takin-VC, which can convert speech into any unseen speaker while preserving the content and emotional features of the source speech.
Taming Language Models for Text-attributed Graph Learning with Decoupled Aggregation
Chuang Zhou (Hong Kong Polytechnic University), Xiao Huang (Hong Kong Polytechnic University)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningTextGraphRetrieval-Augmented Generation
🎯 What it does: In text-attribute graph (TAG) learning, the authors propose the SKETCH framework, which decouples node aggregation from graph convolution. It first retrieves text relevant to the target node through semantic and structural retrieval, then concatenates these retrieved texts into a long context for training a large language model to accomplish the node classification task.
Taming LLMs with Gradient Grouping
Siyuan Li (Zhejiang University), Dan Xu (Hong Kong University of Science and Technology)
OptimizationComputational EfficiencyTextMultimodalityBenchmark
🎯 What it does: This paper proposes an optimizer wrapper named SGG, which utilizes online gradient clustering to group parameters for each layer and computes group-specific learning rate scaling factors, thereby introducing group-level constraints while maintaining parameter-level adaptability; the method can seamlessly integrate with existing optimizers (such as Adam, AdamW, AdaFactor, etc.) and is compatible with parameter-efficient fine-tuning (PEFT) techniques.
TARGA: Targeted Synthetic Data Generation for Practical Reasoning over Structured Data
Xiang Huang (Nanjing University), Yuzhong Qu (Nanjing University)
Data SynthesisTransformerLarge Language ModelPrompt EngineeringGraphTabularRetrieval-Augmented Generation
🎯 What it does: Propose the TARGA framework, which can automatically synthesize highly relevant logical forms and natural language questions for each test problem, serving as a demonstration for unsupervised semantic parsing.
Targeted Syntactic Evaluation for Grammatical Error Correction
Aomi Koyama (Tokyo Metropolitan University), Mamoru Komachi (Hitotsubashi University)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes a grammar error correction (GEC) evaluation paradigm based on Targeted Syntactic Evaluation (TSE), and constructs a minimal pair evaluation set CTSEG corresponding to CEFR-J, enabling fine-grained assessment for each grammatical item.
Task-Specific Information Decomposition for End-to-End Dense Video Captioning
Zhiyue Liu (Guangxi University), Jinyuan Liu (Guangxi University)
GenerationTransformerVision Language ModelContrastive LearningVideoText
🎯 What it does: Propose the DDVC framework, which splits shared event queries into localization queries and description queries, and enhances dense video captioning through contrastive learning and joint label assignment.
TaxoAdapt: Aligning LLM-Based Multidimensional Taxonomy Construction to Evolving Research Corpora
Priyanka Kargupta (University of Illinois at Urbana Champaign), Jiawei Han (University of Illinois at Urbana Champaign)
ClassificationTransformerLarge Language ModelText
🎯 What it does: Proposes the TaxoAdapt framework, which constructs a multi-dimensional, dynamically adaptive literature classification system by leveraging interactive information between LLMs and scientific paper corpora.
TC–RAG: Turing–Complete RAG’s Case study on Medical LLM Systems
Xinke Jiang (Peking University), Yasha Wang (Peking University)
TransformerTextBiomedical DataRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed and implemented TC-RAG, a Turing-complete retrieval-augmented generation framework aimed at enhancing the reliability and accuracy of medical large language models.
TEACH: A Contrastive Knowledge Adaptive Distillation Framework for Classical Chinese Understanding
Yuting Wei (Beijing University of Posts and Telecommunications), Bin Wu (Beijing University of Posts and Telecommunications)
Knowledge DistillationRepresentation LearningSupervised Fine-TuningContrastive LearningTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Designed and implemented the TEACH framework, unifying classical Chinese word sense disambiguation with sentence translation, and constructing a confidence-annotated knowledge base; achieved stepwise reasoning through zero-shot Chain-of-Thought prompting; applied contrastive knowledge self-adaptive distillation to transfer large model reasoning and translation style to small models.
Teaching an Old LLM Secure Coding: Localized Preference Optimization on Distilled Preferences
Mohammad Saqib Hasan (Stony Brook University), Niranjan Balasubramanian (Stony Brook University)
Data SynthesisOptimizationSafty and PrivacyKnowledge DistillationAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: A synthetic dataset named DiSCo, containing unsafe code, secure code, and their secure reasoning pairs, was constructed, along with a localized preference optimization algorithm (LPO) specifically designed to address local differences in security-related code, further enhancing the performance of large language models in generating secure code.
