EMNLP 2025 Papers — Page 15
Conference on Empirical Methods in Natural Language Processing · 1809 papers
Sequential-NIAH: A Needle-In-A-Haystack Benchmark for Extracting Sequential Needles from Long Contexts
Yifei Yu (Tencent Youtu Lab), Xing Sun (Tencent Youtu Lab)
Data SynthesisTransformerLarge Language ModelSupervised Fine-TuningTextGraphBenchmark
🎯 What it does: Designed and released the Sequential-NIAH benchmark to evaluate large language models' ability to extract information in temporal or logical order from long texts, constructed three types of sequential information generation pipelines (synthetic-temporal, real-temporal, real-logical), and trained an evaluation model based on Qwen2.5-Instruct-32B to automatically assess answer completeness and sequential consistency.
SERVAL: Surprisingly Effective Zero-Shot Visual Document Retrieval Powered by Large Vision and Language Models
Thong Nguyen (University of Amsterdam), Andrew Yates (Johns Hopkins University)
RetrievalLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Propose a zero-training visual document retrieval method, first generating document descriptions using a vision-language model, then retrieving using a text encoder.
Shallow Focus, Deep Fixes: Enhancing Shallow Layers Vision Attention Sinks to Alleviate Hallucination in LVLMs
Xiaofeng Zhang (Shanghai Jiao Tong University), Jieping Ye (Alibaba Cloud Computing)
Explainability and InterpretabilityComputational EfficiencyTransformerVideoMultimodalityBenchmark
🎯 What it does: Investigated the relationship between attention sink patterns in image tokens and hallucinations in multimodal large language models (MLLMs), and proposed a training-agnostic plug-in method called EVAS to enhance the density of shallow visual attention, thereby reducing the occurrence of hallucinations.
Shared Path: Unraveling Memorization in Multilingual LLMs through Language Similarities
Xiaoyu Luo (Aalborg University), Qiongxiu Li (Aalborg University)
Explainability and InterpretabilityGraph Neural NetworkLarge Language ModelText
🎯 What it does: This paper systematically studies the memorization behavior of multilingual large language models (MLLM), constructs a language similarity graph, quantifies cross-lingual memorization patterns, and analyzes their relationship with the amount of training data.
SHARP: Steering Hallucination in LVLMs via Representation Engineering
Junfei Wu (New Laboratory of Pattern Recognition, State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences), Tieniu Tan (New Laboratory of Pattern Recognition, State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences)
Representation LearningTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes SHARP, an untrained inference-time intervention method based on internal representation analysis, to reduce hallucination generation in large vision-language models.
Sheaf Discovery with Joint Computation Graph Pruning and Flexible Granularity
Lei Yu (University Of Toronto), Gerald Penn (University Of Toronto)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposed the concept of 'sheaf' and automatically discovered self-contained modules capable of executing specific tasks individually within pre-trained language models through the DiscoGP framework.
SheetDesigner: MLLM-Powered Spreadsheet Layout Generation with Rule-Based and Vision-Based Reflection
Qin Chen (Peking University), Dongmei Zhang (Microsoft)
GenerationData SynthesisLarge Language ModelPrompt EngineeringVision Language ModelTabular
🎯 What it does: Proposes a novel spreadsheet layout generation method called SheetDesigner, which utilizes a multimodal large language model (MLLM) to accomplish component positioning and content filling under zero-training and zero-shot conditions.
SHIFT: Selected Helpful Informative Frame for Video-guided Machine Translation
Boyu Guan (Chinese Academy of Sciences), Chengqing Zong (Chinese Academy of Sciences)
Convolutional Neural NetworkTransformerLarge Language ModelVision Language ModelVideoTextMultimodality
🎯 What it does: Propose the SHIFT framework for video-assisted machine translation, dynamically deciding whether to use video, selecting only one keyframe or using text as input, and adapting to multimodal large language models.
Should I Share this Translation? Evaluating Quality Feedback for User Reliance on Machine Translation
Dayeon Ki (University of Maryland), Marine Carpuat (University of Maryland)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: This paper evaluates the impact of four quality feedback mechanisms (error highlighting, LLM explanations, back-translation, and Q&A tables) on monolingual users' decision accuracy and appropriate reliance when deciding whether to share translations in real-world machine translation scenarios;
SilVar: Speech-Driven Multimodal Model for Reasoning Visual Question Answering and Object Localization
Tan-Hanh Pham (Harvard Medical School, Harvard University), Truong-Son Hy (Alabama University)
RecognitionObject DetectionLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelImageTextMultimodalityAudio
🎯 What it does: Propose SilVar, an end-to-end speech-driven multimodal model that uses speech commands for visual question answering and object localization.
Similarity = Value? Consultation Value-Assessment and Alignment for Personalized Search
Weicong Qin (Renmin University of China), Jun Xu (Renmin University of China)
RetrievalRecommendation SystemTransformerContrastive LearningText
🎯 What it does: Studies how to leverage user-AI assistant consultation logs to enhance personalized search effectiveness on e-commerce platforms, proposing a search model called VAPS based on consultation value assessment.
SimMark: A Robust Sentence-Level Similarity-Based Watermarking Algorithm for Large Language Models
Amirhossein Dabiriaghdam (University of British Columbia), Lele Wang (University of British Columbia)
Safty and PrivacyRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Proposed a sentence similarity-based watermarking algorithm called SimMark, which can add detectable secret marks to LLM-generated text without accessing the model's internal structures.
