ACL 2025 Papers — Page 14
Annual Meeting of the Association for Computational Linguistics · 1699 papers
Revealing the Deceptiveness of Knowledge Editing: A Mechanistic Analysis of Superficial Editing
Jiakuan Xie (University of Chinese Academy of Sciences), Jun Zhao (University of Chinese Academy of Sciences)
Explainability and InterpretabilityTransformerText
🎯 What it does: This paper investigates the 'surface editing' problem in the knowledge editing process, where the model appears to have successfully updated knowledge but still returns old knowledge under specific prompts.
Reversal of Thought: Enhancing Large Language Models with Preference-Guided Reverse Reasoning Warm-up
Jiahao Yuan (East China Normal University), Usman Naseem (Macquarie University)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Proposed the RoT framework, enhancing LLM's logical reasoning ability through reverse reasoning warm-up, preference-guided reverse reasoning, and a cognitive preference manager.
Revisit Self-Debugging with Self-Generated Tests for Code Generation
Xiancai Chen (Peking University), Zhi Jin (Peking University)
AI Code AssistantTransformerLarge Language ModelText
🎯 What it does: This paper systematically studies the self-debugging capability of large language models (LLMs) in the absence of human-generated test cases, proposes and evaluates two self-debugging paradigms: post-execution and in-execution, and verifies their performance on four top-tier LLMs through experiments;
Revisiting Classical Chinese Event Extraction with Ancient Literature Information
Xiaoyi Bao (Hong Kong Polytechnic University), Chu-Ren Huang (Hong Kong Polytechnic University)
Vision Language ModelMultimodality
🎯 What it does: Built a vision-language model based on classical Chinese literary information for classical Chinese event extraction.
Revisiting Common Assumptions about Arabic Dialects in NLP
Amr Keleg (University of Edinburgh), Walid Magdy (University of Edinburgh)
ClassificationTextReview/Survey Paper
🎯 What it does: Systematically quantitatively test four common assumptions about Arabic dialects in natural language processing, expand and multi-label the NADI 2024 dataset, and explore relationships between dialect overlap, lexical discriminability, degree of dialectness in sentences, and multi-label recognition;
Revisiting Compositional Generalization Capability of Large Language Models Considering Instruction Following Ability
Yusuke Sakai (Nara Institute of Science and Technology), Taro Watanabe (Nara Institute of Science and Technology)
GenerationTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Propose the Ordered CommonGen benchmark, which utilizes full permutations of all concept sets and incorporates the phrase 'in the specified order' in prompts to evaluate LLMs' instruction following and combinatorial generation capabilities.
Revisiting Epistemic Markers in Confidence Estimation: Can Markers Accurately Reflect Large Language Models’ Uncertainty?
Jiayu Liu (Hong Kong University of Science and Technology), Yangqiu Song (Hong Kong University of Science and Technology)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Assess the reliability of large language models (LLM) in expressing confidence through semantic markers in question-answering tasks;
Revisiting LLMs as Zero-Shot Time Series Forecasters: Small Noise Can Break Large Models
Junwoo Park (KAIST AI), Jaegul Choo (KAIST AI)
TransformerLarge Language ModelPrompt EngineeringTime SeriesBenchmark
🎯 What it does: Evaluate the performance of large language models (LLMs) in zero-shot time series forecasting, and conduct a comprehensive comparison with state-of-the-art domain-specific prediction models and single-shot linear models.
Revisiting Scaling Laws for Language Models: The Role of Data Quality and Training Strategies
Zhengyu Chen (Meituan Inc), Jingang Wang (Meituan Inc)
Computational EfficiencyData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: Systematically studied the subscaling laws of large language models, analyzing the impact of data quality and training resource allocation on performance.
Revisiting the Test-Time Scaling of o1-like Models: Do they Truly Possess Test-Time Scaling Capabilities?
Zhiyuan Zeng (Fudan University), Xipeng Qiu (Fudan University)
Computational EfficiencyTextBenchmarkChain-of-Thought
🎯 What it does: This paper systematically evaluates the test-time scaling capability of the o1 family of models (QwQ, R1, LIMO), finding that longer Chains-of-Thought do not improve accuracy, and subsequently proposes the Shortest Majority Vote method under parallel scaling to enhance performance.
Revisiting Uncertainty Quantification Evaluation in Language Models: Spurious Interactions with Response Length Bias Results
Andrea Santilli (Sapienza University of Rome), Sinead Williamson (Apple)
Explainability and InterpretabilityTransformerTextBenchmark
🎯 What it does: This paper investigates the issue of evaluation distortion caused by mutual deviation between correctness functions and uncertainty quantification (UQ) methods in UQ assessment, with a focus on the impact of length bias.
Reviving Cultural Heritage: A Novel Approach for Comprehensive Historical Document Restoration
Yuyi Zhang (South China University of Technology), Lianwen Jin (South China University of Technology)
RecognitionRestorationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelDiffusion modelImageTextMultimodalityBenchmark
🎯 What it does: Proposed a complete page historical document restoration system named AutoHDR, consisting of a three-stage pipeline: OCR-assisted damage localization, visual-language context text prediction, and Patch-Autoregressive visual restoration; simultaneously constructed a full-page HDR public dataset called FPHDR.
Rhythm Controllable and Efficient Zero-Shot Voice Conversion via Shortcut Flow Matching
Jialong Zuo (Zhejiang University), Zhou Zhao (Zhejiang University)
GenerationTransformerDiffusion modelFlow-based ModelAudio
🎯 What it does: Proposed an R-VC zero-shot voice conversion system that can efficiently control the rhythm and timbre of the target speaker.
