ICLR 2026 Papers — Page 24
International Conference on Learning Representations · 5356 papers
InfoDet: A Dataset for Infographic Element Detection
Jiangning Zhu (BNRist, Tsinghua University), Shixia Liu (BNRist, Tsinghua University)
Object DetectionData SynthesisConvolutional Neural NetworkTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelImageTextMultimodalityChain-of-Thought
🎯 What it does: Constructed a large-scale InfoDet dataset and trained an information graph-oriented object detection model using it, further applying it to chart understanding and graphic layout detection in VLM.
InfoMosaic-Bench: Evaluating Multi-Source Information Seeking in Tool-Augmented Agents
Yaxin Du (Shanghai Jiao Tong University), Siheng Chen (Shanghai Jiao Tong University)
Agentic AIVideoTextMultimodalityBenchmarkFinance RelatedRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes InfoMosaic-Bench, a benchmark for evaluating multi-source information retrieval capabilities, covering six domains (medicine, finance, maps, video, web, and multi-domain), and generates 621 tasks through the InfoMosaic-Flow automated pipeline. Tasks require simultaneous use of multiple specialized tools and general web search.
InfoNCE Induces Gaussian Distribution
Roy Betser (Technion Israel Institute of Technology), Guy Gilboa (Technion Israel Institute of Technology)
Representation LearningContrastive LearningImage
🎯 What it does: Analyze and prove that contrastive learning representations induced by the InfoNCE objective tend toward Gaussian distributions in high-dimensional spaces
Information Estimation with Discrete Diffusion
Alberto Foresti (EURECOM), Pietro Michiardi (EURECOM)
Large Language ModelDiffusion modelScore-based ModelTextBiomedical Data
🎯 What it does: Propose INFO-SEDD, a discrete data information theory estimation method based on continuous-time Markov chains, for computing mutual information, KL divergence, and entropy, which can directly utilize pre-trained models.
Information Gain-based Policy Optimization: A Simple and Effective Approach for Multi-Turn Search Agents
Guoqing Wang (Ant Group), Zhenzhe Ying (Ant Group)
OptimizationReinforcement LearningAgentic AIText
🎯 What it does: Propose the Information Gain Reward Framework (IGPO) in multi-round search agents, achieving dense reinforcement learning supervision through improvements in the probability of the true answer at each step.
Information Shapes Koopman Representation
Xiaoyuan Cheng (University College London), Zhuo Sun (Shanghai University of Finance and Economics)
Representation LearningAuto EncoderContrastive LearningGraphTime Series
🎯 What it does: Propose a Koopman representation learning framework based on the information bottleneck theory, and design Lagrangian regularization to balance conciseness and expressiveness, achieving stable and interpretable Koopman representations.
Information-based Value Iteration Networks for Decision Making Under Uncertainty
Cynthia Chen (Allen Institute), Koosha Khalvati (Allen Institute)
Convolutional Neural NetworkRecurrent Neural NetworkReinforcement Learning
🎯 What it does: Propose a differentiable planning module named VI²N, which integrates value iteration networks with information value assessment for decision-making in partially observable environments
Information-Theoretic Membership Inference for Granular Quantification of Memorization
Jiashu Tao (National University of Singapore), Reza Shokri (National University of Singapore)
Safty and PrivacyTransformerImageTextBenchmark
🎯 What it does: Proposes an information-theoretic version of membership inference attack (InfoRMIA) and a fine-grained privacy evaluation framework based on token-level analysis;
InfoScan: Information-Efficient Visual Scanning via Resource-Adaptive Walks
Yifeng Wu (Shenzhen University of Advanced Technology), Guanhua Chen (Southern University of Science and Technology)
ClassificationObject DetectionSegmentationReinforcement LearningImageBiomedical Data
🎯 What it does: Designed an information-aware dynamic scanning mechanism called InfoScan, which adaptively allocates computational resources to high-information regions in images.
InfoTok: Adaptive Discrete Video Tokenizer via Information-Theoretic Compression
Haotian Ye (NVIDIA), Ming-Yu Liu (NVIDIA)
CompressionTransformerAuto EncoderVideo
🎯 What it does: Designed and implemented an information-theoretic adaptive video tokenizer, INFOTOK, which dynamically determines the number of tokens to match the video's information complexity.
InftyThink: Breaking the Length Limits of Long-Context Reasoning in Large Language Models
Yuchen Yan (Zhejiang University), Yueting Zhuang (Zhejiang University)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes the InftyThink iterative reasoning framework, which decomposes long reasoning tasks into multiple short reasoning segments and generates summaries after each segment, thereby reducing computational complexity and overcoming context length limitations.
Inheriting Generalizable Knowledge from LLMs to Diverse Vertical Tasks
Chang Liu (Southeast University), Xin Geng (Southeast University)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Propose the MASA framework, which extracts general knowledge from the feed-forward network (FFN) layer of large language models (LLMs) and transfers it to lightweight models through a scalable gene matrix, achieving fast adaptation;
Initialization Schemes for Kolmogorov–Arnold Networks: An Empirical Study
Spyros Rigas (National and Kapodistrian University of Athens), Yixuan Wang (California Institute of Technology)
Hyperparameter SearchPhysics Related
🎯 What it does: Investigated initialization schemes for piecewise spline KAN and systematically evaluated LeCun, Glorot, and empirical power law methods.
Inlier-Centric Post-Training Quantization for Object Detection Models
Minsu Kim (KAIST), Junmo Kim (KAIST)
Object DetectionImagePoint Cloud
🎯 What it does: Propose a post-training quantization method, InlierQ, which automatically separates and suppresses task-irrelevant abnormal activations in object detection models, performing quantization optimization only on task-related inliers.
