EMNLP 2025 Papers — Page 8
Conference on Empirical Methods in Natural Language Processing · 1809 papers
How do autoregressive transformers solve full addition?
Wang Peixu (National University of Singapore), Cheng Xiang (National University of Singapore)
Explainability and InterpretabilityTransformerSupervised Fine-TuningTextSequential
🎯 What it does: This paper systematically investigates the internal mechanisms of autoregressive Transformers in integer addition tasks, using causal analysis to identify attention heads that control carry operations, and improving model accuracy for long-sequence addition through inference intervention and fine-tuning.
How do Language Models Reshape Entity Alignment? A Survey of LM-Driven EA Methods: Advances, Benchmarks, and Future
Zerui Chen (Harbin Institute Of Technology), Bing Qin (Harbin Institute Of Technology)
Representation LearningTransformerLarge Language ModelPrompt EngineeringGraphReview/Survey PaperBenchmark
🎯 What it does: Systematically reviewed language model-driven entity alignment methods and proposed a three-stage classification framework.
How Do Large Vision-Language Models See Text in Image? Unveiling the Distinctive Role of OCR Heads
Ingeol Baek (Chung-Ang University), Hwanhee Lee (Chung-Ang University)
RecognitionRetrievalTransformerVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: This paper analyzes the attention heads in large vision-language models (LVLM) and first identifies OCR heads specialized for recognizing and extracting text information from images, while investigating their differences from traditional retrieval heads.
How Do Social Bots Participate in Misinformation Spread? A Comprehensive Dataset and Analysis
Herun Wan (Xi'an Jiaotong University), Xiang Zhao (National University of Defense Technology)
Anomaly DetectionTransformerLarge Language ModelMixture of ExpertsMultimodalityBenchmark
🎯 What it does: Constructed the MISBOT dataset and analyzed the role of social bots in the spread of misinformation.
How Does DPO Reduce Toxicity? A Mechanistic Neuron-Level Analysis
Yushi Yang (University of Oxford), Adam Mahdi (University of Oxford)
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper provides an in-depth analysis of the internal mechanisms of Direct Preference Optimization (DPO) in reducing toxicity in large language models, revealing that toxicity reduction is primarily driven by distributed activation displacement rather than the suppression of a few toxic neurons; based on this, it proposes an activation editing method without weight updates, achieving more efficient toxicity suppression.
How Far Can LLMs Improve from Experience? Measuring Test-Time Learning Ability in LLMs with Human Comparison
Jiayin Wang (Tsinghua University), Min Zhang (Tsinghua University)
Meta LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: Investigated the test-time learning capability of large language models, designed an evaluation framework based on semantic games, and compared it with human performance.
How Is LLM Reasoning Distracted by Irrelevant Context? An Analysis Using a Controlled Benchmark
Minglai Yang (University of Arizona), Liangming Pan (Peking University)
Explainability and InterpretabilityData-Centric LearningLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Constructed a controllable irrelevant context math reasoning benchmark GSM-DC and systematically evaluated the robustness of LLMs.
How Much Do LLMs Hallucinate across Languages? On Realistic Multilingual Estimation of LLM Hallucination
Saad Obaid Ul Islam, Goran Glavaš (University of Würzburg)
ClassificationAnomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: This paper constructs a hallucination detection model and evaluation set across 30 languages to quantify the hallucination rate of large language models (LLMs) in knowledge-intensive open-ended long-form question-answering scenarios, and compares the performance of 11 open-source LLMs.
How Persuasive Is Your Context?
Tu Nguyen (ETH Zürich), Ryan Cotterell (ETH Zürich)
Explainability and InterpretabilityText
🎯 What it does: This paper proposes a Target Persuasion Score (TPS) based on Wasserstein distance to quantify the impact of context on the answer distribution of language models.
How Private are Language Models in Abstractive Summarization?
Anthony Hughes (University of Sheffield), Ning Ma (University of Sheffield)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextElectronic Health RecordsChain-of-Thought
🎯 What it does: The study evaluates the risk of Personally Identifiable Information (PII) leakage when large language models generate abstract summaries in sensitive domains such as healthcare and law, and explores privacy protection effectiveness by comparing various prompting strategies and instruction fine-tuning methods.
How Sememic Components Can Benefit Link Prediction for Lexico-Semantic Knowledge Graphs?
Hansi Wang (Peking University), Yang Liu (Peking University)
TransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextGraphBenchmark
🎯 What it does: This paper proposes a method to enhance link prediction in lexico-semantic knowledge graphs (Lexico‑Semantic KG) by leveraging semantic units (sememes), and constructs the corresponding sememe prediction dataset SememeDef as well as Chinese link prediction benchmarks HN7 and CWN5.
How to inject knowledge efficiently? Knowledge Infusion Scaling Law for Pre-training Large Language Models
Kangtao Lv (Zhejiang University), Bo Zheng (Taobao & Tmall Group of Alibaba)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper systematically explores how to effectively inject domain knowledge during the pre-training phase of large language models while avoiding memory collapse phenomena.
How to Make Large Language Models Generate 100% Valid Molecules?