Simple Yet Effective: An Information-Theoretic Approach to Multi-LLM Uncertainty Quantification
Maya Kruse (University of Colorado Anschutz Medical Campus), Yanjun Gao (University of Colorado Anschutz Medical Campus)
Large Language ModelTextBiomedical DataElectronic Health RecordsChain-of-Thought
🎯 What it does: Propose a multi-LLM subset aggregation method called MUSE based on Jensen-Shannon Divergence for uncertainty quantification and calibration
SimpleDoc: Multi‐Modal Document Understanding with Dual‐Cue Page Retrieval and Iterative Refinement
Chelsi Jain (Oregon State University), Huazheng Wang (Oregon State University)
RetrievalComputational EfficiencyTransformerLarge Language ModelVision Language ModelMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the SimpleDoc framework to achieve multi-modal document visual question answering by generating answers through dual-clue retrieval and iterative reasoning.
SimulatorArena: Are User Simulators Reliable Proxies for Multi-Turn Evaluation of AI Assistants?
Yao Dou (Georgia Institute of Technology), Jianfeng Gao (Microsoft)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose the SimulatorArena benchmark to evaluate the reliability of user simulators in multi-turn dialogues, and improve the simulator's performance through fine-grained user attributes.
SimVBG: Simulating Individual Values by Backstory Generation
Bangde Du (Tsinghua University), Yiqun Liu (Tsinghua University)
GenerationData SynthesisTransformerLarge Language ModelText
🎯 What it does: This paper proposes the SimVBG framework, which uses personal background story generation to simulate individual value responses.
Single LLM, Multiple Roles: A Unified Retrieval-Augmented Generation Framework Using Role-Specific Token Optimization
Yutao Zhu (Renmin University of China), Ji-Rong Wen (Renmin University of China)
GenerationRetrievalTransformerSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: Propose a unified retrieval-augmented generation framework called RoleRAG, which enables a single frozen LLM to perform multi-tasking by utilizing role-specific tokens;
SinhalaMMLU: A Comprehensive Benchmark for Evaluating Multitask Language Understanding in Sinhala
Ashmari Pramodya (Nara Institute of Science and Technology), Taro Watanabe (Nara Institute of Science and Technology)
Large Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Constructed the first multi-task language understanding benchmark for the low-resource language Sinhala, SinhalaMMLU, containing 7,044 multiple-choice questions covering 6 domains and 30 subjects, all aligned with Sri Lanka's national curriculum, and evaluated the performance of 26 large language models on this benchmark.
Skeletons Matter: Dynamic Data Augmentation for Text-to-Query
Yuchen Ji (Fudan University), Yanghua Xiao (Fudan University)
Data SynthesisData-Centric LearningLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: This paper unifies the text-to-query (Text-to-Query) task, defines a unified task paradigm, and proposes a dynamic data augmentation framework based on query skeletons to enhance model performance.
Sketch-of-Thought: Efficient LLM Reasoning with Adaptive Cognitive-Inspired Sketching
Simon A. Aytes (KAIST), Sung Ju Hwang (KAIST)
Computational EfficiencyTransformerPrompt EngineeringTextMultimodalityChain-of-Thought
🎯 What it does: Propose the Sketch-of-Thought (SoT) framework, which compresses LLM reasoning processes by leveraging three cognitive-inspired reasoning paradigms (concept chains, chunked symbols, expert dictionaries) and dynamically selects paradigms through a lightweight router
Skip-Thinking: Chunk-wise Chain-of-Thought Distillation Enable Smaller Language Models to Reason Better and Faster
Xiaoshu Chen (National University of Defense Technology), Xinwang Liu (National University of Defense Technology)
Computational EfficiencyKnowledge DistillationTransformerTextChain-of-Thought
🎯 What it does: Designed and implemented block-based chained thinking distillation (CWT) and skip thinking training (STT), enabling small language models to better grasp core reasoning logic and significantly improve reasoning speed on inference tasks.
SliceMoE: Routing Embedding Slices Instead of Tokens for Fine-Grained and Balanced Transformer Scaling
Harshil Vejendla (Rutgers University)
ClassificationExplainability and InterpretabilityComputational EfficiencyTransformerMixture of ExpertsText
🎯 What it does: Propose SliceMoE, achieving finer-grained sparse computation by splitting token hidden vectors into continuous segments and routing each segment to a small number of experts.
SlideCoder: Layout-aware RAG-enhanced Hierarchical Slide Generation from Design
Wenxin Tang (Tsinghua University), Michael R. Lyu (Tsinghua University)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: This work proposes the task of automatically generating editable slides from reference images and designs the SlideCoder framework;
Slim-SC: Thought Pruning for Efficient Scaling with Self-Consistency
Colin Hong (Nanyang Technological University Singapore), Dmitrii Ustiugov (Nanyang Technological University Singapore)
Computational EfficiencyTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: To address the computational and latency overhead caused by excessive redundant reasoning chains in Self-Consistency (SC) inference for large language models, Slim-SC introduces a stepwise pruning mechanism during inference based on the similarity of thought hierarchies. It prematurely terminates chains that are similar to previously generated thoughts, significantly reducing inference time and token consumption without compromising, and even slightly improving, accuracy.
