ICLR 2025 Papers — Page 17
International Conference on Learning Representations · 3704 papers
Indirect Gradient Matching for Adversarial Robust Distillation
Hongsin Lee (Korea Advanced Institute of Science and Technology), Changick Kim (Korea Advanced Institute of Science and Technology)
Knowledge DistillationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: An Indirect Gradient Distillation Module (IGDM) is proposed to enhance the effectiveness of adversarial distillation by matching the input gradients of the student model and the teacher model.
INFER: A Neural-symbolic Model For Extrapolation Reasoning on Temporal Knowledge Graph
Ningyuan Li (Beijing University of Posts and Telecommunications), Yifan Zhu (Beijing University of Posts and Telecommunications)
Graph Neural NetworkGraphTime Series
🎯 What it does: A neural symbolic model INFER for temporal knowledge graph extrapolation reasoning is proposed.
Inference Optimal VLMs Need Fewer Visual Tokens and More Parameters
Kevin Li, J Zico Kolter
OptimizationComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageMultimodalityBenchmark
🎯 What it does: This study investigates the optimal trade-off between the number of visual tokens and the size of LLM parameters under a fixed inference computational budget, establishing an optimal scaling law for inference computation.
Inference Scaling for Long-Context Retrieval Augmented Generation
Zhenrui Yue (University of Illinois Urbana-Champaign), Michael Bendersky (Google DeepMind)
GenerationRetrievalComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: A systematic study was conducted on the impact of long-context retrieval-augmented generation (RAG) on the computational expansion during the inference phase, proposing two new strategies, DRAG and IterDRAG, and revealing the linear growth pattern of RAG performance with respect to inference computation from both theoretical and experimental perspectives.
Inference Scaling Laws: An Empirical Analysis of Compute-Optimal Inference for LLM Problem-Solving
Yangzhen Wu (Institute for Interdisciplinary Information Sciences Tsinghua University), Yiming Yang (Carnegie Mellon University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This study investigates the inference scaling laws of large language models, exploring the performance of different model sizes and inference strategies under a given computational budget, and proposes a computationally optimal inference scheme.
Inference-Aware Fine-Tuning for Best-of-N Sampling in Large Language Models
Yinlam Chow (Google Deepmind), Aleksandra Faust (Google Deepmind)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: A reasoning-aware fine-tuning method for the Best-of-N inference strategy is proposed;
Infilling Score: A Pretraining Data Detection Algorithm for Large Language Models
Negin Raoof (University of Texas at Austin), Alex Dimakis
TransformerLarge Language ModelTextBenchmark
🎯 What it does: A pre-training data detection method based on 'Infilling Score' is proposed;
Infinite-Resolution Integral Noise Warping for Diffusion Models
Yitong Deng (Stanford University), Mohammad H. Taghavi (Netflix)
GenerationData SynthesisSuper ResolutionDiffusion modelOptical FlowImageVideo
🎯 What it does: An infinite resolution integral noise distortion algorithm is proposed to transfer pre-trained image diffusion models to the video generation domain without training additional models, achieving spatiotemporal consistent noise through incremental sampling of the Brownian bridge.
Influence Functions for Scalable Data Attribution in Diffusion Models
Bruno Kacper Mlodozeniec (University of Cambridge), Richard E. Turner (University of Cambridge)
OptimizationData-Centric LearningDiffusion modelImage
🎯 What it does: Construct a scalable influence function framework in diffusion models, using (K)E-KFAC to implement data attribution via the Generalised Gauss-Newton approximation;
Influence-Guided Diffusion for Dataset Distillation
Mingyang Chen (Hong Kong University of Science and Technology), Wei Wang (Hong Kong University of Science and Technology)
Data SynthesisKnowledge DistillationDiffusion modelImage
🎯 What it does: View the dataset distillation problem as a guided diffusion generation task, directly generating synthetic samples that meet training effectiveness conditions within the diffusion model.
InfoGS: Efficient Structure-Aware 3D Gaussians via Lightweight Information Shaping
Yunchao Zhang (University of Hong Kong), Yanchao Yang (University of Hong Kong)
Object DetectionSegmentationOptimizationContrastive LearningGaussian SplattingPoint Cloud
🎯 What it does: The mutual information (MI) shaping of the attribute decoding network in the 3D Gaussian Splatting (3DGS) model explicitly encodes the correlation between different Gaussians in the scene, enabling efficient object-level editing (such as segmentation, removal, recoloring, etc.) by fine-tuning network parameters.
