Conference on Empirical Methods in Natural Language Processing Β· 593 papers
Surge: On the Potential of Large Language Models as General-Purpose Surrogate Code Executors
Bohan Lyu (Tsinghua), Jiaming Zhang (Tsinghua)
CodeAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
π― What it does: Designed and constructed the SURGE benchmark to systematically evaluate the ability of large language models (LLMs) to predict code execution outcomes without executing the code; conducted comprehensive experiments on 21 open-source and closed-source LLMs to explore the impact of scale, data volume, and prompting strategies on prediction accuracy.
SwiftKV: Fast Prefill-Optimized Inference with Knowledge-Preserving Model Transformation
Aurick Qiao (Snowflake AI Research), Yuxiong He (Snowflake AI Research)
CodeComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelText
π― What it does: This study proposes SwiftKV, a technique that restructures LLMs to skip subsequent transformer layers during inference (only for prefix/prompt tokens) and directly generate KV caches for the remaining layers using the hidden states from the previous layer, while combining AcrossKV for cross-layer KV cache sharing and lightweight knowledge retention distillation, significantly reducing prefix computational costs and compressing KV caches;
Sycophancy Mitigation Through Reinforcement Learning with Uncertainty-Aware Adaptive Reasoning Trajectories
Mohammad Beigi (University of California Davis), Lifu Huang (University of California Davis)
CodeOptimizationSafty and PrivacyTransformerLarge Language ModelReinforcement LearningText
π― What it does: Propose the SMART framework, which generates high-quality reasoning trajectories in two stages through uncertainty-aware adaptive Monte Carlo Tree Search (UA-MCTS), and employs dense progress rewards for reinforcement learning to alleviate the sycophancy behavior of LLMs.
SynC-LLM: Generation of Large-Scale Synthetic Circuit Code with Hierarchical Language Models
Shang Liu (Hong Kong University of Science and Technology), Zhiyao Xie (Hong Kong University of Science and Technology)
CodeGenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringDiffusion modelTextGraphRetrieval-Augmented Generation
π― What it does: Introduces SynC-LLM, a three-stage hierarchical framework for generating synthetic circuit code at the million-gate level, addressing the scarcity of public circuit data.
Synth-SBDH: A Synthetic Dataset of Social and Behavioral Determinants of Health for Clinical Text
Avijit Mitra (University of Massachusetts Amherst), Hong Yu (U.S. Department of Veterans Affairs)
CodeData SynthesisExplainability and InterpretabilityKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBiomedical DataElectronic Health RecordsBenchmark
π― What it does: Constructed Synth-SBDH, a publicly available synthetic Social and Behavioral Determinants of Health (SBDH) dataset covering 15 SBDH categories, with fine-grained annotations including presence, tense, and rationale.
T2R-BENCH: A Benchmark for Real World Table-to-Report Task
Jie Zhang (Institute of Artificial Intelligence (TeleAI), China Telecom), Xuelong Li (Institute of Artificial Intelligence (TeleAI), China Telecom)
CodeGenerationLarge Language ModelTextTabularBenchmark
π― What it does: Proposed the 'table-to-report' task and constructed a bilingual benchmark T2R-bench for industrial real-world tables (457 tables, 910 questions, and 4,320 report key points), while providing a three-dimensional evaluation system.
TableEval: A Real-World Benchmark for Complex, Multilingual, and Multi-Structured Table Question Answering
Junnan Zhu (Beijing Wenge Technology Co., Ltd.), Nan Xu (State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, CAS)
CodeLarge Language ModelTabularBenchmark
π― What it does: Constructed the TableEval benchmark, covering real-world table question-answering tasks with multi-structural, multi-lingual, and multi-domain characteristics, and proposed the SEAT evaluation framework.
CodeGenerationRetrievalTransformerLarge Language ModelTextTabularBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Proposed TableRAG, an SQL-based retrieval-augmented generation framework specialized for multi-hop question answering over heterogeneous documents (text + tables), and constructed a new HeteQA benchmark dataset.
TactfulToM: Do LLMs have the Theory of Mind ability to understand White Lies?
Yiwei Liu (EPFL), Saku Sugawara (National Institute of Informatics)
CodeTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
π― What it does: Designed and constructed the TactfulToM benchmark to evaluate large language models' understanding and reasoning abilities regarding white lies in real-world dialogues.
Taming Text-to-Image Synthesis for Novices: User-centric Prompt Generation via Multi-turn Guidance
Yilun Liu (Huawei), Osamu Yoshie (Huawei)
CodeGenerationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringDiffusion modelImageText
π― What it does: Propose a text-to-image (prompt) generation framework called DialPrompt based on multi-turn dialogue, helping novice users gradually optimize prompts and generate high-quality images through conversations.