Wen Tao (Nanyang Technological University), Yiwei Wang (University of California Merced)
Drug DiscoveryTransformerLarge Language ModelDiffusion modelTextBiomedical Data
🎯 What it does: This paper proposes a cross-chemical language framework called SmiSelf, which leverages the robustness of SELFIES to convert invalid SMILES generated by LLMs into valid SMILES, achieving 100% valid molecule generation. It also systematically evaluates the performance of SELFIES and SMILES within LLMs, as well as the self-correction capability of LLMs.
How to Protect Yourself from 5G Radiation? Investigating LLM Responses to Implicit Misinformation
Ruohao Guo (Georgia Institute of Technology), Alan Ritter (Georgia Institute of Technology)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This study systematically evaluates the performance of large language models (LLMs) when facing implicit misinformation, constructs the first benchmark dataset for implicit misinformation called ECHOMIST, and proposes two simple mitigation methods (Self-Alert and RAG).
HS-STaR: Hierarchical Sampling for Self-Taught Reasoners via Difficulty Estimation and Budget Reallocation
Feng Xiong (AMAP, Alibaba Group), Xiangxiang Chu (AMAP, Alibaba Group)
OptimizationReinforcement Learning from Human FeedbackSupervised Fine-TuningReinforcement LearningContrastive LearningTextBenchmark
🎯 What it does: Built a hierarchical sampling framework called HS-STAR, which dynamically reallocates the sampling budget using reward-guided difficulty estimation, focusing on mathematical problems at the boundary of model capabilities;
Humanizing Machines: Rethinking LLM Anthropomorphism Through a Multi-Level Framework of Design
Yunze Xiao (Carnegie Mellon University), Mona T. Diab (Carnegie Mellon University)
Large Language ModelReview/Survey Paper
🎯 What it does: Proposes a multi-dimensional anthropomorphic design framework, treating anthropomorphism as an adjustable design parameter rather than a mere risk issue
Humans Hallucinate Too: Language Models Identify and Correct Subjective Annotation Errors With Label-in-a-Haystack Prompts
Georgios Chochlakis (University of Southern California), Shrikanth Narayanan (University of Southern California)
Data-Centric LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: The study proposes utilizing large language models (LLMs) through Label-in-a-Haystack prompts to identify and correct annotation errors in subjective tasks.
HVGuard: Utilizing Multimodal Large Language Models for Hateful Video Detection
Yiheng Jing (Wuhan University), Hongxin Hu (University at Buffalo)
ClassificationTransformerLarge Language ModelMixture of ExpertsVision Language ModelVideoTextMultimodalityChain-of-ThoughtAudio
🎯 What it does: Designed an anti-abuse video detection framework HVGuard based on multi-modal large language models (MLLM), generating reasoning processes via chain-of-thought (CoT) and fusing multi-modal features using Mixture-of-Experts (MoE).
HYDRA: A Multi-Head Encoder-only Architecture for Hierarchical Text Classification
Fabian Karl (University of Ulm), Ansgar Scherp (University of Ulm)
ClassificationTransformerText
🎯 What it does: Propose HYDRA, a multi-head encoder-only architecture for hierarchical text classification, directly treating each level as an independent classification task while sharing the underlying encoder;
HydraOpt: Navigating the Efficiency-Performance Trade-off of Adapter Merging
Taha Ceritli (Samsung R&D Institute UK), Umberto Michieli (Samsung R&D Institute UK)
OptimizationComputational EfficiencyTransformerText
🎯 What it does: Propose HydraOpt, a controllable model merging method that balances storage space and performance across multi-task LoRA adapters.
HydraRAG: Structured Cross-Source Enhanced Large Language Model Reasoning
Xingyu Tan (University of New South Wales), Wenjie Zhang (University of New South Wales)
Explainability and InterpretabilityGraph Neural NetworkLarge Language ModelTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes HydraRAG, an untrained augmented generation framework based on graph structure, document semantics, and source reliability, achieving deep trustworthy reasoning in large language models through proxy-driven multi-hop, multi-entity retrieval and three-factor cross-source verification.
HypER: Literature-grounded Hypothesis Generation and Distillation with Provenance
Rosni Vasu (University of Zurich), Abraham Bernstein (University of Zurich)
GenerationKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningTextBiomedical DataRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed and implemented HypER, a small language model for argument chain verification and evidence-based hypothesis generation in medical literature.
HyperKGR: Knowledge Graph Reasoning in Hyperbolic Space with Graph Neural Network Encoding Symbolic Path
Lihui Liu (Wayne State University)
Representation LearningGraph Neural NetworkGraph
🎯 What it does: Propose a hyperbolic space-based graph neural network framework called HyperKGR, which achieves knowledge graph reasoning using query-specific hyperbolic embeddings and hierarchical message passing.
ICG: Improving Cover Image Generation via MLLM-based Prompting and Personalized Preference Alignment
Zhipeng Bian (Huazhong University of Science and Technology), Zhenhua Dong (Huawei Noah's Ark Lab)
GenerationRecommendation SystemTransformerLarge Language ModelReinforcement LearningPrompt EngineeringDiffusion modelImageTextMultimodality
🎯 What it does: Propose an end-to-end framework called ICG, which leverages multimodal large language models to extract semantic contexts from titles and reference images, generates personalized cover images by combining user embeddings, and achieves unsupervised optimization through multi-reward learning.
