ACL 2026 Papers — Page 15
Annual Meeting of the Association for Computational Linguistics · 2296 papers
Narrative License and Model Sycophancy in LLM Summaries of Scientific Work
Calvin Isch (University of Pennsylvania), Grace Jennings (University of Pennsylvania)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextReview/Survey PaperRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This study evaluates whether large language models (LLMs) exhibit narrative license (NL) when generating academic abstracts, and how the models produce varying degrees of NL under different prompts and user perspectives.
Native Hybrid Attention for Efficient Sequence Modeling
Jusen Du (Tsinghua University), Yu Cheng (Chinese University of Hong Kong)
Computational EfficiencyRepresentation LearningRecurrent Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningTextSequential
🎯 What it does: Propose Native Hybrid Attention (NHA), a hybrid attention architecture that simultaneously integrates linear RNN memory and sliding window Softmax attention at the same level.
NaturalGAIA: A Verifiable Benchmark and Hierarchical Framework for Long-Horizon GUI Tasks
Zihan Zheng (South China Normal University), Qianglong Chen (Zhejiang University)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelVision-Language-Action ModelTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the NaturalGAIA benchmark and the LightManus-Jarvis hierarchical framework for evaluating and enhancing the performance of LLM agents in long-term GUI tasks.
NaturalSloth: Revisiting Denial-of-Service Attacks on Large Language Models
Yiming Chen (National University of Singapore), Haizhou Li (Chinese University of Hong Kong)
Safty and PrivacyComputational EfficiencyAdversarial AttackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: This paper studies denial-of-service (DoS) attacks on large language models (LLMs) in multi-tenant service environments. It proposes a natural instruction-driven DoS attack method and constructs a multi-agent synthesis framework based on manually crafted seeds, generating a large-scale and diverse malicious instruction dataset called NaturalSloth.
Nature-Inspired Population-Based Evolution of Large Language Models
Yiqun Zhang (Northeastern University), Shuyue Hu (Shanghai Artificial Intelligence Laboratory)
OptimizationComputational EfficiencyRepresentation LearningData-Centric LearningTransformerLarge Language ModelReinforcement LearningMixture of ExpertsTextBenchmark
🎯 What it does: Proposed and implemented a population-based evolutionary framework for LLM evolution, named GENOME and GENOME+, which performs gradient-free evolution on multiple models in the LoRA parameter space through genetic operations such as crossover, mutation, selection, succession, and ensemble, to quickly adapt to new tasks.
NavA^3: Understanding Any Instruction, Navigating Anywhere, Finding Anything
Lingfeng Zhang (Tsinghua University), Shanghang Zhang (Peking University)
Depth EstimationAutonomous DrivingOptimizationRobotic IntelligenceReinforcement Learning from Human FeedbackConvolutional Neural NetworkTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelVision-Language-Action ModelSimultaneous Localization and MappingImageTextMultimodalityPoint CloudChain-of-Thought
🎯 What it does: Propose Nav A 3, a hierarchical navigation framework that can parse high-level human instructions and complete long-view navigation tasks in complex indoor environments.
NaviMaster: Learning a Unified Policy for GUI and Embodied Navigation Tasks
Zhihao Luo (East China Normal University), Xin Tan (East China Normal University)
Robotic IntelligenceTransformerLarge Language ModelReinforcement LearningVision Language ModelVision-Language-Action ModelImageTextMultimodality
🎯 What it does: Proposed a unified navigation agent called NaviMaster, which can simultaneously handle GUI navigation and embodied navigation tasks, using a visual-goal trajectory collection pipeline, a unified reinforcement learning framework, and distance-aware dense rewards to achieve joint training and generalization across tasks.
Neo-Classic: A Benchmark for Evaluating Linguistic-Aesthetic Reasoning in Classical Chinese Poetry
Han Zhang (Shanghai Jiao Tong University), Cheng Hua (Shanghai Jiao Tong University)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Proposes the NEO-CLASSIC benchmark, which uses strictly metered poems created by modern experts to evaluate the language aesthetic reasoning ability of LLMs.
NeoAMT: Neologism-Aware Agentic Machine Translation with Reinforcement Learning
Zhongtao Miao (The University of Tokyo), Yoshimasa Tsuruoka (The University of Tokyo)
TransformerLarge Language ModelReinforcement LearningAgentic AITextRetrieval-Augmented Generation
🎯 What it does: Constructed a multilingual neologism machine translation dataset called Neko, and proposed the NeoAMT framework, which uses reinforcement learning and dictionary retrieval to translate sentences containing neologisms.
Nested Browser-Use Learning for Agentic Information Seeking
Baixuan Li (Southeast University), Zhiqiang Gao (Southeast University)
Recommendation SystemAutonomous DrivingComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Designed and implemented the nested browser usage learning framework NestBrowse, which uses four minimal but fully functional browser tools (search, visit, click, fill) and a nested interaction structure, enabling LLMs to efficiently and controllably interact with the browser and complete tasks in deep information retrieval scenarios;
NeuralFSM: Adaptive Multi-Agent Coordination via Learning Finite-State Execution Policy
Jiye Wang (North China Electric Power University), Yuanhe Zhao (North China Electric Power University)
AI Code AssistantRecurrent Neural NetworkGraph Neural NetworkReinforcement LearningPrompt EngineeringContrastive LearningTextGraphChain-of-Thought
🎯 What it does: Propose NeuralFSM, an adaptive multi-agent collaboration framework based on finite state machines, which utilizes temporal graph networks to learn task-conditioned state transitions and sparse communication routing, significantly improving execution efficiency and effectiveness.
