EMNLP 2025 Papers — Page 12
Conference on Empirical Methods in Natural Language Processing · 1809 papers
Not Your Typical Government Tipline: LLM-Assisted Routing of Environmental Protection Agency Citizen Tips
Sharanya Majumder (Stanford University), Daniel E. Ho (Stanford University)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A two-phase LLM-assisted classification pipeline was designed and implemented for the EPA's citizen reporting system, first filtering reports outside the EPA's scope and then routing them to civil or criminal departments.
Not-Just-Scaling Laws: Towards a Better Understanding of the Downstream Impact of Language Model Design Decisions
Emmy Liu (Carnegie Mellon University), Graham Neubig (Carnegie Mellon University)
Explainability and InterpretabilityTransformerLarge Language ModelTextBenchmark
🎯 What it does: Established a database containing 92 open-source models, recording model architectures, data compositions, and generation features, and used regression models to predict performance on multiple benchmark tasks to quantify the impact of model design decisions on downstream performance.
NOVA-63: Native Omni-lingual Versatile Assessments of 63 Disciplines
Jinyang Zhang (School of Computer Science, Peking University), Dayiheng Liu (Alibaba Group)
Large Language ModelTextBenchmark
🎯 What it does: Constructed NOVA-63: a difficulty-controllable, subject-balanced benchmark comprising 89,107 native multilingual multiple-choice questions across 14 languages and 63 subjects, and used it to evaluate the performance of 62 LLMs
NovelHopQA: Diagnosing Multi-Hop Reasoning Failures in Long Narrative Contexts
Abhay Gupta (Algoverse AI Research), Michael Lu (University of California, Berkeley)
TransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Proposed the NOVELHOPQA benchmark for evaluating 1–4 step multi-hop question answering in 64k–128k scale long-form novel texts;
NOVER: Incentive Training for Language Models via Verifier-Free Reinforcement Learning
Wei Liu (King's College London), Yulan He (King's College London)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: Propose the NOVER framework, which uses inference perplexity as a validator-free reward surrogate to perform incentive-based reinforcement learning on large language models, encouraging them to generate intermediate reasoning steps.
Nullspace Disentanglement for Red Teaming Language Models
Yi Han (Harbin Institute of Technology), Ting Liu (Harbin Institute of Technology)
Adversarial AttackTransformerLarge Language ModelText
🎯 What it does: Proposed a black-box red teaming technique called NDR based on nullspace, which can separate and reconstruct success information and semantic information from test cases to generate more aggressive test cases.
NUTMEG: Separating Signal From Noise in Annotator Disagreement
Jonathan Ivey (Johns Hopkins University), David Jurgens (University of Michigan)
Data-Centric LearningTransformerText
🎯 What it does: Propose NUTMEG, a Bayesian model that separates noise and systematic disagreements based on annotator background information, generating subgroup-level true labels; subsequently, these labels are used to learn from disagreement, improving downstream task performance.
OBLIVIATE: Robust and Practical Machine Unlearning for Large Language Models
Xiaoyu Xu (Hong Kong Polytechnic University), Haibo Hu (Hong Kong Polytechnic University)
TransformerLarge Language ModelText
🎯 What it does: Propose the OBLIVIATE framework to achieve machine unlearning for large language models, removing specified data while maintaining model performance and fluency.
OG-RAG: Ontology-grounded retrieval-augmented generation for large language models
Kartik Sharma (Georgia Institute of Technology), Yunqing Li (Lenovo)
RetrievalTransformerLarge Language ModelTextGraphAgriculture RelatedRetrieval-Augmented Generation
🎯 What it does: Propose OG-RAG, an ontology-based retrieval-augmented generation framework that constructs hypergraphs using domain ontologies and retrieves minimal hyperedge contexts during LLM generation;
OmniEval: An Omnidirectional and Automatic RAG Evaluation Benchmark in Financial Domain
Shuting Wang (Renmin University of China), Ji-Rong Wen (Renmin University of China)
Data SynthesisSupervised Fine-TuningPrompt EngineeringTextBenchmarkFinance RelatedRetrieval-Augmented Generation
🎯 What it does: Built a comprehensive RAG evaluation benchmark called OmniEval for the financial domain, covering task-topic matrices, automated data generation, phased evaluation, and multi-layer metrics;
OmniThink: Expanding Knowledge Boundaries in Machine Writing through Thinking
Zekun Xi (Zhejiang University), Ningyu Zhang (Zhejiang University)
GenerationRetrievalTransformerTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the OmniThink framework, which simulates human slow thinking processes in machine writing. It utilizes information trees and concept pools for information acquisition, reflection, and expansion, ultimately generating high-quality long-form articles.
OMS: On-the-fly, Multi-Objective, Self-Reflective Ad Keyword Generation via LLM Agent
Bowen Chen (Sony Group Corporation), Shingo Takamatsu (Sony Group Corporation)
GenerationOptimizationTransformerLarge Language ModelAgentic AIPrompt EngineeringTextTabular
🎯 What it does: Proposes the OMS framework, using LLM agents to achieve training-data-free, real-time multi-objective optimization and self-reflective generation of SSA keywords.