Information Theoretic Text-to-Image Alignment
CHAO WANG, Pietro Michiardi (Eurecom)
GenerationData SynthesisDiffusion modelImageText
🎯 What it does: Self-supervised fine-tuning of text-to-image diffusion models using mutual information (MI) from information theory to enhance the semantic consistency between generated images and prompts.
Injecting Universal Jailbreak Backdoors into LLMs in Minutes
Zhuowei Chen (Guangdong University of Foreign Studies), Shichao Pei (University of Massachusetts Boston)
TransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: The JailbreakEdit method is proposed, which allows for the injection of a universal jailbreak backdoor into a securely aligned LLM with a single model edit;
Injective flows for star-like manifolds
Marcello Massimo Negri (University of Basel), Volker Roth (University of Basel)
OptimizationFlow-based ModelTabularTime SeriesFinance Related
🎯 What it does: This paper studies the construction of injective flows on star-shaped manifolds (codimension 1 manifolds) and derives a formula for computing the Jacobian determinant with exactness and O(d²) complexity, enabling the efficient use of complete probability densities in variational inference (without samples, only knowing the unnormalized target); subsequently, this method is applied to variational inference in target Bayesian penalized models and probabilistic mixture models, demonstrating that the posterior distributions can better align with the priors after imposing manifold constraints across various tasks.
Inner Information Analysis Algorithm for Deep Neural Network based on Community
Guipeng Lan (Tianjin University), Jiachen Yang (Tianjin University)
ClassificationRecognitionConvolutional Neural NetworkImage
🎯 What it does: The InnerSightNet algorithm is proposed, which evaluates the neuronization of learnable layers in deep neural networks and clusters them into communities, revealing the functions and information flow of neuron populations.
Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps
Han Wang (Xi'an Jiaotong University), Hui Wang (Tencent QQ)
GenerationOptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper proposes the LoL framework, which first enhances humor judgment and multi-hop reasoning abilities through instruction tuning, then constructs external knowledge using a teacher-student dialogue loop and GPT-4o reasoning, and finally achieves more creative humor generation through Direct Preference Optimization (DPO).
Input Space Mode Connectivity in Deep Neural Networks
Jakub Vrabel (CEITEC Brno University of Technology), David Krueger (Mila University of Montreal)
Anomaly DetectionExplainability and InterpretabilityAdversarial AttackConvolutional Neural NetworkTransformerGenerative Adversarial NetworkImage
🎯 What it does: This paper extends the concept of pattern connectivity from the parameter space to the input space, proving the existence of low-loss paths in the input space of deep networks, and explores its implications for adversarial sample detection and interpretability.
INS: Interaction-aware Synthesis to Enhance Offline Multi-agent Reinforcement Learning
Yuqian Fu (Chinese Academy of Sciences), Dongbin Zhao (Chinese Academy of Sciences)
Data SynthesisTransformerReinforcement LearningDiffusion model
🎯 What it does: This paper proposes an interactive perception data synthesis method called INS, which utilizes diffusion models to generate high-quality datasets in multi-agent offline reinforcement learning.
InsightBench: Evaluating Business Analytics Agents Through Multi-Step Insight Generation
Gaurav Sahu (ServiceNow Research), Issam H. Laradji
TransformerLarge Language ModelAgentic AITabularBenchmarkFinance Related
🎯 What it does: Created InsightBench, an end-to-end data analysis benchmark with 100 enterprise business use cases, and proposed the AgentPoirot agent and LLaMA-3-Eval evaluation method;
Inspection and Control of Self-Generated-Text Recognition Ability in Llama3-8b-Instruct
Christopher Ackerman, Nina Panickssery
RecognitionGenerationTransformerLarge Language ModelContrastive LearningText
🎯 What it does: This paper experiments with the self-writing recognition ability of Llama3-8b-Instruct, finding that it can distinguish between text generated by itself and that produced by humans or other models across various datasets. By extracting contrastive vectors from the residual flow, a vector that has a causal effect on self-recognition judgment at layer 16 is identified. Subsequently, experiments on steering and coloring using this vector demonstrate that the model can be forced to declare or deny its own writing at both the output and input layers, and its perception can also be adjusted.