Teaching Text Agents to Learn Sequential Decision Making from Failure
Canasai Kruengkrai (Guardian Robot Project, RIKEN), Koichiro Yoshino (Guardian Robot Project, RIKEN)
Reinforcement LearningTextSequential
🎯 What it does: This paper proposes a failure action-aware target and trajectory perturbation method, which reduces the negative impact of failed actions in self-collected trajectories and leverages unsuccessful trajectories to generate new successful ones, thereby improving learning efficiency.
Teaching Vision-Language Models to Ask: Resolving Ambiguity in Visual Questions
Pu Jian (Institute of Automation Chinese Academy of Sciences), Jiajun Zhang (Institute of Automation Chinese Academy of Sciences)
TransformerSupervised Fine-TuningVision Language ModelMultimodalityBenchmark
🎯 What it does: This paper proposes an interactive clarification framework for addressing ambiguity in visual question answering (VQA) and systematically evaluates the performance of vision-language models (VLMs) in this scenario by constructing the ClearVQA benchmark dataset.
TeamLoRA: Boosting Low-Rank Adaptation with Expert Collaboration and Competition
Tianwei Lin (Zhejiang University), Yueting Zhuang (Zhejiang University)
Computational EfficiencyTransformerMixture of ExpertsTextMultimodalityBenchmark
🎯 What it does: Propose TeamLoRA, a parameter-efficient fine-tuning method that treats LoRA experts as a team, enhancing the effectiveness and efficiency of multi-task learning through collaboration and competition mechanisms.
Temporal reasoning for timeline summarisation in social media
Jiayu Song (Queen Mary University Of London), Maria Liakata (Queen Mary University Of London)
Knowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningTextTime Series
🎯 What it does: This paper investigates integrating temporal reasoning capabilities into large language models (LLMs) by enhancing their performance on social media timeline summarization tasks through knowledge distillation.
Temporal Relation Extraction in Clinical Texts: A Span-based Graph Transformer Approach
Rochana Chaturvedi (University of Illinois Chicago), Barbara Di Eugenio (University of Illinois Chicago)
Graph Neural NetworkTransformerLarge Language ModelTextBiomedical DataElectronic Health Records
🎯 What it does: Propose the GRAPHTREX framework, integrating span-based entity and relation prediction with heterogeneous graph transformation to achieve temporal relation extraction in clinical text.
TeRDy: Temporal Relation Dynamics through Frequency Decomposition for Temporal Knowledge Graph Completion
Ziyang Liu (Tsinghua University), Chaokun Wang (Tsinghua University)
Representation LearningGraphTime Series
🎯 What it does: Propose the TeRDy method, which captures long-term and short-term relationship dynamics in temporal knowledge graphs through frequency decomposition to achieve knowledge graph completion.
TESS 2: A Large-Scale Generalist Diffusion Language Model
Jaesung Tae (Yale University), Arman Cohan (Yale University)
GenerationLarge Language ModelSupervised Fine-TuningReinforcement LearningDiffusion modelText
🎯 What it does: Proposed and trained a large-scale general-purpose instruction-following continuous diffusion language model named TESS 2.
TestNUC: Enhancing Test-Time Computing Approaches and Scaling through Neighboring Unlabeled Data Consistency
Henry Peng Zou (University of Illinois Chicago), Philip S. Yu (University of Illinois Chicago)
ClassificationTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Propose the TestNUC method, which leverages the consistency of nearby unlabelled data during testing to enhance LLM predictions.
TETRIS: Optimal Draft Token Selection for Batch Speculative Decoding
Zhaoxuan Wu (Singapore-Mit Alliance For Research And Technology), Bryan Kian Hsiang Low (Singapore-Mit Alliance For Research And Technology)
OptimizationComputational EfficiencyLarge Language ModelText
🎯 What it does: Proposes a method called TETRIS that dynamically optimizes the selection of draft tokens in multi-request batch inference scenarios to improve the throughput of large language models (LLMs).