ICL CIPHERS: Quantifying ”Learning” in In-Context Learning via Substitution Ciphers
Zhouxiang Fang (Johns Hopkins University), Daniel Khashabi (Johns Hopkins University)
ClassificationLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: The study evaluates the 'learning' capability of large language models during reasoning by using reversible substitution ciphers (ICL CIPHERS) in in-context learning;
Icon^2: Aligning Large Language Models Using Self-Synthetic Preference Data via Inherent Regulation
Qiyuan Chen (Zhejiang University), Jian Wu (State Key Laboratory Of Transvascular Implantation Devices And Tidri)
Computational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningContrastive LearningText
🎯 What it does: By extracting hierarchical direction vectors from the LLM representation space, using intrinsic consistency to filter self-synthesized instructions, and generating corresponding acceptance and rejection responses during decoding with bidirectional intrinsic control, thus constructing a high-quality preference dataset without requiring human annotation or multiple generations, improving model alignment with human preferences;
ICR: Iterative Clarification and Rewriting for Conversational Search
Zhiyu Cao (Soochow University), Qiaoming Zhu (Soochow University)
RetrievalLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed the Iterative Clarification and Rewriting (ICR) framework, which incrementally rewrites dialogue queries by generating clarification questions to enhance retrieval performance.
Identification of Multiple Logical Interpretations in Counter-Arguments
Wenzhi Wang (Tohoku University), Kentaro Inui (MBZUAI)
ClassificationExplainability and InterpretabilityLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper proposes a new multi-interpretation counting method, which performs fine-grained splitting and multi-annotation of logical structures in refutation arguments (CA), constructs the CALSA+ dataset, and trains models to identify these multiple logical explanations.
Identifying & Interactively Refining Ambiguous User Goals for Data Visualization Code Generation
Mert Inan, Malihe Alikhani (Northeastern University)
AI Code AssistantTransformerLarge Language ModelPrompt EngineeringMultimodality
🎯 What it does: This paper addresses the task of data visualization code generation by proposing a multi-modal ambiguity classification system, corresponding automatic evaluation metrics, and employing a multi-round dialogue strategy based on pragmatics to reduce ambiguity, thereby improving the accuracy of generated code.
Identifying and Answering Questions with False Assumptions: An Interpretable Approach
Zijie Wang (University of Arizona), Eduardo Blanco (University of Arizona)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes an interpretable method to address question-answering problems containing erroneous assumptions. It first generates and verifies the atomic hypotheses of the question, then provides answers pointing out the erroneous assumptions, combined with retrieval-enhanced evidence to reduce hallucinations in large language models.
Identifying Pre-training Data in LLMs: A Neuron Activation-Based Detection Framework
Hongyi Tang (Hong Kong University of Science and Technology), Yi Yang (Hong Kong University of Science and Technology)
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose a pre-trained data detection algorithm based on neuronal activation differences, NA-PDD, and construct a time-drift-free benchmark, CCNewsPDD.
Identifying Unlearned Data in LLMs via Membership Inference Attacks
Advit Deepak (Stanford University), Diyi Yang (Stanford University)
Adversarial AttackTransformerLarge Language ModelText
🎯 What it does: Proposes an evaluation framework called FUMA for assessing approximate forgetting operations in large language models, detecting whether identifiable traces remain after the model forgets specific samples.
Idiosyncratic Versus Normative Modeling of Atypical Speech Recognition: Dysarthric Case Studies
Vishnu Raja (Stony Brook University), H. Schwartz
RecognitionTransformerLarge Language ModelSupervised Fine-TuningAudio
🎯 What it does: This study compares four adaptive strategies (normalization, personalization, suppressive normalization, and suppressive personalization) to enhance the performance of automatic speech recognition (ASR) for individuals with speech disorders
IG-Pruning: Input-Guided Block Pruning for Large Language Models
Kangyu Qiao (Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences), Yang Feng (Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences)
Computational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes an input-feature-based block-level pruning method called IG-Pruning, which can dynamically select execution paths for Transformer layers during inference, significantly reducing computational costs for large language models.
Igniting Creative Writing in Small Language Models: LLM-as-a-Judge versus Multi-Agent Refined Rewards
Xiaolong Wei (Beihang University), Dawei Yin (Baidu Inc)
GenerationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningGenerative Adversarial NetworkText
🎯 What it does: This paper achieves creative writing of Chinese greetings on a 7-billion parameter small language model using the RLAIF method.
IIET: Efficient Numerical Transformer via Implicit Iterative Euler Method
Xinyu Liu (Northeastern University), JingBo Zhu
Computational EfficiencyKnowledge DistillationTransformerTextOrdinary Differential Equation
🎯 What it does: Proposes an Iterative Implicit Euler Transformer (IIET) that simplifies high-order numerical solutions through iterative implicit Euler methods, further introducing Iteration Influence-Aware Distillation (IIAD) to achieve model compression and inference acceleration.
iKnow-audio: Integrating Knowledge Graphs with Audio-Language Models
Michel Olvera (LTCI, Telecom Paris, Institut Polytechnique de Paris), Slim Essid (NVIDIA)
ClassificationGraph Neural NetworkTransformerPrompt EngineeringGraphAudio
🎯 What it does: Propose the iKnow-audio framework, combining knowledge graphs with the CLAP audio-text model to achieve zero-shot audio classification;
IL-PCSR: Legal Corpus for Prior Case and Statute Retrieval
Shounak Paul (IIT Kharagpur), Ashutosh Modi (IIT Kanpur)
RetrievalGraph Neural NetworkLarge Language ModelTextBenchmark
🎯 What it does: This paper constructs the first parallel corpus in the Indian legal domain, IL-PCSR, simultaneously used for retrieving applicable legal articles and previous case law, and conducts systematic research on retrieval tasks based on this corpus.