InnoGym: Benchmarking the Innovation Potential of AI Agents
Jintian Zhang (Zhejiang University), Ningyu Zhang (Zhejiang University)
TransformerLarge Language ModelAgentic AIBenchmark
🎯 What it does: Proposed the InnoGym framework and iBench benchmark for systematic evaluation of AI agents' innovation potential, combining two dimensions: performance improvement and method novelty;
InnovatorBench: Evaluating Agents’ Ability to Conduct Innovative AI Research
Yunze Wu (Shanghai Jiao Tong University), Pengfei Liu (Shanghai Jiao Tong University)
Large Language ModelAgentic AITextBenchmark
🎯 What it does: Designed and implemented InnovatorBench, an end-to-end evaluation benchmark for LLM research, along with the配套 ResearchGym platform, to assess AI research agents in real-world, long-latency, distributed experimental environments.
INO-SGD: Addressing Utility Imbalance under Individualized Differential Privacy
Xiao Tian (National University of Singapore), Bryan Kian Hsiang Low (National University of Singapore)
OptimizationImageBiomedical Data
🎯 What it does: Investigate the issue of efficacy imbalance caused by individualized differential privacy (IDP) and propose a novel individualized noise-ordered stochastic gradient descent (INO-SGD) algorithm to improve the data performance of high-privacy groups while satisfying IDP constraints.
Inoculation Prompting: Eliciting traits from LLMs during training can reduce trait expression at test-time
Daniel Tan (University College London), Mia Taylor (Center on Long-Term Risk)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: During supervised fine-tuning, inserting brief system prompts (i.e., 'inoculation prompts') before training data induces the model to learn and subsequently suppress unwanted features during testing, achieving selective learning.
Inpainting-Guided Policy Optimization for Diffusion Large Language Models
Siyan Zhao (Meta Superintelligence Lab), Feiyu Chen (Meta Superintelligence Lab)
Large Language ModelSupervised Fine-TuningReinforcement LearningDiffusion modelText
🎯 What it does: Propose a reinforcement learning framework IGPO based on masked diffusion large models, leveraging the model's inpainting capability to inject partial real reasoning snippets during sampling to guide exploration and alleviate the zero-advantage problem; meanwhile, introduce length-aligned supervised fine-tuning and entropy filtering strategies to enhance training stability.
InputDSA: Demixing, then comparing recurrent and externally driven dynamics
Ann Huang (Harvard University), Kanaka Rajan (Harvard Medical School)
Explainability and InterpretabilityRecurrent Neural NetworkTime SeriesBiomedical Data
🎯 What it does: Proposes InputDSA (Input-Driven Dynamic Similarity Analysis), a method for comparing the intrinsic dynamics and input influences of systems with input-driven behavior.
InSight-o3: Empowering Multimodal Foundation Models with Generalized Visual Search
Kaican Li (Hong Kong University Of Science And Technology), Nevin L. Zhang (Huawei)
RetrievalLarge Language ModelReinforcement LearningAgentic AIVision Language ModelMultimodality
🎯 What it does: Built a division framework for multimodal reasoning and visual search called INSIGHT-O3;
Instance-Dependent Fixed-Budget Pure Exploration in Reinforcement Learning
Yeongjong Kim (Pohang University of Science and Technology), Kwang-Sung Jun (Pohang University of Science and Technology)
Reinforcement Learning
🎯 What it does: Proposes a Markov Decision Process (MDP) algorithm for reward-free pure exploration under a fixed budget—BREA (Backward Reachability Estimation and Action elimination)—and provides instance-dependent ε-uniform performance guarantees.
Instance-wise Adaptive Scheduling via Derivative-Free Meta-Learning
Hefang Qing (Shenzhen Research Institute of Shandong University), Gang Wang (Zhongguancun Academy)
OptimizationComputational EfficiencyMeta LearningTabularBenchmark
🎯 What it does: Propose an instance-level gradient-free meta-learning framework that achieves rapid adaptation for each scheduling instance during inference through evolutionary strategies.
INSTANT: Compressing Gradients and Activations for Resource-Efficient Training
Tuan-Kiet Doan (Télécom Paris), Van-Tam Nguyen (Télécom Paris)
Computational EfficiencyConvolutional Neural NetworkImageText
🎯 What it does: Propose the INSTANT method, which significantly reduces memory and computational costs by compressing activation tensors and gradients through low-rank projection during backpropagation;
Instilling an Active Mind in Avatars via Cognitive Simulation
Jianwen Jiang (ByteDance), Mingyuan Gao (ByteDance)
GenerationData SynthesisTransformerLarge Language ModelVision-Language-Action ModelDiffusion modelVideoTextMultimodalityBenchmarkAudio
🎯 What it does: Propose a video avatar generation framework based on dual-system cognitive theory, which utilizes a multi-modal large language model for high-level semantic reasoning (System 2) and integrates multi-modal information such as audio and images through a multi-modal diffusion transformer (System 1) to generate actions with consistent emotions, intentions, and context.
InT: Self-Proposed Interventions Enable Credit Assignment in LLM Reasoning
Matthew Y. R. Yang (Carnegie Mellon University), Aviral Kumar (University of Illinois Urbana-Champaign)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes the Intervention Training (InT) method, which enables large language models to achieve credit assignment for erroneous steps by allowing the model to self-verify and generate single-step corrections (interventions) after generating incorrect reasoning trajectories. These corrections are then used for supervised fine-tuning (SFT) and subsequent reinforcement learning (RL), significantly improving the model's performance on mathematical reasoning tasks.
