ACL 2026 Papers — Page 10
Annual Meeting of the Association for Computational Linguistics · 2296 papers
HalluAudio: A Comprehensive Benchmark for Hallucination Detection in Large Audio-Language Models
Feiyu Zhao (Tianjin University), Jianguo Wei (Tianjin University)
Prompt EngineeringContrastive LearningTextBenchmarkAudio
🎯 What it does: Constructed the HalluxAudio dataset, which covers three major audio domains—speech, environmental sounds, and music—with over 5k human-verified QA pairs, to systematically evaluate the hallucination behavior of large-scale audio-language models.
Hallucination Detection in LLMs with Topological Divergence on Attention Graphs
Alexandra Bazarova (Applied AI Institute), Alexey Zaytsev (Applied AI Institute)
Anomaly DetectionExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Proposes a training-free, attention map topology-based hallucination detection method called TOHA, which uses the topological divergence of attention maps, MTop-Div G(R,P), to measure the structural differences between responses and prompts, thereby determining hallucinatory outputs of LLMs in RAG scenarios.
HalluCitation Matters: Revealing the Impact of Hallucinated References with 300 Hallucinated Papers in ACL Conferences
Yusuke Sakai (Nara Institute of Science and Technology), Taro Watanabe (Nara Institute of Science and Technology)
Anomaly DetectionData-Centric LearningSupervised Fine-TuningPrompt EngineeringTextReview/Survey PaperBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper systematically analyzes hallucinated citations (HalluCitation) appearing in conference papers from ACL, NAACL, and EMNLP 2024-2025, and provides detection methods and their impact on the peer review process.
Hard2Verify: A Step-Level Verification Benchmark for Open-Ended Frontier Math
Shrey Pandit (Salesforce AI Research), Shafiq Joty (Salesforce AI Research)
Large Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed the Hard2Verify benchmark, using manually annotated step-by-step verification data to evaluate the step-by-step verifiers of state-of-the-art mathematical reasoning models.
HarDBench: A Benchmark for Draft-Based Co-Authoring Jailbreak Attacks for Safe Human–LLM Collaborative Writing
EunTae Kim, Buru Chang (Korea University)
Safty and PrivacyAdversarial AttackReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Propose HarDBench benchmark for evaluating large language models in jailbreaking attacks during draft-based collaborative writing, and develop a safety-practical balanced alignment method called SUBA.
Harmonizing the Past, Present, and Future: A Null-Space Constrained Region-Specific Method for Continual Learning in LLMs
Jinhui Chen (Institute of Automation, Chinese Academy of Sciences), Jun Zhao (Institute of Automation, Chinese Academy of Sciences)
Federated LearningExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose a zero-replay continual learning framework called PaRSP, which achieves a triple balance of stability, plasticity, and sustainability for large language models by leveraging task-specific parameter region activation and orthogonal projection in the zero space.
HarmRLVR: Weaponizing Verifiable Rewards for Harmful LLM Alignment
Yuexiao Liu (Tsinghua University), Jing Shao (Shanghai Artificial Intelligence Laboratory)
Adversarial AttackReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Designed and implemented a reinforcement learning attack based on verifiable rewards—HarmRLVR, demonstrating RLVR alignment reversible risk;
Harnessing Negative Signals: Reinforcement Distillation from Teacher Data for LLM Reasoning
Shuyao Xu (INFLY TECH COMPANY LIMITED), Yuan Qi (INFLY TECH COMPANY LIMITED)
Knowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Under the offline distillation setup, training LLMs using positive and negative reasoning trajectories to enhance reasoning performance.
HARPO: Hierarchical Agentic Reasoning for User-Aligned Conversational Recommendation
Subham Raj (Indian Institute of Technology Patna), Sriparna Saha (Indian Institute of Technology Patna)
Recommendation SystemExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringMixture of ExpertsContrastive LearningTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the HARPO framework, redefining conversational recommendation as a structured decision-making problem aimed at user-aligned multi-dimensional quality (relevance, diversity, satisfaction, engagement), and achieving high-quality recommendations through hierarchical preference learning, tree-based reasoning, domain-agnostic tool abstraction, and multi-agent refinement.
HAT: Hallucination Annotation for Translation
Rajen Chatterjee (Apple), Allen Lee (Apple)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: This paper presents HAT, a multilingual translation hallucination annotation dataset covering 38 language pairs and 350k samples.
HCSpec: Two-Tier Horizontal Cascade Speculative Decoding for High-Efficiency Large Language Model Inference
Yizhou Zhang (Shanghai Jiao Tong University), Erhu Feng (Shanghai Jiao Tong University)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed a two-layer horizontal cascading Speculative Decoding framework called HCSpec, which uses a high-capacity dual-layer Transformer for precise prediction in the early stage, and a lightweight single-layer Transformer in the later stage to reduce computational cost, thereby achieving more efficient LLM inference.
HEALing Entropy Collapse: Enhancing Exploration in Few-Shot RLVR via Hybrid-Domain Entropy Dynamics Alignment
Zhanyu Liu (Xiamen University), Jinsong Su (Xiamen University)
Data-Centric LearningMeta LearningReinforcement Learning from Human FeedbackReinforcement LearningPrompt EngineeringTextTabularBiomedical DataBenchmarkPhysics RelatedChain-of-Thought
🎯 What it does: Propose the HEAL framework, combining high-quality general-domain data selection and entropy dynamics alignment (EDA) reward to address the entropy collapse problem in few-shot RLVR;
HeLa-Mem: Hebbian Learning and Associative Memory for LLM Agents
Jinchang Zhu (Hong Kong University of Science and Technology), Menglin Yang (Chinese University of Hong Kong)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningAdversarial AttackReinforcement Learning from Human FeedbackGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextSequentialBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes a bio-inspired memory architecture called HeLa-Mem, modeling the conversation history of LLM agents as a dynamic Hebbian graph to achieve online association and reflective integration.