On LLM-Based Scientific Inductive Reasoning Beyond Equations
Brian S. Lin (Tsinghua University), Maosong Sun (Tsinghua University)
Large Language ModelTextBenchmark
🎯 What it does: Proposed SIRBench-V1 benchmark, focusing on the performance of large language models in scientific induction reasoning (beyond mathematical equations), covering biology and chemistry subtasks, and using synthetic rule-based tasks to test whether models truly reason rather than memorize.
On Pruning State-Space LLMs
Tamer Ghattas (Hebrew University of Jerusalem), Roy Schwartz (Hebrew University of Jerusalem)
CompressionComputational EfficiencyKnowledge DistillationText
🎯 What it does: Investigated and evaluated the adaptation of unstructured pruning (WANDA) and various structured pruning methods (state pruning, head dimension pruning, head merging, SSM-FLAP) to large language models (LLMs) based on state space models (SSMs), comparing their compression effects and performance impacts on four different SSM LLMs (Mamba-2, PHI-Mamba, HLM-3B, SMOL-Mamba).
On Relation-Specific Neurons in Large Language Models
Yihong Liu (LMU Munich), Hinrich Schuetze
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Study neurons specialized in encoding relational information in large language models (LLama-2 7B/13B), and assess their impact on fact recall by suppressing these neurons.
On the Role of Model Prior in Real-World Inductive Reasoning
Zhuo Liu (University of Rochester), Hangfeng He (University of Rochester)
ClassificationTransformerLarge Language ModelPrompt EngineeringMultimodalityBenchmark
🎯 What it does: Investigated the dominant role of model prior in hypothesis generation during inductive reasoning of LLMs in real-world scenarios.
On the Same Wavelength? Evaluating Pragmatic Reasoning in Language Models across Broad Concepts
Linlu Qiu (MIT), Roger P. Levy (MIT)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Evaluate the pragmatic reasoning capabilities of large language models in the Wavelength game, covering both comprehension and generation tasks, and integrate the Rational Speech Act (RSA) framework with language models (LMs);
One Planner To Guide Them All ! Learning Adaptive Conversational Planners for Goal-oriented Dialogues
Huy Quang Dao (Singapore Management University), Lizi Liao (Singapore Management University)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposed a single policy planner PADPP that is adaptable to multi-objective, goal-oriented dialogues, capable of dynamically adjusting goal weights during inference without requiring retraining.
OntologyRAG-Q: Resource Development and Benchmarking for Retrieval-Augmented Question Answering in Qur’anic Tafsir
Sadam Al-Azani (KFUPM), Ahmed Abdelali (Humain)
RetrievalTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Annotated Tafsir ontology, approximately 4,200 QA pairs, and 15 structured Tafsir books were constructed, and an OntologyRAG-Q system based on Ayat-Ontology chunking was proposed.
OpenNER 1.0: Standardized Open-Access Named Entity Recognition Datasets in 50+ Languages
Chester Palen-Michel (Brandeis University), Constantine Lignos (Brandeis University)
ClassificationTransformerSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes OpenNER 1.0, a standardized named entity recognition dataset collection containing 36 manually annotated datasets across 52 languages;
OpenTuringBench: An Open-Model-based Benchmark and Framework for Machine-Generated Text Detection and Attribution
Lucio La Cava (University of Calabria), Andrea Tagarelli (University of Calabria)
ClassificationTransformerContrastive LearningTextBenchmark
🎯 What it does: Constructed the OpenTuringBench benchmark and the OTBDetector framework for training and evaluating the detection and authorship attribution of text generated by open large language models (OLLM).
Orchestrating Audio: Multi-Agent Framework for Long-Video Audio Synthesis
Yehang Zhang (Hong Kong University of Science and Technology), Ying-Cong Chen
GenerationData SynthesisLarge Language ModelAgentic AIVision Language ModelVideoMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-ThoughtAudio
🎯 What it does: Propose the LVAS-Agent multi-agent framework to achieve long video audio synthesis and release the LVAS-Bench dataset.
Order Doesn’t Matter, But Reasoning Does: Training LLMs with Order-Centric Augmentation
Qianxi He (Fudan University), Yanghua Xiao (Ant Group)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkChain-of-Thought
🎯 What it does: By introducing order-centric data augmentation based on logical commutativity, training large language models (LLMs) to be less sensitive to the order of premises and reasoning steps in logical reasoning tasks;
ORPP: Self-Optimizing Role-playing Prompts to Enhance Language Model Capabilities
Yifan Duan (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
OptimizationComputational EfficiencyLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes an ORPP framework based on role-playing prompts, which first iteratively optimizes high-quality role prompts on a small number of samples, and then automatically generates prompts for the remaining questions using few-shot learning, thereby improving the reasoning and generation performance of large language models.
Orthogonal Finetuning Made Scalable
Zeju Qiu, Bernhard Schölkopf (Max Planck Institute for Intelligent Systems)
Computational EfficiencyTransformerSupervised Fine-TuningImageTextMultimodality
🎯 What it does: Proposed OFTv2, which adopts input centralization implementation, Cayley-Neumann parameterization, and supports efficient fine-tuning of quantized base models
Out of Sight, Not Out of Context? Egocentric Spatial Reasoning in VLMs Across Disjoint Frames
Sahithya Ravi (University of British Columbia), Balasaravanan Thoravi Kumaravel (Microsoft Research)
Large Language ModelPrompt EngineeringVision Language ModelMultimodalityBenchmark
🎯 What it does: Propose the DISJOINT-3DQA benchmark to evaluate the cross-frame spatial reasoning ability of Vision-Language Models (VLMs) under front-view scenarios without shared visible objects.