SLlama: Parameter-Efficient Language Model Architecture for Enhanced Linguistic Competence Under Strict Data Constraints
Victor Adelakun Omolaoye (University of Potsdam), Gerard de Melo (University of Potsdam)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes SLlama, a compact variant of Llama-3 trained on the BabyLM dataset with only 10 million tokens without distillation, demonstrating outstanding linguistic capabilities;
SLoW: Select Low-frequency Words! Automatic Dictionary Selection for Translation on Large Language Models
Hongyuan Lu (Chinese University of Hong Kong), Wai Lam (Jilin University)
GenerationComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposed the Automatic Dictionary Selection (ADS) task and developed the SLoW method, which selects dictionaries based on low-frequency words to save tokens and enhance performance in large language model (LLM) translation.
Small Models, Big Results: Achieving Superior Intent Extraction through Decomposition
Danielle Cohen (Google), Anatoly Efros (Google)
ClassificationTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: To extract user intent from mobile and web UI interaction trajectories, a two-phase decomposition method is proposed: first, using Prompt to generate a visual + text summary for each interaction, then aggregating into a complete intent with a fine-tuned model.
SMART: Simulated Students Aligned with Item Response Theory for Question Difficulty Prediction
Alexander Scarlatos (University of Massachusetts Amherst), Andrew Lan (University of Massachusetts Amherst)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Propose the SMART method, which uses LLM to generate and align responses that simulate student answers, thereby predicting the difficulty of open-ended questions;
SmartBench: Is Your LLM Truly a Good Chinese Smartphone Assistant?
Xudong Lu (vivo AI Lab), Fangyuan Li (vivo AI Lab)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Designed and released SmartBench, the first benchmark for Chinese smartphone edge LLM functions, comprising 5 categories of 20 tasks, 2,973 high-quality question-answer pairs, and providing automated evaluation criteria.
SMEC:Rethinking Matryoshka Representation Learning for Retrieval Embedding Compression
Biao Zhang (Taobao & Tmall Group of Alibaba), Bo Zheng (Taobao & Tmall Group of Alibaba)
RetrievalCompressionRepresentation LearningLarge Language ModelSupervised Fine-TuningContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: Propose the Sequential Matryoshka Embedding Compression (SMEC) framework to address gradient fluctuations, information loss, and sample scarcity in high-dimensional embedding compression of large language models.
SNaRe: Domain-aware Data Generation for Low-Resource Event Detection
Tanmay Parekh (University of California Los Angeles), Nanyun Peng (University of California Los Angeles)
GenerationData SynthesisDomain AdaptationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: In low-resource scenarios for event detection, the SNARE framework is proposed, which generates high-quality synthetic data using unlabeled target domain data;
so much depends / upon / a whitespace: Why Whitespace Matters for Poets and LLMs
Sriharsh Bhyravajjula (University of Washington), Maria Antoniak (University of Colorado Boulder)
ClassificationLarge Language ModelVision-Language-Action ModelTextMultimodalityBenchmark
🎯 What it does: This paper introduces WISP fine-grained whitespace symbol classification, systematically analyzing and comparing whitespace usage patterns in 19,000 publicly available poems, 12,000 Reddit unpublished poems, and 51,000 LLM-generated poems.
Social Bias in Multilingual Language Models: A Survey
Lance Calvin Lim Gamboa (University of Birmingham), Mark G. Lee (University of Birmingham)
TransformerReview/Survey Paper
🎯 What it does: Systematic review of social bias in multilingual models, analyzing language diversity, cultural awareness, and evaluation and mitigation methods.
Social Genome: Grounded Social Reasoning Abilities of Multimodal Models
Leena Mathur (Carnegie Mellon University), Louis-Philippe Morency (Massachusetts Institute of Technology)
VideoTextMultimodalityBenchmarkChain-of-ThoughtAudio
🎯 What it does: This paper proposes the SOCIAL GENOME benchmark to evaluate the fine-grained, evidence-based social reasoning capabilities of multi-modal models, and provides human-annotated reasoning trajectories along with a series of semantic and structural evaluation metrics.
Social Good or Scientific Curiosity? Uncovering the Research Framing Behind NLP Artefacts
Eric Chamoun (University of Cambridge), Andreas Vlachos (University of Cambridge)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: This paper proposes an automated framework for extracting knowledge elements from NLP papers and inferring research frameworks through interpretable rules, followed by evaluation in two domains: automatic fact-checking and hate speech detection.
SOCIAL SCAFFOLDS: A Generalization Framework for Social Understanding Tasks
Ritam Dutt (Carnegie Mellon University), Maarten Sap (Carnegie Mellon University)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Propose the SOCIAL SCAFFOLDS framework, which enhances the generalization ability of multi-task dialogue understanding models by utilizing automatically generated social reasoning 'rationales' (intentions, listener responses, presuppositions).
SocioBench: Modeling Human Behavior in Sociological Surveys with Large Language Models
Jia Wang (Shanghai Innovation Institute), Zhongyu Wei (Fudan University)
Large Language ModelPrompt EngineeringTextTabularBenchmark
🎯 What it does: Propose the cross-national and cross-domain SocioBench benchmark, using large language models to simulate human behavior in social surveys and assessing the alignment between the simulated behavior and real survey results.