Image Difference Captioning via Adversarial Preference Optimization
Zihan Huang (UC San Diego), Julian McAuley (UC San Diego)
GenerationLarge Language ModelReinforcement LearningVision Language ModelMultimodalityRetrieval-Augmented Generation
🎯 What it does: Proposes the Adversarial Direct Preference Optimization (ADPO) framework for image difference caption generation, significantly enhancing the model's ability to capture subtle visual differences.
Image Embedding Sampling Method for Diverse Captioning
Sania Waheed (University of Southampton), Na Min An (KAIST)
SegmentationGenerationTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: Proposed the HBoP framework, which samples structured segmentation image embeddings from the last layer of the pre-trained visual model ViT, and generates multi-level (global, regional, fine-grained) diverse descriptions using BLIP;
Implicit Behavioral Alignment of Language Agents in High-Stakes Crowd Simulations
Yunzhe Wang (University of Southern California), Volkan Ustun (USC Institute for Creative Technologies)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AITabular
🎯 What it does: Developed the PEBA-PEvo framework, utilizing LLM-driven generative agents in high-risk scenarios (active shooter events) to iteratively optimize agent personas, enabling group behavior distributions to approach expert-defined real-world distributions.
Implicit Values Embedded in How Humans and LLMs Complete Subjective Everyday Tasks
Arjun Arunasalam (Florida International University), Blase Ur (University of Chicago)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Audit the implicit value performance of six mainstream LLMs across 30 daily tasks, and compare them with 100 American crowdsourced workers
ImpliRet: Benchmarking the Implicit Fact Retrieval Challenge
Zeinab Sadat Taghavi (LMU), Hinrich Schuetze (LMU)
RetrievalTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed the IMPLIRET benchmark to evaluate the performance of retrieval systems in implicit fact retrieval on the document side.
Improbable Bigrams Expose Vulnerabilities of Incomplete Tokens in Byte-Level Tokenizers
Eugene Jang (Northeastern University), Seungwon Shin (KAIST)
Safty and PrivacyAdversarial AttackLarge Language ModelPrompt EngineeringText
🎯 What it does: Investigated the impact of incomplete tokens in byte-level BPE tokenizers on hallucinations generated by large models, and proposed using 'improbable bigrams' constructed from incomplete tokens as an attack method. Experiments demonstrated that even well-trained tokens can cause models to fail in accurately repeating inputs.
Improve LLM-as-a-Judge Ability as a General Ability
Jiachen Yu (Tsinghua University), Xuelong Li (China Telecom)
Data SynthesisTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: Propose a two-phase training framework (SFT Warm-Up + DPO Enhancement) combined with an efficient data synthesis method to train a Judge model capable of evaluating LLM outputs in a generative manner, further enhancing the model's general capabilities.
Improving Chemical Understanding of LLMs via SMILES Parsing
Yunhui Jang (Korea Advanced Institute Of Science And Technology), Sungsoo Ahn (Korea Advanced Institute Of Science And Technology)
Drug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningBiomedical Data
🎯 What it does: This paper proposes the CLEANMOL framework, which enhances LLMs' understanding of molecular structures by designing five SMILES-based parsing tasks to construct an expandable and deterministic molecular pre-training dataset.
Improving Clustering with Positive Pairs Generated from LLM-Driven Labels
Xiaotong Zhang, Ying Li (Zhejiang University)
Representation LearningLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose the PPLL framework: construct positive sample pairs using micro-cluster labels generated by LLMs, train the embedder with BYOL, perform K-means in conjunction with micro-cluster labels, and ultimately generate interpretable cluster labels.
Improving Context Fidelity via Native Retrieval-Augmented Reasoning
Suyuchen Wang (Université de Montréal), Bang Liu (Université de Montréal)
RetrievalLarge Language ModelSupervised Fine-TuningReinforcement LearningTextRetrieval-Augmented Generation
🎯 What it does: Proposes the CARE framework, enabling LLMs to locally retrieve and integrate contextual evidence during reasoning to enhance answer contextual faithfulness.
Improving Cross Lingual Transfer by Pretraining with Active Forgetting
Divyanshu Aggarwal (Microsoft Research India), Sunayana Sitaram (Microsoft Research India)
Domain AdaptationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper introduces an active forgetting mechanism during the pre-training phase of decoder-only large language models, followed by vocabulary expansion and adaptation for new languages, and fine-tuning using only English instructions. Finally, it evaluates the model's cross-lingual transfer capability on multilingual benchmark tasks.
Improving Handshape Representations for Sign Language Processing: A Graph Neural Network Approach
Alessa Carbo (Johns Hopkins University), Eric Nalisnick (Johns Hopkins University)
ClassificationRecognitionGraph Neural NetworkContrastive LearningVideo
🎯 What it does: Propose a dual-network graph neural network (Handshape-GNN) that achieves handshape recognition by separately capturing the temporal dynamics and static configurations of handshapes;
Improving Informally Romanized Language Identification
Adrian Benton (Google Research), Brian Roark (Google Research)
ClassificationTransformerLarge Language ModelText
🎯 What it does: The paper enhances language recognition performance for Indian languages by generating natural spelling variants in Romanized text during training.