Instance-dependent Early Stopping
Suqin Yuan (Sydney AI Centre, University of Sydney), Tongliang Liu (Sydney AI Centre, University of Sydney)
Object DetectionSegmentationComputational EfficiencyImage
🎯 What it does: An instance-level early stopping method called IES is proposed, which determines whether an instance has been 'mastered' by monitoring the second-order difference of the loss for each sample, and stops backpropagation for that instance once it is deemed mastered, thereby reducing the computational load of training.
Instant Policy: In-Context Imitation Learning via Graph Diffusion
Vitalis Vosylius (Imperial College London), Edward Johns (Imperial College London)
Robotic IntelligenceReinforcement Learning from Human FeedbackGraph Neural NetworkReinforcement LearningDiffusion modelPoint Cloud
🎯 What it does: Proposes Instant Policy, which utilizes graph diffusion to complete tasks with a one-time demonstration using a small number of examples.
InstantPortrait: One-Step Portrait Editing via Diffusion Multi-Objective Distillation
Zhixin Lai (Snap Inc), Erli Ding (Snap Inc)
Image TranslationGenerationKnowledge DistillationDiffusion modelImage
🎯 What it does: This study proposes a first-order instruction-driven portrait editing model, IPNet, which can quickly complete edits in various aspects such as background, clothing, and facial features while maintaining facial identity.
InstantSplamp: Fast and Generalizable Stenography Framework for Generative Gaussian Splatting
Chenxin Li (Chinese University of Hong Kong), Yixuan Yuan (Chinese University of Hong Kong)
GenerationData SynthesisCompressionSafty and PrivacyContrastive LearningGaussian SplattingPoint CloudMesh
🎯 What it does: This paper proposes InstantSplamp, a 3D steganography framework that can directly embed copyright or proprietary information, capable of performing steganography and recovery without additional time costs during the generation of Gaussian Splatting 3D assets.
InstantSwap: Fast Customized Concept Swapping across Sharp Shape Differences
Chenyang Zhu (Tsinghua University), Xiu Li (Tsinghua University)
Image TranslationGenerationComputational EfficiencyDiffusion modelScore-based ModelImageBenchmark
🎯 What it does: A custom concept swapping method is proposed to address the issue of customized replacement of concepts in images, which can quickly maintain consistency between the foreground and background.
InstaRevive: One-Step Image Enhancement via Dynamic Score Matching
Yixuan Zhu (Tsinghua University), Jiwen Lu (Tsinghua University)
RestorationGenerationSuper ResolutionDiffusion modelScore-based ModelImage
🎯 What it does: A single-step image enhancement framework based on dynamic score matching and text prompts (InstaRevive) is proposed.
InstaSHAP: Interpretable Additive Models Explain Shapley Values Instantly
James Enouen (University of Southern California), Yan Liu (University of Southern California)
Explainability and InterpretabilityImageTabular
🎯 What it does: This paper proposes the InstaSHAP method by unifying the variational framework of SHAP and GAM, which can instantaneously obtain SHAP values in a single forward pass; it also provides a theoretical analysis of the representational capabilities of SHAP and GAM under correlated features; and demonstrates improvements over traditional FastSHAP and FaithSHAP.
InstaTrain: Adaptive Training via Ultra-Fast Natural Annealing within Dynamical Systems
Chuan Liu (University of Rochester), Tong Geng (University of Rochester)
Time SeriesFinance Related
🎯 What it does: This paper proposes InstaTrain, which achieves ultra-fast temporal prediction training and online updates by transforming the model training process into a natural annealing of an electronic dynamical system.
Instruct-SkillMix: A Powerful Pipeline for LLM Instruction Tuning
Simran Kaur (Princeton University), Sanjeev Arora (Meta)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A fully automated INSTRUCT-SKILLMIX process is proposed, which uses a powerful LLM to first extract a list of 'skills' required for instruction execution, and then randomly combines these skills to generate diverse (instruction, response) training samples.
Instructional Segment Embedding: Improving LLM Safety with Instruction Hierarchy
Tong Wu (Princeton University), Wenxuan Zhou (Zoom Video Communications)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes the Instructional Segment Embedding (ISE) technique, which aims to directly embed instruction-level information into the input representation of large language models, enhancing the model's ability to distinguish and follow different priorities of system instructions and user instructions.