Socratic-MCTS: Test-Time Visual Reasoning by Asking the Right Questions
David Acuna (NVIDIA), Yejin Choi (NVIDIA)
Computational EfficiencyVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: Proposed a test-time algorithm called Socratic-MCTS, which automatically generates long-chain reasoning processes by performing tree search over subquestion-subanswer pairs, transforming non-reasoning visual language models (VLMs).
SOLAR: Towards Characterizing Subjectivity of Individuals through Modeling Value Conflicts and Trade-offs
Younghun Lee (Purdue University), Dan Goldwasser (Purdue University)
TransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Built a retrieval-augmented generation framework called SOLAR, which infers individuals' subjective judgments by leveraging their past Reddit comments.
SolEval: Benchmarking Large Language Models for Repository-level Solidity Smart Contract Generation
Zhiyuan Peng (Shanghai Jiao Tong University), Yuan Luo (Shanghai Jiao Tong University)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes the SolEval benchmark to evaluate the performance of large language models in repository-level code generation for Ethereum Solidity smart contracts.
SoundMind: RL-Incentivized Logic Reasoning for Audio-Language Models
Xingjian Diao (Dartmouth College), Jiang Gui (Dartmouth College)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextMultimodalityChain-of-ThoughtAudio
🎯 What it does: Propose the SoundMind dataset and the SoundMind-RL algorithm for training and evaluating the logical reasoning capabilities of audio-language models.
SPaRC: A Spatial Pathfinding Reasoning Challenge
Lars Benedikt Kaesberg (University of Göttingen), Bela Gipp (University of Göttingen)
Large Language ModelPrompt EngineeringImageBenchmark
🎯 What it does: Propose the SPaRC dataset, containing 1000 2D grid pathfinding puzzles to test spatial and rule-based reasoning;
SPARK: Simulating the Co-evolution of Stance and Topic Dynamics in Online Discourse with LLM-based Agents
Bowen Zhang (Shenzhen Technology University), Hu Huang (University of Science and Technology of China)
Explainability and InterpretabilityTransformerLarge Language ModelAgentic AIText
🎯 What it does: Proposes a multi-agent simulation framework named SPARK based on large language models (LLMs) to jointly model the co-evolutionary process of topic evolution and stance changes in online discussions.
Sparse Activation Editing for Reliable Instruction Following in Narratives
Runcong Zhao (King's College London), Yulan He (King's College London)
Safty and PrivacyRepresentation LearningLarge Language ModelAuto EncoderTextBenchmark
🎯 What it does: This paper proposes a training-free, sparse activation editing framework called Concise‑SAE based on sparse autoencoders (SAE), aimed at enhancing the instruction-following capability of large language models in narrative environments.
Sparse Autoencoder Features for Classifications and Transferability
Jack Gallifant (Harvard University), Danielle Bitterman
ClassificationDomain AdaptationRepresentation LearningLarge Language ModelAuto EncoderImageTextMultimodality
🎯 What it does: Investigate the interpretable classification and cross-modal transfer of sparse autoencoder (SAE) features in large language models, systematically evaluating the impact of model layers, width, pooling, and binarization on performance.
Sparse Neurons Carry Strong Signals of Question Ambiguity in LLMs
Zhuoxuan Zhang (Brown University), Kaidi Xu (Drexel University)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper reveals and utilizes the ambiguous signals encoded by sparse neurons within LLMs through linear probing and neuron activation intervention, thereby achieving controllable regulation of the model's answering and rejection behaviors.
Spatial Layouts in News Homepages Capture Human Preferences
Alexander Spangher (Stanford University), Ben Welsh (University of California, Berkeley)
ClassificationRecommendation SystemData-Centric LearningTransformerImageText
🎯 What it does: Constructed a large-scale news homepage dataset called NewsHomepages, and used the spatial layout (position, size, image) of the homepage as weak labels to train a Transformer model for predicting editors' prioritization of news items.
SPE Attention: Making Attention Equivariant to Semantic-Preserving Permutation for Code Processing
Chengyu Jiao (Southern University of Science and Technology), Yu Zhang (Southern University of Science and Technology)
AI Code AssistantTransformerTextGraph
🎯 What it does: Propose a novel SPE attention mechanism that enables Transformers to be equivariant to semantic-preserving permutations during code processing;
Spec-VLA: Speculative Decoding for Vision-Language-Action Models with Relaxed Acceptance
Songsheng Wang (University of Macau), Derek F. Wong (University of Macau)
Computational EfficiencyRobotic IntelligenceTransformerVision-Language-Action ModelImageTextBenchmark
🎯 What it does: Introduce the Speculative Decoding (SD) framework into Vision-Language-Action (VLA) models, propose the Spec-VLA scheme to address the high computational cost of VLA inference, and design a relaxed acceptance mechanism based on action token distance to improve parallel generation efficiency.
SPECS: Specificity-Enhanced CLIP-Score for Long Image Caption Evaluation
Xiaofu Chen (MBZUAI), Yova Kementchedjhieva (MBZUAI)
Computational EfficiencyRepresentation LearningData-Centric LearningTransformerVision Language ModelContrastive LearningMultimodality
🎯 What it does: Proposed a no-reference evaluation metric for long image captions called SPECS, based on CLIP and incorporating a training objective specifically tailored for details.