HERMES: KV Cache as Hierarchical Memory for Efficient Streaming Video Understanding
Haowei Zhang (Fudan University), Xipeng Qiu (Fudan University)
RecognitionComputational EfficiencyTransformerLarge Language ModelVision Language ModelContrastive LearningVideoMultimodality
🎯 What it does: Propose a training-free, plug-and-play HERMES framework that achieves real-time video stream understanding by restructuring the KV cache into a hierarchical multi-granularity memory; during inference, directly reuse the compressed KV cache without additional retrieval or external computation.
Hetero-Designer: Automated Design of Multi-Agent Systems with Heterogeneous LLMs
Zhiheng Zhang (Chinese Academy of Sciences), Jun Zhao (Chinese Academy of Sciences)
Autonomous DrivingOptimizationFederated LearningMeta LearningDrug DiscoveryAI Code AssistantNeural Architecture SearchGraph Neural NetworkTransformerLarge Language ModelReinforcement LearningPrompt EngineeringMixture of ExpertsAuto EncoderTextGraphRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes a framework for automatically designing heterogeneous LLM multi-agent systems, which can generate collaborative topologies containing different LLM models and roles based on queries;
HeteroCache: A Dynamic Retrieval Approach to Heterogeneous KV Cache Compression for Long-Context LLM Inference
Zhiyuan Shi (Zhejiang University), Wenxiao Wang (Zhejiang University)
CompressionComputational EfficiencyTransformerLarge Language ModelAuto EncoderTextRetrieval-Augmented Generation
🎯 What it does: Proposes HeteroCache, a training-agnostic dynamic KV cache compression framework, which compresses KV cache through head-level heterogeneous hierarchy and dynamic retrieval for long-context LLM inference.
Heterogeneous Adaptive Policy Optimization: Tailoring Optimization to Every Token’s Nature
Zheng Liu (Peking University), Wentao Zhang (Peking University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Proposed a Heterogeneous Adaptive Policy Optimization (HAPO) algorithm, which utilizes entropy information throughout the entire RLHF training process to achieve token-level adaptive optimization, including adaptive sampling temperature, token-level advantage estimation, differential advantage redistribution, and asymmetric adaptive clipping.
HeteroSpec: Leveraging Contextual Heterogeneity for Efficient Speculative Decoding
Siran Liu (Peking University), Yongchao He (ScitiX AI)
Computational EfficiencyAI Code AssistantTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the HeteroSpec framework, which explicitly allocates verification resources by adapting to contextual heterogeneity, thereby improving inference speed.
HiChunk: Evaluating and Enhancing Retrieval Augmented Generation with Hierarchical Chunking
Wensheng Lu (Tencent Youtu Lab), Xing Sun (Tencent Youtu Lab)
RetrievalRecommendation SystemData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose HiCBench, a benchmark for evaluating document chunking quality, and design the HiChunk framework that utilizes fine-tuned LLMs for multi-level document chunking, combined with the Auto-Merge retrieval algorithm to enhance RAG performance.
HiddenGuard: Fine-Grained Safe Generation with Specialized Representation Router
Lingrui Mei (State Key Laboratory Of AI Safety), Xueqi Cheng (State Key Laboratory Of AI Safety)
Safty and PrivacyExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringMixture of ExpertsText
🎯 What it does: Propose the HiddenGuard framework, which achieves token-level safety detection and redaction for LLMs through PRISM (Representation Activators + Router), ensuring that useful information is not entirely rejected.
HiEdit: Lifelong Model Editing with Hierarchical Reinforcement Learning
Yangfan Wang (Harbin Institute of Technology), Jingchi Jiang (Harbin Institute of Technology)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelReinforcement LearningPrompt EngineeringMixture of ExpertsText
🎯 What it does: Propose the HiEdit framework, which utilizes hierarchical reinforcement learning to achieve lifelong editing of LMMs, automatically selecting the most relevant layers and updating parameters;
Hierarchical Acoustic-Semantic Modeling: Modality Separation and Semantic Coherence for Full-Duplex SLMs
Zhenyu Liu (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
RecognitionComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextMultimodalityBenchmarkAudio
🎯 What it does: Proposed a native end-to-end full-duplex speech-language model called Lychee-FD, which addresses the gradient conflicts and semantic dilution issues when sharing deep parameters between speech and semantic modalities.
Hierarchical Policy Optimization for Simultaneous Translation of Unbounded Speech
Siqi Ouyang (Carnegie Mellon University), Lei Li (Carnegie Mellon University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextAudio
🎯 What it does: This paper proposes a post-training framework called Hierarchical Policy Optimization (HPO), which uses reinforcement learning to simultaneously optimize translation quality and latency in the task of simultaneous speech translation.
Hierarchical Reinforcement Learning with Augmented Step-Level Transitions for LLM Agents
Shuai Zhen (Beijing University of Posts and Telecommunications), Yang Deng (Singapore Management University)
Autonomous DrivingRobotic IntelligenceTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringTextSequential
🎯 What it does: Propose the STEP-HRL framework, which utilizes hierarchical reinforcement learning and a local progress module, enabling the LLM agent to make decisions based solely on single-step transitions, eliminating the dependency on long-term history.