InstructRAG: Instructing Retrieval-Augmented Generation via Self-Synthesized Rationales
Zhepei Wei (University of Virginia), Yu Meng (University of Virginia)
GenerationRetrievalExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: A retrieval-augmented generation framework called INSTRUCTRAG is proposed, which explicitly denoises the retrieved text and generates an interpretable answer reasoning process.
Integral Performance Approximation for Continuous-Time Reinforcement Learning Control
Brent A. Wallace (Arizona State University), Jennie Si (Arizona State University)
Reinforcement LearningTime Series
🎯 What it does: A continuous-time reinforcement learning control method based on Integral Performance Approximation (IPA) is proposed;
Integrating Protein Dynamics into Structure-Based Drug Design via Full-Atom Stochastic Flows
Xiangxin Zhou (University of Chinese Academy of Sciences), Jianzhu Ma (Tsinghua University)
Drug DiscoveryGraph Neural NetworkTransformerFlow-based ModelBiomedical DataStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Utilizing generative models to simultaneously convert protein pockets from apo state to holo state and generate corresponding ligand molecules, enhancing the efficiency and effectiveness of structure-based drug design (SBDD).
Integrative Decoding: Improving Factuality via Implicit Self-consistency
Yi Cheng (Hong Kong Polytechnic University), Wayne Xiong (Microsoft Research)
GenerationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A new decoding strategy called Integrative Decoding (ID) is proposed, which implicitly integrates self-consistency during the decoding process to enhance the factuality of large language models.
Intelligence at the Edge of Chaos
Shiyang Zhang (Columbia University), David van Dijk (Yale University)
TransformerLarge Language ModelTextTime Series
🎯 What it does: Pre-trained the GPT-2 model on data generated from elementary cellular automata (ECA) of varying complexity for next-token prediction, and evaluated its transfer performance on ARC-inspired reasoning tasks and Lichess elite chess move prediction tasks to explore the relationship between data complexity and model intelligence.
Intelligent Go-Explore: Standing on the Shoulders of Giant Foundation Models
Cong Lu (University of British Columbia), Jeff Clune (University of British Columbia)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringTextMultimodalityChain-of-Thought
🎯 What it does: By combining the interest judgment capability of large-scale pre-trained foundational models, the Go-Explore algorithm is modified to achieve intelligent state selection, action selection, and archive filtering without relying on manual heuristics, efficiently solving multimodal text/visual search and exploration tasks.
Intent3D: 3D Object Detection in RGB-D Scans Based on Human Intention
Weitai Kang (University of Illinois Chicago), Yan Yan (University of Illinois Chicago)
Object DetectionLarge Language ModelContrastive LearningTextPoint Cloud
🎯 What it does: This paper proposes a 3D intent localization task, constructs the Intent3D dataset, and designs the IntentNet model based on intent understanding, candidate box matching, and cascading adaptive learning, achieving automatic detection of targets in 3D scenes based on free-text intents.
Interaction Asymmetry: A General Principle for Learning Composable Abstractions
Jack Brady (Max Planck Institute for Intelligent Systems), Wieland Brendel (Max Planck Institute for Intelligent Systems)
GenerationRepresentation LearningTransformerAuto EncoderImage
🎯 What it does: Proposes the principle of interactive asymmetry, proving that it can achieve separability and combinatorial generalization of concepts, and implements this theory based on a Transformer VAE model.
Interactive Adjustment for Human Trajectory Prediction with Individual Feedback
Jianhua Sun (Shanghai Jiao Tong University), Cewu Lu (Shanghai Jiao Tong University)
Object TrackingMultimodalityTime Series
🎯 What it does: This paper proposes an Interactive Adjustment Network (IAN) that utilizes individual feedback (i.e., the difference between the model's past predictions and the actual trajectory) to dynamically adjust the current trajectory prediction, thereby improving prediction accuracy.
Interactive Speculative Planning: Enhance Agent Efficiency through Co-design of System and User Interface
Wenyue Hua (Rutgers University), Chi Wang (Google Deepmind)
OptimizationComputational EfficiencyLarge Language ModelAgentic AIMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This paper proposes Interactive Speculative Planning, which reduces the planning latency of large language model agents by using efficient approximate agents and powerful goal agents, combined with a human-computer interaction interface.
Interference Among First-Price Pacing Equilibria: A Bias and Variance Analysis
Luofeng Liao (Columbia University), Congshan Zhang (Meta)
TabularFinance Related
🎯 What it does: A parallel budget-controlled A/B testing scheme based on market segmentation is designed, and theoretical analysis and debiasing estimation of interference are conducted using the FPPE framework in the first-price auction market.