Spectral Scaling Laws in Language Models: emphHow Effectively Do Feed-Forward Networks Use Their Latent Space?
Nandan Kumar Jha (New York University), Brandon Reagen (New York University)
TransformerLarge Language ModelText
🎯 What it does: Investigate the spectral properties of Feed-Forward Network (FFN) width in Transformers regarding utilization of high-dimensional latent spaces, propose a spectral utilization optimization framework for width selection, and conduct hierarchical spectral audits on models such as LLaMA, GPT-2, and nGPT.
Speculating LLMs’ Chinese Training Data Pollution from Their Tokens
Qingjie Zhang (Tsinghua University), Han Qiu (Tsinghua University)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Systematically study and locate polluted Chinese (PoC) tokens in LLM vocabularies, construct PoCDETECT as an automatic detector, and propose POCTRACE method to infer the proportion of contaminated content in training data based on token ID.
Speculative Safety-Aware Decoding
Xuekang Wang (Beijing Institute of Technology), Xueqi Cheng (Chinese Academy of Sciences)
Safty and PrivacyComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Propose a lightweight decoding safety-enhancing method called SSD (Speculative Safety-Aware Decoding), which dynamically switches decoding strategies based on the matching ratio between a small expert model and a large model, thereby improving deep safety alignment while maintaining model usefulness and accelerating inference.
Speculative Streaming: Efficient and Scalable Speculative Decoding with Multi-Stream Attention
Nikhil Bhendawade (Apple), Mahyar Najibi (Apple)
Computational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposes a single-model Speculative Streaming method that achieves accelerated inference by implementing non-autoregressive speculation and verification within the target model through multi-stream attention.
SpecVLM: Enhancing Speculative Decoding of Video LLMs via Verifier-Guided Token Pruning
Yicheng Ji (Zhejiang University), Huan Li (Zhejiang University)
Computational EfficiencyTransformerVision Language ModelVideo
🎯 What it does: Accelerate video reasoning by introducing validator-guided phased video token pruning in video LLMs using untrained speculative decoding.
Speech Discrete Tokens or Continuous Features? A Comparative Analysis for Spoken Language Understanding in SpeechLLMs
Dingdong Wang (Chinese University of Hong Kong), Helen M. Meng (Chinese University of Hong Kong)
ClassificationRecognitionTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningAudio
🎯 What it does: Under the SpeechLLM framework, this paper systematically compares the performance of discrete speech tokens and continuous speech features across six SLU tasks (ASR, ST, KS, IC, ER, PR) based on a unified SSL extraction and LLM fine-tuning process for fair experiments.
Speech Vecalign: an Embedding-based Method for Aligning Parallel Speech Documents
Chutong Meng (George Mason University), Philipp Koehn (Johns Hopkins University)
Representation LearningAudio
🎯 What it does: Propose Speech Vecalign, a parallel speech file alignment method based on paragraph embeddings;
SPIRIT: Patching Speech Language Models against Jailbreak Attacks
Amirbek Djanibekov (Mohamed bin Zayed University of Artificial Intelligence), Nils Lukas (Mohamed bin Zayed University of Artificial Intelligence)
Safty and PrivacyExplainability and InterpretabilityAdversarial AttackTransformerTextAudio
🎯 What it does: Studied and evaluated the vulnerability of speech language models under white-box adversarial attacks, proposing a post-hoc defense method based on activation patches, bias addition, and neuron pruning.
Split-Merge: Scalable and Memory-Efficient Merging of Expert LLMs
Sruthi Gorantla (Amazon AGI), Mahdi Namazifar (Amazon AGI)
Computational EfficiencyTransformerMixture of ExpertsText
🎯 What it does: Propose a Split-Merge zero-shot model merging framework that combines multi-domain expert LLMs into a single model without requiring further training.
Spontaneous Giving and Calculated Greed in Language Models
Yuxuan Li (Carnegie Mellon University), Hirokazu Shirado (Carnegie Mellon University)
Large Language ModelTextFinance RelatedChain-of-Thought
🎯 What it does: This paper compares LLMs with and without reasoning capabilities in various economic games (such as the public goods game, prisoner's dilemma, etc.), studying the impact of reasoning techniques on cooperative and punitive behaviors.
SportReason: Evaluating Retrieval-Augmented Reasoning across Tables and Text for Sports Question Answering
Kaiyue Feng (New York University Shanghai), Chen Zhao (Center for Data Science New York University)
RetrievalLarge Language ModelAgentic AIMultimodalityTabularBenchmarkRetrieval-Augmented Generation
🎯 What it does: Introduce the SPORTREASON benchmark to evaluate the effectiveness of retrieval-augmented reasoning across tables and text in sports question answering.
SQLWOZ: A Realistic Task-Oriented Dialogue Dataset with SQL-Based Dialogue State Representation for Complex User Requirements
Heng-Da Xu (Beijing Institute of Technology), Heyan Huang (Beijing Institute of Technology)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Constructed SQLWOZ—a multi-domain task-oriented dialog dataset that represents dialog states using SQL statements, capable of reflecting multi-value, exclusion, priority, and conditional constraint slots, and conducted dialog state tracking, generation, and end-to-end task-oriented dialog experiments on this dataset.
