EMNLP 2025 Papers — Page 14
Conference on Empirical Methods in Natural Language Processing · 1809 papers
ResFormer: All-Time Reservoir Memory for Long Sequence Classification
Hongbo Liu (Stevens Institute of Technology), Jia Xu (Stevens Institute of Technology)
ClassificationTransformerSupervised Fine-TuningTextMultimodality
🎯 What it does: Proposed a ResFormer architecture combining Reservoir Computing with Transformer for efficiently processing text sequences of arbitrary length.
Reshaping Representation Space to Balance the Safety and Over-rejection in Large Audio Language Models
Hao Yang (Monash University), Gholamreza Haffari (Monash University)
Representation LearningTransformerLarge Language ModelSupervised Fine-TuningTextAudio
🎯 What it does: Propose an unsupervised Reshaping Representation Space (RRS) fine-tuning strategy to enhance the safety alignment of large audio-language models (LALMs), improving safety while maintaining a low rate of over-rejection.
ReSo: A Reward-driven Self-organizing LLM-based Multi-Agent System for Reasoning Tasks
Heng Zhou (University of Science and Technology of China), Lei Bai (Shanghai Artificial Intelligence Laboratory)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Propose the ReSo framework, constructing a reward-driven self-organizing multi-agent system that automatically decomposes task graphs, dynamically selects agents, and collaborates to optimize complex reasoning tasks.
Resource-Rational Noisy-Channel Language Processing: Testing the Effect of Algorithmic Constraints on Inferences
Thomas Hikaru Clark (MIT), Roger P. Levy (MIT)
GenerationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Built and implemented a resource-efficient noise channel language understanding model based on generative models and Sequential Monte Carlo (SMC) inference, evaluating the impact of computational resources such as the number of particles and rejuvenation on human-like reasoning.
ReSURE: Regularizing Supervision Unreliability for Multi-turn Dialogue Fine-tuning
Yiming Du (Chinese University of Hong Kong), Fei Tan (East China Normal University)
OptimizationComputational EfficiencyTransformerSupervised Fine-TuningText
🎯 What it does: Proposes the ReSURE framework, dynamically downscaling unreliable supervision signals in multi-round dialogue fine-tuning to enhance model robustness and response quality.
RETAIL: Towards Real-world Travel Planning for Large Language Models
Bin Deng (Beihang University), Yunhong Wang (Beihang University)
Recommendation SystemTransformerLarge Language ModelAgentic AITextMultimodality
🎯 What it does: This work proposes a tourism planning study for real-world scenarios, constructing the RETAIL dataset that includes explicit and implicit user needs, environmental perception, and detailed itineraries. It also introduces the theme-guided multi-agent framework (TGMA), enhancing decision support for implicit user needs and the ability to generate comprehensive travel plans.
Rethinking Backdoor Detection Evaluation for Language Models
Jun Yan (University of Southern California), Robin Jia (University of Southern California)
Anomaly DetectionTransformerText
🎯 What it does: This paper studies the robustness of backdoor detection methods for language models under different training intensities and proposes an adversarial evaluation protocol based on training intensity;
Rethinking Cross-Subject Data Splitting for Brain-to-Text Decoding
Congchi Yin (Nanjing University of Aeronautics and Astronautics), Piji Li (JD.com)
Graph Neural NetworkTransformerLarge Language ModelBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper re-examines the data partitioning methods for cross-subject brain-text decoding and proposes a correct partitioning criterion to eliminate data leakage.
Rethinking Text-based Protein Understanding: Retrieval or LLM?
Juntong Wu (Peking University), Yu Li (International Digital Economy Academy)
Protein Structure PredictionLarge Language ModelTextBiomedical DataBenchmarkRetrieval-Augmented Generation
🎯 What it does: Analyze and reconstruct the text-protein understanding benchmark, proposing a retrieval-enhanced protein modeling framework
RethinkMCTS: Refining Erroneous Thoughts in Monte Carlo Tree Search for Code Generation
Qingyao Li (Shanghai Jiao Tong University), Weinan Zhang (Shanghai Jiao Tong University)
AI Code AssistantTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposed a framework called RETHINKMCTS for code generation that employs Monte Carlo Tree Search (MCTS) to explore and continuously refine ideas.
Retracing the Past: LLMs Emit Training Data When They Get Lost
Myeongseob Ko (Virginia Tech), Ruoxi Jia (Virginia Tech)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes Confusion-Inducing Attacks (CIA) based on model uncertainty maximization and mismatched Supervised Fine-Tuning (SFT) for aligned models, to extract memorized training data from large language models (LLMs) without requiring training set information.
Retrieval Enhanced Feedback via In-context Neural Error-book
Jongyeop Hyun (Chung-Ang University), Bumsoo Kim (Chung-Ang University)
RetrievalExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringVision Language ModelMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the REFINE framework, constructing a Neural Error-book to provide structured error feedback for multi-modal LLMs; converting errors into retrievable feedback through a teacher-student mechanism.
Retrieval over Classification: Integrating Relation Semantics for Multimodal Relation Extraction
Lei Hei (Northeastern University), Feiliang Ren (Northeastern University)
ClassificationRetrievalTransformerLarge Language ModelVision Language ModelContrastive LearningMultimodality
🎯 What it does: Propose a ROC framework that transforms multimodal relation extraction into a semantic retrieval task, leveraging entity types, positions, and natural language relation descriptions to achieve more refined semantic modeling.