NeuReasoner: Towards Explainable, Controllable, and Unified Reasoning via Mixture-of-Neurons
Haonan Dong (Peking University), Guojie Song (Peking University)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningMixture of ExpertsTextChain-of-Thought
🎯 What it does: This paper conducts a fine-grained white-box analysis of failure modes in large reasoning models, identifies key neuron mixtures (MoN), and builds the NeuReasoner framework based on this, achieving explainability, controllability, and unified reasoning improvement.
Neuro-Symbolic Agentic Reinforcement Learning for Long-Term Original Character Companionship and Interaction
Zhenhan Huang (Neurobo AI)
Autonomous DrivingOptimizationFederated LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIMixture of ExpertsTextSequentialRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: A neural-symbolic reinforcement learning framework called NSARL is proposed for original character companions, modeling long-term interactions as POMDP and decomposing the agent into three sub-policies: Router, Memory, and Persona, which are trained in closed-loop reinforcement learning;
Neuron-Aware Active Few-Shot Learning for LLMs
Zhuowei Chen (University of Pittsburgh), Xiang Lorraine Li (University of Pittsburgh)
ClassificationExplainability and InterpretabilityMeta LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: Proposes NEUFS, an active few-shot learning framework based on dynamic neuronal activation, specifically designed to select the most valuable few-shot examples for annotation from unlabeled data in specialized domains for large language models (LLMs);
New Terms, New Toxicity: Consensus-based Chinese Neologism Toxicity Detection via Search-Augmented LLMs
Shiyao Cui (Tsinghua University), Minlie Huang (Tsinghua University)
ClassificationSafty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: This paper studies toxic neologisms in Chinese online contexts, proposes a classification method based on the origins and consensus validation of toxic neologisms, constructs a lexicon containing 974 toxic neologisms along with structured semantic evolution metadata, and introduces SeTox—a framework that combines static large language models with real-time web search to detect toxicity, capable of calling search tools during identification to obtain the latest public context and generate interpretable safety judgments;
Nirvana: A Specialized Generalist Model With Task-Aware Memory Mechanism
Yuhua Jiang (Shanghai AI Laboratory), Bowen Zhou (Shanghai AI Laboratory)
Recommendation SystemFederated LearningExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationMeta LearningDrug DiscoveryReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextTabularTime SeriesSequentialBiomedical DataMagnetic Resonance ImagingBenchmarkFinance RelatedRetrieval-Augmented Generation
🎯 What it does: Proposed a specialized general-purpose model called Nirvana, which can achieve self-supervised fast fine-tuning during testing through task-aware triggers, and dynamically integrate context via a specialized memory updater.
NiuTrans.LMT: Toward Inclusive and Scalable Multilingual Machine Translation with LLMs
Yingfeng Luo (Northeastern University), JingBo Zhu
Data-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Built and released LMT, a centralized multilingual translation model covering 60 languages and 234 directions, with four scales (0.6B/1.7B/4B/8B), achieving high-quality translation through multi-stage pre-training, supervised fine-tuning, and preference optimization.
NL \Rightarrow Schedule: Evaluate Multitask Scheduling Capability of Large Language Models
Wenrui Liao (Sichuan University), Wenqiang Lei (Sichuan University)
OptimizationTransformerLarge Language ModelAgentic AIPrompt EngineeringTextSequentialBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Established a benchmark for the NL ⇒ Schedule task, proposed a dataset containing 240 real-world scenario natural language descriptions and corresponding feasible scheduling pairs, and systematically evaluated the end-to-end performance of nine large language models (LLMs) on this task; subsequently designed the multi-agent framework MANS, which improved the ability of LLMs in program flow induction and conflict detection through division of labor and collaboration.
No More Stale Feedback: Co-Evolving Critics for Open-World Agent Learning
Zhicong Li (Renmin University of China), Yong Liu (Renmin University of China)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes the ECHO framework, which synchronously co-evolves the Critic and Policy, leveraging natural language diagnostic feedback to achieve dynamic improvement of LLM agents.
No Reader Left Behind: Multi-Agent Summaries Everyone Can Understand
Jimin Jung (Korea University), Heuiseok Lim (Korea University)
GenerationExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelAgentic AIPrompt EngineeringTextReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: Proposed NRLB, a unified multi-agent framework for generating plain language summaries compliant with the U.S. Plain Writing Act.
NoisyCausal: A Benchmark for Evaluating Causal Reasoning Under Structured Noise
Zhi Xu (Northeastern University), Yun Fu (Northeastern University)
Explainability and InterpretabilityGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextGraphBenchmarkChain-of-Thought
🎯 What it does: Proposed the NoisyCausal benchmark to evaluate the causal reasoning ability of large language models under structured noise, and designed a modular graph-oriented reasoning framework;
NOSE: Neural Olfactory-Semantic Embedding with Tri-Modal Orthogonal Contrastive Learning
Yanyi Su (Xiamen University), Jun Cheng (Xiamen University)
Drug DiscoveryGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextMultimodalityGraph
🎯 What it does: Propose a tri-modal (molecular structure, receptor sequence, semantic description) aligned olfactory embedding framework called NOSE, achieving modal information decoupling and fusion through orthogonal injection and weak positive sample contrastive learning.
Not All Directions Matter: Towards Structured and Task-Aware Low-Rank Model Adaptation
Xi Xiao (University of Alabama at Birmingham), Hao Xu (Harvard University)
Computational EfficiencyKnowledge DistillationRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringContrastive LearningImageTextMultimodality
🎯 What it does: Propose STRUCTLORA, which improves LoRA by introducing an information bottleneck filter and a graph neural network coordinator during the training phase;
Not All Tokens Matter: Towards Efficient LLM Reasoning via Token Significance in Reinforcement Learning
Hanbing Liu (Tsinghua University), Dongmei Zhang (Microsoft)
Computational EfficiencyTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Propose the BINGO framework, which improves the efficiency of large model chain-of-thought reasoning by utilizing token importance awareness and dynamic length rewards.