SQUAB: Evaluating LLM robustness to Ambiguous and Unanswerable Questions in Semantic Parsing
Simone Papicchio (Politecnico di Torino), Paolo Papotti (EURECOM)
Explainability and InterpretabilityData-Centric LearningLarge Language ModelPrompt EngineeringTabularBenchmark
🎯 What it does: The paper proposes SQUAB, a framework for automatically generating a semantic parsing test set containing ambiguous and unanswerable questions.
SQUiD: Synthesizing Relational Databases from Unstructured Text
Mushtari Sadia (University of Michigan), Amrita Roy Chowdhury (University of Michigan)
Data SynthesisTransformerLarge Language ModelPrompt EngineeringTextTabularChain-of-Thought
🎯 What it does: This study proposes the SQUiD framework to automatically synthesize structured relational databases from unstructured text, encompassing four stages: schema design, value extraction, table filling, and SQL generation;
SSA-COMET: Do LLMs Outperform Learned Metrics in Evaluating MT for Under-Resourced African Languages?
Senyu Li (Mila Quebec AI Institute), David Ifeoluwa Adelani (Mila Quebec AI Institute)
Data-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Propose the SSA-MTE dataset and SSA-COMET, SSA-COMET-QE evaluation models, focusing on machine translation (MT) quality assessment for African Amazon languages.
SSA: Semantic Contamination of LLM-Driven Fake News Detection
Cheng Xu (University College Dublin), Tahar Kechadi (University College Dublin)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposed an evaluation framework named Semantic Sensitivity Amplifier (SSA) to detect semantic-level benchmark data contamination (BDC) in large language models for fake news detection tasks, assessing model robustness through entity replacement perturbation and the SSA Factor metric.
Stand on The Shoulders of Giants: Building JailExpert from Previous Attack Experience
Xi Wang (National University of Defense and Technology), Wangbaosheng
Adversarial AttackText
🎯 What it does: This study proposes an experience-based automated jailbreak framework called JailExpert, which enhances attack success rate and efficiency by leveraging structured experience, semantic drift grouping, and dynamic updates.
STARE at the Structure: Steering ICL Exemplar Selection with Structural Alignment
Jiaqian Li (Nanyang Technological University), Wenya Wang (Nanyang Technological University)
RetrievalTransformerContrastive LearningText
🎯 What it does: Propose a framework called STARE for ICL example retrieval in semantic parsing, integrating a structure-aware retriever with a pluggable mid-level injection (MLI) module;
STARQA: A Question Answering Dataset for Complex Analytical Reasoning over Structured Databases
Mounica Maddela (Bloomberg), Mausam (Indian Institute of Technology Delhi)
AI Code AssistantLarge Language ModelTabularBenchmark
🎯 What it does: Proposed the STARQA public dataset and designed the TEXT2SQLCODE method that integrates SQL with Python, evaluating the performance of LLMs in complex analytical reasoning question answering.
START: Self-taught Reasoner with Tools
Chengpeng Li (University of Science and Technology of China), Dayiheng Liu (Qwen Team, Alibaba Inc)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Design the START framework, which activates the tool usage capability of LLMs through untrained Hint-infer, and further enhances the model's utilization of tools via self-supervised Hint-RFT (rejection sampling), thereby achieving more accurate reasoning and a more efficient thinking process.
Static or Dynamic: Towards Query-Adaptive Token Selection for Video Question Answering
Yumeng Shi (Nanyang Technological University), Wenya Wang (Nanyang Technological University)
RetrievalTransformerVision Language ModelVideoTextMultimodality
🎯 What it does: Proposed a query-adaptive Token selection framework named EXPLORE-THEN-SELECT to dynamically balance static and dynamic information in video question answering under limited Token budget.
Static Word Embeddings for Sentence Semantic Representation
Takashi Wada (ZOZO Research), Yuki Saito (ZOZO Research)
Knowledge DistillationRepresentation LearningTransformerContrastive LearningText
🎯 What it does: Propose a new static word embedding (SWE) method that extracts word vectors using a pre-trained sentence Transformer, then fine-tunes them via sentence-level PCA (with the first few principal components removed) and knowledge distillation or contrastive learning, resulting in a lightweight model where word vectors can be directly averaged into sentence vectors;
Statistical and Neural Methods for Hawaiian Orthography Modernization
Jaden Kapali (University of Hawai'i at Hilo), Winston Wu (University of Hawai'i at Hilo)
Data SynthesisRecurrent Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Investigated the conversion between traditional (TS) and modern (MS) Hawaiian orthography, comparing the effectiveness of statistical n-gram, Seq2Seq neural networks, and large language models (LLM).
STEER-BENCH: A Benchmark for Evaluating the Steerability of Large Language Models
Kai Chen (University of Southern California), Kristina Lerman (University of Southern California)
Large Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposed the STEER-BENCH benchmark to evaluate the steerability of large language models in adapting to different community norms and perspectives.
Steering Language Models in Multi-Token Generation: A Case Study on Tense and Aspect
Alina Klerings (University of Mannheim), Simone Paolo Ponzetto (University of Mannheim)
GenerationExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Investigate the internal representations of tense and aspect in large language models (Llama-3.1-8B and Qwen-2.5-7B), and achieve syntactic control over multi-word generation through linear concept-driven interventions.