Task-aware Contrastive Mixture of Experts for Quadruple Extraction in Conversations with Code-like Replies and Non-opinion Detection
Chenyuan He (Zhengzhou University), Min Peng (Wuhan University)
CodeTransformerLarge Language ModelMixture of ExpertsContrastive LearningText
π― What it does: Propose a multi-task framework called TaCoMoE based on large language models for extracting quadruples (goal, aspect, opinion, sentiment) from multi-turn dialogues and detecting non-opinion utterances;
Task-Aware Resolution Optimization for Visual Large Language Models
Weiqing Luo (University of North Carolina at Chapel Hill), Tianlong Chen (University of North Carolina at Chapel Hill)
CodeOptimizationLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
π― What it does: This paper addresses the varying resolution requirements of visual large language models across different tasks and proposes a task-aware resolution optimization method.
Teaching Your Models to Understand Code via Focal Preference Alignment
Jie Wu, Scarlett Li (Tsinghua University)
CodeAI Code AssistantReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningText
π― What it does: Studied a fine-grained preference alignment framework called Target-DPO based on iterative debugging, aiming to enable code generation models to more accurately learn error correction strategies.
CodeData SynthesisAdversarial AttackTransformerLarge Language ModelText
π― What it does: Propose the TempParaphraser framework, which uses multiple normal-temperature sampling to simulate high-temperature sampling to evade AI text detection.
π― What it does: This paper proposes a benchmark framework for out-of-distribution (OOD) detection in text-network (TrN) scenarios, named TextTopoOOD, and designs a novel detection model called TNTβOOD based on this framework.
The Enemy from Within: A Study of Political Delegitimization Discourse in Israeli Political Speech
Naama Rivlin-Angert (Hebrew University of Jerusalem), Guy Mor-Lan (Hebrew University of Jerusalem)
CodeClassificationRecognitionTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Studied political legitimacy weakening discourse (PDD), constructed and manually annotated 10,410 Hebrew sentences (sources include speeches from the Israeli parliament, Facebook official account posts, and news media), and proposed a two-stage detection pipeline (first identifying PDD, then predicting its intensity, aggressiveness, target type, etc.), followed by using the model for cross-platform and cross-temporal quantitative analysis.
The Role of Outgoing Connection Heterogeneity in Feedforward Layers of Large Language Models
Felix Stahlberg (Google Research), Shankar Kumar (Google Research)
CodeComputational EfficiencyTransformerLarge Language ModelText
π― What it does: This paper investigates the importance of output connection heterogeneity in feedforward layers of large language models (LLMs) and proposes improving model performance by reducing the entropy of neuron output weights;
The Strawberry Problem: Emergence of Character-level Understanding in Tokenized Language Models
Adrian Cosma (National University of Science and Technology POLITEHNICA), Mihai Dascalu (National University of Science and Technology POLITEHNICA)
CodeRepresentation LearningTransformerLarge Language ModelText
π― What it does: This study investigates the limitations of tokenization on character-level tasks in large language models, proposes 19 synthetic tasks to measure the emergence of character understanding, and subsequently designs a lightweight character-aware module to enhance character-level reasoning capabilities.
The Validation Gap: A Mechanistic Analysis of How Language Models Compute Arithmetic but Fail to Validate It
Leonardo Bertolazzi (University of Trento), Raffaella Bernardi (Free University of Bozen-Bolzano)
CodeExplainability and InterpretabilityTransformerText
π― What it does: This study systematically analyzes the mechanisms by which small LLMs detect errors in simple arithmetic problems, revealing how models identify and correct arithmetic mistakes.
Think and Recall: Layer-Level Prompting for Lifelong Model Editing
Jinke Wang (University of Science and Technology of China), Zhi Zheng (University of Science and Technology of China)
CodeTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextRetrieval-Augmented Generation
π― What it does: Propose the Layer-Level Prompting (LLP) method, achieving lifelong model editing by retrieving prompts in the early layers of LLMs and injecting them in the later layers.
Yuhao Chen (University of Science and Technology of China), Tong Xu (University of Science and Technology of China)
CodeClassificationKnowledge DistillationReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
π― What it does: This paper proposes a two-stage training framework, first obtaining high-quality Chain-of-Thought training data through teacher model distillation for supervised fine-tuning (SFT), and then further enhancing the model's reasoning completeness and consistency in the document-level self-contradiction detection (DSCD) task using GRPO reinforcement learning with custom accuracy, coverage, and format rewards.