Retrieval-Augmented Generation with Estimation of Source Reliability
Jeongyeon Hwang (Pohang University of Science and Technology), Jungseul Ok (Pohang University of Science and Technology)
GenerationRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Proposed a multi-source retrieval-augmented generation framework RA-RAG, which can automatically estimate the reliability of information sources and utilize them in retrieval and answer generation, significantly improving factual accuracy.
Retrieval-augmented GUI Agents with Generative Guidelines
Ran Xu (Emory University), Dong Yu (Tencent AI Lab)
TransformerSupervised Fine-TuningAgentic AIVision Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: We propose a lightweight visual language model adapter, RAG-GUI, which can generate task-related guidance for GUI agents during inference by leveraging online tutorials;
Retrieving Support to Rank Answers in Open-Domain Question Answering
Zeyu Zhang (Amazon AGI), Thuy Vu (Amazon AGI)
RetrievalTransformerTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes a complete QA architecture for open-domain question answering that verifies answers by retrieving and leveraging supporting texts.
Revealing and Mitigating the Challenge of Detecting Character Knowledge Errors in LLM Role-Playing
Wenyuan Zhang (Institute of Information Engineering, Chinese Academy of Sciences), Tingwen Liu (Institute of Information Engineering, Chinese Academy of Sciences)
Anomaly DetectionTransformerLarge Language ModelAgentic AITextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Studied the knowledge error detection task of large language models (LLM) in role-playing (RPA), constructed the RoleKE-Bench benchmark dataset, and proposed a multi-agent reasoning method called S2RD based on self-recollection and self-doubt to enhance error detection capabilities.
Reverse Prompt Engineering: A Zero-Shot, Genetic Algorithm Approach to Language Model Inversion
Hanqing Li (Northwestern University), Diego Klabjan (Northwestern University)
OptimizationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose an untrained reverse prompting engineering (RPE) framework under black-box, zero-shot, and output-restricted scenarios, leveraging the language model's inherent reasoning ability and genetic algorithm iteration optimization to recover the original prompt based on only 5 output texts.
ReviewRL: Towards Automated Scientific Review with RL
Sihang Zeng (University of Washington), Bowen Zhou (Tsinghua University)
Reinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningTextReview/Survey PaperRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed the ReviewRL framework for automatically generating scientific paper review reports and providing paper scores.
Revisiting LLM Value Probing Strategies: Are They Robust and Expressive?
Siqi Shen (University of Michigan), Rada Mihalcea (University of Michigan)
Data SynthesisExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Systematic comparison of three common scoring methods for LLM value detection (token likelihood, sequence perplexity, text generation), and introduction of two evaluation tasks: demographic context inference and value-action consistency testing.
REVIVING YOUR MNEME: Predicting The Side Effects of LLM Unlearning and Fine-Tuning via Sparse Model Diffing
Aly M. Kassem (Mila Quebec AI Institute), Golnoosh Farnadi (Mila Quebec AI Institute)
Safty and PrivacyExplainability and InterpretabilityLarge Language ModelAuto EncoderText
🎯 What it does: Propose the MNEME framework, which utilizes sparse model diffing to detect unexpected side effects caused by LLM fine-tuning or unlearning.
Reward Model Perspectives: Whose Opinions Do Reward Models Reward?
Elle (University of Oxford)
Reinforcement Learning from Human FeedbackTransformerReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper investigates the perspectives and social biases of reward models (RM) in preference learning, quantifying RM's alignment with opinions of different demographic groups, stereotypes, and whether these alignments can be adjusted through contextual prompts;
Reward-Shifted Speculative Sampling Is An Efficient Test-Time Weak-to-Strong Aligner
Bolian Li (Purdue University), Ruqi Zhang (Purdue University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose a reward-offset speculative sampling (SSS) algorithm that achieves weak-to-strong alignment during inference using an unaligned target model and an aligned small draft model, enabling high-quality generation without external reward models.
Reward-Weighted Sampling: Enhancing Non-Autoregressive Characteristics in Masked Diffusion LLMs
Daehoon Gwak (KAIST AI), Jaegul Choo (KAIST AI)
GenerationTransformerLarge Language ModelDiffusion modelTextBenchmark
🎯 What it does: This paper proposes a decoding strategy called Reward-Weighted Sampling (RWS) to enhance the non-autoregressive properties of Masked Diffusion Models (MDMs) and improve generation quality.
RewardDS: Privacy-Preserving Fine-Tuning for Large Language Models via Reward Driven Data Synthesis
Jianwei Wang (South China University of Technology), Ziqian Zeng (South China University of Technology)
Data SynthesisSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextFinance Related
🎯 What it does: Propose the RewardDS framework, which generates high-quality synthetic data through reward-driven data synthesis and self-optimization for privacy-preserving LLM fine-tuning.