OWL: Probing Cross-Lingual Recall of Memorized Texts via World Literature
Alisha Srivastava (University of Massachusetts), Mohit Iyyer (University of Maryland)
Data SynthesisRetrievalTransformerLarge Language ModelTextMultimodalityBenchmarkAudio
🎯 What it does: Constructed the OWL dataset, aligning 20 English novels across 10 languages (including 6 low-resource languages), and evaluated LLMs' multilingual and cross-lingual memory capabilities through three tasks (direct retrieval, name filling, and prefix continuation).
P-MMEval: A Parallel Multilingual Multitask Benchmark for Consistent Evaluation of LLMs
Yidan Zhang (Tongyi Lab, Alibaba Group Inc), Jingren Zhou (Tongyi Lab, Alibaba Group Inc)
ClassificationGenerationPrompt EngineeringTextBenchmark
🎯 What it does: Created a multilingual multitask benchmark P-MMEVAL and evaluated the cross-lingual performance of various LLMs across 10 languages;
PACHAT: Persona-Aware Speech Assistant for Multi-party Dialogue
Dongjie Fu (Zhejiang University), Tao Jin (Zhejiang University)
RecognitionGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningTextAudio
🎯 What it does: Proposed PAChat, integrating LLM with speech encoding and speaker recognition to achieve role-awareness and personalized responses in multi-party speech dialogues;
PAFT: Prompt-Agnostic Fine-Tuning
Chenxing Wei (Shenzhen University), Fei Yu (Carleton University)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: Propose a Prompt-Agnostic Fine-Tuning (PAFT) method, which first constructs a diversified set of synthetic prompts and then dynamically randomly samples these prompts during fine-tuning to enhance the robustness of LLMs to unseen prompts.
Paired by the Teacher: Turning Unpaired Data into High-Fidelity Pairs for Low-Resource Text Generation
Yen-Ju Lu (Johns Hopkins University), Jesus Villalba (Johns Hopkins University)
GenerationData SynthesisKnowledge DistillationLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes a two-stage teacher-student framework called Paired by the Teacher (PbT), where the teacher LLM extracts intermediate representations (IR), and the student model reconstructs the input to generate high-quality input-output pairs in low-resource scenarios without alignment labels;
PakBBQ: A Culturally Adapted Bias Benchmark for QA
Abdullah Hashmat (Lahore University of Management Sciences), Agha Ali Raza (Lahore University of Management Sciences)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Constructed the PakBBQ dataset—a QA bias benchmark tailored for Pakistani culture, containing 214 templates and 17,180 Chinese-English QA pairs.
PAKTON: A Multi-Agent Framework for Question Answering in Long Legal Agreements
Petros Raptopoulos, Giorgos Stamou (National Technical University Of Athens)
RetrievalExplainability and InterpretabilityTransformerAgentic AITextRetrieval-Augmented Generation
🎯 What it does: Proposed the PAKTON multi-agent framework for automated legal contract analysis and question-answering, generating interpretable reports.
PanicToCalm: A Proactive Counseling Agent for Panic Attacks
Jihyun Lee (POSTECH), Gary Lee
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningAgentic AIText
🎯 What it does: The study constructed a psychological first aid dataset for panic attacks called PACE, trained the PACER model, and proposed the PANICEVAL evaluation framework.
Parallel Continuous Chain-of-Thought with Jacobi Iteration
Haoyi Wu (ShanghaiTech University), Kewei Tu (ShanghaiTech University)
OptimizationComputational EfficiencyTransformerTextBenchmarkChain-of-Thought
🎯 What it does: Propose parallel continuous chain-of-thought reasoning (PCCoT), which improves the training and inference efficiency of continuous CoT by parallelizing the update of implicit thought tokens through Jacobi iteration.
Parrot: A Training Pipeline Enhances Both Program CoT and Natural Language CoT for Reasoning
Senjie Jin (Fudan University), Xuanjing Huang (Fudan University)
Reinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Designed the Parrot training pipeline, enhancing the reasoning performance of program chains and natural language chains through three collaborative training steps: information retrieval, program reasoning, and paradigm conversion.
PatentScore: Multi-dimensional Evaluation of LLM-Generated Patent Claims
Yongmin Yoo (Macquarie University), Longbing Cao (Macquarie University)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Developed PatentScore, a multi-dimensional evaluation framework for assessing the structural, legal, and semantic quality of patent claims generated by LLMs.
Path Drift in Large Reasoning Models: How First-Person Commitments Override Safety
Yuyi Huang (Guangzhou Medical University), Derek F. Wong (University of Macau)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This paper identifies and quantifies the Path Drift vulnerability in Large Retrieval Models (LRM) during the Chain-of-Thought reasoning process, and proposes a three-phase attack framework to induce the model to generate non-compliant content.