ThinkEdit: Interpretable Weight Editing to Mitigate Overly Short Thinking in Reasoning Models
Chung-En Sun (University of California San Diego), Tsui-Wei Weng (University of California San Diego)
CodeExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelTextBenchmark
π― What it does: The study finds that large language models suffer performance degradation in chain-of-thought reasoning due to excessively short reasoning chains, and proposes ThinkEdit, a method to alleviate this issue by modifying the output projection weights of a small number of attention heads.
Threading the Needle: Reweaving Chain-of-Thought Reasoning to Explain Human Label Variation
Beiduo Chen (LMU Munich), Barbara Plank (LMU Munich)
CodeExplainability and InterpretabilityData-Centric LearningLarge Language ModelTextChain-of-Thought
π― What it does: This paper proposes a method to automatically extract explanation-label (EL) pairs by leveraging chain-of-thought (CoT) generated by large language models, and enhances explanation quality through discourse segmentation to better capture human annotation differences;
TLUE: A Tibetan Language Understanding Evaluation Benchmark
Fan Gao (University of Electronic Science and Technology of China), Hao Wang (University of Electronic Science and Technology of China)
CodeLarge Language ModelTextBenchmark
π― What it does: This paper constructs and releases the TLUE benchmark, covering multi-domain knowledge tests and safety evaluations to assess the understanding ability of large language models in Tibetan.
Token-Aware Editing of Internal Activations for Large Language Model Alignment
Tianbo Wang (Beihang University), Xianglong Liu (Beihang University)
CodeExplainability and InterpretabilityGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Proposed the Token-Aware Editing (TAE) framework, achieving inference-time alignment for large language models through fine-tuning of internal activations;
TokenSelect: Efficient Long-Context Inference and Length Extrapolation for LLMs via Dynamic Token-Level KV Cache Selection
Wei Wu (University of Science and Technology of China), Hui Xiong (Hong Kong University of Science and Technology (Guangzhou))
CodeComputational EfficiencyTransformerLarge Language ModelTextBenchmark
π― What it does: Proposes TokenSelect, a training-free dynamic token-level KV cache selection method for efficient long-text inference and length extrapolation.
ToM: Leveraging Tree-oriented MapReduce for Long-Context Reasoning in Large Language Models
Jiani Guo (Wuhan University), Yujiu Yang (Tsinghua University)
CodeLarge Language ModelTextBenchmarkChain-of-Thought
π― What it does: Propose a tree-structured MapReduce framework called ToM to enhance the logical coherence and information integration capabilities of large language models in long-text reasoning.
ToneCraft: Cantonese Lyrics Generation with Harmony of Tones and Pitches
Junyu Cheng (South China Normal University), Shuangyin Li (South China Normal University)
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
π― What it does: Propose the ToneCraft framework to achieve Cantonese lyric generation based on large models, ensuring harmony between lyrics and melody in terms of pitch and tone;
Too Consistent to Detect: A Study of Self-Consistent Errors in LLMs
Hexiang Tan (Institute of Computing Technology, Chinese Academy of Sciences), Xueqi Cheng (Institute of Computing Technology, Chinese Academy of Sciences)
CodeAnomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
π― What it does: Studied the phenomenon of self-consistent errors generated by large language models (LLMs) during multi-sample generation, evaluated existing error detection methods, and proposed a cross-model probe to enhance the detection of self-consistent errors.
Toward Efficient Sparse Autoencoder-Guided Steering for Improved In-Context Learning in Large Language Models
Ikhyun Cho (University of Illinois at Urbana-Champaign), Julia Hockenmaier (University of Illinois at Urbana-Champaign)
CodeClassificationLarge Language ModelPrompt EngineeringAuto EncoderText
π― What it does: Proposes a two-step method based on sparse autoencoders: Feature Detection through Prompt Variation (FDPV) for efficiently identifying features usable for steering; SISTER for selectively steering only on label words in ICL classification tasks, thereby enhancing model performance.
Toward Multi-Session Personalized Conversation: A Large-Scale Dataset and Hierarchical Tree Framework for Implicit Reasoning
Xintong Li (University of California San Diego), Jingbo Shang (University of California San Diego)
CodeRetrievalLarge Language ModelTextBenchmark
π― What it does: Proposed a large-scale multi-conversation dialogue dataset called IMPLEXCONV for evaluating implicit reasoning in long-term personalized dialogues; and designed a hierarchical tree retrieval framework named TACITREE to efficiently retrieve hidden implicit information buried in long conversation histories.