NSF-SciFy: Mining the NSF Awards Database for Scientific Claims
Delip Rao (University of Pennsylvania), Chris Callison-Burch (University of Pennsylvania)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Constructed and released the largest dataset of scientific claims and research proposals, NSF-SCIFY, containing approximately 2.8 million scientific claims and research proposals from 400,000 NSF award abstracts;
OASIS: Mitigating Harmful Fine-tuning Attacks on LLMs via Orthogonal and Adaptive Safety Alignment Strategy
Jiayu Tang (University of Electronic Science and Technology of China), Guisong Liu (Southwestern University of Finance and Economics)
Safty and PrivacyComputational EfficiencyAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Proposes the OASIS framework, which actively aligns large language models to resist harmful sample attacks in Fine-Tuning-as-a-Service through orthogonal perturbation and adaptive security layer selection.
OASIS: Online Sample Selection for Continual Instruction Tuning
Minjae Lee (Seoul National University), Jonghyun Choi (Seoul National University)
OptimizationComputational EfficiencyData-Centric LearningTransformerSupervised Fine-TuningPrompt EngineeringContrastive LearningTextMultimodality
🎯 What it does: OASIS proposes an online adaptive sample selection framework in continuous instruction tuning (CIT), which can real-time select the most informative samples from the data stream, significantly shortening training time while maintaining the model's real-time adaptation capability.
Observations and Remedies for Large Language Model Bias in Self-Consuming Performative Loop
Yaxuan Wang (University of California, Santa Cruz), Yang Liu (Ohio State University)
Explainability and InterpretabilityData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes the Self-Consuming Performative Loop (SCPL) for large language models, and investigates the impact of synthetic data on model bias within this framework;
OCP: Outlier-Centric Probing for Dynamic Structured Pruning of LLMs
Yang Ji (Hong Kong University of Science and Technology (Guangzhou)), Ying Sun (National University of Defense Technology)
OptimizationComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsText
🎯 What it does: Propose the OCP framework to achieve input-adaptive structured pruning, using a 'outlier-driven' detection sub-network to determine the number of channels retained at each layer.
OCR-Memory: Optical Context Retrieval for Long-Horizon Agent Memory
Jinze Li (University of Hong Kong), Edith Cheuk-Han Ngai (University of Hong Kong)
RetrievalComputational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose the OCR-Memory framework, which encodes complete interaction trajectories into high-density images and achieves long-term memory through visual retrieval;
OctoBench: Benchmarking Scaffold-Aware Instruction Following in Repository-Grounded Agentic Coding
Deming Ding (Fudan University), Tao Gui (Fudan University)
AI Code AssistantTransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: This paper proposes the OCTOBENCH benchmark to evaluate whether models can follow multi-source, persistent instruction constraints in warehouse-level agent-based coding.
Octopus: Gated Selective Attention for Memory-Bounded Long-Context Inference in Large Language Models
Chien Van Nguyen (University of Oregon), Thien Huu Nguyen (University of Oregon)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose the OCTOPUS framework, introducing Gated Selective Attention (GSA) on a pre-trained Transformer to achieve fixed KV cache inference.
OctoTools: A Multi-Agent Framework with Extensible Tools for Complex Reasoning
Pan Lu, James Zou
Recommendation SystemAnomaly DetectionAutonomous DrivingOptimizationFederated LearningSafty and PrivacyComputational EfficiencyDrug DiscoveryAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringMixture of ExpertsImageTextMultimodalityTabularTime SeriesBiomedical DataBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes OctoTools, a training-agnostic, scalable multi-agent framework that enables LLMs to accomplish complex reasoning tasks through multi-step tool calls using standardized tool cards and planning-execution processes.
ODL-TempLLM: Ontology-Guided and Description Logic-Reasoned Temporal Reasoning with LLMs
Jinshuo Liu (Wuhan University), Jeff Z. Pan (Wuhan University)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose ODL-TempLLM, constructing an event-time ontology, symbolic reasoning, and logical constraint retrieval to explicitly model temporal reasoning and enhance the temporal reasoning capability of LLMs.
ODUTQA-MDC: A Task for Open-Domain Underspecified Tabular QA with Multi-turn Dialogue-based Clarification
Zhensheng Wang (Beijing Normal University), Weijia Jia (Beijing Normal University)
RetrievalRecommendation SystemData-Centric LearningTransformerLarge Language ModelAgentic AIPrompt EngineeringTextTabularBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the ODUTQA-MDC task, construct the ODUTQA dataset, fine-grained fuzzy detection labels, dynamic clarification interface, and implement the MAIC-TQA multi-agent framework.
OLA: Output Language Alignment in Code-Switched LLM Interactions
Juhyun Oh (KAIST), Alice Oh (KAIST)
Explainability and InterpretabilityData-Centric LearningAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: The study investigates the failure of large language models (LLMs) to implicitly output language alignment during code-switching interactions, constructs the OLA benchmark, and proposes the CS-DPO method based on preference alignment.
OMIBench: Benchmarking Olympiad-Level Multi-Image Reasoning in Large Vision-Language Models
Qiguang Chen (Harbin Institute of Technology), Wanxiang Che (Harbin Institute of Technology)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningDrug DiscoveryTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkPhysics RelatedRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose OMIBench, a multi-graph Olympiad-level reasoning benchmark, to evaluate the ability of large vision-language models in multi-graph reasoning.