Rewarding the Unlikely: Lifting GRPO Beyond Distribution Sharpening
Andre Wang He (Carnegie Mellon University), Sean Welleck (Carnegie Mellon University)
AI Code AssistantReinforcement LearningTextBenchmark
🎯 What it does: Investigate the diversity decay problem of GRPO in formal theorem proving, discovering that its reinforcement of high-probability solutions leads to distribution narrowing, and propose the 'Unlikeliness Reward' method to alleviate this bias by providing greater rewards to rare correct solutions; meanwhile, explore the impact of PPO iteration counts on bias, propose an improved GRPO training scheme, and implement an open-source pipeline on the Lean prover.
reWordBench: Benchmarking and Improving the Robustness of Reward Models with Transformed Inputs
Zhaofeng Wu (Cancer), Marjan Ghazvininejad (Virgo)
Reinforcement Learning from Human FeedbackLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes the reWordBench benchmark to evaluate the robustness of reward models under input transformations that preserve semantics or rankings, and improves robustness by incorporating sentence synonym regularization during training;
RICO: Improving Accuracy and Completeness in Image Recaptioning via Visual Reconstruction
Yuchi Wang (Peking University), Xu Sun (Kuaishou Technology)
GenerationTransformerLarge Language ModelReinforcement LearningVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Developed a vision-reconstruction-based image recaptioning framework called RICO, which generates reference images using a text-to-image model and iteratively refines captions by comparing the original image with the reconstructed image via a large model; to reduce computational costs, a single-step version called RICO-Flash was also introduced through Direct Preference Optimization (DPO) learning.
RiTTA: Modeling Event Relations in Text-to-Audio Generation
Yuhang He (Microsoft Research), Vibhav Vineet (Microsoft Research)
GenerationTransformerPrompt EngineeringDiffusion modelFlow-based ModelTextBenchmarkRetrieval-Augmented GenerationAudio
🎯 What it does: This paper addresses the insufficient modeling of audio event relationships in text-to-audio (TTA) generation by constructing a relation corpus containing 11 relationships and 25 audio event classes, and proposing a multi-stage relation-aware evaluation metric. Subsequently, a Gated Prompt Tuning strategy is introduced, significantly enhancing the relationship modeling performance of existing TTA models while keeping the parameter increment extremely low.
RJE: A Retrieval-Judgment-Exploration Framework for Efficient Knowledge Graph Question Answering with LLMs
Can Lin (University Of Science And Technology Of China), Wangqiu Zhou (Hefei University Of Technology)
RetrievalComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextGraphBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the Retrieval-Judgment-Exploration (RJE) framework, divided into three stages: retrieval, judgment, and exploration, combining small-scale LLMs for knowledge graph question answering.
RLAE: Reinforcement Learning-Assisted Ensemble for LLMs
Yuqian Fu (Institute of Automation Chinese Academy of Sciences), Dongbin Zhao (Institute of Automation Chinese Academy of Sciences)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposed a reinforcement learning-based dynamic weight allocation framework named RLAE to adjust the contribution ratios of individual models in real-time during span-level integration of multi-model LLMs.
Robust Adaptation of Large Multimodal Models for Retrieval Augmented Hateful Meme Detection
Jingbiao Mei (University of Cambridge), Bill Byrne (University of Cambridge)
ClassificationDomain AdaptationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose the RA-HMD framework, which fine-tunes large multimodal models (LMM) using a two-stage LoRA + contrastive learning approach, combined with a retrieval-enhanced KNN classifier, to improve the accuracy, robustness, and interpretability of hate meme detection.
Robust Native Language Identification through Agentic Decomposition
Ahmet Yavuz Uluslu (University of Zurich), Rico Sennrich (University of Zurich)
ClassificationTransformerLarge Language ModelAgentic AIPrompt EngineeringText
🎯 What it does: Propose a NLI process based on agent splitting, where specialized agents first identify various linguistic features such as syntax, lexicon, and idioms, and then a coordinator agent synthesizes this evidence to make the final native language judgment.
RoDEval: A Robust Word Sense Disambiguation Evaluation Framework for Large Language Models
Luyang Zhang (Qilu University of Technology), Wenpeng Lu (Qilu University of Technology)
ClassificationTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Propose the RoDEval framework for systematic evaluation of the word sense disambiguation (WSD) capabilities of large language models (LLMs).
RoT: Enhancing Table Reasoning with Iterative Row-Wise Traversals
Xuanliang Zhang (Harbin Institute of Technology), Wanxiang Che (Harbin Institute of Technology)
Large Language ModelPrompt EngineeringTabularBenchmarkChain-of-Thought
🎯 What it does: This paper proposes the Row-of-Thought (ROT) method, which uses an iterative row-level traversal approach for table reasoning, avoiding hallucination issues in traditional Long CoT, and requires no training, directly applicable to non-reasoning large language models.
Route Sparse Autoencoder to Interpret Large Language Models
Wei Shi (University Of Science And Technology Of China), Xiangnan He (University Of Science And Technology Of China)
Explainability and InterpretabilityLarge Language ModelAuto EncoderText
🎯 What it does: Propose a novel sparse autoencoder framework named RouteSAE, which utilizes a routing mechanism to extract interpretable features from multi-layer activations.
Router-Tuning: A Simple and Effective Approach for Dynamic Depth
Shwai He (University of Maryland), Dong Yu (Tencent AI Lab)
Computational EfficiencyTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Propose Router-Tuning, a method that achieves Mixture of Depths (dynamic depth) by fine-tuning only a few router parameters.
RPDR: A Round-trip Prediction-Based Data Augmentation Framework for Long-Tail Question Answering
Yiming Zhang (Zhejiang University), Chen Zhao (NYU Shanghai)
Data SynthesisRetrievalTransformerLarge Language ModelContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Propose the RPDR framework, which enhances long-tail QA performance by generating synthetic long-tail question-answer pairs, using round-trip prediction to filter easy-to-learn samples, and training a dense retriever.