RiOT: Efficient Prompt Refinement with Residual Optimization Tree
Chenyi Zhou (Zhejiang University), Qiang Zhang (Zhejiang University)
Large Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes an automatic prompt optimization framework based on residual optimization tree (RIOT), which iteratively improves prompts using text gradients and achieves diverse exploration through multi-candidate prompt generation combined with perplexity selection;
Robust Estimation of Population-Level Effects in Repeated-Measures NLP Experimental Designs
Alejandro Benito-Santos (Universidad Nacional de Educación a Distancia), Víctor Fresno (Universidad Nacional de Educación a Distancia)
Explainability and InterpretabilityData-Centric LearningText
🎯 What it does: Linear mixed-effects model analysis of multi-source variations in NLP experiments, evaluating group effects of language models in English-Spanish bilingual gender bias detection tasks.
Robust Utility-Preserving Text Anonymization Based on Large Language Models
Tianyu Yang (Technical University of Darmstadt), Iryna Gurevych (Technical University of Darmstadt)
OptimizationSafty and PrivacyKnowledge DistillationData-Centric LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposed a robust text anonymization framework RUPTA based on large language models (LLMs), which jointly iteratively rewrites text using a privacy evaluator, utility evaluator, and optimizer to defend against LLM re-identification attacks while preserving the utility for downstream tasks.
RoCoFT: Efficient Finetuning of Large Language Models with Row-Column Updates
Md Kowsher (Nokia Bell Labs), Niloofar Yousefi (University of Central Florida)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This study proposes the RoCoFT method, which achieves parameter-efficient fine-tuning by updating only a small number of rows or columns in the Transformer weight matrix;
RolePlot: A Systematic Framework for Evaluating and Enhancing the Plot-Progression Capabilities of Role-Playing Agents
Pinyi Zhang (East China Normal University), Kai Zhang (Fudan University)
Data-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelTextBenchmark
🎯 What it does: Investigated the ability of role-playing agents (RPA) in story progression, proposed the RolePlot framework, and enhanced user interaction experience.
Rolling the DICE on Idiomaticity: How LLMs Fail to Grasp Context
Maggie Mi (University of Sheffield), Nafise Sadat Moosavi (University of Sheffield)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposed a new dataset called DICE to assess the ability of large language models to understand idioms in context.
Root Defense Strategies: Ensuring Safety of LLM at the Decoding Level
Xinyi Zeng (University of Chinese Academy of Sciences), Yu Tian (East China Normal University)
Safty and PrivacyComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposed a root defense strategy RDS based on the LLM decoding layer, leveraging the model's inherent discriminative capability during the decoding process to achieve progressive safe generation, combined with Speculative Decoding to improve speed.
RoToR: Towards More Reliable Responses for Order-Invariant Inputs
Soyoung Yoon (Seoul National University), Seung-won Hwang (Seoul National University)
Large Language ModelTextBenchmark
🎯 What it does: Investigated the position bias problem in language models with list inputs, proposing the RoToR method and Selective Routing to adapt to mixed order-sensitive and order-insensitive inputs.
RPO: Retrieval Preference Optimization for Robust Retrieval-Augmented Generation
Shi-Qi Yan (University of Science and Technology of China), Zhen-Hua Ling (University of Science and Technology of China)
GenerationRetrievalOptimizationSupervised Fine-TuningReinforcement LearningTextRetrieval-Augmented Generation
🎯 What it does: Propose the RPO algorithm to enhance the evaluation and utilization of retrieval context in Retrieval-Augmented Generation (RAG), addressing knowledge conflicts.
RSCF: Relation-Semantics Consistent Filter for Entity Embedding of Knowledge Graph
Junsik Kim (Gwangju Institute of Science and Technology), Kangil Kim (Gwangju Institute of Science and Technology)
Representation LearningGraphBenchmark
🎯 What it does: Introduced the Relation Semantic Consistency Filter (RSCF) in knowledge graph embeddings to enhance the consistency of entity transformations and model performance.
RSVP: Reasoning Segmentation via Visual Prompting and Multi-modal Chain-of-Thought
Yi Lu (Opus AI Research), Wenbo Zhu
SegmentationLarge Language ModelPrompt EngineeringVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: Propose RSVP, a two-stage reasoning segmentation framework based on multi-modal chain-of-thought visual prompts, enabling end-to-end inference from complex implicit queries to fine-grained segmentation.
Rubrik’s Cube: Testing a New Rubric for Evaluating Explanations on the CUBE dataset
Diana Galvan-Sosa (University of Cambridge), Paula Buttery (University of Cambridge)
Explainability and InterpretabilityTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose the Rubrik evaluation framework and the CUBE dataset to systematically assess the quality of explanations generated by large language models (LLMs);
RUBY: An Effective Framework for Multi-Constraint Multi-Hop Question Generation
Wenzhuo Zhao (South China Normal University), Shuangyin Li (South China Normal University)
GenerationTransformerLarge Language ModelText
🎯 What it does: Developed the RUBY framework for multi-constraint multi-hop question generation tasks. It first compresses high-dimensional constraints such as multi-hop types and intentions into low-dimensional multi-hop skeletons via a High-Dimensional Semantic Constraint Reduction (HDR) module. Then, it employs a divide-and-conquer strategy with Sub-Question Answer Pair Generation (SQAG) and Multi-Hop Question Generation (MHQG) processes. During decoding, it introduces Reasoning Dynamic Projection (RD-Projection) technology to enhance generation accuracy and constraint consistency.