Omni-Embed-Audio: Leveraging Multimodal LLMs for Robust Audio-Text Retrieval
HaeJun Yoo (Sogang University), Du-Seong Chang (Sogang University)
RetrievalTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningTextMultimodalityAudio
🎯 What it does: This study proposes the Omni-Embed‑Audio (OEA) model, which utilizes a multi-modal large language model (LLM) with native audio understanding as a unified encoder, and achieves audio-text alignment through LoRA adaptation; meanwhile, it introduces the User Intent Query (UIQ) benchmark and hard negative sample discrimination metrics to comprehensively evaluate the robustness of the retrieval system in real search scenarios.
Omni-I2C: A Holistic Benchmark for High-Fidelity Image-to-Code Generation
Jiawei Zhou (Wuhan University), Jing Zhang (Wuhan University)
GenerationAI Code AssistantTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose Omni-I2C, a large-scale, cross-lingual, and multi-domain Image-to-Code benchmark, to evaluate the ability of large multimodal models to convert complex digital graphics into executable code.
Omni-RewardBench: Toward a Comprehensive Evaluation of Generative Reward Models Across Modalities
Chi-Min Chan (Hong Kong University of Science and Technology), Yike Guo (Hong Kong University of Science and Technology)
Reinforcement Learning from Human FeedbackLarge Language ModelTextMultimodalityBenchmark
🎯 What it does: Proposes Omni-RewardBench, the first benchmark for evaluating multimodal reward models.
On Emergent Social World Models — Evidence for Functional Integration of Theory of Mind and Pragmatic Reasoning in Language Models
Polina Tsvilodub (University of Tübingen), Michael Franke (University of Tübingen)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringWorld ModelText
🎯 What it does: Investigate whether large language models possess social world models and explore whether theory of mind (ToM) and pragmatic reasoning share the same computational mechanisms.
On the (In-)Security of the Shuffling Defense in the Transformer Secure Inference
Zhengyi Li (Shanghai Jiao Tong University), Minyi Guo (Shanghai Jiao Tong University)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelContrastive LearningText
🎯 What it does: In Transformer secure inference, the authors propose an alignment attack against the commonly used shuffled activation defense. By aligning approximate activation vectors obtained from multiple queries and solving a linear system, the attack can recover model weights.
On the Continued Value of Universal Dependencies in the Era of Large Language Models
Wenxi Li (Minzu University of China), Jingyu Peng (City University of Hong Kong)
ClassificationRecognitionExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextMultimodalityBenchmark
🎯 What it does: Explore the continued value of Universal Dependencies (UD) in the era of large language models (LLMs), propose a cross-lingual adversarial paraphrase identification task (CAPI) to evaluate the impact of UD on LLMs, and systematically compare three UD integration strategies (Prompt, Fine-Tuning, Attention).
On the Effect of Hyperparameters in Language Modeling for Computational Linguistics
Ruoxi Ning (University of Waterloo), Freda Shi (MBZUAI)
Explainability and InterpretabilityComputational EfficiencyHyperparameter SearchData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: Reproduce and systematically evaluate three linguistics experiments based on language models (lexical acquisition, syntactic generalization, and word order preference), investigating the impact of hyperparameter variations on experimental conclusions.
On the Emergence and Test-Time Use of Structural Information in Large Language Models
Michelle Chao Chen (ETH Zurich), Siyuan Guo (Max Planck Institute for Intelligent Systems)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: This paper investigates how large language models learn abstract structural information during training, and how this structural information is utilized during testing, with a particular focus on the generation and combination of structural transformations.
On the Emotion Understanding of Synthesized Speech
Yuan Ge (Northeastern University), Tong Xiao (Kunming University of Science and Technology)
RecognitionDomain AdaptationTransformerLarge Language ModelContrastive LearningTextAudio
🎯 What it does: Systematically evaluate the generalization ability of synthetic speech emotion recognition models, and analyze the gap between human speech and synthetic speech in emotional understanding;
On the Hidden Objective Biases of Group-based Reinforcement Learning
Aleksandar Fontana (Scuola Superiore Sant'Anna), Andrea Saracino (Scuola Superiore Sant'Anna)
OptimizationTransformerLarge Language ModelReinforcement LearningContrastive LearningTextReview/Survey Paper
🎯 What it does: This paper proposes a unified agent objective framework for group-based reinforcement learning methods (such as GRPO), and systematically analyzes gradient bias, the AdamW optimizer with reward scaling, and pruning out-of-bound issues caused by momentum.
On the Proper Treatment of Units in Surprisal Theory
Samuel Kiegeland, Ryan Cotterell (ETH Zürich)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose a general framework that explicitly separates the 'unit inventory (U)' from the language model symbol set (Σ), and provides a derivation method to convert the probability of a pre-trained language model into any unit inventory.
On the Rejection Criterion for Proxy-based Test-time Alignment
Ayoub Hammal (Université Paris-Saclay), Caio Corro (INSA Rennes)
Domain AdaptationExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningReinforcement Learning from Human FeedbackGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes a new probabilistic graphical model to describe proxy-based test-time alignment, and based on this model, it introduces a new rejection criterion;
On the Role of Discriminative Models in Generative Relation Extraction
Guozheng Li (Southeast University), Ziyu Shang (Southeast University)
ClassificationGenerationRepresentation LearningTransformerSupervised Fine-TuningPrompt EngineeringContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Propose the D2G framework, which injects the top-k relational candidates from discriminative models into generative relation extraction, thereby improving the extraction performance.
On-policy Reinforcement Fine-tuning with Offline reward for Multi-step Embodied Planning
Di Wu (Tongji University), Bo Jin (Tongji University)
Robotic IntelligenceTransformerSupervised Fine-TuningReinforcement LearningVision-Language-Action ModelImageTextMultimodality
🎯 What it does: Propose the ORBIT framework, which uses on-policy reinforcement fine-tuning (RFT) with offline expert trajectory rewards to address the high-cost interaction and sparse reward problems in multi-step embodied planning.