RRInf: Efficient Influence Function Estimation via Ridge Regression for Large Language Models and Text-to-Image Diffusion Models
Zhuozhuo Tu (University of Sydney), Yuxuan Du (Nanyang Technological University)
Explainability and InterpretabilityComputational EfficiencyTransformerImageText
🎯 What it does: Propose the RRInf method, which reconstructs influence functions using ridge regression and solves them with normalized stochastic gradients, achieving efficient influence function estimation on LLMs and diffusion models.
RTE-GMoE: A Model-agnostic Approach for Relation Triplet Extraction via Graph-based Mixture-of-Expert Mutual Learning
Aziguli Wulamu (University of Science and Technology Beijing), Bowen Xing (University of Science and Technology Beijing)
Graph Neural NetworkLarge Language ModelMixture of ExpertsTextBenchmark
🎯 What it does: Propose the RTE-GMoE framework, leveraging graph-based Mixture of Experts (MoE) with mutual learning to enhance relation triple extraction.
RTQA : Recursive Thinking for Complex Temporal Knowledge Graph Question Answering with Large Language Models
Zhaoyan Gong (Zhejiang University), Wen Zhang (Zhejiang University)
TransformerLarge Language ModelTextGraphRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Developed a training-free, plug-and-play framework called RTQA, which recursively decomposes complex temporal knowledge graph question answering problems and performs bottom-up reasoning using large language models (LLMs) and temporal knowledge graphs.
RuCCoD: Towards Automated ICD Coding in Russian
Alexandr Nesterov, Elena Tutubalina (AIRI Sber AI Skoltech Sber AI Lab)
ClassificationRecognitionTransformerLarge Language ModelSupervised Fine-TuningTextElectronic Health RecordsRetrieval-Augmented Generation
🎯 What it does: Constructed the first Russian ICD-10 coding dataset RuCCoD, proposed an encoding pipeline based on entity recognition + linking, and further used this pipeline to automatically re-annotate large EHR collections to enhance diagnostic prediction performance.
Rule Discovery for Natural Language Inference Data Generation Using Out-of-Distribution Detection
Juyoung Han (Kangwon National University), Changki Lee (Kangwon National University)
GenerationData SynthesisData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: By combining OOD detection with BERT clustering, new NLI sentence transformation rules are discovered and generated from the SNLI dataset, and model performance is improved by utilizing LLM+CoT to generate training samples.
s1: Simple test-time scaling
Niklas Muennighoff (Stanford University), Tatsunori Hashimoto (Stanford University)
Computational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: For reasoning tasks on language models, an extremely simple test-time scaling method called 'budget forcing' is proposed, along with the construction of a high-quality reasoning dataset s1K containing only 1,000 questions. Supervised fine-tuning of Qwen2.5-32B-Instruct on this dataset yields the model s1-32B, which can achieve a performance improvement trend with increased test-time computation by controlling the length of the thinking process (maximum token count or multiple additions of 'Wait').
s3: You Don’t Need That Much Data to Train a Search Agent via RL
Pengcheng Jiang (University of Illinois Urbana Champaign), Jiawei Han (University of Illinois Urbana Champaign)
RetrievalOptimizationTransformerLarge Language ModelReinforcement LearningAgentic AITextBiomedical DataRetrieval-Augmented Generation
🎯 What it does: Train an LLM agent that only searches, optimized via reinforcement learning to enhance the generation quality of RAG systems
SABER: Uncovering Vulnerabilities in Safety Alignment via Cross-Layer Residual Connection
Maithili Joshi (Indian Institute of Technology Delhi), Tanmoy Chakraborty (Indian Institute of Technology Delhi)
Safty and PrivacyTransformerLarge Language ModelTextBenchmark
🎯 What it does: The study proposes a white-box jailbreak method called SABER, which bypasses the safety alignment mechanism of LLMs by inserting cross-layer residual connections between two layers.
SAE-SSV: Supervised Steering in Sparse Representation Spaces for Reliable Control of Language Models
Zirui He (NJIT), Mengnan Du (NJIT)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningAuto EncoderText
🎯 What it does: Proposes a supervised steering framework called SAE-SSV based on sparse autoencoders, designed to regulate the behavior of large language models within a sparse interpretable subspace.
SAEs Are Good for Steering – If You Select the Right Features
Dana Arad (Technion Israel Institute of Technology), Yonatan Belinkov
Explainability and InterpretabilityRepresentation LearningAdversarial AttackAuto EncoderTextBenchmark
🎯 What it does: This paper studies how to select the most effective steering features by analyzing input/output scores of sparse autoencoder (SAE) features to achieve unsupervised control over language model generation.
SAFE-SQL: Self-Augmented In-Context Learning with Fine-grained Example Selection for Text-to-SQL
Jimin Lee (Chung-Ang University), Hwanhee Lee (New York University)
AI Code AssistantTransformerLarge Language ModelText
🎯 What it does: Proposes the SAFE-SQL framework, which leverages LLM to self-generate and filter high-quality examples for unsupervised text-to-SQL generation.