Hierarchical Token Prepending: Enhancing Information Flow in Decoder-based LLM Embeddings
Xueying Ding (Carnegie Mellon University), Tong Zhao (Snap Inc)
RetrievalRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Propose a training-free hierarchical token prepending method (Hierarchical Token Prepending, HTP), which inserts sentence-level and document-level summary tokens into the input to construct multi-level reversed information flow, and uses mean pooling as the readout function to enhance the embedding quality of the decoder LLM.
HiGoE: Hierarchical Graph of Evidence to Enhance Retrieval-Augmented Generation for Long-context Summarization
Long Yuan (Wuhan University of Technology), Chuan Ma (Chongqing University)
GenerationRetrievalRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextGraphBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper addresses the problem of long text summarization by proposing the Hierarchical Graph of Evidence (HiGoE) framework, which converts documents into a verifiable proposition-evidence hierarchical graph and employs retrieval-augmented generation;
HiKEY: Hierarchical Multimodal Retrieval for Open-Domain Document Question Answering
Joongmin Shin (Korea University), Heuiseok Lim (Korea University)
RetrievalRecommendation SystemGraph Neural NetworkTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityGraphTabularRetrieval-Augmented Generation
🎯 What it does: Propose the HiKEY framework, which utilizes document hierarchy for hierarchical coarse-to-fine retrieval and constructs a multi-modal retrieval subgraph to support open-domain question answering on industrial-scale multi-page PDFs.
HistLens: Mapping Idea Change across Concepts and Corpora
Yi Jing (Tsinghua University), Zhifang Sui (Peking University)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelAuto EncoderContrastive LearningText
🎯 What it does: Propose the HistLens framework, which decomposes concepts into interpretable features through sparse autoencoders (SAE), supporting cross-time and cross-corpus concept historical analysis with multiple concepts and corpora, and enabling the detection and comparison of implicit concepts.
HiSVD: Principled Low-Rank Approximation of LLMs via Hierarchical Modeling of Information Capacity and Spectral Structure
Zhuo Chen (Shanghai Jiao Tong University), Weihang Chen (Xiaohongshu Inc)
CompressionComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This study proposes a low-rank compression method called HiSVD based on hierarchical rank allocation, specifically designed for compressing large language models.
HOPE: Hybrid Optimized Parallel Encoding with Supervised and Unsupervised Semantic Fusion for Depression Symptom Detection
Tu-Phuong Mai (VNU University of Engineering and Technology), Hoang-Quynh Le (VNU University of Engineering and Technology)
ClassificationRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextElectronic Health RecordsRetrieval-Augmented Generation
🎯 What it does: Proposed the HOPE framework, using hybrid parallel encoding for multi-label detection of depressive symptoms.
HopWeaver: Cross-Document Synthesis of High-Quality and Authentic Multi-Hop Questions
Zhiyu Shen (Sun Yat-sen University), Jianxing Yu (Hong Kong Metropolitan University)
Data SynthesisTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Propose the HopWeaver framework, which automatically synthesizes high-quality, multi-hop cross-document questions (bridge-type and comparison-type) from raw corpora and provides an evaluation system;
How Can Synthetic Data Improve Multilingual Language Model Pretraining? A Data Quality Perspective
Tongyao Zhu (National University of Singapore), Shiqi Chen (City University of Hong Kong)
Data-Centric LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Propose two methods to improve the quality of pre-training data for low-resource languages: SynTrans (translating high-quality English knowledge data into the target language directly for pre-training) and SynRank (training a classifier using the synthetic data generated by SynTrans to rank and filter real noisy data by quality), as well as an untrained quality evaluation metric called DA-QwenScore; these methods are applied to pre-train 1B TinyLLaMA with 20B tokens, and the effectiveness is evaluated on multilingual benchmarks (XARC-E, XARC-C, XCOPA, XHellaswag).
How Context Shapes Truth: Geometric Transformations of Statement-level Truth Representations in LLMs
Shivam Adarsh (University of Copenhagen), Christina Lioma (University of Copenhagen)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: Studied how adding context to LLMs changes the geometric properties of sentence-level truth value vectors.
How Controllable Are Large Language Models? A Unified Evaluation across Behavioral Granularities
Ziwen Xu (Zhejiang University), Shumin Deng (Zhejiang University)
Data SynthesisExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes SteerEval, a hierarchical benchmark for evaluating the controllability of large language models (LLMs), covering three domains: linguistic features, emotion, and personality. It sets three levels of specifications (L1, L2, L3) in each domain to systematically assess current steering methods.
How Do Inpainting Artifacts Propagate to Language?
Pratham Yashwante (UC San Diego), Sukruth Rao (UC San Diego)
RestorationExplainability and InterpretabilityTransformerPrompt EngineeringVision Language ModelDiffusion modelContrastive LearningImageTextMultimodalityTime SeriesBenchmarkAudio
🎯 What it does: This paper proposes a two-stage diagnostic framework, first restoring masked images through diffusion-based image inpainting, and then inputting the original image and the restored image into a frozen vision-text model for comparison, exploring how the quality of restoration affects the generated image descriptions.