Inbar Pendzel (University of Haifa), Einat Minkov (University of Haifa)
CodeClassificationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningTextGraph
π― What it does: Built a user embedding space (SocialVec) based on the large-scale Twitter social network, and combined it with text representations for author-driven social NLP tasks such as stance detection and toxicity detection, further analyzing the association between author stance and socio-demographic features.
Towards General-Domain Word Sense Disambiguation: Distilling Large Language Model into Compact Disambiguator
Liqiang Ming (Shenzhen University), Yuncong Li (EnCopilot Inc.)
CodeClassificationKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringText
π― What it does: Propose to utilize large language models (LLMs) for distilling or annotating silver-standard word sense disambiguation (WSD) data, training small models to achieve general-domain WSD.
π― What it does: Investigate the feasibility of infinite-length prefixes in prefix learning, prove their convergence using NTK, and propose NTK-Attention to achieve efficient prefix approximation.
Towards Transferable Personality Representation Learning based on Triplet Comparisons and Its Applications
Kai Tang (Zhejiang University), Haobo Wang (Zhejiang University)
CodeRepresentation LearningTransformerLarge Language ModelContrastive LearningText
π― What it does: Constructed a triplet-based personality tendency comparison dataset (PTCD) and obtained measurable sentence-level personality embeddings through triplet learning;
CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
π― What it does: This paper created and made publicly available the LENS (Longitudinal English Nonnative Speaker) corpus, which includes 15 academic writing samples from graduate students with different first languages (L1), along with fine-grained error annotations, particularly L1 interference tags.
Train It and Forget It: Merge Lists are Unnecessary for BPE Inference in Language Models
Tomohiro Sawada, Kartik Goyal (Georgia Institute Of Technology)
CodeSafty and PrivacyComputational EfficiencyTransformerLarge Language ModelText
π― What it does: Explores inference algorithms for large language models that do not rely on BPE merge lists, and systematically evaluates their impact on model performance and safety.
CodeExplainability and InterpretabilityRepresentation LearningAuto EncoderText
π― What it does: Proposed the HierarchicalTopK training objective, enabling a single sparse autoencoder to maintain high interpretability and reconstruction quality under various sparsity budgets.
CodeClassificationRecognitionData SynthesisTransformerLarge Language ModelPrompt EngineeringText
π― What it does: Propose LMTransplant, a novel text data augmentation paradigm that embeds seed text into a bidirectional context generated by large language models (LLMs) and regenerates text within this context.
TreeRare: Syntax Tree-Guided Retrieval and Reasoning for Knowledge-Intensive Question Answering
Boyi Zhang (University of Rochester), Hangfeng He (University of Rochester)
CodeRetrievalLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Propose the TreeRare framework, which utilizes syntactic trees for semantic decomposition and performs sub-component retrieval and answer generation at each node, eventually aggregating to produce the final answer.
TrojanStego: Your Language Model Can Secretly Be A Steganographic Privacy Leaking Agent
Dominik Meier (University of GΓΆttingen), Bela Gipp (Bocconi University)
CodeSafty and PrivacyAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Propose the TrojanStego threat model, which leverages LLMs to embed sensitive contextual information into natural outputs through linguistic steganography, achieving data leakage without prompts or explicit control;
π― What it does: Proposed a framework that combines prompt learning and small model fine-tuning, enhancing data by retrieving similar context ICE (In-Context Examples) to achieve better context awareness and preference alignment in machine translation (MT) and text style transfer (TST) tasks.
TurBLiMP: A Turkish Benchmark of Linguistic Minimal Pairs
Ezgi BaΕar (University of Groningen), Arianna Bisazza (University of Groningen)
CodeTextBenchmark
π― What it does: This paper proposes TurBLiMP, a Turkish language model evaluation benchmark covering 16 grammatical phenomena, with 1,000 minimal sentence pairs for each phenomenon, and expands lexical diversity through manual and semi-automated methods, further incorporating experimental paradigms for word order and subclauses.
π― What it does: Propose TurboRAG by precomputing KV caches for each retrieved segment and concatenating them during inference, reducing prefill computation and latency in RAG systems.
Turning Logic Against Itself: Probing Model Defenses Through Contrastive Questions
Rachneet Singh Sachdeva (Technical University of Darmstadt), Iryna Gurevych (Technical University of Darmstadt)
CodeAdversarial AttackLarge Language ModelPrompt EngineeringTextChain-of-Thought
π― What it does: Propose the POATE two-phase adversarial generation method, which utilizes opposite-polarity questions and adversarial templates to induce LLMs to generate harmful responses, and designs two chain-of-thought (CoT) defenses: Intent-Aware CoT and Reverse Thinking CoT.