SAFE: Schema-Driven Approximate Distance Join for Efficient Knowledge Graph Querying
Sangoh Lee (Pohang University of Science and Technology), Wook-Shin Han (Pohang University of Science and Technology)
Computational EfficiencyLarge Language ModelGraphRetrieval-Augmented Generation
🎯 What it does: Propose the SAFE framework, which leverages the schema graph of knowledge graphs (KGs) for query graph generation and approximate distance Join (ADJ), enhancing the performance of question answering when combining large language models (LLMs) with KGs.
SafeKey: Amplifying Aha-Moment Insights for Safety Reasoning
Kaiwen Zhou (University Of California Santa Cruz), Xin Eric Wang (University Of California Santa Cruz)
Safty and PrivacyLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed the SafeKey framework, which enhances the security of large reasoning models against jailbreak attacks by reinforcing safe 'aha-moment'.
SAFENUDGE: Safeguarding Large Language Models in Real-time with Tunable Safety-Performance Trade-offs
Joao Fonseca (New York University), Julia Stoyanovich (New York University)
Safty and PrivacyLarge Language ModelText
🎯 What it does: Propose SAFENUDGE, a real-time safety protection mechanism for text generation, which suppresses jailbreak attacks by controlling text generation and guidance.
SafeScientist: Enhancing AI Scientist Safety for Risk-Aware Scientific Discovery
Kunlun Zhu (University of Illinois Urbana Champaign), Jiaxuan You (University of Illinois Urbana Champaign)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose the SafeScientist framework and SciSafetyBench benchmark to enhance the safety and ethics of AI scientists in high-risk scientific tasks.
SAKI-RAG: Mitigating Context Fragmentation in Long-Document RAG via Sentence-level Attention Knowledge Integration
Wenyu Tao (South China University of Technology), Xiangmin Xu (South China University of Technology)
RetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Propose the SAKI-RAG framework, utilizing sentence-level attention linking and dual-axis retrieval technology to address the context fragmentation problem in long document RAG.
Sali4Vid: Saliency-Aware Video Reweighting and Adaptive Caption Retrieval for Dense Video Captioning
MinJu Jeon (Hanyang University), Dong-Jin Kim (Hanyang University)
GenerationRetrievalTransformerVision Language ModelVideoTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Proposed the Sali4Vid framework, combining significance video reweighting and semantic adaptive caption retrieval to achieve dense video caption generation and event localization.
Same evaluation, more tokens: On the effect of input length for machine translation evaluation using Large Language Models
Tobias Domhan (Amazon AGI), Dawei Zhu (Amazon AGI)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: This paper investigates the length dependency of large language models in long-text machine translation evaluation and proposes a length-invariant evaluation scheme.
Same Question, Different Words: A Latent Adversarial Framework for Prompt Robustness
Tingchen Fu (Renmin University of China), Fazl Barez (University of Oxford)
Adversarial AttackReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Proposed the Latent Adversarial Paraphrasing (LAP) framework, which uses dual-cycle adversarial training to search for the worst semantically equivalent paraphrases in the latent space and optimize the robustness of large language models (LLMs).
SAMULE: Self-Learning Agents Enhanced by Multi-level Reflection
Yubin Ge (AWS AI Labs), Yi Zhang (AWS AI Labs)
Large Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextBenchmark
🎯 What it does: Designed and implemented the SAMULE framework, leveraging multi-level reflective synthetic training backtracking to enable self-learning and continuous improvement of LLM agents in both non-interactive and interactive scenarios.
SAND: Boosting LLM Agents with Self-Taught Action Deliberation
Yu Xia (University of California San Diego), Julian McAuley (University of California San Diego)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIText
🎯 What it does: Proposed and implemented the Self-taught Action Deliberation (SAND) framework, enabling LLM agents to proactively deliberate, evaluate, and select the optimal action among candidate actions before making decisions.
SATBench: Benchmarking LLMs’ Logical Reasoning via Automated Puzzle Generation from SAT Formulas
Anjiang Wei (Stanford University), Alex Aiken (Stanford University)
TransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: This paper introduces SATBench, a benchmark that evaluates the logical reasoning capabilities of large language models (LLMs) by automatically converting Boolean satisfiability (SAT) formulas into natural language puzzles.
SATER: A Self-Aware and Token-Efficient Approach to Routing and Cascading
Yuanzhe Shen (Fudan University), Xuanjing Huang (Fudan University)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: The study proposes the SATER two-stage training method for achieving efficient routing between small and large models.
Scalable and Culturally Specific Stereotype Dataset Construction via Human-LLM Collaboration
Weicheng Ma (Georgia Institute of Technology), Soroush Vosoughi (Dartmouth College)
Data-Centric LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Leveraged large language models to generate potential stereotypes, which were then verified and instantiated by local annotators from various countries, constructing a cross-national, cross-cultural Spanish stereotype dataset named EspanStereo.
Scalable Data Synthesis through Human-like Cognitive Imitation and Data Recombination
Zhongyi Ye (University of Science and Technology of China), Enhong Chen (University of Science and Technology of China)
Data SynthesisLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes the Cognitive Composition Synthesis (CCS) framework, which generates large-scale, scalable pre-training data through three steps: knowledge-skill mapping, solution refinement, and quality filtering, utilizing multi-source data.