Paths Not Taken: Understanding and Mending the Multilingual Factual Recall Pipeline
Meng Lu (Brown University), Ellie Pavlick (Brown University)
RetrievalExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: Investigate the mechanisms of multilingual large language models in fact recall tasks, revealing that they first retrieve facts in an English-centric space and then translate the answer back to the target language; two vector interventions are proposed to address retrieval and translation errors to improve cross-lingual fact consistency.
Pathway to Relevance: How Cross-Encoders Implement a Semantic Variant of BM25
Meng Lu (Brown University), Carsten Eickhoff (University of Tübingen)
RetrievalExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerText
🎯 What it does: Provide a fine-grained explanation of the internal mechanisms of BERT-based cross-encoders, identifying and locating BM25 components such as soft TF, IDF, and document length, and verifying their functions by constructing a diagnostic dataset and path patches;
PathwiseRAG: Multi-Dimensional Exploration and Integration Framework
Hengrui Zhang (Chongqing University of Posts and Telecommunications), Yulu Du (Chongqing University of Posts and Telecommunications)
GenerationRetrievalGraph Neural NetworkTransformerTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the PathwiseRAG framework, reimagining Retrieval-Augmented Generation (RAG) as a multi-dimensional dynamic exploration process, achieving higher quality answer generation through intent-aware retrieval strategies, dynamic reasoning networks, and parallel path exploration.
PBI-Attack: Prior-Guided Bimodal Interactive Black-Box Jailbreak Attack for Toxicity Maximization
Ruoxi Cheng (Alibaba Group), Xiaojun Jia (Nanyang Technological University)
Adversarial AttackVision Language ModelImageTextMultimodality
🎯 What it does: Studied a PBI-Attack method that maximizes toxicity in a black-box environment by leveraging prior features and dual-modal interactive iterative optimization, successfully bypassing the security mechanisms of large vision-language models.
Permutative Preference Alignment from Listwise Ranking of Human Judgments
Yang Zhao (University of Texas), Mingzhang Yin (University of Florida)
Reinforcement Learning from Human FeedbackTransformerReinforcement LearningTextBenchmark
🎯 What it does: Propose the Permutative Preference Alignment (PPA) method, which performs offline list-wise alignment for LLMs by maximizing NDCG.
PERSEVAL: A Framework for Perspectivist Classification Evaluation
Soda Marem Lo (University of Turin), Davide Bernardi (Amazon)
ClassificationTransformerPrompt EngineeringMixture of ExpertsText
🎯 What it does: Propose the PERSEVAL framework for unified evaluation of perspectivist text classification models, focusing on the separation between annotators during training and test users, and evaluating at four levels: global, text, user, and feature.
Persona-Augmented Benchmarking: Evaluating LLMs Across Diverse Writing Styles
Kimberly Truong, Steven Wu
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes a persona-based LLM rephrasing process that utilizes personas to diversify writing styles in existing benchmarks and evaluates the impact of different writing styles on model performance.
Personality Matters: User Traits Predict LLM Preferences in Multi-Turn Collaborative Tasks
Sarfaroz Yunusov (Brock University), Ali Emami (Emory University)
TransformerLarge Language ModelText
🎯 What it does: Conducted multi-round collaborative tasks with 32 university students categorized into Keirsey's four types to assess their preferences between GPT-4 and Claude 3.5.
Personality Vector: Modulating Personality of Large Language Models by Model Merging
Seungjong Sun (Sungkyunkwan University), Jang Hyun Kim (Sungkyunkwan University)
Representation LearningLarge Language ModelTextMultimodality
🎯 What it does: Achieving personalized control in large language models through model merging.
Personalization up to a Point: Why Personalized Content Moderation Needs Boundaries, and How We Can Enforce Them
Emanuele Moscato (Bocconi University), Debora Nozza (Bocconi University)
ClassificationAnomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Studied the conflicts between achieving personalized content moderation on social media and legal boundaries, and proposed a framework that ensures legal compliance through boundary constraint mechanisms.
Personalized Language Models via Privacy-Preserving Evolutionary Model Merging
Kyuyoung Kim (KAIST AI), Jaehyung Kim (KAIST AI)
Safty and PrivacyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose a privacy-preserving language model fusion method called PriME based on evolutionary algorithms, aiming to reduce privacy leakage risks from data sharing while retaining user preferences.
Personalized LLM Decoding via Contrasting Personal Preference
Hyungjune Bu (Yonsei University), Jaehyung Kim (Yonsei University)
Recommendation SystemTransformerLarge Language ModelContrastive LearningTextBenchmark
🎯 What it does: After personalizing large language models using parameter-efficient fine-tuning (LoRA) for each user, we propose a decoding strategy called COPE, which maximizes implicit user rewards by comparing the likelihood ratio between user models and baseline models during decoding. We further enhance personalization by combining DPO with synthetic negative samples.
PerspectiveMod: A Perspectivist Resource for Deliberative Moderation
Eva Maria Vecchi (University of Stuttgart), Gabriella Lapesa (Heinrich-Heine University of Düsseldorf)
Text
🎯 What it does: Introduce the PerspectiveMod dataset, which records whether moderation is needed in online discussions and the reasons for moderation, and collects personal backgrounds and attitudes of expert and non-expert annotators.