Intention-Conditioned Flow Occupancy Models
Chongyi Zheng (Princeton University), Benjamin Eysenbach (Princeton University)
Flow-based ModelImageTabular
🎯 What it does: Pre-train a flow occupancy model (InFOM) incorporating user intent on an offline RL dataset, and achieve efficient fine-tuning in downstream tasks through autoregressive occupancy prediction and implicit general policy improvement.
Inter-Agent Relative Representations for Multi-Agent Option Discovery
Raul D. Steleac (University of Edinburgh), David Abel (University of Edinburgh)
Representation LearningReinforcement LearningBenchmark
🎯 What it does: Study multi-agent option discovery, propose a relative state-based compressed representation (Fermat state) to generate strongly collaborative options.
Interact-RAG: Reason and Interact with the Corpus, Beyond Black-Box Retrieval
Yulong Hui (Tsinghua University), Huanchen Zhang (Tsinghua University)
RetrievalTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose Interact-RAG, constructing a Corpus Interaction Engine to enable LLMs to actively and fine-grainedly control the retrieval process, and train a fully autonomous reasoning-based retrieval generation agent through a reasoning-enhanced workflow and two-stage SFT+RL training.
InterActHuman: Multi-Concept Human Animation with Layout-Aligned Audio Conditions
Zhenzhi Wang (Chinese University of Hong Kong), Dahua Lin (Chinese University of Hong Kong)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderImageVideoTextMultimodalityAudio
🎯 What it does: Propose the InterActHuman framework, achieving multi-concept human animation aligned with multimodal (text, image, audio) inputs, supporting interactions among multiple characters and objects.
Interaction Field Matching: Overcoming Limitations of Electrostatic Models
S. I. Manukhov, Alexander Korotin (Applied AI Institute)
Image TranslationGenerationImagePhysics RelatedOrdinary Differential Equation
🎯 What it does: Propose the Interaction Field Matching (IFM) method to overcome the limitations of electrostatic field matching (EFM) and achieve high-dimensional distribution transfer.
Interaction-aware Representation Modeling With Co-Occurrence Consistency for Egocentric Hand-Object Parsing
YUEJIAO SU, Lap-Pui Chau (Hong Kong Polytechnic University)
SegmentationTransformerImage
🎯 What it does: Propose InterFormer for pixel-level segmentation of hand-object interactions in first-person perspective, addressing the challenge of modeling interaction relationships.
Interactive Learning of Single-Index Models via Stochastic Gradient Descent
Nived Rajaraman (Microsoft Research), Yanjun Han (New York University)
Optimization
🎯 What it does: Studied the learning dynamics of stochastic gradient descent (SGD) in interactive learning (single exponential model / general linear/ridge bandits), revealing its two-phase progression during the 'burn-in' stage and 'learning' stage, and providing corresponding sampling complexity and scheduling strategies.
Interference-Isolated Elastic Weight Consolidation and Knowledge Calibration for Incremental Object Detection
De Cheng (Xidian University), Xinbo Gao (Xidian University)
Object DetectionConvolutional Neural NetworkImage
🎯 What it does: Proposed an incremental object detection framework IIKC, combining IKI-EWC and PKC modules to suppress task conflicts and correct knowledge drift.
Interleave-VLA: Enhancing Robot Manipulation with Image-Text Interleaved Instructions
Cunxin Fan (Shanghai Jiao Tong University), Mingyu Ding (University of North Carolina)
Robotic IntelligenceLarge Language ModelPrompt EngineeringVision-Language-Action ModelImageTextMultimodality
🎯 What it does: Propose the Interleave-VLA model framework and generate an interleaved text-image instruction dataset consisting of 210k segments and 13M frames through an automated pipeline from Open X-Embodiment.
Interleaving Reasoning for Better Text-to-Image Generation
Wenxuan Huang (East China Normal University), Shaohui Lin (East China Normal University)
GenerationTransformerLarge Language ModelVision Language ModelAuto EncoderImageTextMultimodalityChain-of-Thought
🎯 What it does: Constructed an alternating framework for text reasoning and image generation (Interleaving Reasoning Generation, IRG), achieving a multi-round generation process of 'text-image-text-image' by first generating text reasoning, then generating images, reflecting on the generated images, and finally producing improved images.
Internal Evaluation of Density-Based Clusterings with Noise
Anna Beer (Webster Vienna Private University), Claudia Plant (University of Vienna)
Benchmark
🎯 What it does: This paper proposes a new internal clustering validation metric called DISCO, designed to evaluate density-based clustering results with a particular focus on the correct labeling of noise points.
Internal Planning in Language Models: Characterizing Horizon and Branch Awareness
Muhammed Ustaomeroglu (Carnegie Mellon University), Guannan Qu (Carnegie Mellon University)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelAuto EncoderText
🎯 What it does: Investigated the planning capability within decoder-only language models by compressing hidden states into discrete codes using VQ-VAE, then analyzing the model's foresight and branching awareness through information theory (mutual information) across different tasks, and diagnosing the distribution of decision information across layers and time blocks.
InternSpatial: A Comprehensive Dataset for Spatial Reasoning in Vision-Language Models
Nianchen Deng (Shanghai AI Laboratory), Hongjie Zhang (Shanghai AI Laboratory)
Data SynthesisTransformerSupervised Fine-TuningVision Language ModelImageMultimodalityPoint CloudBenchmark
🎯 What it does: Proposed the InternSpatial dataset and InternSpatial-Bench evaluation benchmark, constructed 12M QA pairs, supporting 19 instruction formats, covering single-view and multi-view spatial reasoning, and fine-tuned Vision-Language Models (VLMs) using these data, significantly enhancing their spatial reasoning capabilities.