RuleArena: A Benchmark for Rule-Guided Reasoning with LLMs in Real-World Scenarios
Ruiwen Zhou (University of California, Santa Barbara), William Yang Wang (University of California, Santa Barbara)
TransformerTextBenchmarkChain-of-Thought
🎯 What it does: Designed and released the RULEARENA benchmark, covering three real-world scenarios: airline baggage fees, NBA trades, and taxation. Constructed 95 rules averaging approximately 400 characters each, along with 816 reasoning test questions based on these rules, to evaluate the performance of large language models (LLMs) in rule-driven reasoning.
S-RAG: A Novel Audit Framework for Detecting Unauthorized Use of Personal Data in RAG Systems
Zhirui Zeng (University of Auckland), Zijian Zhang (Beijing Institute of Technology)
Anomaly DetectionSafty and PrivacyTransformerLarge Language ModelTextBiomedical DataRetrieval-Augmented Generation
🎯 What it does: Propose a black-box auditing framework named S-RAG, which can detect whether user text data is used by the external database of a retrieval-augmented generation (RAG) system, thereby achieving traceability of personal data usage.
S^2R: Teaching LLMs to Self-verify and Self-correct via Reinforcement Learning
Ruotian Ma (Tencent), Jia Li (Hong Kong University of Science and Technology (Guangzhou))
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextSequential
🎯 What it does: Enable large language models to self-verify and self-correct during reasoning through supervised fine-tuning and reinforcement learning, thereby enhancing their deep reasoning capabilities.
S^3 - Semantic Signal Separation
Márton Kardos (Aarhus University), Roberta Rocca (Aarhus University)
Computational EfficiencyRepresentation LearningTransformerText
🎯 What it does: Propose a topic modeling method based on Independent Component Analysis (ICA) called Semantic Signal Separation (S3), which decomposes document vectors into topic axes in the sentence embedding space, enabling topic discovery without preprocessing and directly using contextual embeddings.
S2WTM: Spherical Sliced-Wasserstein Autoencoder for Topic Modeling
Suman Adhya (Indian Association for the Cultivation of Science), Debarshi Kumar Sanyal (Indian Association for the Cultivation of Science)
Representation LearningAuto EncoderText
🎯 What it does: Propose a spherical sliced-Wasserstein autoencoder (S2WTM), which trains topic models on a spherical surface to avoid posterior collapse in VAEs.
Safe: Enhancing Mathematical Reasoning in Large Language Models via Retrospective Step-aware Formal Verification
Chengwu Liu (Peking University), Ming Zhang (Peking University)
Explainability and InterpretabilityTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Proposed the Safe framework, which uses retrospective step-by-step formal verification of each step in LLM mathematical reasoning to reduce hallucinations;
Safer or Luckier? LLMs as Safety Evaluators Are Not Robust to Artifacts
Hongyu Chen (Cohere), Seraphina Goldfarb-Tarrant (Cohere)
Safty and PrivacyTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Evaluated and compared the robustness and human consistency of 11 large language models in safety assessments, focusing on the impact of input disguises such as humility, authority, halo, verbosity, and position, and attempted to enhance robustness through an expert-selected 'jury'
SafeRAG: Benchmarking Security in Retrieval-Augmented Generation of Large Language Model
Xun Liang (Renmin University of China), Mengwei Wang (Renmin University of China)
Safty and PrivacyLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed the SafeRAG benchmark for systematic evaluation of the security of retrieval-augmented generation (RAG) when facing data injection attacks.
Safety Alignment via Constrained Knowledge Unlearning
Zesheng Shi (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
Safty and PrivacyTransformerLarge Language ModelText
🎯 What it does: Propose a secure alignment method CKU, which evaluates the scores of neurons in the LLM MLP layer, locks important neurons, and performs gradient ascent only on the remaining neurons when unlearning harmful knowledge, thereby eliminating vulnerable harmful information while retaining most of the original knowledge.
SAKE: Steering Activations for Knowledge Editing
Marco Scialanga (Sorbonne Université), Marcin Detyniecki (EPFL)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose SAKE, a knowledge editing method that achieves robust updates of factual information in large language models through activation guides and optimal transport.
SAM Decoding: Speculative Decoding via Suffix Automaton
Yuxuan Hu (Renmin University of China), Jing Zhang (Renmin University of China)
GenerationRetrievalComputational EfficiencyLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose a retrieval-based speculative decoding method called SAM-Decoding, which utilizes a suffix automaton (SAM) to dynamically and statically retrieve the optimal suffix from the generated text and corpus to generate drafts, and can be adaptively integrated with generative methods (e.g., Token Recycling, EAGLE-2);
Sample-Efficient Human Evaluation of Large Language Models via Maximum Discrepancy Competition
Kehua Feng (Zhejiang University), Huajun Chen (Zhejiang University)
Reinforcement Learning from Human FeedbackLarge Language ModelText
🎯 What it does: Proposes a sample-efficient human evaluation method based on Maximum Difference Competition (MAD Competition) for reliable ranking of large language models with minimal human annotations.
Sandcastles in the Storm: Revisiting the (Im)possibility of Strong Watermarking
Fabrice Y Harel-Canada (University of California Los Angeles), Amit Sahai (University of California Los Angeles)
Safty and PrivacyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Conduct large-scale experiments and manual evaluations on random walk attacks against text watermarks to test the validity of theoretical assumptions (fast mixing and quality preservation).