Steering LLM Reasoning Through Bias-Only Adaptation
Viacheslav Sinii, Daniil Gavrilov (T Tech)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper inserts trainable steering vectors in each layer and freezes the original weights, achieving performance comparable to full model fine-tuning on mathematical reasoning tasks using reinforcement learning.
Step Guided Reasoning: Improving Mathematical Reasoning using Guidance Generation and Step Reasoning
Lang Cao (Huawei Technologies Co., Ltd.), Yitong Li (Huawei Technologies Co., Ltd.)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes an untrained step-by-step guided reasoning (SGR) framework, which leverages LLMs to self-generate 'step guides' and 'step answers' during the reasoning process, thereby significantly enhancing mathematical reasoning capabilities.
Step-level Verifier-guided Hybrid Test-Time Scaling for Large Language Models
Kaiyan Chang (Northeastern University), JingBo Zhu
Computational EfficiencyTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes a step-level verifier-guided hybrid test-time scaling method to enhance the reasoning performance of large language models.
StepER: Step-wise Knowledge Distillation for Enhancing Reasoning Ability in Multi-Step Retrieval-Augmented Language Models
Kyumin Lee (POSTECH), Hwanjo Yu (POSTECH)
RetrievalKnowledge DistillationTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the STEPER framework, a multi-step retrieval-augmented language model based on progressive knowledge distillation, which enhances the multi-step reasoning ability of small models by utilizing step-by-step training data generated by a teacher model and combining difficulty-adaptive training.
StepSearch: Igniting LLMs Search Ability via Step-Wise Proximal Policy Optimization
Xuhui Zheng (ImVision Innovation), Yichao Wu (ImVision Innovation)
GenerationRetrievalTransformerLarge Language ModelReinforcement LearningTextRetrieval-Augmented Generation
🎯 What it does: Propose STEPSEARCH, a reinforcement learning framework based on token-level stepping rewards for training LLMs to reason through search interactions in multi-hop question answering.
Stepwise Informativeness Search for Improving LLM Reasoning
Siyuan Wang (University of Southern California), Xiang Ren (University of Southern California)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose a method that, during reasoning, uses a tree search framework to guide large language models (LLMs) to actively reference previously underutilized steps when generating multi-step reasoning processes, reducing redundancy between steps and generating more accurate and concise reasoning paths.
Stepwise Reasoning Checkpoint Analysis: A Test Time Scaling Method to Enhance LLMs’ Reasoning
Zezhong Wang (Chinese University of Hong Kong), Kam-Fai Wong (Huawei Noah's Ark Lab)
Explainability and InterpretabilityComputational EfficiencyTransformerTextChain-of-Thought
🎯 What it does: Proposes a framework SRCA that inserts checkpoints between reasoning steps, improving the mathematical reasoning accuracy of LLMs through Answer Clustering Search (ACS) and Checkpoint Candidate Amplification (CCA).
Sticker-TTS: Learn to Utilize Historical Experience with a Sticker-driven Test-Time Scaling Framework
Jie Chen (Gaoling School of Artificial Intelligence Renmin University of China), Ji-Rong Wen (Gaoling School of Artificial Intelligence Renmin University of China)
Knowledge DistillationLarge Language ModelTextChain-of-Thought
🎯 What it does: Propose the Sticker-TTS framework, which employs three collaborative modules—Sticker Extractor, Sticker Modifier, and Sticker Utilizer—to perform multi-round iterative reasoning during inference by leveraging historical experience.
Stimulate the Critical Thinking of LLMs via Debiasing Discussion
Ruiyu Xiao (Harbin Institute of Technology), Ting Liu (Harbin Institute of Technology)
Explainability and InterpretabilityData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIText
🎯 What it does: Proposed and implemented a multi-stage training framework MDTA that leverages multi-perspective discussions and truth alignment to enhance the critical thinking and discussion capabilities of large language models (LLMs).
Stop Looking for “Important Tokens” in Multimodal Language Models: Duplication Matters More
Zichen Wen (Shanghai Jiao Tong University), Linfeng Zhang (Shanghai Jiao Tong University)
Computational EfficiencyTransformerVision Language ModelImageVideoTextMultimodality
🎯 What it does: Propose a training-agnostic, plug-and-play acceleration method called DART based on the visual token duplication rate, which directly retains tokens with the largest differences from the Pivot token, thereby significantly compressing the number of visual tokens.
STRICT: Stress-Test of Rendering Image Containing Text
Tianyu Zhang (University of Montreal), Suyuchen Wang (University of Montreal)
GenerationData SynthesisPrompt EngineeringDiffusion modelTextBenchmark
🎯 What it does: Proposed the STRICT benchmark for systematically evaluating the ability of diffusion models to render readable text in images;
Stronger Baselines for Retrieval-Augmented Generation with Long-Context Language Models
Alex Laitenberger (Stanford University), Nelson F. Liu (Stanford University)
GenerationRetrievalTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper evaluates whether single-stage retrieval-generation (RAG) can still rival multi-stage retrieval-augmented generation pipelines in long-context language models, particularly through comparisons with ReadAgent and RAPTOR in multi-round retrieval and text compression.