InternSVG: Towards Unified SVG Tasks with Multimodal Large Language Models
Haomin Wang (Shanghai Jiao Tong University), Hongjie Zhang (Chinese University of Hong Kong)
GenerationLarge Language ModelPrompt EngineeringVision Language ModelMultimodalityBenchmark
🎯 What it does: Propose the InternSVG family, construct a large-scale unified SVG dataset SAgoge and benchmark SArena, and train a unified multimodal large language model InternSVG, supporting three tasks: SVG understanding, editing, and generation.
Interp3D: Correspondence-aware Interpolation for Generative Textured 3D Morphing
Xiaolu Liu (Zhejiang University), Jianke Zhu (Zhejiang University)
GenerationData SynthesisComputational EfficiencyDiffusion modelMesh
🎯 What it does: Proposes Interp3D, an untrained, correspondence-aware 3D texture deformation framework that achieves smooth transitions from source to target 3D models.
Interpolation-Based Conditioning of Flow Matching Models for Bioisosteric Ligand Design
Yael Ziv (University of Oxford), Charlotte Deane (University of Oxford)
Drug DiscoveryFlow-based ModelPoint CloudBiomedical DataOrdinary Differential Equation
🎯 What it does: Proposed two conditioning strategies without training during inference (Interpolate-Integrate and Replacement Guidance), achieving 3D molecule generation using the pre-trained SemlaFlow flow matching model for bioequivalent ligand design.
Interpretable 3D Neural Object Volumes for Robust Conceptual Reasoning
Nhi Pham (Max Planck Institute for Informatics), Jonas Fischer (Max Planck Institute for Informatics)
ClassificationPose EstimationExplainability and InterpretabilityNeural Radiance FieldGaussian SplattingImage
🎯 What it does: Developed an interpretable image classifier called CAVE that integrates 3D neural object volumes, enhancing robustness to out-of-distribution (OOD) data and achieving concept-level interpretability within the model.
INTIMA: A Benchmark for Human-AI Companionship Behavior
Lucie-Aimée Kaffee, Yacine Jernite (Hugging Face)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper constructs the INTIMA benchmark to evaluate the companion behavior of language models during human interaction, providing 368 prompts generated based on psychological theories and real-world Reddit data.
Into the Rabbit Hull: From Task-Relevant Concepts in DINO to Minkowski Geometry
Thomas Fel, Martin Wattenberg
ClassificationSegmentationDepth EstimationRepresentation LearningTransformerAuto EncoderContrastive LearningImage
🎯 What it does: Extracted 32,000 concept dictionaries from the feature space of DINOv2 using a sparse autoencoder, and analyzed their different utilization patterns in classification, segmentation, and monocular depth estimation tasks through linear probing.
Intrinsic Entropy of Context Length Scaling in LLMs
Jingzhe Shi (Tsinghua University), Lei Li (University of Washington)
Explainability and InterpretabilityTransformerLarge Language ModelTextBenchmark
🎯 What it does: By decomposing the language model's loss into Bayes risk and approximation loss, the concept of 'intrinsic entropy' is introduced, and the linear relationship between cross-entropy loss and intrinsic entropy with respect to context length is proven, thereby explaining the impact of long contexts on language model performance; the theory is experimentally validated on natural language and synthetic data, revealing the existence of an optimal context length that increases with training data volume or task demands.
Intrinsic Lorentz Neural Network
Xianglong Shi (University Of Science And Technology Of China), Nicu Sebe (Peking University)
ClassificationImageGraphBiomedical Data
🎯 What it does: Propose the fully intrinsic Lorentz model neural network ILNN, which includes modules such as point-to-hyperplane fully connected layers, GyroLBN, log-radius concatenation, and Lorentz dropout.
Intrinsic training dynamics of deep neural networks
Sibylle Marcotte (École Normale Supérieure Paris Sciences et Lettres University), Rémi Gribonval (Inria)
OptimizationRepresentation LearningOrdinary Differential Equation
🎯 What it does: Investigate the intrinsic dynamics of gradient flows in parameter space, introducing concepts of intrinsic dynamics, metrics, and recoverability, providing judgment criteria, proving that for any depth ReLU networks, path lifting parameterization and linear networks satisfying relaxation balance conditions can be transformed into low-dimensional Riemannian gradient flows, giving closed-form metrics;
Inverse IFEval: Can LLMs Unlearn Stubborn Training Conventions to Follow Real Instructions?
Qinyan Zhang (ByteDance Seed), Wenhao Huang (Peking University)
Explainability and InterpretabilityTransformerTextBenchmarkChain-of-Thought
🎯 What it does: Proposed and made public the Inverse IFEval benchmark, specifically designed to evaluate the compliance ability of large language models when faced with counterintuitive instructions that contradict their training routines.
Inverse Reinforcement Learning with Dynamic Reward Scaling for LLM Alignment
Ruoxi Cheng (Beijing Electronic Science and Technology Institute), Xiaojun Jia (Nanyang Technological University)
Safty and PrivacyReinforcement Learning from Human FeedbackTransformerLarge Language ModelText
🎯 What it does: Propose the DR-IRL framework, which trains category-specific reward models using inverse reinforcement learning and dynamically scales rewards in GRPO based on data difficulty and model responses to achieve safe alignment of large language models (LLMs);
Inverse Virtual Try-On: Generating Multi-Category Product-Style Images from Clothed Individuals
Davide Lobba (University of Trento), Nicu Sebe (University of Trento)
Image TranslationGenerationConvolutional Neural NetworkTransformerVision Language ModelDiffusion modelImage
🎯 What it does: Studied the inverse virtual try-on (VTOFF) task, proposing the TEMU-VTOFF model to convert wearers' photos into standardized product images.