SARA: Salience-Aware Reinforced Adaptive Decoding for Large Language Models in Abstractive Summarization
Nayu Liu (Tiangong University), Kaiwen Wei (Tiangong University)
GenerationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Propose a Salience-Aware Reinforced Adaptive Decoding (SARA) method that utilizes key contextual information and reinforcement learning to dynamically adjust the weights of context and prior knowledge in large language models (LLMs) during the summarization process, thereby reducing hallucination and improving summary quality.
Scalable Vision Language Model Training via High Quality Data Curation
Hongyuan Dong (ByteDance Douyin Content Group), Jiao Ran (ByteDance Douyin Content Group)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodality
🎯 What it does: Proposed and trained the SAIL-VL series of scalable vision-language models, constructed a high-quality SAIL-Caption dataset, and adopted a multi-stage curriculum-based SFT strategy.
SCALE: Towards Collaborative Content Analysis in Social Science with Large Language Model Agents and Human Intervention
Chengshuai Zhao (Arizona State University), Huan Liu (Arizona State University)
ClassificationTransformerLarge Language ModelAgentic AIPrompt EngineeringTextChain-of-Thought
🎯 What it does: Developed the SCALE framework, which uses multi-agent systems to simulate the content analysis process in social sciences and combines human expert intervention to achieve automated annotation.
ScaleBiO: Scalable Bilevel Optimization for LLM Data Reweighting
Rui Pan (University of Illinois Urbana-Champaign), Tong Zhang (University of Illinois Urbana-Champaign)
OptimizationLarge Language ModelText
🎯 What it does: Proposed a scalable first-order bi-level optimization algorithm called ScaleBiO for data reweighting in large-scale LLMs.
Scaling Laws and Efficient Inference for Ternary Language Models
Tejas Vaidhya (Nolano AI), Irina Rish (Nolano AI)
Computational EfficiencyTransformerText
🎯 What it does: This paper studies the ternary quantized language model (TriLM), first analyzing its sensitivity to parameters and training data through scaling laws, then training the TriTera model on 1.2 trillion tokens; it proposes efficient 1.6-bit and 2-bit weight quantization packaging schemes and implements the GPU kernel TriRun, significantly accelerating inference;
Scaling Text-Rich Image Understanding via Code-Guided Synthetic Multimodal Data Generation
Yue Yang (University of Pennsylvania), Christopher Clark (Allen Institute for Artificial Intelligence)
GenerationData SynthesisLarge Language ModelPrompt EngineeringVision Language ModelMultimodality
🎯 What it does: By letting large language models automatically generate rendering code and render various text-rich images, then using the code as context to generate corresponding visual language instructions, thus constructing a large-scale synthetic dataset;
Scaling up the State Size of RNN LLMs for Long-Context Scenarios
Kai Liu (Tongji University), Kai Chen (Tongji University)
RetrievalComputational EfficiencyRecurrent Neural NetworkTransformerLarge Language ModelText
🎯 What it does: Propose an efficient method to expand the state size of RNN models, matching the state size of Transformers at 2k context length, thereby enhancing long-text memory and retrieval capabilities.
ScanEZ: Integrating Cognitive Models with Self-Supervised Learning for Spatiotemporal Scanpath Prediction
Ekta Sood (University of Colorado Boulder), Sidney K. D’Mello
Data SynthesisRepresentation LearningTransformerSupervised Fine-TuningTextSequential
🎯 What it does: Designed and implemented ScanEZ, a self-supervised framework that utilizes synthetic eye movement data generated by cognitive models and masked learning to simultaneously predict spatial positions and dwell times during the reading process.
SCAR: Data Selection via Style Consistency-Aware Response Ranking for Efficient Instruction-Tuning of Large Language Models
Zhuang Li (RMIT University), Gholamreza Haffari (Monash University)
Computational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: This paper proposes the SCAR (Style Consistency-Aware Response Ranking) method, which automatically filters training samples with consistent linguistic forms and instruction surprisal to enhance the instruction fine-tuning of large language models (LLMs) using only a small amount of data.
SceneGenAgent: Precise Industrial Scene Generation with Coding Agent
Xiao Xia (Tsinghua University), Yuxiao Dong (Tsinghua University)
GenerationAI Code AssistantLarge Language ModelSupervised Fine-TuningAgentic AIText
🎯 What it does: Propose SceneGenAgent, an LLM-based encoding agent capable of generating C# code from text descriptions and rendering industrial scenes in Tecnomatix software.
SciVer: Evaluating Foundation Models for Multimodal Scientific Claim Verification
Chengye Wang (Yale NLP Lab), Yilun Zhao (Yale NLP Lab)
Large Language ModelImageTextMultimodalityTabularBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the SCIVER benchmark to evaluate the ability of proposition verification in multimodal scientific papers;
SConU: Selective Conformal Uncertainty in Large Language Models
Zhiyuan Wang (University of Electronic Science and Technology of China), Kaidi Xu (Drexel University)
Explainability and InterpretabilityComputational EfficiencyLarge Language ModelText
🎯 What it does: Proposes the SConU framework, which eliminates confidence outliers that violate the exchangeability assumption through significance testing and calculates the minimum manageable risk level to achieve reliable uncertainty estimation for large language models.
SCOP: Evaluating the Comprehension Process of Large Language Models from a Cognitive View
Yongjie Xiao (Sichuan University), Wenqiang Lei (Sichuan University)
Explainability and InterpretabilityLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the SCOP framework to evaluate five key skills of large language models (LLMs) in the reading comprehension process from a cognitive perspective.