Structure-Conditional Minimum Bayes Risk Decoding
Bryan Eikema (University of Amsterdam), Mario Giulianelli (University College London)
GenerationData-Centric LearningLarge Language ModelSupervised Fine-TuningTextSequentialBenchmark
🎯 What it does: Propose a structure-aware Minimum Bayes Risk (MBR) decoding method and conduct systematic evaluation on dialogue and instruction-following tasks.
Structured Moral Reasoning in Language Models: A Value-Grounded Evaluation Framework
Mohna Chakraborty (University of Michigan), David Jurgens (University of Michigan)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationTransformerSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: A structured prompting framework based on values and ethical theories was constructed to improve the accuracy and consistency of language models in moral reasoning tasks, and the moral capabilities of large models were transferred to small models through reasoning-level distillation.
Structured Preference Optimization for Vision-Language Long-Horizon Task Planning
Xiwen Liang (Sun Yat-sen University Shenzhen Campus), Xiaodan Liang (Sun Yat-sen University Shenzhen Campus)
OptimizationReinforcement Learning from Human FeedbackVision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose the Structured Preference Optimization (SPO) framework, which enhances the reasoning and decision-making quality of vision-language models in long-horizon task planning through structured preference evaluation and progressive learning.
Structuring Radiology Reports: Challenging LLMs with Lightweight Models
Johannes Moll (Stanford University), Jean-Benoit Delbrouck (Stanford University)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringBiomedical DataElectronic Health Records
🎯 What it does: The study uses a lightweight task-specific model (<300M parameters) to convert chest X-ray reports into standard structured templates.
Studying Rhetorically Ambiguous Questions
Oghenevovwe Ikumariegbe (University of Arizona), Ellen Riloff (University of Arizona)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Investigated questions that can be interpreted as rhetorical or informative in different contexts, constructing and annotating a new dataset named SRAQ;
Studying the Role of Input-Neighbor Overlap in Retrieval-Augmented Language Models Training Efficiency
Ehsan Doostmohammadi (Linköping University), Marco Kuhlmann (Linköping University)
RetrievalComputational EfficiencyTransformerSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: Investigate the impact of query-retrieval context overlap degree on training efficiency and test performance in retrieval-augmented language models, and accelerate model activation through manually synthesized overlapping contexts.
SUA: Stealthy Multimodal Large Language Model Unlearning Attack
Xianren Zhang (Pennsylvania State University), Suhang Wang (Pennsylvania State University)
Adversarial AttackTransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: Investigate whether multimodal large language models (MLLMs) truly forget sensitive information after unlearning, and propose a general adversarial attack framework named SUA that uses noise to activate the model and recover the forgotten knowledge.
Subjective Behaviors and Preferences in LLM: Language of Browsing
Sai Sundaresan (Adobe Research), N Anushka (Adobe Research)
Recommendation SystemData-Centric LearningTransformerLarge Language ModelReinforcement LearningContrastive LearningSequentialBenchmark
🎯 What it does: A language model was developed to better capture users' subjective behaviors and preferences based on 'browse language' (i.e., page visit sequences) generated by users on websites or applications;
Subtle Risks, Critical Failures: A Framework for Diagnosing Physical Safety of LLMs for Embodied Decision Making
Yejin Son (Yonsei University), Chan Young Park (University of Washington)
Safty and PrivacyExplainability and InterpretabilityRobotic IntelligenceLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes the SAFEL framework to systematically evaluate the physical safety of large language models (LLMs) in embodied decision-making, and constructs the EMBODYGUARD benchmark, which includes two types of scenarios: malicious instructions (mal) and situational risks (sit); meanwhile, 13 LLMs are evaluated, revealing their weaknesses in modules such as safe rejection, goal understanding, transfer modeling, and action ordering;
SUE: Sparsity-based Uncertainty Estimation via Sparse Dictionary Learning
Tamás Ficsor (University of Szeged), Gábor Berend (University of Szeged)
ClassificationExplainability and InterpretabilityTransformerText
🎯 What it does: Proposed a model uncertainty estimation method called SUE based on sparse dictionary learning, and evaluated it on various NLP tasks.
Summarizing Speech: A Comprehensive Survey
Fabian Retkowski (Karlsruhe Institute of Technology), Alexander Waibel (Karlsruhe Institute of Technology)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextMultimodalityReview/Survey PaperRetrieval-Augmented GenerationAudio
🎯 What it does: This paper provides a systematic review of speech summarization (Speech Summarization, SSum) research since 2020, organizing the scope of research, challenges, data resources, evaluation methods, and major technical approaches, and conducting quantitative comparisons and future outlooks based on these.
Superficial Self-Improved Reasoners Benefit from Model Merging
Xiangchi Yuan (Georgia Institute of Technology), Wenke Lee (Georgia Institute of Technology)
Knowledge DistillationLarge Language ModelText
🎯 What it does: Studied the 'shallow self-improvement reasoner' phenomenon during the self-improvement process of LLMs, and proposed an iterative model merging (IMM) method to alleviate model collapse and enhance the model's generalization capability.
Superpose Task-specific Features for Model Merging
Haiquan Qiu (Tsinghua University), Quanming Yao (Tsinghua University)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerText
🎯 What it does: Propose a model fusion method based on the linear representation hypothesis—STF, which achieves multi-task model fusion without training by performing singular value decomposition (SVD) on the task matrix and solving a linear system in the singular space.