Text is All You Need: LLM-enhanced Incremental Social Event Detection
Zitai Qiu (Macquarie University), Jian Yang (Macquarie University)
ClassificationGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: This paper proposes a novel social event detection framework, LSED, which utilizes large language models (LLMs) to normalize, expand, and translate short texts. The processed texts are then encoded with SBERT, time-vectorized, and projected into hyperbolic space for clustering, with the entire process not relying on graph structures.
Text-to-ES Bench: A Comprehensive Benchmark for Converting Natural Language to Elasticsearch Query
Dongge Xue (East China University of Science and Technology), Tong Ruan (East China University of Science and Technology)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposed and implemented the text-to-ES task of converting natural language into Elasticsearch queries, and constructed a corresponding large-scale benchmark dataset;
That doesn’t sound right: Evaluating speech transcription quality in field linguistics corpora
Eric Le Ferrand, Emily Prud’hommeaux
RecognitionData-Centric LearningTransformerTextAudio
🎯 What it does: Evaluate and filter the quality of speech transcriptions in language documentation work to enhance ASR training effectiveness.
That is Unacceptable: the Moral Foundations of Canceling
Soda Marem Lo (Università degli Studi di Torino), Marco Antonio Stranisci (Università degli Studi di Torino)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Constructed and analyzed a manually annotated dataset named CADE, investigating the impact of different moral perspectives on judgments in cancel culture (canceling) comments, and compared the alignment of large language models (LLMs) with human annotators in assessing non-acceptance.
The AI Gap: How Socioeconomic Status Affects Language Technology Interactions
Elisa Bassignana (IT University of Copenhagen Pioneer Center for AI), Dirk Hovy (Bocconi University)
TransformerLarge Language ModelPrompt EngineeringTextTabular
🎯 What it does: Surveyed 1,000 respondents from diverse socioeconomic status (SES) backgrounds to collect data on the frequency, scenarios, and specific tasks related to their use of language technologies and large language models (LLMs), along with 6,482 real interaction prompts. Conducted quantitative and qualitative analyses of prompt length, abstraction level, topic distribution, and human-like expressions.
The Alternative Annotator Test for LLM-as-a-Judge: How to Statistically Justify Replacing Human Annotators with LLMs
Nitay Calderon (Technion), Rotem Dror (University of Haifa)
Data-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextMultimodalityChain-of-Thought
🎯 What it does: Propose a statistical testing-based 'alt-test' method to verify whether large language models (LLMs) can replace human annotations;
The Cross-linguistic Role of Animacy in Grammar Structures
Nina Gregorio (University of Edinburgh), Edoardo Ponti (University of Edinburgh)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper trains a multilingual LLM to perform semantic 'animacy' classification of nouns across 11 languages, and uses this data to explore cross-linguistic correlations between animacy and grammatical structures (numerical marking, word order, semantic roles, relative clauses);
The Distracting Effect: Understanding Irrelevant Passages in RAG
Chen Amiraz (Technology Innovation Institute), Zohar Karnin (Technology Innovation Institute)
RetrievalLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: This paper studies the 'dispersion effect' of irrelevant retrieved passages in Retrieval-Augmented Generation (RAG), proposing a quantifiable measurement method and evaluating its robustness;
The Efficiency vs. Accuracy Trade-off: Optimizing RAG-Enhanced LLM Recommender Systems Using Multi-Head Early Exit
Huixue Zhou (Meta Platforms), Tianlong Chen (University of North Carolina at Chapel Hill)
RetrievalRecommendation SystemComputational EfficiencyGraph Neural NetworkLarge Language ModelTextTabularRetrieval-Augmented Generation
🎯 What it does: Studied an LLM recommendation framework OptiRAG-Rec combining retrieval-augmented generation (RAG) and multi-head early stopping for CTR prediction, balancing accuracy and efficiency.
The Esethu Framework: Reimagining Sustainable Dataset Governance and Curation for Low-Resource Languages
Jenalea Rajab (Lelapa AI), Benjamin Rosman (Lelapa AI)
Data-Centric LearningTransformerSupervised Fine-TuningTextAudio
🎯 What it does: Propose the Esethu framework to build a sustainable, community-driven low-resource language corpus and release the ViXSD isiXhosa speech dataset;
The Essence of Contextual Understanding in Theory of Mind: A Study on Question Answering with Story Characters
Chulun Zhou (Chinese University of Hong Kong), Wai Lam (Chinese University of Hong Kong)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Constructed the CHARTOM-QA benchmark to evaluate the Theory of Mind (ToM) question-answering capabilities of large language models (LLMs) in the global context of novels.