Persuasion Dynamics in LLMs: Investigating Robustness and Adaptability in Knowledge and Safety with DuET-PD
Bryan Chen Zhengyu Tan (Singapore University of Technology and Design), Roy Ka-Wei Lee (Singapore University of Technology and Design)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose the DuET-PD framework to evaluate stance shifts of LLMs in multi-round persuasive dialogues, and develop the Holistic DPO training method to enhance robustness and adaptability against misleading and corrective persuasion.
pFedGPT: Hierarchically Optimizing LoRA Aggregation Weights for Personalized Federated GPT Models
Zhanming Shen (Zhejiang University), Miao Pan (Zhejiang University)
Federated LearningTransformerLarge Language ModelText
🎯 What it does: Under the federated learning framework, personalized aggregation of LoRA parameters is achieved through hierarchical Bayesian optimization, enabling efficient personalization fine-tuning of large language models.
Phi: Preference Hijacking in Multi-modal Large Language Models at Inference Time
Yifan Lan (Pennsylvania State University), Jinghui Chen (Pennsylvania State University)
OptimizationAdversarial AttackTransformerLarge Language ModelReinforcement LearningVision Language ModelImageTextMultimodality
🎯 What it does: This paper studies attack methods that achieve preference hijacking by optimizing images during inference in multimodal large language models.
PhoniTale: Phonologically Grounded Mnemonic Generation for Typologically Distant Language Pairs
Sana Kang (KAIST), Rita Singh (Carnegie Mellon University)
GenerationRecurrent Neural NetworkTransformerLarge Language ModelContrastive LearningTextAudio
🎯 What it does: Propose a cross-lingual keyword mnemonic generation system called PHONITALE, specifically designed for language pairs with significant phonetic structural differences such as English-Korean. The system can automatically generate L1 keyword sequences similar to L2 word forms through IPA transcription, speech segmentation, and keyword matching, and utilize LLM to generate corresponding oral cues.
PhonoThink: Improving Large Language Models’ Reasoning on Chinese Phonological Ambiguities
Jianfei Ma (Hong Kong Polytechnic University), Ziqi Zhang (Hong Kong Polytechnic University)
RecognitionTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: Build a multi-stage training framework (SFT+RL), enhancing the reasoning capabilities of large language models (LLMs) for Chinese speech ambiguity identification and error correction tasks through three subtask datasets and synthetic reasoning chain data.
Pierce the Mists, Greet the Sky: Decipher Knowledge Overshadowing via Knowledge Circuit Analysis
Haoming Huang (Hong Kong University of Science and Technology), Xuming Hu (Hong Kong University of Science and Technology)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Proposed and implemented the PHANTOMCIRCUIT framework for systematic analysis and detection of knowledge obscuring phenomena in large language models.
PIIvot: A Lightweight NLP Anonymization Framework for Question-Anchored Tutoring Dialogues
Matthew Zent (Eedi), Simon Woodhead
Safty and PrivacyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Developed the PIIvot framework, which leverages potential PII tagging and context-aware LLM alternatives to achieve lightweight PII anonymization, and released the largest real-world Q&A tutoring dialogue dataset QATD 2k based on this.
Pixels Versus Priors: Controlling Knowledge Priors in Vision-Language Models through Visual Counterfacts
Michal Golovanevsky (Brown University), Carsten Eickhoff (University of Tübingen)
Explainability and InterpretabilityPrompt EngineeringVision Language ModelMultimodality
🎯 What it does: This paper proposes the Visual CounterFact dataset and designs the Pixels Versus Priors (PvP) activation-level intervention method to control the dependency of multimodal large language models between visual information and world knowledge priors.
Plan Dynamically, Express Rhetorically: A Debate-Driven Rhetorical Framework for Argumentative Writing
Xueguan Zhao (Qilu University of Technology), Deyu Zhou (Southeast University)
GenerationTransformerLarge Language ModelAgentic AITextChain-of-Thought
🎯 What it does: Proposed a debate-driven argument writing framework named DARE, which first dynamically constructs hierarchical outline trees through multi-agent debate, then performs rhetorical optimization based on Bitzer's rhetorical situation theory, ultimately generating high-quality argumentative essays.
PLAN-TUNING: Post-Training Language Models to Learn Step-by-Step Planning for Complex Problem Solving
Mihir Parmar (Google), Tomas Pfister (Google)
Knowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes PLAN-TUNING, a post-training method that enhances the performance of small open-source LLMs on complex reasoning tasks by leveraging synthetic natural planning trajectories (i.e., problem decomposition steps).
PlanGEN: A Multi-Agent Framework for Generating Planning and Reasoning Trajectories for Complex Problem Solving
Mihir Parmar (Google), Hamid Palangi (Google)
OptimizationLarge Language ModelAgentic AITextBenchmarkChain-of-Thought
🎯 What it does: Propose a multi-agent framework named PlanGEN for generating natural language planning and reasoning trajectories.
Planning-Aware Code Infilling via Horizon-Length Prediction
Yifeng Ding (University of Illinois Urbana Champaign), Zijian Wang (Meta)
GenerationAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTextSequential
🎯 What it does: This paper studies the code completion task, proposing to enhance the model's planning ability by predicting the remaining number of tokens.