How Do LLMs and VLMs Understand Viewpoint Rotation Without Vision? An Interpretability Study
Zhen Yang (Beijing Institute of Technology), Wenpeng Lu (Qilu University of Technology)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Investigated the perspective rotation understanding (VRU) capability of large language models (LLMs) and vision-language models (VLMs) under purely textual descriptions, and constructed the VRUBench dataset for evaluation;
How do Role Models Shape Collective Morality? Exemplar-Driven Moral Learning in Multi-Agent Simulation
Junjie Liao (Beijing Normal University), Fangwei Zhong (Beijing Normal University)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringWorld ModelTextSequentialReview/Survey PaperChain-of-Thought
🎯 What it does: Construct a multi-agent simulation framework based on large language models, design four role model-driven moral learning experiments (Alignment, Collapse, Conflict, Construction), and investigate the mechanism by which role models shape collective morality through a four-stage cognitive cycle of observation, reflection, etc.
How Good is Your Wikipedia? Auditing Data Quality for Low-resource and Multilingual NLP
Kushal Tatariya (KU Leuven), Miryam de Lhoneux (KU Leuven)
Data-Centric LearningPrompt EngineeringContrastive LearningTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Systematically evaluate the quality of non-English Wikipedia using conventional data cleaning methods such as script filtering, Exact-match deduplication, and MinHash deduplication, and build a four-tier quality ranking based on the filtering ratio;
How Instruction and Reasoning Data shape Post-Training: Data Quality through the Lens of Layer-wise Gradients
Ming Li (University of Maryland), Tianyi Zhou (University of Maryland)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Conducts gradient spectrum analysis of large language models trained on instructions and reasoning data of varying quality, investigating the impact of data quality on the dynamics of fine-tuning.
How Long Reasoning Chains Influence LLMs’ Judgment of Answer Factuality
Minzhu Tu (State Key Laboratory of AI Safety), Keping Bi (State Key Laboratory of AI Safety)
Explainability and InterpretabilityTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: This paper systematically studies the impact of adding reasoning chains on the accuracy judgment of answers when large language models (LLMs) act as evaluators;
How Memory Management Impacts LLM Agents: An Empirical Study of Experience-Following Behavior
Zidi Xiong (Harvard University), Zhen Xiang (University of Minnesota-Twin Cities)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelReinforcement LearningAgentic AITextTabularTime SeriesRetrieval-Augmented Generation
🎯 What it does: An empirical study on memory management (memory addition and deletion) of large language model (LLM) agents, exploring its impact on long-term performance.
How Much Would a Clinician Edit This Draft? Evaluating LLM Alignment for Patient Message Response Drafting
Parker Seegmiller (Dartmouth), Sarah Masud Preum (Dartmouth)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBiomedical DataElectronic Health RecordsRetrieval-Augmented Generation
🎯 What it does: This paper investigates the alignment issues of large language models (LLMs) in drafting responses to patient portal information, constructs a clinically oriented thematic system and an editing load evaluation framework, releases an expert-annotated dataset of patient portal responses, and conducts large-scale comparative experiments on various local and commercial LLMs and multiple adaptation strategies.
How Should We Enhance the Safety of Large Reasoning Models: An Empirical Study
Zhexin Zhang (Tsinghua University), Minlie Huang (Tsinghua University)
Safty and PrivacyKnowledge DistillationTransformerSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: This study targets large-scale reasoning models (LRMs) and systematically explores empirical methods to enhance their safety through supervised fine-tuning (SFT). The authors first find that distilling safe responses directly from a pre-trained LRM (DeepSeek-R1) does not significantly reduce the attack success rate. Subsequently, by analyzing five types of high-risk reasoning patterns in the distillation data (lack of safety awareness, over-reasoning, repeated hesitation, inconsistent responses, etc.), they propose improved prompting methods, 'RealSafe CoT' and 'Improved CoT', which significantly enhance safety. They then verify that short or template-based reasoning (Short CoT, Template CoT) can also achieve similar safety effects, indicating that long reasoning chains are not essential. Finally, the authors conduct comprehensive ablation experiments to systematically evaluate the impact of distillation sources, sample size, benign samples, learning rates, and other hyperparameters on safety, reasoning ability, and over-rejection rates.
How to Improve LLMs’ Performance on Specific Languages: A Perspective on LLM-Derived Language Similarity
Xinhe Shi, Linchao Zhu (Zhejiang University)
OptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningTextMultimodalityReview/Survey PaperBenchmark
🎯 What it does: This paper proposes a language similarity measurement based on LLM, and uses this measurement to guide the selection of languages for fine-tuning, thereby improving the performance of the target language.
How Tokenization Limits Phonological Knowledge Representation in Language Models and How to Improve Them
Disen Liao (University of Waterloo), Freda Shi (University of Waterloo)
Explainability and InterpretabilityRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningText
🎯 What it does: Studied the impact of subword tokenization on language models' phonological knowledge (such as prosody, syllable segmentation, etc.), and proposed a lightweight IPA fine-tuning method to enhance the model's phonological reasoning ability.
How Training Data Shapes the Use of Parametric and In-Context Knowledge in Language Models
Minsung Kim (Seoul National University), Meeyoung Cha (Max Planck Institute for Security and Privacy)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: The study investigates how the attributes of training data influence the use of parameter knowledge and contextual knowledge by large language models, as well as the resolution of conflicts, through controlled experiments using synthetic biography data.
HoWToBench: Holistic Evaluation for LLM’s Capability in Human-level Writing using Tree of Writing
Andrew Zhuoer Feng (Tsinghua University), Minlie Huang (Tsinghua University)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the Tree-of-Writing evaluation framework and the HOWTOBENCH Chinese writing benchmark for comprehensive assessment of writing generated by LLMs.