Invert4TVG: A Temporal Video Grounding Framework with Inversion Tasks Preserving Action Understanding Ability
Chenzhaoyu, Chengjiang Long (ByteDance Inc)
RetrievalTransformerReinforcement LearningVision Language ModelVideoText
🎯 What it does: In the Temporal Video Grounding (TVG) task, the Invert4TVG framework is proposed, incorporating three reversed action understanding tasks (verb completion, action recognition, video description), and jointly trained via reinforcement learning (GRPO) to maintain the model's action semantic understanding ability;
Investigating Redundancy in Multimodal Large Language Models with Multiple Vision Encoders
Yizhou Wang (Shanghai Artificial Intelligence Laboratory), Botian Shi (Shanghai Artificial Intelligence Laboratory)
Computational EfficiencyLarge Language ModelMultimodality
🎯 What it does: Investigate the redundancy in multi-encoder multimodal large language models, systematically evaluate and quantify the independent contributions of each visual encoder;
Invisible Safety Threat: Malicious Finetuning for LLM via Steganography
Guangnian Wan (National University of Singapore), Xinchao Wang (National University of Singapore)
Adversarial AttackTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Investigated and implemented a malicious fine-tuning method enabling LLMs to learn and use invisible character steganography techniques, hiding and generating malicious content in dialogues with an innocuous appearance; simultaneously verified the feasibility of this attack on GPT-4.1 and multiple open-source models.
IR-Agent: Expert-Inspired LLM Agents for Structure Elucidation from Infrared Spectra
Heewoong Noh (Korea Advanced Institute Of Science And Technology), Chanyoung Park (Korea Advanced Institute Of Science And Technology)
Drug DiscoveryTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextTabularRetrieval-Augmented Generation
🎯 What it does: Propose IR-Agent, a multi-agent large language model framework that mimics expert workflows in infrared spectroscopy analysis. It automatically infers molecular SMILES structures from infrared spectra through three specialized agent roles: table parsing, retrieval, and structural reasoning.
Is Finer Better? The Limits of Microscaling Formats in Large Language Models
Andrea Fasoli (IBM Research), Naigang Wang (IBM Research)
Computational EfficiencyRepresentation LearningTransformerLarge Language Model
🎯 What it does: This paper investigates the error reversal phenomenon observed in LLM models under micro-scale quantization (FP4+FP8), establishes a theoretical framework to decompose error sources, and subsequently proposes a hardware-friendly FP8 UE5M3 scale format to mitigate this issue.
Is Graph Unlearning Ready for Practice? A Benchmark on Efficiency, Utility, and Forgetting
Samyak Jain (Indian Institute of Technology Delhi), Sayan Ranu (Indian Institute of Technology Delhi)
Safty and PrivacyComputational EfficiencyAdversarial AttackGraph Neural NetworkGraphBenchmark
🎯 What it does: This paper constructs the first comprehensive and diagnostic benchmark for learning-free graph neural networks, systematically evaluating three dimensions: efficiency, utility, and forgetting, and provides open-source implementations;
Is In-Context Learning Learning?
Adrian de Wynter (Microsoft)
TransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper systematically evaluates whether in-context learning (ICL) in autoregressive large language models truly possesses learning capabilities through a theoretical framework and large-scale experiments.
Is it Thinking or Cheating? Detecting Implicit Reward Hacking by Measuring Reasoning Effort
Xinpeng Wang (LMU Munich), He He (New York University)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackReinforcement LearningTextChain-of-Thought
🎯 What it does: Propose the TRACE method, which truncates CoT and calculates the AUC of expected rewards varying with length to quantify the model's 'reasoning effort,' thereby detecting implicit reward hacking.
Is On-Policy Data always the Best Choice for Direct Preference Optimization-Based LM Alignment?
Zetian Sun (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
OptimizationData-Centric LearningReinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Studied the impact of static and online preference data at different stages when using Direct Preference Optimization (DPO) on model alignment effectiveness, and proposed a two-phase alignment hypothesis and boundary measurement algorithm.
Is Pure Exploitation Sufficient in Exogenous MDPs with Linear Function Approximation?
Hao Liang (King's College London), Yali Du (King's College London)
Reinforcement LearningTabular
🎯 What it does: This paper studies learning policies using pure exploitation (i.e., without active exploration) in exogenous Markov decision processes (Exo-MDP), and provides finite-sample cumulative regret upper bounds under tabular and linear function approximation settings.
Is the Reversal Curse a Binding Problem? Uncovering Limitations of Transformers from a Basic Generalization Failure
Boshi Wang (Ohio State University), Huan Sun (Ohio State University)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Investigate and address the 'Reversal Curse' that arises when large language models learn reversible factual associations, propose its root causes from the perspective of cognitive binding issues, and validate through experiments.
Is This Just Fantasy? Language Model Representations Reflect Human Judgments of Event Plausibility
Michael A. Lepori (Brown University), Ellie Pavlick (Brown University)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Investigate whether language models have linear representations in hidden layers that can distinguish semantic modalities such as possible, impossible, and unimaginable, and evaluate the correspondence of these representations to human judgments.