Scaling Low-Resource MT via Synthetic Data Generation with LLMs
Ona de Gibert (University of Helsinki), Jörg Tiedemann (University of Cambridge)
Data SynthesisTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Leveraged GPT-4o for large-scale forward translation to construct a document-level synthetic parallel corpus containing 7 low-resource languages, and extended it to 147 language pairs through a pivot mechanism; conducted automatic and manual quality assessments of the corpus, trained and fine-tuned multiple MT models on it, and ultimately released the public synthetic corpus SynOPUS.
Scaling Rich Style-Prompted Text-to-Speech Datasets
Anuj Diwan (University of Texas at Austin), Eunsol Choi (New York University)
GenerationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextAudio
🎯 What it does: This paper constructs a large-scale, richly labeled speech dataset called ParaSpeechCaps, proposes two automatic annotation processes, and subsequently trains a style-driven text-to-speech model on this dataset;
Scaling Up Temporal Domain Generalization via Temporal Experts Averaging
Aoming Liu (Boston University), Bryan A. Plummer (Boston University)
Domain AdaptationConvolutional Neural NetworkTransformerSupervised Fine-TuningMixture of ExpertsImageTextBenchmark
🎯 What it does: Propose a framework called Temporal Experts Averaging (TEA) that uses full-model weight averaging to address the Temporal Domain Generalization (TDG) problem.
SciEvent: Benchmarking Multi-domain Scientific Event Extraction
Bofu Dong (Indiana University), Ming Jiang (Indiana University)
Large Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Propose the SciEvent benchmark, which annotates multi-domain scientific abstracts using a unified event extraction (EE) scheme, covering five disciplines.
SciNLP: A Domain-Specific Benchmark for Full-Text Scientific Entity and Relation Extraction in NLP
Decheng Duan (Nanjing University of Science and Technology), Chengzhi Zhang (Nanjing University of Science and Technology)
RecognitionTransformerLarge Language ModelTextGraphBenchmark
🎯 What it does: Constructed SciNLP, the first fine-grained entity and relation extraction benchmark dataset tailored for the NLP domain, covering full-text documents.
SciRIFF: A Resource to Enhance Language Model Instruction-Following over Scientific Literature
David Wadden (Allen Institute for AI), Arman Cohan (Yale University)
ClassificationLarge Language ModelSupervised Fine-TuningTextBiomedical DataBenchmark
🎯 What it does: This paper constructs SciRIFF, a science literature-oriented instruction-following dataset containing 137K expert-written instructions, covering 54 tasks across five capabilities: information extraction, summarization, question answering, statement verification, and classification.
SCRIBE: Structured Chain Reasoning for Interactive Behaviour Explanations using Tool Calling
Fares Fawzi (École Polytechnique Fédérale de Lausanne), Tanja Käser (École Polytechnique Fédérale de Lausanne)
Data SynthesisExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningAgentic AITextChain-of-Thought
🎯 What it does: Propose the SCRIBE framework, enabling small language models to generate personalized, verifiable student behavior feedback explanations through multi-hop tool calls;
SDGO: Self-Discrimination-Guided Optimization for Consistent Safety in Large Language Models
Peng Ding (Nanjing University), Shujian Huang (Nanjing University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes a self-reinforcement learning framework (SDGO) based on the discriminative capabilities of LLMs, achieving enhanced generation safety during training by using the model's safety discrimination results during the generation phase as rewards;
SEA: Supervised Embedding Alignment for Token-Level Visual-Textual Integration in MLLMs
Yuanyang Yin (MoE Key Lab of BIPC University of Science and Technology of China), Feng Zhao (MoE Key Lab of BIPC University of Science and Technology of China)
Representation LearningTransformerSupervised Fine-TuningVision Language ModelContrastive LearningMultimodalityBenchmark
🎯 What it does: Enhancing the alignment quality between vision and text in the LLM embedding space through the SEA method, which introduces token-level supervision during the pre-training phase of multi-modal large language models.
SEAL: Structure and Element Aware Learning Improves Long Structured Document Retrieval
Xinhao Huang (HKUST (Guangzhou)), Zeyi Wen (HKUST (Guangzhou))
RetrievalTransformerSupervised Fine-TuningContrastive LearningTextBenchmark
🎯 What it does: Propose the SEAL framework, leveraging structure-aware contrastive learning and element-level alignment to enhance long-structured document retrieval; and release the StructDocRetrieval dataset with structural annotations.
Search Wisely: Mitigating Sub-optimal Agentic Searches By Reducing Uncertainty
Peilin Wu (University of Texas at Dallas), Zhiyu Chen (University of Texas at Dallas)
RetrievalOptimizationTransformerSupervised Fine-TuningReinforcement LearningAgentic AITextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper first systematically defines and quantifies two types of suboptimal search behaviors in Agentic RAG systems—over-search and under-search—and conducts fine-grained step-level analysis on multi-hop QA datasets, proving these behaviors are widespread and correlated with the model's uncertainty about its own knowledge boundaries. Subsequently, β-GRPO is proposed, an improved policy gradient method that incorporates a confidence threshold for search queries into the reinforcement learning reward, training the model to favor high-confidence search decisions when generating correct answers.
Search-o1: Agentic Search-Enhanced Large Reasoning Models
Xiaoxi Li (Renmin University of China), Zhicheng Dou (Renmin University of China)
RetrievalTransformerLarge Language ModelAgentic AITextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes the Search-o1 framework, combining agentic retrieval-augmented generation with Reason-in-Documents, enabling large reasoning models (LRM) to actively retrieve and precisely utilize external knowledge during the reasoning process, thereby improving reasoning coherence and accuracy.