HowToNarrate: A General-Domain Benchmark for Synchronized Video Narration with External Knowledge
Xueyan Wang (Renmin University of China), Qin Jin (Renmin University of China)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed the HowToNarrate large-scale general-domain synchronized video narration dataset.
HSCodeComp: A Realistic and Expert-level Agent Benchmark for Hierarchical Rule Application
Tian Lan (Alibaba Group), Longyue Wang (Alibaba Group)
TransformerLarge Language ModelAgentic AIVision Language ModelTextMultimodalityTabularBenchmarkFinance RelatedRetrieval-Augmented Generation
🎯 What it does: Constructed a real expert-level e-commerce benchmark called HSCODECOMP for hierarchical rule application, and evaluated various LLMs, VLMs, and agents.
HSGraphAgent: Knowledge-Graph-Guided Large Language Models for Harmonized System Code Classification
Qiang Xia (Tongji University), Jian Li (Tongji University)
ClassificationGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextGraphBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed the HSGraphAgent framework, which uses a knowledge graph to guide LLMs in hierarchical and regulation-constrained HS code classification.
HTMR: Hybrid Token Masking Reinforcement Learning with Verifiable Rewards for Event Argument Extraction with Multi-Perspective Reasoning
Jianwen Luo (Soochow University), Jianmin Yao (Soochow University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: Proposes the HTMR (Hybrid Token Masking Reinforcement Learning with Verifiable Rewards) framework to improve the sequence generation task in event argument extraction (EAE).
Human or LLM as Standardized Patients? A Comparative Study in Medical Education
Bingquan Zhang (Wuhan University), Benyou Wang (Chinese University of Hong Kong Shenzhen)
TransformerLarge Language ModelAgentic AIPrompt EngineeringTextBiomedical DataBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the EasyMED multi-agent framework to achieve controllable, contextualized virtual standardized patients (VSP), and create SPBench, an artificially annotated dialogue benchmark for quantitatively evaluating the behavioral similarity between VSP and human standardized patients.
HumanLLM: Benchmarking and Improving LLM Anthropomorphism via Human Cognitive Patterns
Xintao Wang (Fudan University), Yanghua Xiao (Fudan University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the HumanLLM framework, which achieves realistic simulation of multi-modal interactive behaviors by training LLMs through psychology-pattern-driven dialogues.
Hybrid Autoregressive-Diffusion Model for Real-Time Sign Language Production
Maoxiao Ye (University of Auckland), Sathiamoorthy Manoharan
GenerationTransformerDiffusion modelFlow-based ModelAuto EncoderVideoSequential
🎯 What it does: Proposed HybridSign, a hybrid autoregressive-diffusion framework for low-latency sign language generation.
HybridKV: Hybrid KV Cache Compression for Efficient Multimodal Large Language Model Inference
Bowen Zeng (Zhejiang University), Huan Li (Zhejiang University)
CompressionComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsImageVideoTextMultimodality
🎯 What it does: Propose HYBRIDKV, a hybrid KV cache compression framework, to significantly reduce GPU memory usage and improve decoding speed in multi-modal large language model inference.
Hyperion: Private Token Sampling with Homomorphic Encryption
Lawrence Lim (University of California, Santa Barbara), Amr El Abbadi (University of California, Santa Barbara)
Safty and PrivacyComputational EfficiencyLarge Language ModelText
🎯 What it does: Developed the Hyperion algorithm, achieving non-interactive, randomized Token sampling in a homomorphic encryption environment.
HyperMem: Hypergraph Memory for Long-Term Conversations
Juwei Yue (Institute of Information Engineering Chinese Academy of Sciences), Yafeng Deng (EverMind AI)
RetrievalFederated LearningSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextGraphSequentialRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Constructed the HyperMem hypergraph hierarchical memory system, which includes dialogue segmentation (episode detection), topic aggregation, fact extraction, and links topics, chapters, and facts through hyperedges to achieve explicit modeling of high-order associations; meanwhile, a coarse-to-fine retrieval process was designed to sequentially filter topics → chapters → facts.
HypoEval: Hypothesis-Guided Evaluation for Natural Language Generation
Mingxuan Li (University of Chicago), Chenhao Tan (University of Chicago)
GenerationExplainability and InterpretabilityData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Designed and implemented a parameter-free tuning, sample-efficient HypoEval framework, which guides LLMs to perform multi-dimensional evaluation of text generation and aggregate scores through a small number of human evaluations and hypotheses generated from literature abstracts.
I²B-LPO: Latent Policy Optimization via Iterative Information Bottleneck
Huilin Deng (University of Science and Technology of China), Yu Kang (University of Science and Technology of China)
OptimizationTransformerLarge Language ModelReinforcement LearningAuto EncoderTextBenchmark
🎯 What it does: By leveraging conditional variational autoencoder (CVAE) to generate diverse latent variables in high-entropy states and injecting them into the attention mechanism of LLM to form a 'latent branch,' while using information bottleneck (IB) to filter and self-reward generated trajectories, this approach solves the exploration collapse problem in RLVR.
I2E: From Image Pixels to Actionable Interactive Environments for Text-Guided Image Editing
Jinghan Yu (Huazhong University Of Science And Technology), Zhiyuan Ma (Huazhong University Of Science And Technology)
Image TranslationRestorationSegmentationGenerationOptimizationTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsVision-Language-Action ModelDiffusion modelImageTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the I2E framework, which treats image editing as an interactive structured environment, achieving precise text-guided editing through a decompose-execute approach.