TVQACML: Benchmarking Text-Centric Visual Question Answering in Multilingual Chinese Minority Languages
Sha Jiu (Minzu University of China), Jialedongzhu (Minzu University of China)
CodeTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmark
π― What it does: Constructed a text-centric visual question answering benchmark TVQACML for multilingual Chinese minority languages, containing 8,000 real images and 32,000 high-quality QA pairs, covering 8 languages and 30 scenarios.
Type-Less yet Type-Aware Inductive Link Prediction with Pretrained Language Models
Alessandro De Bellis, Eugenio Di Sciascio (Politecnico di Bari)
CodeRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringGraph
π― What it does: Propose an inductive link prediction model called TyleR that captures implicit types without explicit type information, enhancing node semantic representations through pre-trained language models.
UICOMPASS: UI Map Guided Mobile Task Automation via Adaptive Action Generation
Yuanzhang Lin (Beihang University), Hailong Sun (Beihang University)
CodeTransformerLarge Language ModelAgentic AIGraph
π― What it does: Proposes UICOMPASS, a framework that automatically generates UI maps using static analysis and LLM, and achieves mobile task automation through adaptive path planning
UltraIF: Advancing Instruction Following from the Wild
Kaikai An (Peking University), Baobao Chang (Peking University)
CodeData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― What it does: Designed the ULTRAIF method, which first uses LLM to decompose real user instructions into simplified instructions, constraints, and corresponding evaluation questions; then trains UltraComposer to generate complex instructions with constraints in one go; subsequently uses a generate-evaluate process combined with preference learning (SFT+DPO/NC) to build high-quality instruction-following data; finally achieves performance comparable to the Instruct version using an 8B base model.
UNCERTAINTY-LINE: Length-Invariant Estimation of Uncertainty for Large Language Models
Roman Vashurin (Mohamed bin Zayed University of Artificial Intelligence), Maxim Panov (Mohamed bin Zayed University of Artificial Intelligence)
CodeGenerationTransformerLarge Language ModelText
π― What it does: Propose a post-hoc debiasing method called UNCERTAINTY-LINE to eliminate the impact of length bias in the generated text of large language models on uncertainty estimation.
UNComp: Can Matrix Entropy Uncover Sparsity? β A Compressor Design from an Uncertainty-Aware Perspective
Jing Xiong (University of Hong Kong), Ngai Wong (University of Hong Kong)
CodeCompressionComputational EfficiencyTransformerLarge Language ModelText
π― What it does: Propose the UNCOMP framework, which employs an uncertainty-aware method based on truncated matrix entropy to dynamically perform two-stage compression on the KV cache and hidden states of LLMs, achieving significant acceleration and memory savings for long-context reasoning.
Understanding and Mitigating Overrefusal in LLMs from an Unveiling Perspective of Safety Decision Boundary
Licheng Pan (Zhejiang University), Zhixuan Chu (Zhejiang University)
CodeSafty and PrivacyExplainability and InterpretabilityRepresentation LearningLarge Language ModelReinforcement LearningPrompt EngineeringMixture of ExpertsTextBenchmark
π― What it does: This paper investigates the overrefusal phenomenon of large language models (LLMs) on legitimate queries, proposing the RASS framework based on representation learning. By analyzing the safety decision boundary, it generates and filters overrefusal prompts closest to the boundary, further constructing a multilingual overrefusal benchmark, MORBENCH, to evaluate models' rejection behaviors across different languages and safety categories.
Understanding the Information Propagation Effects of Communication Topologies in LLM-based Multi-Agent Systems
Xu Shen (Jilin University), Xin Wang (Jilin University)
CodeOptimizationExplainability and InterpretabilityGraph Neural NetworkLarge Language ModelReinforcement LearningAgentic AIText
π― What it does: A causal analysis-based LLM multi-agent system communication topology optimization framework named EIB-LEARNER is proposed and studied.
Understanding the Thinking Process of Reasoning Models: A Perspective from Schoenfeldβs Episode Theory
Ming Li (University of Maryland), Tianyi Zhou (University of Maryland)
CodeExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
π― What it does: Analyze the chain-of-thought processes of large-scale reasoning models, construct a fine-grained annotated dataset based on Schoenfeld's theory of mathematical problem-solving, and publicly release the annotation protocol.