Interleaved Scene Graphs for Interleaved Text-and-Image Generation Assessment
Dongping Chen (University of Washington), Ranjay Krishna (University of Washington)
GenerationData SynthesisGraph Neural NetworkLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Proposed the Interleaved Scene Graph (ISG) evaluation framework and its corresponding ISG-BENCH benchmark for multi-granularity evaluation of text-image interleaved generation.
InterMask: 3D Human Interaction Generation via Collaborative Masked Modeling
Muhammad Gohar Javed (University of Alberta), Xingyu Li (University of Alberta)
GenerationData SynthesisPose EstimationTransformerDiffusion modelVideoTextMesh
🎯 What it does: This work proposes a framework called InterMask based on discrete spatial mask generation, which can generate high-quality 3D interactive actions for two people based on text descriptions and seamlessly support reactive generation.
Intermediate Layer Classifiers for OOD generalization
Arnas Uselis (University of Tübingen), Seong Joon Oh (University of Tübingen)
ClassificationDomain AdaptationConvolutional Neural NetworkTransformerSupervised Fine-TuningImage
🎯 What it does: This paper proposes the use of Intermediate Layer Linear Classifiers (ILC) in deep networks for OOD generalization.
Internet of Agents: Weaving a Web of Heterogeneous Agents for Collaborative Intelligence
Weize Chen (Tsinghua University), Maosong Sun (Tsinghua University)
Large Language ModelAgentic AITextBenchmark
🎯 What it does: This paper proposes the Internet of Agents (IoA), a distributed multi-agent collaboration framework based on an instant messaging architecture, supporting heterogeneous third-party agents, cross-device collaboration, and dynamic team and dialogue process management.
Interpretable Bilingual Multimodal Large Language Model for Diverse Biomedical Tasks
Lehan Wang (Hong Kong University of Science and Technology), Xiaomeng Li (Sun Yat-Sen University)
ClassificationRecognitionObject DetectionGenerationTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBiomedical DataMagnetic Resonance ImagingComputed TomographyChain-of-Thought
🎯 What it does: A bilingual interpretable medical multimodal large language model MedRegA is proposed, supporting image-level and region-level visual-language tasks.
Interpretable Causal Representation Learning for Biological Data in the Pathway Space
Jesus de la Fuente Cedeño (University of Navarra), Mikel Hernaez (University of Navarra)
Explainability and InterpretabilityRepresentation LearningDrug DiscoveryAuto EncoderBiomedical Data
🎯 What it does: This paper proposes a model called SENA-discrepancy-VAE that integrates biological pathways as prior knowledge into causal representation learning (CRL). It can learn interpretable causal latent factors from Perturb-seq data and predict the effects of unseen gene/drug perturbations.
Interpretable Unsupervised Joint Denoising and Enhancement for Real-World low-light Scenarios
Li Huaqiu, Haoqian Wang (Tsinghua University)
RestorationExplainability and InterpretabilityTransformerImage
🎯 What it does: A zero-reference, interpretable joint denoising and low-light enhancement framework is proposed, which can simultaneously recover noise and brightness without using reference images.
Interpretable Vision-Language Survival Analysis with Ordinal Inductive Bias for Computational Pathology
Pei Liu (University of Electronic Science and Technology of China), Mao Ye (University of Electronic Science and Technology of China)
Explainability and InterpretabilityTransformerVision Language ModelImageBiomedical Data
🎯 What it does: This study proposes a survival analysis framework for pathological slides based on a visual-language model, VLSA, which utilizes language-encoded prognostic priors to assist multi-instance aggregation for predicting survival risk in panoramic slides.
Interpreting and Editing Vision-Language Representations to Mitigate Hallucinations
Nicholas Jiang, Yossi Gandelsman (University of California)
SegmentationExplainability and InterpretabilityTransformerVision Language ModelImage
🎯 What it does: The paper decodes the intermediate image representations in visual-language models (VLM) to a language vocabulary, using their internal confidence to distinguish between real objects and hallucinations, and suppresses hallucinations through linear orthogonalization of image embeddings (PROJECTAWAY);
Interpreting Emergent Planning in Model-Free Reinforcement Learning
Thomas Bush (University of Cambridge), David Krueger (Mila, University of Montreal)
Explainability and InterpretabilityRecurrent Neural NetworkReinforcement LearningAgentic AISequential
🎯 What it does: This paper explores whether model-agnostic reinforcement learning agents (DRC) implement planning internally through a conceptual interpretability approach and provides a non-behavioral proof.