ICDAGENT: Empowering Agentic Large Language Models for Explainable Medical Coding
Ziyi Yin (Pennsylvania State University), Fenglong Ma (Pennsylvania State University)
Explainability and InterpretabilityAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIPrompt EngineeringTextBiomedical DataElectronic Health RecordsRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Designed and implemented ICDAGENT, a multi-agent framework composed of two large language models (encoding agent and critical agent), for automatically assigning ICD codes to clinical texts and providing evidence-based interpretable explanations;
Identifying Bias in Machine-generated Text Detection
Kevin Stowe (Pindrop), Kailash Patil (Pindrop)
Anomaly DetectionExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningTextBenchmark
🎯 What it does: Examined bias in machine-generated text detection systems across dimensions such as gender, race/ethnicity, English language learner (ELL) status, and economic status. Systematically evaluated 16 publicly available detection models and compared them with human assessment results.
Identifying the Periodicity of Information in Natural Language
Yulin OU, Hendrik Buschmeier (Bielefeld University)
RecognitionText
🎯 What it does: Proposed the AutoPeriod of Surprisal (APS) algorithm, which directly identifies information periodicity in a corpus at the single-document level;
Identity-Robust Language Model Generation via Content Integrity Preservation
Miao Zhang (New York University), Rumi Chunara (New York University)
GenerationSafty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed an identity-robust, zero-training, lightweight generation framework called IRG. It first detects and determines whether identity expressions in user prompts are critical to the answer, then neutralizes non-critical identity information, generates core answers using neutralized queries, and optionally presents personalized results while maintaining content consistency.
IEvoAgent: Evolving Conversational Agent based on User Implicit Feedback
Yichen Cai (Zhejiang University), Pei Chen (Zhejiang University)
OptimizationRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextSequentialBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed a two-stage IEvoAgent framework that utilizes user implicit feedback to achieve real-time evolution and multi-turn alignment in dialogue systems
If an LLM Were a Character, Would It Know Its Own Story? Evaluating Lifelong Learning in LLMs
Siqi Fan (University of Electronic Science and Technology of China), Yequan Wang (Beijing Academy of Artificial Intelligence)
Explainability and InterpretabilityMeta LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructing the LIFESTATE-BENCH benchmark to evaluate the lifelong learning capabilities of LLMs
IF-CRITIC: Towards a Fine-Grained LLM Critic for Instruction-Following Evaluation
Bosi Wen (Conversational Artificial Intelligence Group Tsinghua University), Minlie Huang (Conversational Artificial Intelligence Group Tsinghua University)
OptimizationExplainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Developed a fine-grained LLM evaluator called IF-CRITIC, which evaluates all constraints in the instruction at once through checklist-driven critical generation;
IF-RewardBench: Benchmarking Judge Models for Instruction-Following Evaluation
Bosi Wen (Tsinghua University), Minlie Huang (Tsinghua University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes IF-RewardBench, a comprehensive meta-evaluation benchmark for assessing discriminative models in instruction-following tasks, constructing 842 multi-type instructions and 6011 multi-model responses, and generating complete preference graphs through human annotation;
IGenBench: Benchmarking the Reliability of Text-to-Infographic Generation
Yinghao Tang (Zhejiang University), Wei Chen (Zhejiang University)
GenerationTransformerLarge Language ModelPrompt EngineeringImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes the IGENBENCH benchmark, specifically designed to evaluate the reliability of text-to-infographic generation models.
Illusions of Confidence? Diagnosing LLM Truthfulness via Neighborhood Consistency
Haoming Xu (Zhejiang University), Ningyu Zhang (Zhejiang University)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the Neighborhood Consistency Belief (NCB) metric and the corresponding cognitive pressure test to evaluate the credibility of LLMs under contextual interference; and design Structure-Aware Training (SAT) to enhance model robustness based on this.
Illusions of the Gold Standard: A Large-scale Analysis of Human Evaluation Protocols for Long-form Text Generation
Katelyn X. Mei (University of Washington), Lucy Lu Wang (University of Washington)
GenerationTransformerLarge Language ModelPrompt EngineeringTextReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: Conducted a large-scale analysis of human evaluation protocols for long text generation in *CL conference papers from 2023 to 2025, constructing and applying 20 reportable criteria to statistically analyze and visualize manual and LLM-assisted annotation results.
ImCoref-CeS: An Improved Lightweight Pipeline for Coreference Resolution with LLM-based Checker-Splitter Refinement
Kangyang Luo (Tsinghua University), Maosong Sun (Tsinghua University)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposed the ImCoref-CeS framework, first improving the Maverick core model ImCoref (introducing a lightweight bridging module LBM, biaffine scorer, and hybrid mention regularization), and then verifying and splitting mentions and clusters through the LLM Checker-Splitter, achieving lightweight high-precision coreference resolution.
ImF: Embedding an Implicit Fingerprint in Your Large Language Models
Jiaxuan Wu (China Agricultural University), Juan Wen (China Agricultural University)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose an implicit fingerprint technique named ImF for embedding verifiable ownership information in large language models, which can be validated in black-box environments;
Immediate Inference: The Missing Foundation in Large Language Model Logical Reasoning
Sihang Jiang (Fudan University), Jiansheng Wei (Huawei)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This study systematically explores the fundamental role of Immediate Inference (II) in large language models (LLMs), constructs the first benchmark for II called IIBench, and demonstrates that II can explain and enhance the performance of various downstream reasoning tasks through zero-shot evaluation, error type analysis, causal mediation models, and training intervention experiments.