SCOPE: Optimizing Key-Value Cache Compression in Long-context Generation
Jialong Wu (Southeast University), Deyu Zhou (Southeast University)
CompressionOptimizationTransformerLarge Language ModelText
🎯 What it does: To address KV cache compression for long-text generation in large language models, the SCOPE framework is proposed, which separates the prefill and decoding stages, compresses KV caches separately, and effectively selects reusable heavy-hitters during decoding through strategies such as sliding windows, adaptive widening, and non-continuous updates.
SCULPT: Systematic Tuning of Long Prompts
Shanu Kumar (Microsoft Corporation), Manish Gupta (Microsoft Corporation)
Reinforcement LearningPrompt EngineeringText
🎯 What it does: This paper proposes the SCULPT framework, which systematically improves long prompts by utilizing a hierarchical tree structure and an actor-critic mechanism;
SDBench: A Survey-based Domain-specific LLM Benchmarking and Optimization Framework
Cheng Guo (Zhejiang University), Wei Dong (Zhejiang University)
OptimizationLarge Language ModelSupervised Fine-TuningTextReview/Survey PaperBenchmark
🎯 What it does: Proposed the SDBench framework, which automatically constructs knowledge architectures and evaluation datasets by leveraging domain reviews, and implemented BridgeBench and BridgeGPT in the field of bridge engineering.
SDD: Self-Degraded Defense against Malicious Fine-tuning
ZiXuan Chen, Ziqian Zeng (South China University Of Technology)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Propose a new self-degradation defense framework (SDD), which defends against malicious fine-tuning attacks by letting LLMs generate high-quality but irrelevant answers under harmful prompts;
SDPO: Segment-Level Direct Preference Optimization for Social Agents
Aobo Kong (Nankai University), Fei Huang (Tongyi Lab)
OptimizationReinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextBenchmark
🎯 What it does: Proposed Segment-Level Direct Preference Optimization (SDPO), performing direct preference optimization on key segments in multi-turn social dialogues;
SEA: Low-Resource Safety Alignment for Multimodal Large Language Models via Synthetic Embeddings
Weikai Lu (South China University of Technology), Ziqian Zeng (South China University of Technology)
Data SynthesisOptimizationRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVideoTextMultimodalityBenchmarkAudio
🎯 What it does: Propose the SEA (Synthetic Embedding Augmented Safety Alignment) method, which utilizes gradient optimization to generate additional modal embeddings to replace real multi-modal data for achieving low-resource multi-modal safety alignment, while simultaneously constructing the VA-SafetyBench video/audio safety evaluation benchmark.
SeaKR: Self-aware Knowledge Retrieval for Adaptive Retrieval Augmented Generation
Zijun Yao (Tsinghua University), Juanzi Li (Tsinghua University)
GenerationRetrievalTransformerLarge Language ModelTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose SEAKR—a self-aware knowledge retrieval framework that quantifies uncertainty through LLM internal states to determine when to retrieve and dynamically integrate retrieval results, achieving adaptive retrieval-augmented generation.
SEAL: Scaling to Emphasize Attention for Long-Context Retrieval
Changhun Lee (Pohang University of Science and Technology), Eunhyeok Park (Pohang University of Science and Technology)
RetrievalTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Propose a method to enhance the performance of large language models in long-text retrieval tasks by weighting attention heads or channels (SEAL).
Second Language (Arabic) Acquisition of LLMs via Progressive Vocabulary Expansion
Jianqing Zhu (King Abdullah University of Science and Technology), Jinchao Xu (King Abdullah University of Science and Technology)
TransformerLarge Language ModelText
🎯 What it does: Developed AraLLaMA, an Arabic LLM, achieving efficient Arabic subword tokenizer through progressive vocabulary expansion (I-BPE), and demonstrating excellent performance in multi-task and dialogue evaluations;
SECRET: Semi-supervised Clinical Trial Document Similarity Search
Trisha Das (University of Illinois Urbana-Champaign), Jimeng Sun (University of Illinois Urbana-Champaign)
RetrievalRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextBiomedical Data
🎯 What it does: Propose a semi-supervised clinical trial protocol similarity search framework called SECRET, which uses LLM to generate question-answer pairs to compress long documents, followed by contrastive learning at the question-answer level and trial level to obtain high-quality embeddings for retrieving similar trials.
SEE: Strategic Exploration and Exploitation for Cohesive In-Context Prompt Optimization
Wendi Cui (Intuit), Sricharan Kumar (Intuit)
OptimizationLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Propose the SEE (Strategic Exploration and Exploitation) framework to achieve collaborative optimization of instructions and examples, generating optimal prompts with zero/few samples;
SeedBench: A Multi-task Benchmark for Evaluating Large Language Models in Seed Science
Jie Ying (Shanghai Artificial Intelligence Laboratory), Nanqing Dong (Shanghai Artificial Intelligence Laboratory)
TransformerLarge Language ModelTextBenchmarkAgriculture Related
🎯 What it does: Constructed SeedBench—a multi-task LLM evaluation benchmark for seed science, covering three stages: gene information retrieval, gene function and regulation, and variety breeding with agronomic traits. It includes 2,264 expert-verified question-answer pairs, supporting 0-shot and 1-shot testing.