UnifiedVisual: A Framework for Constructing Unified Vision-Language Datasets
Pengyu Wang (Fudan University), Xipeng Qiu (Fudan University)
CodeData SynthesisTransformerLarge Language ModelVision Language ModelDiffusion modelContrastive LearningImageTextMultimodalityChain-of-Thought
π― What it does: Proposed the UnifiedVisual framework for constructing datasets that simultaneously support multi-modal understanding and generation, and built the UnifiedVisual-240K dataset based on this, helping to mutually reinforce VLLM in visual understanding and generation tasks;
Unilaw-R1: A Large Language Model for Legal Reasoning with Reinforcement Learning and Iterative Inference
Hua Cai (UniDT), Tianke Ban (Fudan University)
CodeExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextBenchmarkChain-of-Thought
π― What it does: Built a 7B-parameter legal reasoning large language model called Unilaw-R1, enhancing legal reasoning accuracy and interpretability through two-stage training (supervised fine-tuning + reinforcement learning) and multi-agent iterative reasoning.
UnitCoder: Scalable Code Synthesis from Pre-training Corpora
Yichuan Ma (Fudan University), Kai Chen (Shanghai AI Laboratory)
CodeData SynthesisAI Code AssistantTransformerLarge Language ModelAgentic AIText
π― What it does: Propose the UnitCoder framework, which uses automatically generated unit tests to supervise the quality of pre-trained code libraries, and generates over 500K executable high-quality code data through an iterative repair and refinement process.
Unlearning vs. Obfuscation: Are We Truly Removing Knowledge?
Guangzhi Sun (University of Cambridge), Mark Gales (University of Cambridge)
CodeSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper proposes a method to distinguish LLM's 'true unlearning' from 'obfuscation' from an uncertainty perspective, and designs probing questions (open-ended, Yes/No, MCQ) automatically generated to evaluate the differences between the two. Subsequently, a new unlearning technique called DF-MCQ is proposed, which achieves the effect of knowledge 'erasure' by normalizing the answer distribution of multiple-choice questions (MCQ) to a uniform distribution using KL-divergence.
Unmasking Deceptive Visuals: Benchmarking Multimodal Large Language Models on Misleading Chart Question Answering
Zixin Chen (Hong Kong University of Science and Technology), Huamin Qu (Hong Kong University of Science and Technology)
CodeLarge Language ModelMultimodalityTabularBenchmarkChain-of-Thought
π― What it does: Created the Misleading ChartQA benchmark dataset (3,026 multimodal QA instances) and evaluated 24 mainstream multimodal large language models on this dataset; proposed the Region-Aware Misleader Reasoning (RAMR) pipeline to enhance models' reasoning capabilities for misleading charts.
V-SEAM: Visual Semantic Editing and Attention Modulating for Causal Interpretability of Vision-Language Models
Qidong Wang (Tongji University), Ming Jiang (University of Wisconsin-Madison)
CodeExplainability and InterpretabilityTransformerPrompt EngineeringVision Language ModelContrastive LearningImageMultimodality
π― What it does: Propose the V-SEAM framework, combining visual semantic editing and attention modulation to perform causal interpretability analysis on vision-language models and enhance VQA performance
VC4VG: Optimizing Video Captions for Text-to-Video Generation
Yang Du (Renmin University of China), Qin Jin (Renmin University of China)
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoTextMultimodalityBenchmark
π― What it does: Propose the VC4VG framework, which decomposes video subtitles into five dimensions (subject attributes, environment, actions, camera parameters, style) and optimizes them to enhance the training effectiveness of text-to-video models.
VerIF: Verification Engineering for Reinforcement Learning in Instruction Following
Hao Peng (Tsinghua University), Juanzi Li (Tsinghua University)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
π― What it does: Propose VERIF, a hybrid approach combining rule-based code verification with large inference models (e.g., QwQ-32B) verification, applied to instruction-following reinforcement learning;
Vision-and-Language Navigation with Analogical Textual Descriptions in LLMs
Yue Zhang (Michigan State University), Parisa Kordjamshidi (Michigan State University)
CodeAutonomous DrivingTransformerLarge Language ModelPrompt EngineeringVision Language ModelVision-Language-Action ModelImageTextMultimodalityBenchmark
π― What it does: Introduce an analogy reasoning module in vision-and-language navigation (VLN) driven by large language models, enhancing contextual understanding for navigation decisions by comparing multi-perspective images to generate distinctive scene and spatial descriptions.
VistaWise: Building Cost-Effective Agent with Cross-Modal Knowledge Graph for Minecraft
Honghao Fu (Hong Kong University of Science and Technology (Guangzhou)), Hao Wang (Hong Kong University of Science and Technology (Guangzhou))
CodeObject DetectionConvolutional Neural NetworkLarge Language ModelSupervised Fine-TuningVision-Language-Action ModelVideoTextMultimodalityGraphRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Built a cost-effective agent framework named VistaWise in Minecraft, leveraging cross-modal knowledge graphs, lightweight object detection, and desktop-level mouse/keyboard skill libraries to accomplish complex tasks using only hundreds of frames of data.