On-the-Fly VLA Adaptation via Test-Time Reinforcement Learning
Changyu Liu (University of Missouri-Kansas City), Cheng Han (University of Missouri-Kansas City)
Robotic IntelligenceTransformerReinforcement LearningPrompt EngineeringVision Language ModelVision-Language-Action ModelImageTextMultimodality
🎯 What it does: This paper proposes a framework based on test-time reinforcement learning (TT-VLA), which can perform online adaptation of Vision-Language-Action (VLA) models during the inference phase;
One Battle After Another: Probing LLMs’ Limits on Multi-Turn Instruction Following with a Benchmark Evolving Framework
Qi Jia (Shanghai Artificial Intelligence Laboratory), Guangtao Zhai (Shanghai Jiao Tong University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringFlow-based ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed an scalable multi-turn instruction-following evaluation framework and the EvolIF benchmark, utilizing a three-layer tracking mechanism and a query synthesis agent driven by large language models (LLMs) to dynamically generate dialogues, and introducing a patience threshold based on Flow theory and process-oriented evaluation metrics;
One Cognitive Loop Is Enough: SODA unlocks Pure-Text Spatial Reasoning in Large Language Models
Shunwen Bai (Zhejiang University), Mingjun Cheng (Zhejiang University)
Computational EfficiencyKnowledge DistillationRepresentation LearningData-Centric LearningRobotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringMixture of ExpertsTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the SODA framework, which integrates the OODA cognitive cycle into LLMs to enhance spatial reasoning capabilities in pure text environments, constructs the SPOD-143k large-scale spatial dataset and the SPOD-Bench benchmark, and achieves model fine-tuning and reinforcement learning through a two-stage training approach of SFT+GRPO.
One Pair Suffices: Unlocking Universal Zero-Shot Translation via Cross-Architecture Alignment
Hao Zong (Dalian University of Technology), Degen Huang (Dalian University of Technology)
Domain AdaptationComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningTextMultimodalityBenchmark
🎯 What it does: Propose a parameter-efficient Hybrid Cross-Alignment (HCA) framework, connecting a frozen NLLB encoder with a frozen Qwen decoder through bidirectional adapters, achieving zero-shot translation across architectures;
One Persona, Many Cues, Different Results: How Sociodemographic Cues Impact LLM Personalization
Franziska Weeber (University of Stuttgart), Sebastian Padó (University of Stuttgart)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Investigated the impact of six commonly used persona prompts (name, explicit description, conversation history) on the personalization bias of large language models, and constructed a scalable multi-task evaluation benchmark.
One Refiner to Unlock Them All: Inference-Time Reasoning Elicitation via Reinforcement Query Refinement
Yixiao Zhou (Zhejiang University), Hehe Fan (Zhejiang University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Propose a reasoning-time query refinement framework called ReQueR, which utilizes reinforcement learning to train a Refiner to activate the reasoning capabilities of frozen large language models by rewriting user questions.
One Single Hub Text Breaks CLIP: Identifying Vulnerabilities in Cross-Modal Encoders via Hubness
Hiroyuki Deguchi (NTT, Inc.), Yusuke Sakai (Nara Institute of Science and Technology)
RetrievalExplainability and InterpretabilityComputational EfficiencyAdversarial AttackTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: By analyzing hub embeddings in high-dimensional embedding spaces, a method is proposed to identify hub text in cross-modal encoders (e.g., CLIP), and further optimize it using beam local search;
One Tokenizer To Rule Them All: Emergent Language Plasticity via Multilingual Tokenizers
Diana Abagyan (Cohere Labs), Sara Hooker (Cohere Labs)
Federated LearningComputational EfficiencyKnowledge DistillationRepresentation LearningData-Centric LearningMeta LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextMultimodality
🎯 What it does: This paper proposes a 'Universal Tokenizer,' which covers more languages than the main training language during the pre-training phase, thereby significantly enhancing the model's adaptability to new languages during subsequent training (linguistic plasticity).
One-step Nonautoregressive Natural Language Generation with Shortcut Flow Matching Models
Jędrzej Warczyński (Poznan University of Technology), Mateusz Lango (Poznan University of Technology)
GenerationTransformerDiffusion modelScore-based ModelFlow-based ModelRectified FlowText
🎯 What it does: This paper proposes a first-order non-autoregressive natural language generation method that completes text generation in a single step using the shortcut flow matching model.
OneRec-Think: In-Text Reasoning for Generative Recommendation
Zhanyu Liu, Guorui Zhou (Kuaishou Inc)
Recommendation SystemExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextSequentialChain-of-Thought
🎯 What it does: Propose the OneRec-Think framework, which combines large language models with generative recommendation systems to achieve text-based chain reasoning and personalized recommendations.
Open Schrödinger’s Closed Box: Identifying Retrieval Augmented Generation in API-Accessible Large Language Model Services
Yukun Jiang (CISPA Helmholtz Center for Information Security), Yang Zhang (CISPA Helmholtz Center for Information Security)
Anomaly DetectionSafty and PrivacyTransformerPrompt EngineeringContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Proposed a framework named RAG-ID to detect whether API-accessible large language model services employ retrieval-augmented generation (RAG) technology, and further infer the models and knowledge bases they use, providing six attack methods that can efficiently identify RAG attributes in both white-box and black-box scenarios.