Seeking Rational Demonstrations for Large Language Models: A Domain Generalization Approach to Unsupervised Cross-Domain Keyphrase Generation
Guangzhen Zhao (Nanjing University of Posts and Telecommunications), Zhenjiang Dong (Nanjing University of Posts and Telecommunications)
GenerationDomain AdaptationTransformerLarge Language ModelContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Proposes the Seeking Rational Demonstrations (SRD) method, which leverages source domain labeled samples to provide effective examples for large language models (LLMs) in unsupervised cross-domain keyword generation tasks.
Segment First or Comprehend First? Explore the Limit of Unsupervised Word Segmentation with Large Language Models
Zihong Zhang (Wuhan University), Bo Du (Wuhan University)
SegmentationTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper utilizes large language models for unsupervised word segmentation, proposing the LLM-WS framework and the LLACA method to evaluate and enhance segmentation performance.
Segment-Based Attention Masking for GPTs
Shahar Katz (Tel Aviv University), Lior Wolf (Tel Aviv University)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Designed and verified a new attention mask mechanism (Masked Attention by Segment, MAS), allowing tokens within the same input block to access each other during the prefill stage of the GPT model to leverage information from future tokens; restoring traditional causal masks during the generation stage; directly applying MAS to existing public GPT models such as Llama, Qwen, and Mistral via lightweight LoRA fine-tuning.
Segment-Level Diffusion: A Framework for Controllable Long-Form Generation with Diffusion Language Models
Xiaochen Zhu (University of Cambridge), Andreas Vlachos (University of Cambridge)
GenerationTransformerDiffusion modelContrastive LearningText
🎯 What it does: This study proposes the Segment-Level Diffusion (SLD) framework, which leverages text segmentation, adversarial and contrastive learning to enhance latent representations, and achieves controllable long-text generation through latent diffusion models and autoregressive decoders.
Selecting and Merging: Towards Adaptable and Scalable Named Entity Recognition with Large Language Models
Zhuojun Ding (Huazhong University of Science and Technology), Chenghao Fan (Huazhong University of Science and Technology)
RecognitionTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose the SaM framework, which dynamically selects and merges multi-domain expert models to generate task-specific NER models tailored for the target domain.
Self-Error-Instruct: Generalizing from Errors for LLMs Mathematical Reasoning
Erxin Yu (Hong Kong Polytechnic University), Lifeng Shang (Huawei Noah's Ark Lab)
Data SynthesisData-Centric LearningLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Propose the Self-Error-Instruct (SEI) framework, which first extracts error cases from the target large language model (LLM), uses GPT-4o to generate error key phrases and clusters them into error types, then generates targeted training samples via self-instruction conditioned on each error type. Subsequently, a round of learning filters high-quality data and iteratively fine-tunes the LLM to enhance mathematical reasoning ability.
Self-Instructed Derived Prompt Generation Meets In-Context Learning: Unlocking New Potential of Black-Box LLMs
Zhuo Li (Shenzhen International Center for Industrial and Applied Mathematics), Anningzhe Gao (Shenzhen International Center for Industrial and Applied Mathematics)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: To improve the quality of prompts for black-box LLMs, this paper proposes a self-instruction reinforcement learning framework that automatically generates semantically consistent derivative prompts. It enhances the answer quality under the original prompt by constructing contextual examples using the derivative prompts and their responses.
SELF-PERCEPT: Introspection Improves Large Language Models’ Detection of Multi-Person Mental Manipulation in Conversations
Danush Khanna (Manipal University Jaipur), Kripabandhu Ghosh (IISER Kolkata)
ClassificationAnomaly DetectionTransformerPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Propose the MultiManip dataset and develop the SELF-PERCEPT two-stage prompting framework for detecting multi-round, multi-person psychological manipulation dialogues.
Self-supervised Quantized Representation for Seamlessly Integrating Knowledge Graphs with Large Language Models
Qika Lin (National University Of Singapore), Mengling Feng (National University Of Singapore)
CompressionKnowledge DistillationRepresentation LearningGraph Neural NetworkLarge Language ModelSupervised Fine-TuningAuto EncoderGraph
🎯 What it does: Proposes a self-supervised quantization representation (SSQR) method, compressing the structural and semantic information of knowledge graph (KG) entities into discrete codes (tokens), and seamlessly inputs them into large language models (LLMs) to complete KG tasks through instruction tuning.
Self-Taught Agentic Long Context Understanding
Yufan Zhuang (AMD), Emad Barsoum (AMD)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningAgentic AITextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the AgenticLU framework, which achieves self-clarification and context retrieval for LLMs in long-text scenarios through Chain-of-Clarifications (CoC), thereby improving the accuracy of multi-step reasoning and long-form question answering.
SelfElicit: Your Language Model Secretly Knows Where is the Relevant Evidence
Zhining Liu (University of Illinois Urbana Champaign), Hanghang Tong (University of Illinois Urbana Champaign)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose SELFELICIT, a method that automatically identifies and highlights key evidence in text during inference by leveraging the attention scores of language models themselves, thereby improving the factual accuracy of answers.
Semantic Exploration with Adaptive Gating for Efficient Problem Solving with Language Models
Sungjae Lee (POSTECH), Jungseul Ok (POSTECH)
Computational EfficiencyLarge Language ModelTextChain-of-Thought
🎯 What it does: Proposed a framework named SEAG, integrating adaptive gating, semantic exploration, and early stopping, to enhance the efficiency and accuracy of large language models in multi-step reasoning tasks.