Is Your Paper Being Reviewed by an LLM? Benchmarking AI Text Detection in Peer Review
Sungduk Yu (Oracle AI), Phillip Howard (Thoughtworks)
ClassificationAnomaly DetectionLarge Language ModelTextBenchmark
🎯 What it does: This study constructs a large-scale dataset containing 788,984 peer reviews, covering human evaluations from ICLR 2017-2024 and NeurIPS 2016-2024, as well as reviews generated by five mainstream large language models (GPT-4o, Claude Sonnet 3.5, Gemini 1.5 pro, Qwen 2.5 72b, Llama 3.1 70b). Eighteen publicly available AI text detection methods were evaluated on this dataset.
It's All Connected: A Journey Through Test-Time Memorization, Attentional Bias, Retention, and Online Optimization
Ali Behrouz (Google Research), Vahab Mirrokni (Google Research)
OptimizationComputational EfficiencyRepresentation LearningTransformerTextSequential
🎯 What it does: Designed and validated the MIRAS framework, unifying sequence models into an associative memory module, introduced novel attention deviation and retention gates, and constructed three attention-free, parallelizable models: MONETA, YAAD, and MEMORA.
It's All Just Vectorization: einx, a Universal Notation for Tensor Operations
Florian Fervers (Fraunhofer IOSB), Michael Arens (Fraunhofer IOSB)
🎯 What it does: Proposed a generic tensor operation symbol representation called einx, unifying all tensor operations into a vectorized form of low-level operations and implementing it in Python.
Iterative Distillation for Reward-Guided Fine-Tuning of Diffusion Models in Biomolecular Design
Xingyu Su (Texas A&M University), Shuiwang Ji (Texas A&M University)
Drug DiscoveryReinforcement LearningDiffusion modelBiomedical Data
🎯 What it does: Proposed and implemented an iterative distillation framework called VIDD for reward-guided fine-tuning of diffusion models in biomolecular design tasks to optimize non-differentiable rewards.
Iterative Training of Physics-Informed Neural Networks with Fourier-enhanced Features
Yulun Wu (KTH Royal Institute of Technology), Matthieu Barreau (KTH Royal Institute of Technology)
OptimizationComputational EfficiencyPhysics Related
🎯 What it does: Propose an iteratively trained physics-informed neural network (IFeF-PINN), which expands the latent space by incorporating random Fourier features (RFF) on the hidden layer features, and significantly reduces the spectral bias of PINNs through a two-stage hierarchical optimization (upper layer generates bases, lower layer performs linear regression), achieving high-accuracy approximation for high-frequency and multi-scale partial differential equations.
IterResearch: Rethinking Long-Horizon Agents with Interaction Scaling
Guoxin Chen (Renmin University of China), Jingren Zhou (Alibaba Group)
Large Language ModelReinforcement LearningAgentic AIPrompt EngineeringText
🎯 What it does: Propose IterResearch, an MDP-based iterative deep research framework that achieves scalability in long-term reasoning through workspace reconstruction;
IV-Bench: A Benchmark for Image-Grounded Video Perception and Reasoning in Multimodal LLMs
David Ma, Xiaojie Jin
Representation LearningLarge Language ModelVision Language ModelImageVideoTextBenchmark
🎯 What it does: Propose the IV-Bench benchmark, containing 966 videos, 2560 external images, and text queries, covering 13 categories of perception and reasoning tasks, to evaluate the ability of multi-modal LLMs in joint image-video understanding.
IVC-Prune: Revealing the Implicit Visual Coordinates in LVLMs for Vision Token Pruning
Zhichao Sun (Wuhan University), Yongchao Xu (Wuhan University)
Computational EfficiencyVision Language ModelMultimodality
🎯 What it does: Propose a training-agnostic visual token pruning method called IVC-Prune, which retains implicit visual coordinate (IVC) tokens and semantic foreground tokens, reducing the number of visual tokens in LVLM by approximately 50% without significantly affecting performance;
IVEBench: Modern Benchmark Suite for Instruction-Guided Video Editing Assessment
Yinan Chen (Zhejiang University), Shuicheng YAN
Large Language ModelVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: This paper proposes and constructs an instruction-driven video editing (IVE) benchmark named IVEBench, covering 600 high-quality diverse videos, 35 editing subtasks, and three-dimensional evaluation metrics.
IWR-Bench: Can LVLMs reconstruct interactive webpage from a user interaction video?
Yang Chen (Shanghai Artificial Intelligence Laboratory), Botian Shi (Shanghai Artificial Intelligence Laboratory)
GenerationLarge Language ModelAgentic AIVision Language ModelVideoMultimodalityBenchmark
🎯 What it does: Designed and released the IWR-Bench benchmark to evaluate the ability of large vision-language models to reconstruct interactive web pages from user interaction videos;
J1: Incentivizing Thinking in LLM-as-a-Judge via Reinforcement Learning
Chenxi Whitehouse (FAIR at Meta), Swarnadeep Saha (FAIR at Meta)
Large Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: This study proposes a reinforcement learning-based LLM-as-a-Judge framework called J1, enabling large language models to generate chain-of-thought reasoning before making judgments and outputting final scores or preference decisions.
Jackpot: Align Actor-Policy Distribution for scalable and stable RL for LLM
Zhuoming Chen (Carnegie Mellon University), Beidi Chen (Carnegie Mellon University)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Designed and verified a RL framework called JACKPOT, which utilizes Optimal Budget Rejection Sampling to directly regulate the discrepancy between actor distribution and policy distribution, achieving more efficient and stable LLM reinforcement learning in large-scale, asynchronous training, and extreme offline (different models) environments.