Improving Instruct Models for Free: A Study on Partial Adaptation
Ozan Irsoy (Bloomberg), Duccio Pappadopulo (Bloomberg)
Domain AdaptationComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Studied the regulation of instruction fine-tuning intensity in large language models through training irrelevant parts using partial adaptation methods, examining performance changes from baseline models to instruction models;
Improving Large Language Model Safety with Contrastive Representation Learning
Samuel Simko (ETH Zurich), Zhijing Jin (MPI)
Safty and PrivacyRepresentation LearningAdversarial AttackTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Propose a contrastive learning-based LLM security defense framework that uses triplet loss and adversarial hard negative mining to separate and cluster harmful and benign representations in the embedding space, thereby enhancing robustness against jailbreak attacks.
Improving Large Language Models Function Calling and Interpretability via Guided-Structured Templates
Hy Dang (University Of Notre Dame), Meng Jiang (University Of Notre Dame)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Propose a reasoning framework and dataset (ToolGT) based on structured templates to enhance the accuracy and interpretability of large language models in function calls.
Improving Low-Resource Sequence Labeling with Knowledge Fusion and Contextual Label Explanations
Peichao Lai (Peking University), Bin Cui (Peking University)
Data SynthesisExplainability and InterpretabilityRepresentation LearningData-Centric LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: This paper proposes a framework that combines the knowledge enhancement process of large language models (LLMs) with the span-based KnowFREE model to improve sequence labeling tasks under low-resource conditions. The framework generates extended entity labels, part-of-speech labels, and tokenization labels using LLMs, and enhances the model's understanding of entity boundaries and semantic distributions by generating contextualized entity explanations through interpretive prompts.
Improving Multilingual Retrieval-Augmented Language Models through Dialectic Reasoning Argumentations
Leonardo Ranaldi (University of Edinburgh), Alexandra Birch (University of Edinburgh)
RetrievalExplainability and InterpretabilityTransformerSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: What was done: Proposed the Dialectic-RAG (D-RAG) framework, which integrates dialectical reasoning-based argumentative explanations to improve the answer quality of multilingual retrieval-augmented generation (RAG) models.
Improving Neutral Point-of-View Generation with Data- and Parameter-Efficient RL
Jessica Hoffmann, Lucas Dixon (Google DeepMind)
GenerationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This study develops and evaluates a parameter-efficient reinforcement learning (PE-RL) method to improve the quality of neutral viewpoint responses generated by large language models on controversial topics.
Improving Online Job Advertisement Analysis via Compositional Entity Extraction
Kai Krüger (Bundesinstitut für Berufsbildung), Alan Akbik (Humboldt-Universität zu Berlin)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose a compositional entity modeling framework for extracting job requirements from online job advertisements and construct the GOJA dataset
Improving Reasoning Capabilities in Small Models through Mixture-of-layers Distillation with Stepwise Attention on Key Information
Yao Chen (Chinese Academy of Sciences), Tingwen Liu (Chinese Academy of Sciences)
Knowledge DistillationTextChain-of-Thought
🎯 What it does: Enhancing the inference capability of small models by progressively transferring the teacher model's attention to key information during inference to the student model.
Improving Rule-based Reasoning in LLMs using Neurosymbolic Representations
Varun Dhanraj (University of Waterloo), Chris Eliasmith (University of Waterloo)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: This paper proposes a neuro-symbolic method that encodes LLM hidden states into symbolic vectors, solves problems in the symbolic space, and returns results to enhance rule-based reasoning capabilities.
Improving Task Diversity in Label Efficient Supervised Finetuning of LLMs
Abhinav Arabelly (University of Wisconsin-Madison), Jifan Zhang (University of Wisconsin-Madison)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposes two sampling strategies based on task diversity and model uncertainty (Task Diversity and Weighted Task Diversity) for label-efficient supervised fine-tuning (SFT) of large language models (LLMs);
Improving the Quality of Web-mined Parallel Corpora of Low-Resource Languages using Debiasing Heuristics
Aloka Fernando (University of Moratuwa), Surangika Ranathunga (Massey University)
Representation LearningData-Centric LearningTransformerText
🎯 What it does: Investigate the bias of different multilingual models (LASER3, XLM-R, LaBSE) in parallel corpus filtering during web mining, and propose and systematically evaluate multiple rule-based heuristic filtering combinations to reduce cross-model discrepancies and improve NMT performance.
Improving Zero-shot Sentence Decontextualisation with Content Selection and Planning
Zhenyun Deng (University of Cambridge), Andreas Vlachos (University of Cambridge)
GenerationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose a zero-shot sentence decontextualization framework ECSP, which first identifies ambiguous EDUs through EDU segmentation and selects relevant discourse content, followed by content planning to rewrite sentences
In Benchmarks We Trust ... Or Not?