Interpreting Language Reward Models via Contrastive Explanations
Junqi Jiang (Imperial College London), Manuela Veloso (J.P. Morgan AI Research)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes using contrastive explanations (CF and SF) to interpret language reward models (RM), generating comparisons that perturb the original responses on high-level attributes, and analyzing the local and global behavior of RM through these comparisons.
Interpreting the Second-Order Effects of Neurons in CLIP
Yossi Gandelsman (University of California Berkeley), Jacob Steinhardt (University of California Berkeley)
SegmentationExplainability and InterpretabilityAdversarial AttackTransformerLarge Language ModelVision Language ModelImageText
🎯 What it does: This paper automatically explains individual neurons in the CLIP-ViT model by constructing a 'second-order lens', using text descriptions to reveal neuron functions, and applies this explanation to zero-shot semantic segmentation and large-scale semantic adversarial sample generation.
IntersectionZoo: Eco-driving for Benchmarking Multi-Agent Contextual Reinforcement Learning
Vindula Jayawardana (Massachusetts Institute of Technology), Cathy Wu (Massachusetts Institute of Technology)
Autonomous DrivingReinforcement LearningTabularTime SeriesBenchmark
🎯 What it does: A multi-agent contextual reinforcement learning benchmark package, IntersectionZoo, based on real intersections has been established to evaluate the generalization of eco-driving algorithms.
Intervening Anchor Token: Decoding Strategy in Alleviating Hallucinations for MLLMs
Feilong Tang (Hong Kong University of Science and Technology), Ser-Nam Lim (University of Central Florida)
GenerationData SynthesisTransformerLarge Language ModelVision Language ModelTextMultimodality
🎯 What it does: This paper explores the causes of hallucination phenomena in multimodal large language models (MLLMs) and proposes a decoding strategy called TAME (Dynamic Intervention of Anchor Tag Propagation) to mitigate the hallucination problem.
Intrinsic Dimension Correlation: uncovering nonlinear connections in multimodal representations
Lorenzo Basile (University of Trieste), Alex Rodriguez (University of Trieste)
Anomaly DetectionRepresentation LearningImageTextMultimodality
🎯 What it does: A nonlinear correlation measure based on intrinsic dimensionality, ID Cor, is proposed to detect nonlinear associations between high-dimensional datasets and the hidden layers of neural networks.
Intrinsic User-Centric Interpretability through Global Mixture of Experts
Vinitra Swamy (École Polytechnique Fédérale de Lausanne), Tanja Käser (École Polytechnique Fédérale de Lausanne)
Explainability and InterpretabilityLarge Language ModelMixture of ExpertsTextTabularTime SeriesBiomedical Data
🎯 What it does: We propose InterpretCC, an interpretable neural network with adaptive sparse feature/concept gating and expert mixture, providing locally interpretable and easily operable prediction explanations.
Inverse Attention Agents for Multi-Agent Systems
Qian Long (University of California Los Angeles), Demetri Terzopoulos (University of California Los Angeles)
Robotic IntelligenceReinforcement LearningAgentic AISequential
🎯 What it does: Design and train an inverse attention agent to infer the attention targets of similar agents through an inverse attention network, thereby achieving adaptive cooperation in multi-agent environments.
Inverse Constitutional AI: Compressing Preferences into Principles
Arduin Findeis (University of Cambridge), Robert D. Mullins
CompressionExplainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Proposes the Inverse Constitutional AI (ICAI) method, which compresses preference data into an interpretable set of natural language principles (the constitution) and reconstructs the original preferences through LLM.
Inverse decision-making using neural amortized Bayesian actors
Dominik Straub (Technical University of Darmstadt), Constantin A. Rothkopf (Technical University of Darmstadt)
OptimizationRobotic IntelligenceReinforcement LearningTabular
🎯 What it does: A reverse decision-making framework based on a neural network approximate Bayesian actor has been developed, capable of efficiently inferring perception and action parameters, priors, and cost functions in continuous action tasks.