Playpen: An Environment for Exploring Learning From Dialogue Game Feedback
Nicola Horst (University of Potsdam), Alessandro Suglia (University of Edinburgh)
Reinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: Proposed the PLAYPEN environment for large language models to self-play in dialogue games and learn from game feedback;
Please Translate Again: Two Simple Experiments on Whether Human-Like Reasoning Helps Translation
Di Wu (Language Technology Lab University of Amsterdam), Christof Monz (Language Technology Lab University of Amsterdam)
GenerationTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Investigated the role of human-like chain-of-thought (CoT) decomposition in large language model (LLM) translation, and compared the Step-by-step prompting method with the 'Translate-again' multi-round self-refinement approach.
PLLuM-Align: Polish Preference Dataset for Large Language Model Alignment
Karolina Seweryn (National Research Institute), Arkadiusz Janz (Wrocław University of Science and Technology)
Safty and PrivacyData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Proposed and constructed the first purely manually annotated, culturally detailed Polish alignment preference dataset, PLLuM-Align, specifically designed for training and evaluating the safety and accuracy of Polish language models.
Pluralistic Alignment for Healthcare: A Role-Driven Framework
Jiayou Zhong (University of Waterloo), Usman Naseem (Macquarie University)
TransformerLarge Language ModelAgentic AIPrompt EngineeringMixture of ExpertsTextChain-of-Thought
🎯 What it does: Proposed and implemented the ETHOSAGENTS framework, achieving multi-value alignment in medical scenarios using dynamic role generation technology.
Plutus: Benchmarking Large Language Models in Low-Resource Greek Finance
Xueqing Peng (Fin AI), Sophia Ananiadou (Athena Research Center)
Domain AdaptationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkFinance Related
🎯 What it does: This paper proposes Plutus-ben, the first Greek financial evaluation benchmark, and trains Plutus-8B, the first large language model for Greek financial tasks, based on this benchmark;
Pointing to a Llama and Call it a Camel: On the Sycophancy of Multimodal Large Language Models
Renjie Pi (Hong Kong University of Science and Technology), Xiaofang Zhou (Hong Kong University of Science and Technology)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: This paper studies the sycophancy behavior of multi-modal large language models when receiving image inputs and proposes Sycophantic Reflective Tuning (SRT) to alleviate this issue.
POINTS-Reader: Distillation-Free Adaptation of Vision-Language Models for Document Conversion
Yuan Liu (Tencent Inc), Jie Zhou (Tencent Inc)
GenerationData SynthesisDomain AdaptationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityTabularBenchmark
🎯 What it does: Propose a completely distillation-free two-stage method, first preheating the model using synthetic data generated with a unified output format, then using the model to self-annotate, filter, and iteratively retrain on real documents, thereby improving end-to-end document conversion performance.
Polysemantic Dropout: Conformal OOD Detection for Specialized LLMs
Ayush Gupta (Computer Science Lab, SRI), Susmit Jha (Computer Science Lab, SRI)
Anomaly DetectionTransformerBiomedical Data
🎯 What it does: Studied a method for detecting out-of-distribution (OOD) inputs in specialized large language models (LLMs) during inference, utilizing the model's dropout tolerance as a non-consistency measure (NCM) and implementing it within the Inductive Conformal Anomaly Detection (ICAD) framework.
PORTS: Preference-Optimized Retrievers for Tool Selection with Large Language Models
Lorenzo Molfetta (University of Bologna), Gianluca Moro (University of Bologna)
RetrievalTransformerLarge Language ModelContrastive LearningTextBenchmark
🎯 What it does: Developed a training framework for tool retrieval based on LLM called PORTS, designed to pre-select the most relevant tool documents for queries.
PoseStitch-SLT: Linguistically Inspired Pose-Stitching for End-to-End Sign Language Translation
Abhinav Joshi (Indian Institute of Technology Kanpur), Ashutosh Modi (Indian Institute of Technology Kanpur)
Data SynthesisPose EstimationTransformerSupervised Fine-TuningTextSequential
🎯 What it does: Construct a pose concatenation dataset based on language templates, using keypoint sequences to achieve sign language translation without gloss.
POSITION BIAS MITIGATES POSITION BIAS: Mitigate Position Bias Through Inter-Position Knowledge Distillation
Yifei Wang (State Key Laboratory of Multimodal Artificial Intelligence Systems Institute of Automation Chinese Academy of Sciences), Daniel Dajun Zeng (State Key Laboratory of Multimodal Artificial Intelligence Systems Institute of Automation Chinese Academy of Sciences)
RetrievalKnowledge DistillationTransformerTextChain-of-Thought
🎯 What it does: Propose the Pos2Distill framework, which mitigates position bias (PB) in large models through inter-positional knowledge distillation, and designs two instances, Pos2Distill-R1 and Pos2Distill-R2, for retrieval and reasoning tasks, respectively.