Semantic-Eval : A Semantic Comprehension Evaluation Framework for Large Language Models Generation without Training
Shusheng Li (Northeastern University), Wenjun Tan (Northeastern University)
Computational EfficiencyRepresentation LearningLarge Language ModelTextBenchmark
🎯 What it does: Proposed Semantic-Eval, a training-free semantic understanding evaluation framework for assessing the quality of text generated by large language models.
SEOE: A Scalable and Reliable Semantic Evaluation Framework for Open Domain Event Detection
Yi-Fan Lu (Beijing Institute of Technology), Heyan Huang (Beijing Institute of Technology)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Constructed an scalable and reliable semantic evaluation framework named SEOE, which includes a more representative benchmark (564 event types covering 7 domains) and a semantic F1 evaluation method based on large language models.
Separating Tongue from Thought: Activation Patching Reveals Language-Agnostic Concept Representations in Transformers
Clément Dumas (ENS Paris-Saclay), Robert West (EPFL)
Explainability and InterpretabilityRepresentation LearningTransformerPrompt EngineeringText
🎯 What it does: Through activation patching techniques in multilingual translation tasks, causal analysis of Transformer internal representations demonstrates the existence of language-agnostic concept representations, and verifies that averaging concept vectors across different languages can improve translation quality and definition generation performance.
Serial Lifelong Editing via Mixture of Knowledge Experts
YuJu Cheng, Yu-Chiang Frank Wang (National Taiwan University)
Large Language ModelMixture of ExpertsTextBenchmark
🎯 What it does: Propose a serial lifelong knowledge editing (sLKE) framework to address the issue of knowledge conflicts in LLMs caused by multiple updates to the same fact;
SEUF: Is Unlearning One Expert Enough for Mixture-of-Experts LLMs?
Haomin Zhuang (University of Notre Dame), Xiangliang Zhang (University of Notre Dame)
OptimizationComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsTextBenchmark
🎯 What it does: Propose SEUF, a machine-unlearning framework for sparse Mixture-of-Experts (MoE) Large Language Models (LLMs), achieving targeted unlearning through expert attribution and routing anchoring.
SGIC: A Self-Guided Iterative Calibration Framework for RAG
Guanhua Chen (University of Macau), Derek F. Wong (University of Macau)
GenerationRetrievalTransformerSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: Propose a self-guided iterative calibration framework named SGIC, which dynamically corrects answers in retrieval-augmented generation (RAG) by leveraging the context reasoning capability of large language models and uncertainty scores.
Shaping the Safety Boundaries: Understanding and Defending Against Jailbreaks in Large Language Models
Lang Gao (Mohamed bin Zayed University of Artificial Intelligence), Xiuying Chen (University of Notre Dame)
Safty and PrivacyExplainability and InterpretabilityComputational EfficiencyAdversarial AttackTransformerLarge Language ModelText
🎯 What it does: Conducted a large-scale analysis of over 30,000 samples, revealing that adversarial attacks bypass security boundaries by exploiting activation shifts, and proposed a lightweight defense method called ABD based on activation boundaries.
SHARE: An SLM-based Hierarchical Action CorREction Assistant for Text-to-SQL
Ge Qu (University of Hong Kong), Reynold Cheng (University of Hong Kong)
AI Code AssistantTransformerSupervised Fine-TuningTextBenchmark
🎯 What it does: Built a hierarchical action error correction assistant called SHARE based on small language models, helping LLMs achieve more precise error localization and self-correction.
SHARE: Shared Memory-Aware Open-Domain Long-Term Dialogue Dataset Constructed from Movie Script
Eunwon Kim (Sogang University), Buru Chang (Korea University)
GenerationData SynthesisTransformerLarge Language ModelText
🎯 What it does: Constructed a long-term dialogue dataset named SHARE, extracting character features, personal events, and shared memories from movie scripts, and proposed the EPISODE framework to generate more coherent and engaging long-term dialogues by leveraging shared memories.
Sharper and Faster mean Better: Towards More Efficient Vision-Language Model for Hour-scale Long Video Understanding
Daoze Zhang (Alibaba Group), Yingda Chen (Alibaba Group)
Computational EfficiencyConvolutional Neural NetworkTransformerLarge Language ModelVideoMultimodalityBenchmark
🎯 What it does: Proposed Sophia, a vision-language model for hour-level long videos, combining two-stage shot-based frame pruning and hierarchical sparse attention to achieve efficient long video understanding.
Sheep’s Skin, Wolf’s Deeds: Are LLMs Ready for Metaphorical Implicit Hate Speech?
Jingjie Zeng, Hongfei Lin (Dalian University of Technology)
ClassificationRecognitionGenerationTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextStochastic Differential Equation
🎯 What it does: Investigate the vulnerability of large language models (LLMs) in detecting implicit metaphorical hate speech, and evaluate their performance by generating metaphorical hate texts through adversarial jailbreaking and energy-constrained decoding methods.
ShifCon: Enhancing Non-Dominant Language Capabilities with a Shift-based Multilingual Contrastive Framework
Hengyuan Zhang (Tsinghua University), Furu Wei (Tsinghua University)
Representation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: Propose the ShifCon framework, which enhances language model capabilities for minority languages by projecting non-main language representations into the main language subspace and back, along with multilingual contrastive learning.