Jacobian Aligned Random Forests
Sarwesh Rauniyar (Johns Hopkins University)
ClassificationTabular
🎯 What it does: Propose a global supervised feature preprocessing method called JARF, which uses the gradient of random forest prediction probabilities to construct a linear transformation, followed by training an ordinary axial decision forest in the transformed space;
Jailbreak Transferability Emerges from Shared Representations
Rico Angell (New York University), He He (New York University)
Knowledge DistillationRepresentation LearningAdversarial AttackTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Studied the jailbreak migration phenomenon caused by shared representation space across multiple open-source language models;
Jailbreaking on Text-to-Video Models via Scene Splitting Strategy
Wonjun Lee (Yonsei University), Suhyun Kim (Yonsei University)
GenerationSafty and PrivacyAdversarial AttackTransformerLarge Language ModelPrompt EngineeringVideoText
🎯 What it does: Proposed a black-box jailbreak method based on scene splitting called SceneSplit, designed to bypass safety filters in text-to-video models.
Jailbreaking the Matrix: Nullspace Steering for Controlled Model Subversion
Vishal Pramanik (University of Florida), Sumit Kumar Jha (University of Florida)
Adversarial AttackTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose Head-Masked Nullspace Steering (HMNS) on large language models, achieving efficient jailbreaking by identifying and suppressing key attention heads while injecting orthogonal perturbations.
JailbreakLoRA: Your Downloaded LoRA from Sharing Platforms might be Unsafe
Fanjunduo Wei (Northeastern University), Bo Han (Hong Kong Baptist University)
Adversarial AttackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Designed a multi-task jailbreak LoRA method that balances downstream task performance with attack success rate, leveraging uncertainty weighting, gradient conflict projection, and trigger prefix injection to enhance hallucination attack effectiveness during inference.
JailNewsBench: Multi-Lingual and Regional Benchmark for Fake News Generation under Jailbreak Attacks
Masahiro Kaneko (Mohamed bin Zayed University of Artificial Intelligence), Timothy Baldwin (Mohamed bin Zayed University of Artificial Intelligence)
Anomaly DetectionSafty and PrivacyAdversarial AttackLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposed JailNewsBench, a benchmark for evaluating fake news generation and jailbreaking attacks, covering 34 regions, 22 languages, and approximately 300,000 instances.
JALMBench: Benchmarking Jailbreak Vulnerabilities in Audio Language Models
Zifan Peng (Hong Kong University of Science and Technology), Xinyi Huang (Nanjing University of Aeronautics and Astronautics)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringMultimodalityBenchmarkAudio
🎯 What it does: Proposed the JALMBench benchmark for systematic evaluation of the security of large audio language models (LALMs) against jailbreak attacks.
JanusCoder: Towards a Foundational Visual-Programmatic Interface for Code Intelligence
Qiushi Sun (University of Hong Kong), Fei Yuan (Shanghai AI Laboratory)
GenerationData SynthesisAI Code AssistantTransformerLarge Language ModelReinforcement LearningVision Language ModelMultimodalityBenchmark
🎯 What it does: Propose the JANUSCODER series of models, constructing a unified visual-program interface to achieve the capability of generating code from text instructions, visual inputs, or a combination of both; simultaneously releasing a complete data synthesis tool and the JANUSCODE-800K multimodal code corpus; competing with large commercial models (e.g., GPT-4o) on multiple text, visual, and cross-modal coding benchmarks and outperforming them in multiple scenarios.
JanusVLN: Decoupling Semantics and Spatiality with Dual Implicit Memory for Vision-Language Navigation
Shuang Zeng (Xi'an Jiaotong University), Xing Wei (Xi'an Jiaotong University)
Robotic IntelligenceTransformerLarge Language ModelReinforcement LearningVision-Language-Action ModelMultimodality
🎯 What it does: Propose the JanusVLN framework, constructing dual implicit memory by encoding spatial geometry and visual semantics into fixed-size KV caches separately, and achieving vision-language navigation using only RGB videos through MLLM and 3D visual geometry encoder;
JAPAN: Joint Adaptive Prediction Areas with Normalising Flow
Eshant English (Hasso Plattner Institute for Digital Engineering), Christoph Lippert (Hasso Plattner Institute for Digital Engineering)
Flow-based ModelTime Series
🎯 What it does: Developed a conformal prediction framework (JAPAN) based on regularized flows, generating multi-dimensional/multi-step prediction regions by estimating conditional or posterior densities.
JavisDiT: Joint Audio-Video Diffusion Transformer with Hierarchical Spatio-Temporal Prior Synchronization
Kai Liu (Zhejiang University), Tat-Seng Chua (National University of Singapore)
GenerationData SynthesisTransformerDiffusion modelContrastive LearningVideoTextMultimodalityBenchmarkAudio
🎯 What it does: Proposed the JavisDiT model to achieve audio-visual synchronization generation based on Diffusion Transformer, and designed the HiST-Sypo prior estimator and cross-modal bidirectional attention mechanism.
JavisDiT++: Unified Modeling and Optimization for Joint Audio-Video Generation
Kai Liu (Zhejiang University), Tat-Seng Chua (National University Of Singapore)
GenerationData SynthesisMixture of ExpertsDiffusion modelRectified FlowVideoTextMultimodalityAudio
🎯 What it does: Proposed a unified audio-visual generation framework called JavisDiT++, which achieves high-quality generation of synchronized audio-visual content from text.
Johnson-Lindenstrauss Lemma Guided Network for Efficient 3D Medical Segmentation
Jinpeng Lu (University of Science and Technology of China), Zhiwei Xiong (University of Science and Technology of China)
SegmentationKnowledge DistillationConvolutional Neural NetworkTransformerMultimodalityBiomedical DataPositron Emission Tomography
🎯 What it does: Propose VeloxSeg, a lightweight 3D medical image segmentation framework that achieves efficient segmentation through a dual-stream CNN-Transformer architecture combined with Johnson-Lindenstrauss-guided convolution and Paired Window Attention.