Position: LLMs Can be Good Tutors in English Education
Jingheng Ye (Squirrel Ai Learning), Qingsong Wen (Squirrel Ai Learning)
TransformerLarge Language ModelSupervised Fine-TuningAgentic AITextMultimodality
🎯 What it does: This paper proposes using large language models (LLM) as mentors in English education, outlining three core functions: data augmentation, task prediction, and intelligent agents, aiming to enhance personalized learning and scalability;
PoSum-Bench: Benchmarking Position Bias in LLM-based Conversational Summarization
Xu Sun (Orange Research), Anastasia Shimorina (Orange Research)
GenerationTransformerLarge Language ModelTextSequentialBenchmark
🎯 What it does: Construct the PoSum-Bench benchmark dataset for no-reference evaluation of displacement bias (lead/recency) in LLMs during dialogue summarization, and systematically analyze multilingual, multi-domain dialogues.
Power doesn’t reside in size: A Low Parameter Hybrid Language Model (HLM) for Sentiment Analysis in Code-mixed data
Pavan Sai Balaga (Indian Institute of Technology Roorkee), Ashish Mittal (IBM Research)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a low-parameter hybrid language model (HLM) for sentiment analysis on code-mixed text.
PPC-GPT: Federated Task-Specific Compression of Large Language Models via Pruning and Chain-of-Thought Distillation
Tao Fan (Hong Kong University of Science and Technology), Qiang Yang (WeBank Co., Ltd)
CompressionFederated LearningSafty and PrivacyKnowledge DistillationLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Constructed the PPC-GPT framework, using a federated approach to compress large language models into task-specific small models while achieving differential privacy protection during the process;
PPTAgent: Generating and Evaluating Presentations Beyond Text-to-Slides
Hao Zheng (Chinese Academy of Sciences), Le Sun (Chinese Academy of Sciences)
GenerationTransformerLarge Language ModelVision Language ModelTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose an edit-based two-phase method, PPTAGENT, which automatically generates high-quality presentations using a large language model by referencing existing presentations, and introduce an evaluation framework, PPTEVAL, to comprehensively assess content, design, and coherence;
Pragmatic Inference Chain (PIC) Improving LLMs’ Reasoning of Authentic Implicit Toxic Language
Xi Chen (Nanyang Technological University), Shuo Wang (University of Macau)
ClassificationRecognitionTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Developed a 'Pragmatic Inference Chain (PIC)' prompting method based on pragmatics and linguistics, and created a real-world dataset containing 3,097 Chinese examples of implicit harmful language to train and evaluate large language models (LLMs) in reasoning about implicit harmful language.
Pre-trained Language Models Learn Remarkably Accurate Representations of Numbers
Marek Kadlčík (Faculty of Informatics, Masaryk University), Michal Spiegel (Faculty of Informatics, Masaryk University)
Representation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Investigate the precision of numerical embeddings in pre-trained language models and propose a sine-based probe to achieve near-perfect numerical decoding.
Pre-trained Models Perform the Best When Token Distributions Follow Zipf’s Law
Yanjin He (Peking University), Meng Jiang (University of Notre Dame)
Representation LearningTransformerLarge Language ModelTextMultimodalityBiomedical Data
🎯 What it does: Investigated the impact of vocabulary size on pre-trained model performance and proposed using Zipf's Law to measure word distribution for determining the optimal vocabulary size.
Pre-training CLIP against Data Poisoning with Optimal Transport-based Matching and Alignment
Tong Zhang (Zhejiang University), Wenzhi Chen (Zhejiang University)
Adversarial AttackTransformerVision Language ModelContrastive LearningMultimodality
🎯 What it does: Propose OTCCLIP, a CLIP pre-training defense framework based on optimal transport, which reconstructs image-text pairs by leveraging fine-grained visual-textual feature matching to disrupt associations in attacked samples;
Precise In-Parameter Concept Erasure in Large Language Models
Yoav Gur-Arieh (Tel Aviv University), Mor Geva (Tel Aviv University)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelAuto EncoderText
🎯 What it does: Proposes the PISCES method, which directly locates and precisely removes knowledge of specified concepts in the parameter space of language models.
Predicate-Guided Generation for Mathematical Reasoning
Jiajun Chen (New York University), Yik-Cheung Tam (New York University)
GenerationData SynthesisAI Code AssistantTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Constructed the Prolog-MATH dataset using a two-phase automated process to first generate mathematical predicates and then synthesize complete Prolog programs, further enhancing problem-solving coverage through GRPO reinforcement learning.
Predicting Prosodic Boundaries for Children’s Texts
Mansi Dhamne (Sardar Patel Institute Of Technology), Preeti Rao (Indian Institute Of Technology)
ClassificationText
🎯 What it does: This paper proposes a text-based model to predict prosodic pause positions in children's reading materials, focusing on the occurrence of speech boundaries and avoiding unnatural pauses;
Preemptive Detection and Correction of Misaligned Actions in LLM Agents
Haishuo Fang (TU Darmstadt), Iryna Gurevych (TU Darmstadt)
Anomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmark
🎯 What it does: Propose InferAct, which leverages the Theory-of-Mind (ToM) reasoning capability of LLMs to predict critical actions executed by LLM agents, detect and alert users to potential misoperations, and prevent adverse consequences;
PricingLogic: Evaluating LLMs Reasoning on Complex Tourism Pricing Tasks
Yunuo Liu (Hunan University), Xiaoyu Shen (EIT)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkFinance Related
🎯 What it does: Proposed and made public the PricingLogic benchmark to systematically evaluate the reasoning and computational capabilities of large language models in handling real-world tourism pricing tasks (including multiple overlapping discount rules and combination packages).