Searching for the Most Human-like Emergent Language
Brendon Boldt (Carnegie Mellon University), David R. Mortensen (Carnegie Mellon University)
GenerationHyperparameter SearchTextBenchmark
🎯 What it does: Generate emergent languages with the highest similarity to human language statistics by performing hyperparameter search in a signaling game environment, and evaluate them on XferBench deep transfer learning tasks.
Section-Level Simplification of Biomedical Abstracts
Jan Bakker (University of Amsterdam), Jaap Kamps (University of Amsterdam)
ClassificationRecurrent Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningBiomedical Data
🎯 What it does: Proposed a two-step method: first align PLS sentences with technical abstracts and mark corresponding sections, then classify unaligned sentences into sections; finally generated the COCHRANE-SECTIONS dataset, achieving section-level correspondence between technical abstracts and PLS.
Seeing Culture: A Benchmark for Visual Reasoning and Grounding
Burak Satar (Singapore Management University), Chong-Wah Ngo (Singapore Management University)
SegmentationVision Language ModelMultimodalityBenchmark
🎯 What it does: Proposed the Seeing Culture Benchmark (SCB), evaluating cultural reasoning abilities of multimodal vision-language models through two-stage visual question answering and cultural object segmentation
Seeing is Believing, but How Much? A Comprehensive Analysis of Verbalized Calibration in Vision-Language Models
Weihao Xuan (University of Tokyo), Naoto Yokoya (University of Tokyo)
Explainability and InterpretabilityPrompt EngineeringVision Language ModelImageVideoTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This paper evaluates the calibration of visual-language models (VLM) in expressing confidence in natural language, and conducts systematic experiments on three input modalities and four task domains.
Seeing More, Saying More: Lightweight Language Experts are Dynamic Video Token Compressors
Xiangchen Wang (Southern University of Science and Technology), Feng Zheng (Southern University of Science and Technology)
CompressionComputational EfficiencyTransformerLarge Language ModelVision Language ModelVideoTextMultimodality
🎯 What it does: Propose a lightweight, language expert-driven dynamic video token compression method called LangDC.
Seeing the Same Story Differently: Framing‐Divergent Event Coreference for Computational Framing Analysis
Jin Zhao (Brandeis University), Nianwen Xue (Brandeis University)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: Proposed and implemented the FRECO task, which identifies event pairs presenting the same real-world event with contrasting frameworks in different news texts.
Seeing Through Words, Speaking Through Pixels: Deep Representational Alignment Between Vision and Language Models
Zoe Wanying He (University of California San Diego), Meenakshi Khosla (University of California San Diego)
Representation LearningVision Language ModelMultimodality
🎯 What it does: Systematically analyze cross-modal alignment in large monomodal vision and language models under hierarchical structures, semantic perturbations, human preferences, and aggregation.
Select-Then-Decompose: From Empirical Analysis to Adaptive Selection Strategy for Task Decomposition in Large Language Models
Shuodi Liu (Beijing University of Posts and Telecommunications), Zhaofeng He (Beijing University of Posts and Telecommunications)
OptimizationComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This study systematically organizes methods of task decomposition in large language models (LLMs), proposes six classification dimensions, selects five representative methods, conducts large-scale experiments across different decomposition methods, task types, and model scales, summarizes three core insights, and based on these insights designs an adaptive 'Select-Then-Decompose (S&D)' closed-loop strategy. The strategy dynamically selects the most suitable decomposition method and suppresses errors through a validation module, ultimately achieving a Pareto optimal balance between performance and cost.
Selective Preference Optimization via Token-Level Reward Function Estimation
Kailai Yang (University of Manchester), Sophia Ananiadou (University of Manchester)
OptimizationReinforcement Learning from Human FeedbackTransformerReinforcement LearningContrastive LearningTextBenchmark
🎯 What it does: Proposes Selective Preference Optimization (SePO), a framework that uses Direct Preference Optimization (DPO) to estimate token-level reward functions and efficiently selects key tokens for alignment based on these estimates;
Self-Adjust Softmax
Chuanyang Zheng (Chinese University of Hong Kong), Yu Li (Chinese University of Hong Kong)
OptimizationTransformerText
🎯 What it does: Propose Self-Adjust Softmax (SA-Softmax), replacing traditional softmax with z·softmax(z) or its normalized variants in the attention calculation of Transformers to enhance gradient propagation and eliminate gradient vanishing.
Self-Augmented Preference Alignment for Sycophancy Reduction in LLMs
Chien Hung Chen, Hsin-Hsi Chen (Academia Sinica)
Data-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Investigate the 'flattery bias' in open-source large language models during the alignment process, and propose an alignment method that reduces this bias by having the model evaluate user suggestions.
Self-Critique and Refinement for Faithful Natural Language Explanations
Yingming Wang (University of Copenhagen), Pepa Atanasova (University of Copenhagen)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Enhance the credibility of natural language explanations (NLE) generated by models through an SR-NLE framework that realizes self-criticism and iterative improvement within a single LLM.