Inverse Rendering using Multi-Bounce Path Tracing and Reservoir Sampling
Yuxin Dai (Nanyang Technological University), Ying He (SenseTime Research and Tetras.AI)
RestorationOptimizationNeural Radiance FieldMesh
🎯 What it does: A two-stage physics-based inverse rendering framework MIRReS is proposed, which jointly optimizes explicit triangular meshes, materials, and environmental lighting.
InverseBench: Benchmarking Plug-and-Play Diffusion Priors for Inverse Problems in Physical Sciences
Hongkai Zheng (California Institute of Technology), Yisong Yue (California Institute of Technology)
RestorationOptimizationComputational EfficiencyDiffusion modelImageMagnetic Resonance ImagingBenchmarkPhysics Related
🎯 What it does: The INVERSEBENCH framework is proposed, providing a systematic evaluation of Plug-and-Play diffusion priors (PnPDP) for inverse problems in five scientific fields (optical scattering, compressed sensing MRI, black hole imaging, full waveform inversion, and Navier–Stokes initial field reconstruction).
InversionGNN: A Dual Path Network for Multi-Property Molecular Optimization
Yifan Niu (Hong Kong University of Science and Technology), Jia Li (Hong Kong University of Science and Technology)
OptimizationDrug DiscoveryGraph Neural NetworkGraph
🎯 What it does: A dual-path graph neural network called InversionGNN is proposed, which learns chemical knowledge through predictive paths and utilizes reverse generation paths for molecular optimization via gradients, thereby seeking Pareto optimal solutions in multi-objective molecular optimization.
InvestESG: A multi-agent reinforcement learning benchmark for studying climate investment as a social dilemma
Xiaoxuan Hou (University of Washington), Natasha Jaques (Google Deepmind)
Reinforcement LearningTabularBenchmarkFinance Related
🎯 What it does: A multi-agent reinforcement learning benchmark, InvestESG, has been designed and implemented to simulate the long-term interactions between companies and investors under mandatory ESG disclosure policies, and to study the incentive effects of this policy on corporate emission reduction investments.
Investigating Pattern Neurons in Urban Time Series Forecasting
Chengxin Wang (National University of Singapore), Gary Tan (National University of Singapore)
TransformerTime Series
🎯 What it does: This study investigates neurons related to low-frequency events (such as holidays) in urban time series models and proposes a training method (PN-Train) to enhance the prediction accuracy of low-frequency events by detecting and fine-tuning these neurons.
Investigating the Pre-Training Dynamics of In-Context Learning: Task Recognition vs. Task Learning
Xiaolei Wang (Renmin University of China), Ji-Rong Wen (Renmin University of China)
ClassificationTransformerLarge Language ModelText
🎯 What it does: This study investigates the competitive relationship between task recognition (TR) and task learning (TL) during the pre-training phase of large language models, and quantifies their impact on in-context learning (ICL) performance using metrics.
IPDreamer: Appearance-Controllable 3D Object Generation with Complex Image Prompts
Bohan Zeng (Peking University), Shuicheng YAN
GenerationData SynthesisDiffusion modelScore-based ModelImageTextMesh
🎯 What it does: IPDreamer provides a controllable 3D object generation framework based on complex image prompts, capable of transferring image details with high fidelity to 3D meshes.
IRIS: LLM-Assisted Static Analysis for Detecting Security Vulnerabilities
Ziyang Li (University of Pennsylvania), Mayur Naik (University of Pennsylvania)
AI Code AssistantTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Combining large language models (LLM) with static analysis for vulnerability detection across entire Java projects, automatically inferring taint specifications and filtering false positives through contextual analysis.
Is Factuality Enhancement a Free Lunch For LLMs? Better Factuality Can Lead to Worse Context-Faithfulness
Baolong Bi (Chinese Academy of Sciences), Xueqi Cheng (Chinese Academy of Sciences)
TransformerLarge Language ModelText
🎯 What it does: Evaluated the impact of various factual enhancement methods (such as DoLa, ICD, ITI, TruthX, etc.) on the contextual fidelity of LLAMA2-CHAT in knowledge editing tasks, revealing their negative mechanisms through logit, hidden states, and interpretability algorithms.
Is In-Context Learning Sufficient for Instruction Following in LLMs?
Hao Zhao (École Polytechnique Fédérale de Lausanne), Nicolas Flammarion (École Polytechnique Fédérale de Lausanne)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: This study investigates the effect of in-context learning (ICL) on instruction alignment for large language models and conducts a systematic comparison with traditional instruction fine-tuning (IFT).