Open Your Model’s Eyes: Video and Context-Aware Multimodal Backchannel Prediction
Min-Jae Kim (Korea University), Gyeong-Moon Park (Korea University)
ClassificationTransformerSupervised Fine-TuningVision Language ModelContrastive LearningVideoTextMultimodalityAudio
🎯 What it does: A novel multimodal behind-channel prediction framework called CAMA-BC was studied, which integrates audio, text, and video information and achieves modality alignment through hierarchical cross-attention, addressing issues such as visual information bias, temporal deviation, and imbalance between context and response.
OpenRubrics: Towards Scalable Synthetic Rubric Generation for Reward Modeling and LLM Alignment
Tianci Liu (Purdue University), Haoyu Wang (University at Albany)
Data SynthesisReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningContrastive LearningTextBenchmark
🎯 What it does: Propose the OpenRubrics dataset and the Rubric-based Reward Model (RUBRIC-RM), which automatically construct high-quality rubrics and use them for reward modeling and policy optimization through contrastive rubric generation and consistency filtering.
OPeRA: A Dataset of Observation, Persona, Rationale, and Action for Evaluating LLMs on Human Online Shopping Behavior Simulation
Ziyi Wang (Northeastern University), Dakuo Wang (Northeastern University)
Recommendation SystemExplainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringVision-Language-Action ModelTextSequentialBenchmark
🎯 What it does: This paper proposes and constructs the OPeRA dataset, which records real users' observations, personalized information, reasons, and operations during online shopping, aiming to evaluate the ability of large language models to simulate user behavior.
OptiCo: Adaptive Distributed Training Optimization via Collaborative Agent Reasoning
Sheng Chen (Zhejiang Lab), Yang Liu (Nanyang Technological University)
OptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelAgentic AIPrompt EngineeringTextChain-of-Thought
🎯 What it does: Proposes OptiCo, a distributed training strategy automation optimization framework based on multi-agent and large language models, which can automatically tune parameters such as tensor parallelism, pipeline parallelism, and data parallelism through interactive reasoning during the training process.
Optimizing Length Compression in Large Reasoning Models
Zhengxiang Cheng (Huazhong University of Science and Technology), Tianyi Zhou (MBZUAI)
OptimizationComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: Propose a post-training RL method called LC-R1, which utilizes GRPO and combines length reward with compression reward to specifically compress the ineffective reasoning of large inference models;
Optimizing RAG Rerankers with LLM Feedback via Reinforcement Learning
Yuhang Wu (Nanjing University of Science and Technology), Rui Xia (Nanjing University)
RetrievalOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextRetrieval-Augmented Generation
🎯 What it does: In Retrieval-Augmented Generation systems, end-to-end training of the document re-ranker is performed using the reinforcement learning framework RRPO, directly optimizing the quality of LLM generation;
Optimizing User Profiles via Contextual Bandits for Retrieval-Augmented LLM Personalization
Linfeng Du (McGill University), Haolun Wu (McGill University)
RetrievalRecommendation SystemOptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose a user profiling optimization framework called PURPLE based on contextual bandits, improving the personalization effectiveness of retrieval-enhanced LLMs.
OS-Sentinel: Towards Safety-Enhanced Mobile GUI Agents via Hybrid Validation in Realistic Workflows
Qiushi Sun (University of Hong Kong), Lingpeng Kong (University of Hong Kong)
Safty and PrivacyTransformerPrompt EngineeringVision Language ModelTextSequential
🎯 What it does: Built a new platform for mobile device security research called MobileRisk-Live and the corresponding frozen trajectory dataset MobileRisk, and proposed the OS-Sentinel hybrid security detection framework;
OS-Symphony: A Holistic Framework for Robust and Generalist Computer-Using Agents
Bowen Yang (University of Science and Technology of China), Zichen Ding (Shanghai AI Laboratory)
Domain AdaptationAutonomous DrivingOptimizationComputational EfficiencyRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringVision Language ModelVision-Language-Action ModelImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the OS-SYMPHONY framework, combining Orchestrator, Reflection-Memory Agent (RMA), and multi-tool agents to address long-term robustness and domain generalization issues.
OSCBench: Benchmarking Object State Change in Text-to-Video Generation
Xianjing Han (National University of Singapore), Jingjing Chen (Fudan University)
GenerationData SynthesisTransformerLarge Language ModelVision Language ModelVideoTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed OSCBench, a benchmark specifically designed to evaluate object state changes (OSC) in text-to-video generation, constructing three types of scenarios: conventional, innovative, and composite, and assessing the OSC capabilities of six models.
Outcome Accuracy is Not Enough: Aligning the Reasoning Process of Reward Models
Binghai Wang (Fudan University), Junyang Lin (Alibaba Group)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This paper investigates the 'deceptive alignment' problem of reward models (RMs) during the RLHF process, proposes a fine-grained rationale consistency metric, and uses the METAJUDGE framework to perform strict one-to-one semantic matching on model-generated rationales. Subsequently, a hybrid reward combining rationale consistency and outcome accuracy is introduced in the training of the generation reward model (GenRM), proving that the multiplicative form can effectively prevent models from 'cutting corners' based solely on outcome accuracy, ultimately improving the performance of reward models on multiple benchmarks and enhancing the generation quality of RLHF.
Outcome-Grounded Advantage Reshaping for Fine-Grained Credit Assignment in Mathematical Reasoning
Ziheng Li (Fudan University), Hongcheng Guo (Fudan University)
OptimizationExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelReinforcement LearningPrompt EngineeringMixture of ExpertsTextBenchmarkChain-of-Thought
🎯 What it does: Proposes a fine-grained credit assignment framework called OAR based on the impact of final answers, aimed at improving reward propagation in Group Relative Policy Optimization (GRPO) for long reasoning tasks.