SelfRACG: Enabling LLMs to Self-Express and Retrieve for Code Generation
Qian Dong (Tsinghua University), Shaoping Ma (Tsinghua University)
GenerationRetrievalAI Code AssistantLarge Language ModelContrastive LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the SelfRACG framework, enabling LLMs to self-express information needs, thereby enhancing generation quality in retrieval-augmented code generation;
Semantic Inversion, Identical Replies: Revisiting Negation Blindness in Large Language Models
Jinsung Kim (Korea University), Heuiseok Lim (Korea University)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Investigate the phenomenon of negation blindness in large language models when handling negation queries and propose a verification framework.
Semantic Networks Extracted from Students’ Think-Aloud Data are Correlated with Students’ Learning Performance
Pingjing Yang (University of Illinois at Urbana-Champaign), Jana Diesner (University of Illinois at Urbana-Champaign)
TransformerLarge Language ModelTextGraph
🎯 What it does: Construct a semantic network integrating textbook materials and students' think-aloud data, and investigate the correlation between network structure and student academic performance.
SemCSE: Semantic Contrastive Sentence Embeddings Using LLM-Generated Summaries For Scientific Abstracts
Marc Felix Brinner (Bielefeld University), Sina Zarrieß (Bielefeld University)
Representation LearningTransformerLarge Language ModelContrastive LearningTextBenchmark
🎯 What it does: Perform unsupervised semantic embedding learning on scientific papers using LLM-generated scientific abstracts to build the SemCSE model;
SEMMA: A Semantic Aware Knowledge Graph Foundation Model
Arvindh Arun (Institute for AI University of Stuttgart), Steffen Staab
Representation LearningGraph Neural NetworkLarge Language ModelTextGraph
🎯 What it does: Propose the SEMMA model, which achieves zero-shot link prediction by leveraging dual-modal information generated by LLMs, including relational text semantics and graph structure.
SeMob: Semantic Synthesis for Dynamic Urban Mobility Prediction
Runfei Chen (Tongji University), Wei Huang (Tongji University)
Autonomous DrivingGraph Neural NetworkTransformerLarge Language ModelAgentic AITextMultimodalityTime Series
🎯 What it does: Propose the SeMob framework, which automatically extracts event texts using multi-agent LLMs and fuses them with spatiotemporal traffic data to enhance event-driven urban mobility prediction.
SemVink: Advancing VLMs’ Semantic Understanding of Optical Illusions via Visual Global Thinking
Sifan Li (University of California, Merced), Yiwei Wang (University of California, Merced)
RecognitionVision Language ModelDiffusion modelImageBenchmark
🎯 What it does: Proposed the HC-Bench benchmark to evaluate VLM's ability to identify visual illusions and hidden content, and significantly improved recognition accuracy through low-resolution scaling (SemVink).
SenDetEX: Sentence-Level AI-Generated Text Detection for Human-AI Hybrid Content via Style and Context Fusion
Lei Jiang, Xiaolong Zheng (Institute Of Automation Chinese Academy Of Sciences)
ClassificationData SynthesisAnomaly DetectionConvolutional Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Propose the AutoFill-Refine synthesis strategy to construct a sentence-level human-AI hybrid text dataset, and introduce the SenDetEX framework to integrate sentence style and context for AI-generated text detection.
SensorLLM: Aligning Large Language Models with Motion Sensors for Human Activity Recognition
Zechen Li (University of New South Wales), Flora D. Salim (University of New South Wales)
RecognitionTransformerLarge Language ModelSupervised Fine-TuningTime Series
🎯 What it does: Proposed SensorLLM, a two-stage framework that aligns motion sensor time series with natural language and utilizes large language models (LLM) for human activity recognition (HAR).
Sentence Smith: Controllable Edits for Evaluating Text Embeddings
Hongji Li (University of Zurich), Juri Opitz (University of Zurich)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextGraphBenchmark
🎯 What it does: This paper proposes the SENTENCESMITH framework, which utilizes semantic graphs (AMR) for controllable and transparent transformation of sentence meaning, and applies this framework to generate challenging negative samples for text embedding models, thereby achieving fine-grained model evaluation.
Separate the Wheat from the Chaff: Winnowing Down Divergent Views in Retrieval Augmented Generation
Song Wang (University of Virginia), Jundong Li (University of Virginia)
GenerationRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Propose the WinnowRAG framework, which groups retrieved documents into perspective-based clusters through query-aware clustering, and iteratively filters noise documents while preserving valuable information via a multi-agent winnowing stage with a Critic LLM, ultimately improving the accuracy of retrieval-augmented generation.
SEPS: A Separability Measure for Robust Unlearning in LLMs
Wonje Jeung (Yonsei University), Albert No (Yonsei University)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: This paper investigates the robustness of machine unlearning in large language models, proposing a new evaluation metric called SEPS to measure the model's ability to separately handle forgetting (forget) and retaining (retain) queries within the same prompt; after discovering that existing methods fail under mixed prompts, we propose a Mixed Prompt (MP) training strategy, developing two approaches, MP-ME (target-free) and MP-IDK (targeted), to enhance the model's forgetting and retaining performance in mixed query scenarios.
seqBench: A Tunable Benchmark to Quantify Sequential Reasoning Limits of LLMs
Mohammad Ramezanali (Salesforce AI), Paolo Santi (Massachusetts Institute of Technology)
TransformerLarge Language ModelPrompt EngineeringTextSequentialBenchmark
🎯 What it does: Proposed a tunable seqBench benchmark to precisely probe large language models (LLMs) sequential reasoning capabilities through dimensions such as logical depth, backtracking counts, and noise ratio.