ImmersiveTTS: Environment-Aware Text-to-Speech with Multimodal Diffusion Transformer and Domain-Specific Representation Alignment
Jun-Hak Yun (Korea University), Seong-Whan Lee (Korea University)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderContrastive LearningMultimodalityAudio
🎯 What it does: Proposed an environment-aware text-to-speech generation model called ImmersiveTTS, which can simultaneously generate natural speech and corresponding background sounds given text and environmental descriptions.
IMPACT: Importance-Aware Activation Space Reconstruction
Md Mokarram Chowdhury (Iowa State University), Yang Li (Iowa State University)
CompressionComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Designed an IMPACT framework for low-rank compression of large language models (LLMs), focusing on activation reconstruction and gradient importance, explicitly linking reconstruction error with model performance during the compression process.
Impatient Users Confuse AI Agents: High-fidelity Simulations of Human Traits for Testing Agents
Muyu He, Nazneen Rajani (Collinear AI Inc)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextSequentialBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the TraitBasis method, which uses activation vectors to high-fidelity simulate user traits (such as impatience, confusion, skepticism, and incoherence) in dialogue models, and builds the τ-Trait benchmark based on this;
Implicit Representations of Grammaticality in Language Models
Yingshan Susan Wang (Massachusetts Institute of Technology), Yoon Kim (Massachusetts Institute of Technology)
Explainability and InterpretabilityRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: Trained a linear probe to extract syntactic acceptability information from the hidden layers of a pre-trained language model, using synthetic acceptable/unacceptable sentence pairs (generated by inserting/deleting/local shuffling noise) for training, and evaluated on multiple English and cross-lingual benchmarks.
ImplicitMemBench: Measuring Unconscious Behavioral Adaptation in Large Language Models
Chonghan Qin (University of Hong Kong), Lingpeng Kong (Harbin Institute of Technology)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes IMPLICITMEMBENCH, a benchmark for evaluating implicit memory (i.e., unconscious behavioral adaptation) in large language models (LLMs), focusing on three cognitive mechanisms: procedural memory, priming effect, and classical conditioning.
ImpRIF: Stronger Implicit Reasoning Leads to Better Complex Instruction Following
Yuancheng Yang (ByteDance China), Haihua Yang (ByteDance China)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningData-Centric LearningAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningMixture of ExpertsTextGraphTabularBenchmarkChain-of-Thought
🎯 What it does: This paper proposes the ImpRIF method, which significantly improves the execution capability of large models under complex instructions by constructing a verifiable implicit reasoning graph (ERG) and using graph-based Chain-of-Thought training.
Improving Autoformalization Using Direct Dependency Retrieval
Shaoqi Wang (Zhejiang University), Jun Zhou (Ant Group)
RetrievalOptimizationComputational EfficiencyAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: Developed an automated formalization framework based on Direct Dependency Retrieval (DDR), which improves the quality of formalizing mathematical statements by generating and verifying candidate dependencies.
Improving Long-Context Translation via Self-Supervised Dual Learning
Shanbo Cheng (Nanjing University), Shujian Huang (Nanjing University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningContrastive LearningTextBenchmark
🎯 What it does: Propose the LongDu self-supervised post-training framework, enhancing the reliability of long document translation through loop consistency.
Improving Retrieval-Augmented Generation without Taxonomy-based Error Categorization
Gongbo Zhang (Columbia University), Chunhua Weng
RetrievalOptimizationExplainability and InterpretabilityTransformerSupervised Fine-TuningReinforcement LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed a retrieval-augmented generation (RAG) improvement framework called RePAIR based on response-action learning, which directly maps erroneous responses to corrective actions without requiring explicit error classification;
Improving the Distributional Alignment of LLMs using Supervision
Gauri Kambhatla (University of Texas at Austin), Matthew Lease (University of Texas at Austin)
Recommendation SystemFederated LearningExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Research and improve the alignment between the opinion distribution generated by large language models (LLMs) and the distribution from human survey responses across different sociodemographic groups, proposing and verifying a distribution calibration method based on supervised regression.
ImReasoner: Improving Memory-based Language Models for Reasoning-in-a-Haystack Tasks
Ching-Yun Ko (IBM Research), Pin-Yu Chen (IBM Research)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningRecurrent Neural NetworkTransformerLarge Language ModelAuto EncoderContrastive LearningTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes an improved memory-enhanced language model called ImReasoner, designed to address the 'needle-in-a-haystack' task in long-context reasoning, combining memory updates during inference with a staged training approach;
In-Context Representation Hijacking
Itay Yona, Yossi Gandelsman
Explainability and InterpretabilityAdversarial AttackTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Proposed and verified Doublespeak: an attack method that replaces harmful keywords with harmless synonyms in context, enabling large language models' internal representations to gradually shift from harmless to harmful without altering the surface semantics of the input.
In-depth Research Impact Summarization through Fine-Grained Temporal Citation Analysis
Hiba Arnaout, Iryna Gurevych (TU Darmstadt)
GenerationRecommendation SystemExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextTabularTime SeriesBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose a research impact summary generation task based on fine-grained citation intent analysis and time awareness, automatically extracting two types of impacts—recognition and criticism—from massive citation contexts and narrating the evolution of paper impacts over time periods.