Ine Gevers (CLiPS University of Antwerp), Walter Daelemans (CLiPS University of Antwerp)
TextReview/Survey PaperBenchmark
🎯 What it does: This study conducts a systematic review of various issues present in natural language processing (NLP) benchmarks and proposes two specific checklists for benchmark creators and evaluators based on existing remedial methods in the literature. Meanwhile, the authors assess three major mainstream text benchmarks (SuperGLUE, WinoGrande, ARC-AGI) using these checklists and further suggest introducing small downstream test sets beyond standard benchmarks to enhance models' transferability and robustness in real-world applications.
In-Context Learning Boosts Speech Recognition via Human-like Adaptation to Speakers and Language Varieties
Nathan Roll (Stanford University), Dan Jurafsky (Stanford University)
RecognitionDomain AdaptationTransformerLarge Language ModelPrompt EngineeringTextMultimodalityAudio
🎯 What it does: This paper proposes and implements a scalable context learning framework that leverages the Phi-4-MM model to rapidly adapt to different speakers and language variants in speech recognition through alternating audio-text examples during inference;
Inceptive Transformers: Enhancing Contextual Representations through Multi-Scale Feature Learning Across Domains and Languages
Asif Shahriar (Bangladesh University of Engineering and Technology), M Saifur Rahman (Bangladesh University of Engineering and Technology)
ClassificationConvolutional Neural NetworkTransformerTextBiomedical Data
🎯 What it does: Add lightweight Inception-style 1D convolution modules and post-self-attention layers to pre-trained Transformer encoders to enhance multi-scale local features and dynamically assign weights.
Incorporating Diverse Perspectives in Cultural Alignment: Survey of Evaluation Benchmarks Through A Three-Dimensional Framework
Meng-Chen Wu (Amazon.com), Narayanan Sadagopan (Amazon.com)
TextReview/Survey PaperBenchmark
🎯 What it does: Investigated the cultural alignment evaluation benchmark, proposing a three-dimensional framework encompassing cultural groups, cultural elements, and awareness scope, and conducted a systematic analysis on 105 text-based datasets.
IndiGEC: Multilingual Grammar Error Correction for Low-Resource Indian Languages
Ujjwal Sharma (Indian Institute of Technology Bombay), Pushpak Bhattacharyya (Indian Institute of Technology Bombay)
GenerationData SynthesisTransformerText
🎯 What it does: Propose the Mask-Translate&Fill (MTF) method, which generates high-quality synthetic grammar error correction data by leveraging monolingual data, machine translation, and masked language models; simultaneously release the multilingual IndiGEC corpus
IndoSafety: Culturally Grounded Safety for LLMs in Indonesian Languages
Muhammad Falensi Azmi (Universitas Indonesia), Fajri Koto (MBZUAI)
Data SynthesisSafty and PrivacyLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Propose the IndoSafety dataset, covering safety evaluation in Indonesian official language, colloquial language, and regional languages such as Javanese, Sundanese, and Minangkabau;
Infini-gram mini: Exact n-gram Search at the Internet Scale with FM-Index
Hao Xu (University of Washington), Hannaneh Hajishirzi (University of Washington)
RetrievalComputational EfficiencyText
🎯 What it does: Built a system named INFINI-GRAM MINI, which achieves efficient exact string search and counting on internet-scale (multi-petabyte) text corpora using FM-Index.
InfiniBench: A Benchmark for Large Multi-Modal Models in Long-Form Movies and TV Shows
Kirolos Ataallah (King Abdullah University of Science and Technology), Mohamed Elhoseiny (King Abdullah University of Science and Technology)
TransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningVideoTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes InfiniBench, a video understanding benchmark for long-duration films and TV programs, containing over 1,000 hours of video, 87.7K question-answer pairs, and eight core cognitive skills (localization and reasoning), with public sharing;
InfoGain-RAG: Boosting Retrieval-Augmented Generation through Document Information Gain-based Reranking and Filtering
Zihan Wang (Kuaishou Technology), Han Li (Kuaishou Technology)
RetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Proposes the InfoGain-RAG framework, which selects the most valuable retrieved documents by measuring their contribution to enhancing the confidence of LLM-generated responses, thereby improving the quality of answers in Retrieval-Augmented Generation.
Information Integration in Large Language Models is Gated by Linguistic Structural Markers
Wei Liu (Zhejiang University), Nai Ding (Zhejiang University)
TransformerLarge Language ModelText
🎯 What it does: This paper proposes a window-based prediction method to quantify the information integration window of large language models (LLMs) and humans at sentence and clause boundaries, measuring the difference between local window predictions and full context predictions using Jensen-Shannon divergence.
InMind: Evaluating LLMs in Capturing and Applying Individual Human Reasoning Styles
Zizhen Li (Nankai University), Kaipeng Zhang (Shanghai AI Laboratory)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes the InMind framework, which collects dual-layer cognitive annotations from two modes—observers and participants—in the social reasoning game Avalon, to evaluate whether large language models (LLMs) can capture and apply individualized reasoning styles.
Instructing Large Language Models for Low-Resource Languages: A Systematic Study for Basque
Oscar Sainz (University of the Basque Country), Aitor Soroa (University of the Basque Country)
Data SynthesisData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: For resource-scarce languages (taking Basque as an example), systematically explored how to perform instruction fine-tuning using only target language corpora, existing multilingual foundation models or instruction-tuned models, and synthetic instruction data, without manual instruction data, ultimately releasing multi-scale (8B, 70B) Basque instruction-tuned LLMs;
Integral Transformer: Denoising Attention, Not Too Much Not Too Little
Ivan Kobyzev (Huawei Noah's Ark Lab), Boxing Chen (Huawei Noah's Ark Lab)
Representation LearningTransformerLarge Language ModelText
🎯 What it does: Propose Integral Transformer, which de-noises attention logits through integration while retaining focus on special tokens.