Overcoming Copyright Barriers in Corpus Distribution Through Non-Reversible Hashing
Arthur Amalvy (Academia Sinica), Hen-Hsen Huang (Academia Sinica)
Safty and PrivacyPrompt EngineeringText
🎯 What it does: Propose a method that utilizes non-reversible hashing to share copyrighted text annotations, allowing users with the original text to legally obtain annotations.
P-Check: Advancing Personalized Reward Model via Learning to Generate Dynamic Checklist
Kwangwook Seo (Yonsei University), Dongha Lee (Yonsei University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelContrastive LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the P-CHECK framework, training a pluggable checklist generator to dynamically generate user-specific evaluation criteria to guide personalized reward prediction.
PACE: Predictive Adaptive Context Extraction for Long-Horizon LLM Agents
Lei Wei (Alibaba International Digital Commerce Group), Bin Wang (Peking University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringAuto EncoderTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Designed a framework called PACE to dynamically manage the context of long-sequence LLM agents, adaptively extracting and compressing historical interactions based on the idea of next-step prediction;
PaCoRe: Learning to Scale Test-Time Compute with Parallel Coordinated Reasoning
Jingcheng Hu (Tsinghua University), Heung-Yeung Shum (StepFun)
Computational EfficiencyAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextSequentialRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed a parallel coordination reasoning framework called PaCoRe, which can achieve computational scaling up to millions of tokens during testing by performing multi-round parallel exploration and message compression within a fixed context window;
PAM: Enhancing General Alignment of Large Reasoning Models through Priority-Aware Metacognition
Zhihao Xu (Renmin University of China), Xiting Wang (Renmin University)
Safty and PrivacyExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper proposes the PRIORITY-AWARE METACOGNITION (PAM) method, which enhances the model's performance on general alignment tasks such as safety, helpfulness, and instruction following by first allowing large reasoning models to identify task priorities (such as harmfulness and helpfulness) and generate metacognition (self-monitoring and strategy planning), thus maintaining reasoning depth.
Paper Circle: An Open-source Multi-agent Research Discovery and Analysis Framework
Komal Kumar (Mohamed bin Zayed University of Artificial Intelligence), Hisham Cholakkal (Aws Generative Ai Innovation Center)
RetrievalRecommendation SystemExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationData-Centric LearningTransformerLarge Language ModelAgentic AIPrompt EngineeringContrastive LearningTextGraphReview/Survey PaperBenchmarkRetrieval-Augmented Generation
🎯 What it does: Paper Circle provides a multi-agent collaborative research literature retrieval and analysis platform, automatically completing tasks such as retrieval, sorting, knowledge graph construction, and peer review.
Paper2Rebuttal: A Multi-Agent Framework for Transparent Author Response Assistance
Qianli Ma (Shanghai Jiao Tong University), Zhipeng Zhang (Shanghai Jiao Tong University)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose REBUTTALAGENT, a multi-agent framework that helps authors efficiently and verifiably write rebuttals through a 'write after evidence' process involving focus splitting, internal and external evidence construction, plan verification, and separation of the final draft.
PaperRegister: Boosting Flexible-grained Paper Search via Hierarchical Register Indexing
Zhuoqun Li (Chinese Academy of Sciences), Le Sun (Chinese Academy of Sciences)
RetrievalTransformerLarge Language ModelReinforcement LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: A hierarchical index tree based on a hierarchical registry was constructed to support different granularity levels in paper retrieval queries, and an online adaptive retrieval system was implemented with a view recognizer.
Par-ITA: Benchmarking Seq2Seq and LLMs on a Human-Supervised Parallel Corpus for Italian Hyperpartisan Neutralization
Michele Joshua Maggini (Universidade de Santiago de Compostela), Pablo Gamallo (Universidade de Santiago de Compostela)
Data-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Created the first fully human-supervised Italian extremist neutralization parallel corpus, Par-ITA, and proposed a three-stage annotation process;
PAR: Training-Free Positional Perturbation and Attention Recycling for Faithful OCR
Yao Yao (Shanghai Jiao Tong University), Hai Zhao (Shanghai Jiao Tong University)
RecognitionTransformerPrompt EngineeringVision Language ModelDiffusion modelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: Propose a training-agnostic, inference-time intervention framework called PAR, designed to suppress hallucinations caused by language priors in visual language models during OCR tasks, thereby improving the visual consistency of text recognition.
ParaCodex: A Profiling-Guided Autonomous Coding Agent for Reliable Parallel Code Generation and Translation
Erel Kaplan (Technion Israel Institute of Technology), Gal Oren (Technion Israel Institute of Technology)
OptimizationAI Code AssistantTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose PARACODEX, an autonomous coding agent based on Codex, designed for the automatic translation and migration of OpenMP GPU offload code in high-performance computing;
Parallel Test-Time Scaling for Latent Reasoning Models
Runyang You, Wenjie Li (Shandong Jianzhu University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsContrastive LearningTextChain-of-Thought
🎯 What it does: This paper studies how to achieve parallel test-time scaling (TTS) in reasoning models with continuous latent spaces, and proposes two uncertainty-based sampling strategies as well as a latent reward model to support multi-path reasoning and aggregation.
Parallel Universes, Parallel Languages: A Comprehensive Study on LLM-based Multilingual Counterfactual Example Generation
Qianli Wang (Technische Universität Berlin), Vera Schmitt (Technische Universität Berlin)
GenerationData SynthesisAdversarial AttackTransformerLarge Language ModelPrompt EngineeringTextMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper addresses adversarial causal example generation in multilingual environments, using large language models (LLMs) to conduct systematic evaluations across six languages. It compares adversarial examples generated directly versus those generated through translation, and automatically evaluates their effectiveness, fluency, and minimality. Furthermore, it analyzes cross-lingual edit similarity and error types, and explores the impact of adversarial data augmentation (CDA) on model performance and robustness.