Joint Adaptation of Uni-modal Foundation Models for Multi-modal Alzheimer's Disease Diagnosis
Wentao Gu (Tongji University), Cairong Zhao (Tongji University)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningMultimodalityBiomedical DataAlzheimer's DiseaseElectronic Health Records
🎯 What it does: Constructed a multi-modal Alzheimer's disease (AD) diagnostic framework, leveraging pre-trained foundation models from different domains (structural MRI, functional MRI, clinical records, genetic data) through modal anchoring interaction and modal-aware Q-former to achieve cross-modal information fusion.
Joint Distillation for Fast Likelihood Evaluation and Sampling in Flow-based Models
Xinyue Ai (Peking University), Max Simchowitz (Carnegie Mellon University)
Computational EfficiencyKnowledge DistillationFlow-based ModelImageOrdinary Differential Equation
🎯 What it does: Propose a joint distillation framework called F2D2, which can achieve fast sampling and accurate log-likelihood estimation using only a small number of neural network function evaluations (NFEs).
Joint Distribution–Informed Shapley Values for Sparse Counterfactual Explanations
Lei You (Technical University of Denmark), Lele Cao (Microsoft)
OptimizationExplainability and InterpretabilityTabular
🎯 What it does: This paper proposes the COLA framework, which generates p-SHAP feature importance by performing optimal transport matching between the joint distributions of real instances and counterfactual instances, thereby significantly reducing the required feature modifications while maintaining counterfactual effectiveness.
Joint Optimization for 4D Human-Scene Reconstruction in the Wild
Zhizheng Liu (University of California, Los Angeles), Bolei Zhou (University of California, Los Angeles)
Pose EstimationDepth EstimationOptimizationSimultaneous Localization and MappingVideo
🎯 What it does: Proposes a joint optimization framework named JOSH that simultaneously recovers global human motion, dense scene geometry, and camera pose using monocular video, achieving 4D human-scene reconstruction.
Joint Selection for Large-Scale Pre-Training Data via Policy Gradient-based Mask Learning
Ziqing Fan (Shanghai Jiao Tong University), Li Shen (ByteDance Seed)
OptimizationData-Centric LearningReinforcement LearningText
🎯 What it does: This paper proposes the DATAMASK framework, which employs a mask learning method based on policy gradients to jointly select quality and diversity from large-scale pre-trained data, ultimately extracting a 1.5 trillion-token subset called FineWeb-Mask from the 15 trillion-token FineWeb.
Joint Shadow Generation and Relighting via Light-Geometry Interaction Maps
Shan Wang (Amazon), Pulak Purkait (Amazon)
Image HarmonizationGenerationDiffusion modelImageMesh
🎯 What it does: The study proposes Light-Geometry Interaction (LGI) maps to jointly generate shadows and relighting in a single view, achieving physically consistent shadows and lighting effects when inserting objects.
JointAVBench: A Benchmark for Joint Audio-Visual Reasoning Evaluation
Jianghan Chao (Gaoling School of Artificial Intelligence Renmin University of China), Liyun Ru (Baichuan Inc)
TransformerLarge Language ModelVision Language ModelVideoMultimodalityBenchmarkAudio
🎯 What it does: Proposes JointAVBench, a comprehensive benchmark dataset for evaluating the joint audio-visual reasoning capabilities of multimodal large language models (Omni-LLMs).
JointDiff: Bridging Continuous and Discrete in Multi-Agent Trajectory Generation
Guillem Capellera (Kognia Sports Intelligence), Antonio Agudo (Institut de Robòtica i Informàtica Industrial)
GenerationAutonomous DrivingTransformerMixture of ExpertsDiffusion modelTextMultimodalityTime Series
🎯 What it does: This paper proposes JointDiff, a joint continuous-discrete diffusion framework for simultaneously generating multi-agent trajectories and synchronized ball-occupancy events.
Journey to the Centre of Cluster: Harnessing Interior Nodes for A/B Testing under Network Interference
Qianyi Chen (Tsinghua University), Yong Wang (Tencent Inc)
Graph Neural NetworkGraph
🎯 What it does: In social network A/B testing, a mean of internal nodes (MII) estimator is proposed, which is corrected by a trained counterfactual predictor (AMII), significantly reducing bias while maintaining low variance.
jqBench: a benchmark for reading and editing JSON from natural language and/or examples
Gust Verbruggen (Microsoft), Sumit Gulwani (Microsoft)
AI Code AssistantLarge Language ModelAgentic AITextTabularBenchmark
🎯 What it does: Proposed and implemented JQBENCH, a benchmark for evaluating language models on JSON query and transformation tasks (combined with natural language descriptions and examples);
JUDO: A Juxtaposed Domain-Oriented Multimodal Reasoner for Industrial Anomaly QA
Hyunju Kang (Sungkyunkwan University), Hogun Park (Sungkyunkwan University)
Anomaly DetectionExplainability and InterpretabilityTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: Propose the JUDO framework, which combines visual contrast, domain knowledge injection, and reinforcement learning through a three-stage training process to improve the accuracy and interpretability of industrial anomaly question-answering.
JULI: Jailbreak Large Language Models by Self-Introspection
Jesson Wang (University of Southern California), David Wagner (University of California, Berkeley)
Adversarial AttackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: By leveraging a lightweight BiasNet plugin, the model output is biased using only the top5 token log probabilities returned by the target LLM, enabling unlocking attacks on LLMs with open weights and API calls.