IntentionFrame: A Semi-Structured, Multi-Aspect Framework for Fine-Grained Conversational Intention Understanding
Jinggui Liang, Lizi Liao (Singapore Management University)
ClassificationExplainability and InterpretabilityLarge Language ModelReinforcement LearningTextSequential
🎯 What it does: Propose the IntentionFrame semi-structured framework and the WeRG weakly supervised reinforcement generation method for fine-grained conversational intent understanding.
Inter-sentence Context Modeling and Structure-aware Representation Enhancement for Conversational Sentiment Quadruple Extraction
Yu Zhang (Jiangsu Ocean University), Huihui Lv (Jiangsu Ocean University)
RecognitionTransformerLarge Language ModelText
🎯 What it does: This paper proposes the ICMSR framework, which utilizes inter-sentence context modeling and structure-aware enhancement to achieve end-to-end extraction of target-aspect-opinion-sentiment quadruples in multi-turn dialogues.
Interdisciplinary Research in Conversation: A Case Study in Computational Morphology for Language Documentation
Enora Rice (University of Colorado Boulder), Alexis Palmer (University of Colorado Boulder)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: A case study on user-centered design for the multilingual IGT generation model GlossLM, evaluating its usability in real language document workflows
InterIDEAS: Philosophical Intertextuality via LLMs
Yue Yang (Monash University), Hao Wang (Monash University)
ClassificationTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This study constructs the InterIDEAS dataset, automatically extracting and annotating intertextual references and their content, functional, and emotional attributes in modern philosophical texts from 1750 to 1950 using LLM and retrieval-augmented generation (RAG) technologies.
Internal Chain-of-Thought: Empirical Evidence for Layer‐wise Subtask Scheduling in LLMs
Zhipeng Yang (Hong Kong University of Science and Technology), Xuming Hu (Southeast University)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Investigate the internal chain-of-thought in large language models, demonstrating that they learn and execute subtasks at different hierarchical levels.
Interpretability Analysis of Arithmetic In-Context Learning in Large Language Models
Gregory Polyakov (University of Tübingen), Seyed Ali Bahrainian (University of Tübingen)
Explainability and InterpretabilityTransformerText
🎯 What it does: This paper systematically studies the in-context learning (ICE) mechanisms of large language models (LLMs) in three arithmetic tasks through multiple mechanism explanation techniques.
Interpretable Mnemonic Generation for Kanji Learning via Expectation-Maximization
Jaewook Lee (University of Massachusetts Amherst), Andrew Lan (University of Massachusetts Amherst)
GenerationExplainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: A interpretable generative framework is proposed for keyword memorization of Japanese kanji, learning rules via the EM algorithm and generating mnemonic sentences based on these rules.
Interpretable Text Embeddings and Text Similarity Explanation: A Survey
Juri Opitz (University of Zurich), Simon Clematide (University of Zurich)
Explainability and InterpretabilityTextReview/Survey Paper
🎯 What it does: This paper reviews and systematizes explainable text embeddings and text similarity explanation methods, proposing four categories of explainable embeddings: spatial shaping, sparse representation, structured objects, and set representations, as well as post-hoc explanation frameworks such as interaction attribution, global explanation, and surrogate models.
Interpretation Meets Safety: A Survey on Interpretation Methods and Tools for Improving LLM Safety
Seongmin Lee (Georgia Tech), Duen Horng Chau (Georgia Tech)
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelTextReview/Survey Paper
🎯 What it does: This paper reviews research on combining explanation methods with LLM security, proposing a unified framework and taxonomy, summarizing nearly 70 works, and discussing strategies and tools for safety enhancement as well as future directions.
Intrinsic Test of Unlearning Using Parametric Knowledge Traces
Yihuai Hong (New York University), Mor Geva (Tel Aviv University)
Explainability and InterpretabilityTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper locates 'concept vectors' in the MLP layer of large language models using a vocabulary projection method, constructing the CONCEPTVECTORS benchmark to simultaneously evaluate internal parameters and external behaviors, assessing the effectiveness of different unlearning methods in eliminating conceptual knowledge.
Introducing Spotlight: A Novel Approach for Generating Captivating Key Information from Documents
Ankan Mullick (Indian Institute of Technology Kharagpur), Niloy Ganguly (Indian Institute of Technology Kharagpur)
GenerationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Introduce and formally define the concept of 'Spotlight,' clarifying its distinctions and characteristics compared to other text condensation forms such as abstracts, titles, and previews;
Investigating How Pre-training Data Leakage Affects Models’ Reproduction and Detection Capabilities
Masahiro Kaneko (Mohamed bin Zayed University of Artificial Intelligence), Timothy Baldwin (Mohamed bin Zayed University of Artificial Intelligence)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Systematically evaluate the pre-training data leakage rate of large language models (LLMs), explore how the proportion of leaked instances affects model reproduction rate and leakage detection rate, and propose a supervised detection method based on a few examples to mitigate the impact of low leakage rates.