VLASCD: A Visual Language Action Model for Simultaneous Chatting and Decision Making
Zuojin Tang (Zhejiang University), Bin Liu (E-surfing Digital Life Technology Co., Ltd.)
CodeAutonomous DrivingTransformerLarge Language ModelSupervised Fine-TuningVision-Language-Action ModelVideoText
π― What it does: Propose a multi-input multi-output visual-language-action model called MIMO-VLA, capable of simultaneously performing natural language dialogue and real-time driving decision-making.
π― What it does: Built the SimbaBench benchmark, aggregating and unifying 8,605+ hours of speech data across 61 African languages, followed by fine-tuning the Simba series of ASR, TTS, and SLID models on this benchmark, and systematically evaluating their performance.
VoiceBBQ: Investigating Effect of Content and Acoustics in Social Bias of Spoken Language Model
Junhyuk Choi (Chung-Ang University), Bugeun Kim (Chung-Ang University)
CodeTransformerLarge Language ModelTextBenchmarkAudio
π― What it does: This study proposes the VoiceBBQ dataset to evaluate social bias in speech language models from both content and acoustic perspectives;
VRoPE: Rotary Position Embedding for Video Large Language Models
Zikang Liu (Institute of Automation Chinese Academy of Sciences), Jing Liu (Institute of Automation Chinese Academy of Sciences)
CodeRepresentation LearningTransformerLarge Language ModelVision Language ModelVideoText
π― What it does: Proposed a Video Rotation Position Encoding (VRoPE) for video large language models to better capture the spatiotemporal structure of videos and avoid position bias.
CodeComputational EfficiencyLarge Language ModelVision Language ModelVision-Language-Action ModelMultimodality
π― What it does: This paper proposes a multi-modal token pruning framework called Navigation-Aware Pruning (NAP), which integrates Background View Pruning (BGP), Backtracking Node Pruning (BTP), and Vocabulary-based Instruction Pruning (VPP). It significantly reduces the number of visual, textual, and historical node tokens in Vision-and-Language Navigation (VLN) tasks while maintaining or even improving navigation success rates.
Weaver: Interweaving SQL and LLM for Table Reasoning
Rohit Khoja (Arizona State University), Vivek Gupta (Arizona State University)
CodeExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringMultimodalityTabularFinance Related
π― What it does: Proposed Weaver, a modular and interpretable framework that dynamically switches between SQL and LLM to address table QA tasks involving mixed structured and unstructured data.
WebEvolver: Enhancing Web Agent Self-Improvement with Co-evolving World Model
Tianqing Fang (Tencent AI Lab), Dong Yu (Tencent AI Lab)
CodeTransformerLarge Language ModelSupervised Fine-TuningAgentic AIWorld ModelTextBenchmark
π― What it does: By constructing a self-improving framework called WebEvolver, which co-trains an agent model with a co-evolving world model, enabling Web agents to continuously enhance their performance with the help of self-generated trajectories and multi-step simulation reasoning.
Weight-Aware Activation Sparsity with Constrained Bayesian Optimization Scheduling for Large Language Models
Ming Wang (Harbin Institute of Technology), Liqiang Nie (Harbin Institute of Technology)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelText
π― What it does: Propose an untrained activation sparsity framework WAS, which uses weight information and activation values together to determine thresholds, and employs constrained Bayesian optimization to assign different sparsity rates to each Transformer block, while achieving a GPU kernel that supports non-uniform sparsity, significantly accelerating LLM inference.
What Do Indonesians Really Need from Language Technology? A Nationwide Survey
Muhammad Dehan Al Kautsar (Mohamed bin Zayed University of Artificial Intelligence), Fajri Koto (Mohamed bin Zayed University of Artificial Intelligence)
CodeRetrievalSafty and PrivacyTextReview/Survey Paper
π― What it does: A nationwide questionnaire was conducted, collecting 861 responses from users of over 700 languages in Indonesia regarding their needs and concerns about language technologies.
When Big Models Train Small Ones: Label-Free Model Parity Alignment for Efficient Visual Question Answering using Small VLMs
Abhirama Subramanyam Penamakuri (Indian Institute of Technology Jodhpur), Anand Mishra (Indian Institute of Technology Jodhpur)
CodeKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringVision Language ModelMultimodality
π― What it does: Developed a Model Parity Aligner (MPA) framework that leverages unlabeled images and high-quality pseudo-labels generated by a large-scale visual language model (L-VLM) to specifically enhance the visual question answering (VQA) performance of small-scale visual language models (S-VLM).