The Harmonic Structure of Information Contours
Eleftheria Tsipidi (ETH Zürich), Mario Giulianelli (ETH Zürich)
Representation LearningTransformerText
🎯 What it does: This paper proposes the harmonic surprise hypothesis, utilizing time-scaled harmonic regression to detect and quantify periodic variations in surprise rates across texts in different languages.
The Hidden Attention of Mamba Models
Ameen Ali Ali, Lior Wolf (Tel Aviv University)
ClassificationSegmentationExplainability and InterpretabilityTransformerImageText
🎯 What it does: This paper reinterprets the Mamba (Selective State Space) model as an implicit self-attention mechanism and proposes a new interpretable method based on this reinterpretation;
The Impact of Auxiliary Patient Data on Automated Chest X-Ray Report Generation and How to Incorporate It
Aaron Nicolson (CSIRO), Bevan Koopman (CSIRO)
GenerationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningImageTextMultimodalityTabularElectronic Health Records
🎯 What it does: In automatic chest X-ray report generation, the authors integrated multimodal patient data from MIMIC-CXR and MIMIC-IV-ED (such as demographics, medications, medical history, etc.) and injected them into a language model in the form of embeddings, achieving more accurate report generation.
The Impact of Token Granularity on the Predictive Power of Language Model Surprisal
Byung-Doh Oh (New York University), William Schuler (Ohio State University)
Large Language ModelText
🎯 What it does: Investigate the impact of subword tokenization granularity on the surprisal prediction of reading difficulty, train language models (LMs) with different vocabulary sizes, and evaluate their predictions of natural reading time and syntax traps.
The Impossibility of Fair LLMs
Jacy Reese Anthis (University of Chicago), Chenhao Tan (University of Chicago)
Explainability and InterpretabilityTransformerLarge Language ModelTextReview/Survey Paper
🎯 What it does: Theoretically analyze the applicability of existing fairness frameworks on general large language models (LLMs), systematically elaborate the inherent challenges of unattainability of fairness in LLMs, and propose future research directions.
The Invisible Hand: Unveiling Provider Bias in Large Language Models for Code Generation
Xiaoyu Zhang (Xi'an Jiaotong University), Yang Liu (Nanyang Technological University)
Data-Centric LearningAI Code AssistantLarge Language ModelPrompt EngineeringText
🎯 What it does: Investigate provider bias in code generation by large language models (LLMs), build and publicly release an automated dataset construction pipeline, and conduct systematic experiments on seven mainstream LLMs;
The Knowledge Microscope: Features as Better Analytical Lenses than Neurons
Yuheng Chen (Chinese Academy of Sciences), Jun Zhao (Chinese Academy of Sciences)
Safty and PrivacyExplainability and InterpretabilityAuto EncoderText
🎯 What it does: This paper decomposes activation vectors in neural networks into finer-grained features using sparse autoencoders (SAE), explores the mechanisms by which these features store factual knowledge in language models, and proposes the FeatureErase method to achieve the deletion of private information.
The Lawyer That Never Thinks: Consistency and Fairness as Keys to Reliable AI
Dana R Alsagheer (University of Houston), Weidong Shi (University of Houston)
Explainability and InterpretabilityTransformerLarge Language ModelTextBenchmark
🎯 What it does: Systematically evaluate the consistency of six mainstream large language models (LLMs) in legal and rational tasks, proposing and applying metrics such as test-retest consistency score (TRCS) and ICC, revealing significant instability in models during repeated reasoning.
The Male CEO and the Female Assistant: Evaluation and Mitigation of Gender Biases in Text-To-Image Generation of Dual Subjects
Yixin Wan (University of California, Los Angeles), Kai-Wei Chang (University of California, Los Angeles)
GenerationLarge Language ModelPrompt EngineeringDiffusion modelImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes the Paired Stereotype Test (PST) framework to systematically evaluate gender bias in text-to-image models under dual-agent scenarios, and designs 1952 prompts based on occupation and organizational power, using Stereotype Score to quantify bias; significant male stereotypes are found on DALLE-3, and a FairCritic method driven by LLM is proposed to eliminate bias through judgment and feedback.