Investigating Neurons and Heads in Transformer-based LLMs for Typographical Errors
Kohei Tsuji (NAIST), Tomoya Iwakura (NAIST)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Studied how neurons and attention heads in Transformer-based LLMs detect and repair spelling errors in inputs, and proposed methods to identify 'typo neurons' and 'typo heads'; further validated their roles through ablation experiments using activation differences and KL divergence screening.
Investigating Pedagogical Teacher and Student LLM Agents: Genetic Adaptation Meets Retrieval-Augmented Generation Across Learning Styles
Debdeep Sanyal (Birla AI Labs), Murari Mandal (Pennsylvania State University)
OptimizationTransformerLarge Language ModelAgentic AITextTabularRetrieval-Augmented Generation
🎯 What it does: Propose a multi-agent teaching simulation framework based on LLM, integrating adaptive student agents and teacher agents evolved via genetic algorithms, and achieving personalized retrieval through Persona-RAG;
Investigating the interaction of linguistic and mathematical reasoning in language models using multilingual number puzzles
Antara Raaghavi Bhattacharya (Harvard University), David Alvarez-Melis (Harvard University)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper investigates the interaction between language and mathematical reasoning in large language models for cross-lingual numerical puzzles, revealing through a series of experiments the critical role of explicit arithmetic symbols in model performance.
Investigating Value-Reasoning Reliability in Small Large Language Models
Xia Du (Beijing Language and Culture University), Dong Yu (Beijing Language and Culture University)
Adversarial AttackData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: Evaluate the reliability of small large language models in value reasoning tasks, proposing a multi-dimensional evaluation framework and constructing an enhanced dataset.
Invisible Entropy: Towards Safe and Efficient Low-Entropy LLM Watermarking
Tianle Gu (Mohamed bin Zayed University of Artificial Intelligence), Xiuying Chen (Mohamed bin Zayed University of Artificial Intelligence)
Safty and PrivacyComputational EfficiencyAI Code AssistantTransformerLarge Language ModelTextSequential
🎯 What it does: Proposed Invisible Entropy (IE), a low-entropy LLM watermarking scheme that achieves secure and efficient watermark detection without requiring the original LLM, utilizing a lightweight feature extractor, entropy labeler, and threshold navigator.
IPIGuard: A Novel Tool Dependency Graph-Based Defense Against Indirect Prompt Injection in LLM Agents
Hengyu An (Zhejiang University), Shouling Ji (Zhejiang University)
Safty and PrivacyLarge Language ModelAgentic AITextBenchmark
🎯 What it does: Proposes IPIGUARD, a defense framework for LLM agents based on Tool Dependency Graph (TDG). The framework generates a TDG through LLM planning before task execution, then executes tool calls in DAG topological order, ensuring resistance to injected instructions during execution via three key mechanisms (Argument Estimation, Node Expansion, Fake Tool Invocation), effectively preventing malicious tool calls caused by Indirect Prompt Injection (IPI) attacks.
Is Cognition Consistent with Perception? Assessing and Mitigating Multimodal Knowledge Conflicts in Document Understanding
Zirui Shao (Zhejiang University), Jiajun Bu (Zhejiang University)
RecognitionLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelMultimodalityBenchmark
🎯 What it does: Proposed a multi-modal knowledge consistency fine-tuning approach to reduce conflicts between cognitive (VQA) and perceptual (OCR) tasks in document understanding for multi-modal large language models.
Is the Top Still Spinning? Evaluating Subjectivity in Narrative Understanding
Melanie Subbiah (Columbia University), Kathleen McKeown (Columbia University)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Redefine the authenticity evaluation of narrative summaries by shifting from traditional binary support/not support judgments to detecting ambiguity and subjectivity through LLM-based claim rewriting; introduce the Ambiguity Rewrite Metric (ARM) as an automated assessment tool.
It’s All About In-Context Learning! Teaching Extremely Low-Resource Languages to LLMs
Yue Li (University of Sheffield), Carolina Scarton (University of Sheffield)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper investigates the learning capabilities of large language models (LLMs) on ultra-low-resource languages (especially rare scripts), systematically comparing the performance of zero-shot and few-shot in-context learning (ICL) and parameter-efficient fine-tuning (PEFT) methods across 20 low-resource languages.
Iterative Multilingual Spectral Attribute Erasure
Shun Shao (University of Cambridge), Anna Korhonen
Representation LearningTransformerLarge Language ModelText
🎯 What it does: Proposed a multilingual spectral attribute elimination method (IMSAE) to identify and remove shared bias subspaces in multilingual embedding spaces, supporting zero-shot field debiasing when the target language is missing.
Iterative Prompt Refinement for Safer Text-to-Image Generation
Jinwoo Jeon (Korea University), Byung-Jun Lee (Korea University)
GenerationSafty and PrivacySupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: Propose an iterative prompt refinement algorithm (IPR) that leverages vision-language models to repeatedly adjust prompts based on generated images, thereby enhancing the safety of text-to-image models.
iTool: Reinforced Fine-Tuning with Dynamic Deficiency Calibration for Advanced Tool Use
Yirong Zeng (Harbin Institute of Technology), Ting Liu (Harbin Institute of Technology)
TransformerSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: Proposes iTool, which combines easy-to-difficult SFT, MCTS path exploration, and iterative ReFT to enhance LLM tool usage capabilities, especially in complex scenarios.