When Truthful Representations Flip Under Deceptive Instructions?
Xianxuan Long (Case Western Reserve University), Pan Li (Case Western Reserve University)
CodeExplainability and InterpretabilityRepresentation LearningLarge Language ModelPrompt EngineeringAuto EncoderText
π― What it does: Studied changes in the internal representations of LLMs under deceptive instructions, and analyzed differences between honest and deceptive patterns through linear probes and sparse autoencoders.
Where Confabulation Lives: Latent Feature Discovery in LLMs
Thibaud Ardoin (Freie UniversitΓ€t Berlin), Gerhard Wunder (Freie UniversitΓ€t Berlin)
CodeExplainability and InterpretabilityTransformerPrompt EngineeringText
π― What it does: This paper extracts and re-projects a sparse, interpretable 'confabulation' direction from the LLM activation space by comparing dialog prompts of known and unknown entities, using Sparse Principal Component Analysis (SPCA), and injects this vector into intermediate layers to achieve controllable regulation of the model's confabulation behavior.
Why Do Some Inputs Break Low-Bit LLM Quantization?
Ting-Yun Chang (University of Southern California), Robin Jia (University of Southern California)
CodeExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
π― What it does: Investigate why 3-4 bit weight quantization in large language models leads to significant errors on certain inputs and systematically analyze the sources and mechanisms of these errors.
WildScore: Benchmarking MLLMs in-the-Wild Symbolic Music Reasoning
Gagan Mundada (University of California, San Diego), Junda Wu (University of California, San Diego)
CodeLarge Language ModelVision Language ModelMultimodalityBenchmark
π― What it does: Proposed the WildScore benchmark, constructing a multimodal symbolic music reasoning dataset using real sheet music and Reddit community questions, and evaluated the performance of large multimodal language models.
Will It Still Be True Tomorrow? Multilingual Evergreen Question Classification to Improve Trustworthy QA
Sergey Pletenev (Skoltech), Viktor Moskvoretskii (EPFL)
CodeClassificationExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
π― What it does: This paper constructs a multilingual sustainability issue classification dataset called EverGreenQA, trains and evaluates a lightweight multilingual classifier EG-E5, and applies sustainability discrimination to self-knowledge estimation, QA dataset filtering, and GPT-4o retrieval behavior explanation.
π― What it does: Proposes the XAutoLM framework, which achieves efficient and resource-friendly model and hyperparameter search for fine-tuning language models through meta-learning-driven AutoML;
XLQA: A Benchmark for Locale-Aware Multilingual Open-Domain Question Answering
Keonwoo Roh, Seong-Whan Lee (Korea University)
CodeLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
π― What it does: Constructed a multilingual open-domain QA benchmark named XLQA tailored for regional differences, containing 3,000 English seed questions and their expanded versions in eight languages (English, Korean, Arabic, Hebrew, Japanese, Russian, Vietnamese, Simplified Chinese), with human-validated annotations of answer regional sensitivity.
You Are What You Train: Effects of Data Composition on Training Context-aware Machine Translation Models
PaweΕ MΔ ka (Maastricht University), Gerasimos Spanakis (Maastricht University)
CodeData-Centric LearningTransformerText
π― What it does: Investigated the impact of the sparsity of context-rich examples in training data on the performance of context-aware machine translation models and systematically verified the sparse hypothesis.
Your Language Model Can Secretly Write Like Humans: Contrastive Paraphrase Attacks on LLM-Generated Text Detectors
Hao Fang (Tsinghua University), Min Zhang (Harbin Institute of Technology)
CodeAdversarial AttackTransformerLarge Language ModelContrastive LearningText
π― What it does: Propose a training-free adversarial rewriting method called CoPA, which utilizes contrastive word distribution adjustment to generate text that better aligns with human writing characteristics, thereby evading LLM text detectors.
ZERA: Zero-init Instruction Evolving Refinement Agent β From Zero Instructions to Structured Prompts via Principle-based Optimization
Seungyoun Yi, Sungrae Park (Upstage AI Research)
CodeOptimizationTransformerLarge Language ModelPrompt EngineeringText
π― What it does: Propose the ZERA framework to achieve zero-initialization automatic prompt optimization, evaluating and improving system prompts, user prompts, and task descriptions through eight principles;
π― What it does: Propose a tree search algorithm called Zoom Eye, enabling multi-modal large language models to simulate human zooming operations during reasoning to acquire image details;