PRIM: Towards Practical In-Image Multilingual Machine Translation
Yanzhi Tian (Beijing Institute of Technology), Yuhang Guo (Beijing Institute of Technology)
Image TranslationTransformerVision Language ModelImageTextMultimodality
🎯 What it does: Studied multilingual image translation in real-world scenarios, proposed the first real-world multilingual IIMT dataset PRIM, and designed an end-to-end model called VisTrans;
PRIME: Large Language Model Personalization with Cognitive Dual-Memory and Personalized Thought Process
Xinliang Frederick Zhang (University of Michigan), Lu Wang (University of Michigan)
Explainability and InterpretabilityKnowledge DistillationRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Built and evaluated a personalized LLM framework PRIME based on a cognitive dual memory model, introducing a personalized thinking mechanism
PrimeX: A Dataset of Worldview, Opinion, and Explanation
Rik Koncel-Kedziorski (Apple), Tim Paek (Apple)
Data-Centric LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: Constructed and released the PRIMEX dataset, containing public opinions, free-text explanations, and Primal World Beliefs survey results from 858 American respondents.
Primus: A Pioneering Collection of Open-Source Datasets for Cybersecurity LLM Training
Yao-Ching Yu (TrendMicro), Wen-Kwang Tsao (TrendMicro)
Knowledge DistillationRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsTextBenchmark
🎯 What it does: This study constructs a complete dataset system named PRIMUS, covering three stages of LLM training in the field of cybersecurity: pre-training, instruction fine-tuning, and reasoning distillation. Continuous pre-training, instruction fine-tuning, model fusion, and reasoning distillation were implemented based on Llama-3.1-8B-Instruct, ultimately releasing a series of LLMs tailored for cybersecurity.
Principled Personas: Defining and Measuring the Intended Effects of Persona Prompting on Task Performance
Pedro Henrique Luz de Araujo (University of Vienna), Benjamin Roth (University of Vienna)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes and validates three normative objectives for persona prompting (Expertise Advantage, Robustness, Fidelity), and constructs evaluation metrics based on these; subsequently, systematic experiments are conducted on 27 objective tasks involving 9 open-source LLMs (Gemma2, Llama3, Qwen2.5) to investigate the impact of different persona attributes (expert level, name, color, education level, professional specialization) on model performance; finally, three prompt improvement strategies are attempted to enhance Robustness.
Prior Prompt Engineering for Reinforcement Fine-Tuning
Pittawat Taveekitworachai (SCB 10X R&D), Kunat Pipatanakul (SCB 10X R&D)
AI Code AssistantTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Investigates the use of prior prompt engineering (pPE) in reinforcement learning fine-tuning (RFT) to guide language models in learning diverse generation behaviors, systematically evaluating the impact of five pPE strategies on model performance and behavior.
Priority on High-Quality: Selecting Instruction Data via Consistency Verification of Noise Injection
Hong Zhang (Huazhong University of Science and Technology), Kangzheng Liu (Huazhong University of Science and Technology)
Data-Centric LearningSupervised Fine-TuningTextBenchmark
🎯 What it does: This paper proposes a noise-injection-based instruction data consistency verification method to select high-quality instruction data that aligns with the preferences of pre-trained models, avoiding reliance on external models or rules;
PRISM: Efficient Long-Range Reasoning With Short-Context LLMs
Dulhan Jayalath (University of Oxford), Beliz Gunel (Google DeepMind)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes PRISM, an efficient incremental method that uses short-context large language models (LLMs) for long-range reasoning; it reduces the required context length and computational cost through structured memory and incremental reasoning.
Proactive Assistant Dialogue Generation from Streaming Egocentric Videos
Yichi Zhang (Meta), Seungwhan Moon (Meta)
GenerationData SynthesisTransformerLarge Language ModelVision Language ModelVideoTextMultimodality
🎯 What it does: Proposes a framework for generating proactive assistant dialogues from real-time first-person perspective videos.
Proactive Hearing Assistants that Isolate Egocentric Conversations
Guilin Hu (University of Washington), Shyamnath Gollakota (University of Washington)
RecognitionRecurrent Neural NetworkTransformerAudio
🎯 What it does: This paper proposes a proactive hearing assistant that can real-time identify and isolate the conversation partner of the wearer without user manual commands, suppress other interfering speech, and support multi-speaker scenarios.
Probabilistic Soundness Guarantees in LLM Reasoning Chains
Weiqiu You (University of Pennsylvania), Eric Wong (University of Pennsylvania)
Explainability and InterpretabilityTransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Proposes Autoregressive Reasoning Entailment Stability (ARES), a probabilistic reasoning framework that evaluates the credibility of each step in the LLM's generated reasoning chain and provides statistical safety guarantees.
Probability Distribution Collapse: A Critical Bottleneck to Compact Unsupervised Neural Grammar Induction
Jinwook Park (Gwangju Institute of Science and Technology), Kangil Kim (Gwangju Institute of Science and Technology)
Representation LearningText
🎯 What it does: Propose the probability distribution collapse (PDC) problem and design a collapse-relaxing neural parameterization (CRNP) to alleviate this bottleneck, achieving more compact and accurate unsupervised neural syntax induction;