Shifting from Ranking to Set Selection for Retrieval Augmented Generation
Dahyun Lee (LG AI Research), Moontae Lee (LG AI Research)
GenerationRetrievalTransformerSupervised Fine-TuningTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose a set-based retrieval method called SETR, which leverages chain-of-thought reasoning to identify query information needs and select a complete subset of documents from the candidate list, replacing traditional top-k sorting;
Should I Believe in What Medical AI Says? A Chinese Benchmark for Medication Based on Knowledge and Reasoning
Yue Wu (Xunfei Healthcare Technology Company Limited), Xiaodong Tao (Xunfei Healthcare Technology Company Limited)
Drug DiscoveryLarge Language ModelTextBiomedical DataBenchmark
🎯 What it does: Constructed the Chinese drug knowledge and reasoning benchmark ChiDrug to evaluate large language models' knowledge and reasoning capabilities across six dimensions: drug indications, dosage, contraindications, mechanism of action, drug recommendations, and interactions.
SHuBERT: Self-Supervised Sign Language Representation Learning via Multi-Stream Cluster Prediction
Shester Gueuwou (Toyota Technological Institute Chicago), Alexander H. Liu (Massachusetts Institute of Technology)
Representation LearningTransformerSupervised Fine-TuningVideo
🎯 What it does: Propose a self-supervised multi-stream Transformer model named SHuBERT for learning contextualized sign language representations from visual features of hands, facial expressions, and body postures;
SIFT-50M: A Large-Scale Multilingual Dataset for Speech Instruction Fine-Tuning
Prabhat Pandey (Amazon AGI), Andreas Schwarz (Amazon AGI)
Data SynthesisTransformerLarge Language ModelSupervised Fine-TuningTextMultimodalityBenchmarkAudio
🎯 What it does: Proposed the SIFT-50M (50 million QA pairs) multilingual instruction dataset, trained the SIFT-LLM (speech-text LLM), and released the EvalSIFT evaluation benchmark.
Sightation Counts: Leveraging Sighted User Feedback in Building a BLV-aligned Dataset of Diagram Descriptions
Wan Ju Kang (KAIST), James Thorne (KAIST)
Data SynthesisRetrievalTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningMultimodality
🎯 What it does: Construct the Sightation dataset by evaluating the chart descriptions generated by VLM using the feedback from users who have seen the charts, forming multi-task BLV alignment data.
Silencing Empowerment, Allowing Bigotry: Auditing the Moderation of Hate Speech on Twitch
Prarabdh Shukla (Indian Institute of Science), Arjun Bhagoji (University of Chicago)
Anomaly DetectionLarge Language ModelText
🎯 What it does: This study conducts a large-scale audit of Twitch's automatic content moderation tool AutoMod by building a chatbot and using the Twitch API to evaluate its effectiveness in detecting hate speech.
SimulS2S-LLM: Unlocking Simultaneous Inference of Speech LLMs for Speech-to-Speech Translation
Keqi Deng (University of Cambridge), Phil Woodland
GenerationTransformerLarge Language ModelFlow-based ModelTextAudio
🎯 What it does: This paper proposes SimulS2S-LLM, a large language model (LLM) framework capable of achieving real-time speech-to-speech translation after offline training, leveraging offline-trained LLMs and a streaming acoustic encoder, combined with CIF to extract boundary-aware speech prompts, and employing a wait-k strategy for synchronous inference.
SINCon: Mitigate LLM-Generated Malicious Message Injection Attack for Rumor Detection
Mingqing Zhang (Chinese Academy of Sciences), Liang Wang (Chinese Academy of Sciences)
Adversarial AttackGraph Neural NetworkLarge Language ModelContrastive LearningText
🎯 What it does: This paper proposes a robustness enhancement method called SINCon based on contrastive learning, aiming to improve the defense capability of rumor detection models based on message propagation trees (MPT) when facing malicious message injection attacks generated by large language models (LLMs).
Single-to-mix Modality Alignment with Multimodal Large Language Model for Document Image Machine Translation
Yupu Liang (Chinese Academy of Sciences), Yu Zhou (Chinese Academy of Sciences)
Image TranslationTransformerLarge Language ModelImageTextMultimodality
🎯 What it does: Propose a single-modal to hybrid-modal alignment framework called M4Doc, which leverages large multimodal language models (MLLMs) during the training phase to enhance the generalization and performance of lightweight document image machine translation (DIMT) models, and removes the MLLM during inference to maintain efficiency;
Sinhala Encoder-only Language Models and Evaluation
Tharindu Ranasinghe (Lancaster University), Ruslan Mitkov (Lancaster University)
ClassificationRecognitionTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper collects and publicly releases the largest Sinhala monolingual corpus with 1.5B words, and trains three mainstream encoder-only Transformer models (BERT, RoBERTa, Electra) on this corpus. Meanwhile, the authors design and release the first Sinhala natural language understanding benchmark, SINHALA-GLUE, covering six tasks (sentiment analysis, offensive language detection, news headline prediction, semantic text similarity, named entity recognition, offensive term detection).
SkillAggregation: Reference-free LLM-Dependent Aggregation
Guangzhi Sun (University of Cambridge), Mark Gales (University of Cambridge)
ClassificationTransformerLarge Language ModelText
🎯 What it does: Proposed a no-reference multi-LLM evaluation aggregation method called SkillAggregation, which can learn to dynamically weight and fuse judgments from multiple LLMs into more accurate decisions.
SkillVerse : Assessing and Enhancing LLMs with Tree Evaluation
Yufei Tian (University of California Los Angeles), Zizhao Zhang (Google Cloud AI)
Explainability and InterpretabilityTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposes the SKILLVERSE framework, which utilizes critical feedback generated by LLM reviewers, refines it into atomic judgments, clusters them into a hierarchical tree, and thereby obtains a fine-grained model capability assessment.