Paramanu: Compact and Competitive Monolingual Language Models for Low-Resource Morphologically Rich Indian Languages
Mitodru Niyogi (Université Grenoble Alpes), Arnab Bhattacharya (Indian Institute of Technology Kanpur)
GenerationData SynthesisComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningText
🎯 What it does: Trained a 400M parameter monolingual generative language model called PARAMANU for five Indian languages (Bengali, Hindi, Marathi, Tamil, Telugu), paired with a publicly available low-fidelity morphology-aligned tokenizer; also created a Bengali instruction set and translated it into the remaining languages for instruction fine-tuning.
Parameter Importance is Not Static: Evolving Parameter Isolation for Supervised Fine-Tuning
Zekai Lin (Tencent Hunyuan), Minlong Peng (Tencent Hunyuan)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Designed and implemented a dynamic parameter isolation method called EPI, which can update the parameter protection mask in real-time during supervised fine-tuning based on gradient information, in order to reduce task interference and catastrophic forgetting.
Paraphrasing as Zero-shot Translation with Feature-guided Diversity Enhancement
Ziyue Yan (Zhengzhou University), Hongfei Xu (Zhengzhou University)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: Propose a framework that utilizes multilingual neural machine translation (MNMT) for zero-shot paraphrasing, directly generating diverse paraphrases in the same language within a single model;
PARASITE: Conditional System Prompt Poisoning to Hijack LLMs
Viet Pham (Indiana University Bloomington), Thai Le (Indiana University Bloomington)
Safty and PrivacyExplainability and InterpretabilityAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: Developed a black-box framework called PARASITE, which injects a 'sleeping agent' into system prompts, generating misleading answers only when triggered by specific questions, while maintaining the model's performance on regular inputs.
ParaSuite: Boosting LLM Reasoning via Paradox Resolution
Bin Chen (University of Chinese Academy of Sciences), Hongyang Chen (University of Chinese Academy of Sciences)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the ParaSuite pipeline, including data synthesis, evaluation, and training, and construct the PARADOX paradox dataset. Design two tasks, PCI and MEG, train ParadoxBreaker-7B to enhance the ability of LLMs in paradox identification and resolution, and achieve better performance in general STEM reasoning.
PARIF: Pushing the Pareto Frontier of Instruction Following and Reasoning with Curriculum Reinforcement Learning
Rongchuan Mu (Harbin Institute of Technology), Bing Qin (Harbin Institute of Technology)
OptimizationComputational EfficiencyRepresentation LearningData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringDiffusion modelAuto EncoderContrastive LearningTextChain-of-Thought
🎯 What it does: Propose a two-stage RLVR curriculum learning framework called PARIF to enhance the comprehensive performance of large reasoning models in instruction following and reasoning capabilities.
Parity-Aware Byte-Pair Encoding: Improving Cross-lingual Fairness in Tokenization
Negar Foroutan (EPFL), Rico Sennrich (University of Zurich)
CompressionRepresentation LearningData-Centric LearningTransformerLarge Language ModelContrastive LearningTextMultimodality
🎯 What it does: Proposed Parity-Aware BPE, improving the byte pair encoding (BPE) algorithm by prioritizing languages with the lowest compression rate, reducing the discrepancy in the number of tokens across languages.
PASs-MoE: Mitigating Misaligned Co-drift among Router and Experts via Pathway Activation Subspaces for Continual Learning
ZhiYan Hou, Jinqiao Wang (Southeast University)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerSupervised Fine-TuningMixture of ExpertsImageTextMultimodalityBenchmark
🎯 What it does: To address continuous instruction fine-tuning for multi-modal large models, this paper proposes a Mixture-of-Experts LoRA method based on path activation subspaces, significantly alleviating the co-drift problem between routers and experts, and enhancing the model's memory and generalization capabilities on task sequences.
PaT: Planning-after-Trial for Efficient Test-Time Code Generation
Youngsik Yoon (POSTECH), Jungseul Ok (POSTECH)
Computational EfficiencyAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Proposed the Planning-after-Trial (PaT) strategy, which avoids unnecessary planning costs by first experimenting and then planning, significantly reducing computational costs during testing;
PaTaRM: Bridging Pairwise and Pointwise Signals via Preference-Aware Task-Adaptive Reward Modeling
Ai Jian (Meituan), Xunliang Cai (Meituan)
OptimizationExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringMixture of ExpertsGenerative Adversarial NetworkTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose PaTaRM, a unified framework capable of training pointwise generative reward models using pairwise preference data.
Patches of Nonlinearity: Instruction Vectors in Large Language Models
Irina Bigoulaeva (Technical University of Darmstadt), Iryna Gurevych (Technical University of Darmstadt)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper systematically studies how large language models internally process instructions, identifies and analyzes the geometric structure and causal roles of instruction vectors (Instruction Vectors, IVs), and further reveals their evolution across different training stages (SFT and DPO).
PDR: A Plug-and-Play Positional Decay Framework for LLM Pre-training Data Detection
Jinhan Liu, Dandan Guo (Jilin University)
Safty and PrivacyTransformerLarge Language ModelText
🎯 What it does: Proposes an untrained, plug-and-play Positional Decay Reweighting (PDR) framework to enhance the probabilistic-based methods of detecting membership inference attacks on LLM pre-training data.
PDTrim: Targeted Pruning for Prefill-Decode Disaggregation in Inference
Hao Zhang (Beijing Academy of Artificial Intelligence), Yonghua Lin (Beijing Academy of Artificial Intelligence)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: Proposes a block-level pruning method called PDTrim for the Prefill-Decode (PD) disaggregation inference scenario, which iteratively removes blocks using a pruning set and a distillation set, and performs pruning optimization separately for the prefill and decode stages;