Incentivizing Parametric Knowledge via Reinforcement Learning with Verifiable Rewards for Cross-Cultural Entity Translation
Jiang Zhou (Tianjin University), Deyi Xiong (Tianjin University)
Computational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: Propose the EA-RLVR framework, which activates cross-cultural entity knowledge in LLM parameters through reinforcement learning and verifiable rewards, achieving entity translation without the need for an external knowledge base.
Incomplete In-context Learning
Wenqiang Wang (Sun Yat-sen University), Xiaochun Cao (Lanzhou University)
ClassificationData SynthesisRetrievalTransformerLarge Language ModelPrompt EngineeringImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Proposes Incomplete In-context Learning (IICL) for incomplete retrieval datasets and designs the Iterative Judgments and Integrated Prediction (IJIP) framework.
Incorporating Temporal Coherence to Cross-Document Event Coreference Resolution
Xinyu Chen (Soochow University), Qiaoming Zhu (Soochow University)
RecognitionRepresentation LearningGraph Neural NetworkTransformerPrompt EngineeringContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes the CohTP framework, which achieves cross-document event coreference resolution by constructing and refining event time graphs, dividing events into time periods, and performing coreference inference within the same time period.
Inertia in Moral and Value Judgments of Large Language Models
Bruce W. Lee (University of Pennsylvania), Hyunsoo Cho (Ewha Womans University)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringTextReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: By injecting diverse social attributes (age, gender, religion, education, occupation) into large-scale random role-playing scenarios and letting seven large language models answer questions based on human values and morality questionnaires (PVQ-RR, MFQ-30) to examine their value orientations and 'inertia' phenomena.
INFACT: A Diagnostic Benchmark for Induced Faithfulness and Factuality Hallucinations in Video-LLMs
Yangjunqi, Xilin Chen (Chinese Academy of Sciences)
Large Language ModelPrompt EngineeringVision Language ModelDiffusion modelScore-based ModelFlow-based ModelRectified FlowNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkContrastive LearningGaussian SplattingOptical FlowVideoTextMultimodalityBenchmarkRetrieval-Augmented GenerationStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper constructs the INFACT benchmark, which contains 9,800 video question-answer instances, and designs a fine-grained illusion classification of faithfulness and factuality. It introduces two measurable model reliability metrics, RR and TSS, under four evaluation modes (Base, Visual Degradation, Evidence Corruption, Temporal Intervention).
InferenceDynamics: Adaptive LLM Routing through Structured Capability and Knowledge Profiling
Haochen Shi (Hong Kong University of Science and Technology), Yangqiu Song (Hong Kong University of Science and Technology)
Recommendation SystemComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextBenchmark
🎯 What it does: Propose an expandable LLM routing framework called InferenceDynamics, which intelligently assigns queries based on the model's capabilities and knowledge characteristics;
Inferring Events from Time Series using Language Models
Mingtian Tan (University of Virginia), Thomas Hartvigsen (University of Virginia)
Data SynthesisKnowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextTime SeriesBenchmarkChain-of-Thought
🎯 What it does: Investigated whether large language models can infer natural language events corresponding to time series, and constructed a benchmark dataset and automated task generation method based on sports matches.
InfiniteWeb: Scalable Web Environment Synthesis for GUI Agent Training
Ziyun Zhang (Peking University), Yan Lu (Microsoft Research Asia)
Data SynthesisReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Designed and implemented the INFINITEWEB system, which can automatically generate functionally complete multi-page websites, corresponding user tasks, and verifiable evaluators, providing a large-scale and scalable environment for GUI agent training.
Inflated Excellence or True Performance? Rethinking Medical Diagnostic Benchmarks with Dynamic Evaluation
Xiangxu Zhang (Renmin University of China), Xian Wu (Tencent)
TransformerLarge Language ModelPrompt EngineeringTextBiomedical DataElectronic Health RecordsBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes the DyReMe framework, which includes dynamic generation of diagnostic questions (DyGen) and multi-dimensional evaluation (EvalMed), to assess the reliability of large language models in medical diagnosis.
Influence-based Online Experience Selection for Effective RLHF
Yifan Gong (Renmin University of China), Xing Xie (Renmin University of China)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Propose an online experience screening method called InfOES based on influence, aimed at improving alignment performance in RLHF
Inhibitory Attacks on Backdoor-based Fingerprinting for Large Language Models
Hang fu, Xue Yiming
Federated LearningSafty and PrivacyExplainability and InterpretabilityAdversarial AttackTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose two anti-fingerprinting attack methods for LLM integration scenarios: Token Filter Attack (TFA) and Sentence Verification Attack (SVA), and demonstrate that they can effectively suppress backdoor fingerprints without modifying model parameters.
Injecting Context via Situation Working Memory for Logical Reasoning with LLMs
Jieun Kim (Yonsei University), Sung-Bae Cho (Yonsei University)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Proposes SituW, a situation working memory based on a situational model, which guides the logical reasoning of large language models by dynamically constructing and updating situational memory.
InquireMobile: Teaching VLM-based Mobile Agent to Request Human Assistance via Reinforcement Fine-Tuning
Qihang Ai (Alibaba Group), Yuning Jiang (Alibaba)
Safty and PrivacyExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningAgentic AIPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes a new safety interaction benchmark called InquireBench, and a mobile intelligent agent based on VLM called InquireMobile, aiming to let mobile agents proactively request human confirmation at critical decision points to improve system safety and explainability.