Is Large-scale Pretraining the Secret to Good Domain Generalization?
Piotr Teterwak (Boston University), Kate Saenko (Boston University)
Domain AdaptationTransformerVision Language ModelContrastive LearningImageText
🎯 What it does: This paper proposes the Alignment Hypothesis by analyzing the image-text alignment of large pre-trained models such as CLIP, and based on this, divides the DomainBed target data into two categories: In-Pretraining (IP) and Out-of-Pretraining (OOP). It systematically evaluates the performance of existing Domain Generalization (DG) methods on these two types of data, revealing that most methods rely on pre-trained features, approximate oracle performance on IP, but significantly fail on OOP.
Is uniform expressivity too restrictive? Towards efficient expressivity of GNNs
Sammy Khalife (Cornell University), Josué Tonelli-Cueto (Johns Hopkins University)
Graph Neural NetworkGraph
🎯 What it does: This paper studies the expressive power of Graph Neural Networks (GNNs) for GC2 (two-variable constrained model) queries under both uniform and non-uniform expressiveness. It proves that GNNs with Pfaffian activation functions such as Sigmoid and tanh cannot achieve uniform expressiveness. Additionally, it proposes an 'approximate uniform expressiveness' scheme, which utilizes step-like activation functions that require only O(log log ∆) parameters to express any GC2 query at a maximum degree ∆.
Is Your Model Really A Good Math Reasoner? Evaluating Mathematical Reasoning with Checklist
Zihao Zhou (University of Liverpool), Kaizhu Huang (Duke Kunshan University)
TransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: The MATHCHECK checklist is proposed to systematically evaluate the mathematical reasoning abilities of LLMs, covering four dimensions: task generalization (solving, answer feasibility, result judgment, process judgment) and reasoning robustness (problem understanding, irrelevant interference, scene understanding);
Is Your Multimodal Language Model Oversensitive to Safe Queries?
Xirui Li (University of California), Cho-Jui Hsieh (University of California)
Safty and PrivacyTransformerLarge Language ModelVision Language ModelTextMultimodalityBenchmark
🎯 What it does: This paper constructs and evaluates the over-sensitivity behavior of multimodal large language models (MLLMs), proposing MOSSBench, a benchmark specifically designed to measure the misrejection of benign queries, and conducts systematic experiments on 20 models.
Is Your Video Language Model a Reliable Judge?
Ming Liu (Iowa State University), Wensheng Zhang (Iowa State University)
Large Language ModelAgentic AIVision Language ModelVideoText
🎯 What it does: This study evaluates the methods for assessing video language models (VLM) in video understanding tasks, comparing the reliability and accuracy of individual VLM assessments, LLM collaborative discussions (Agent-Debate), and collective thinking aggregation assessments.
Isometric Regularization for Manifolds of Functional Data
Hyeongjun Heo (Seoul National University), Yonghyeon Lee (Korea Institute for Advanced Study)
Auto EncoderImagePoint Cloud
🎯 What it does: This paper proposes an isometric regularization method for infinite-dimensional function data, constructing a latent variable model as a Riemannian manifold, and maintaining the geometric consistency between the latent space and the data manifold through regularization, thereby enhancing the reconstruction and interpolation performance of function data (such as SDF, BRDF, and neural operators).
It Helps to Take a Second Opinion: Teaching Smaller LLMs To Deliberate Mutually via Selective Rationale Optimisation
Sohan Patnaik (Adobe), Balaji Krishnamurthy (Adobe)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: The COALITION framework is proposed, which generates and refines high-quality reasoning chains by training two different variants of the same model without relying on external large language models and manually annotated reasoning paths, thereby enhancing the performance of small language models on complex tasks.
Iterative Label Refinement Matters More than Preference Optimization under Weak Supervision
Yaowen Ye (University of Hong Kong), Jacob Steinhardt (University of California)
OptimizationReinforcement Learning from Human FeedbackSupervised Fine-TuningReinforcement LearningText
🎯 What it does: The study investigates the post-training of language models under weak supervision (small models or time-constrained humans) and finds that traditional SFT+DPO cannot improve model performance in this scenario; it proposes an Iterative Label Refinement (ILR) method that improves training data through feedback comparison and re-conducts SFT, thereby enhancing model performance.
Iterative Nash Policy Optimization: Aligning LLMs with General Preferences via No-Regret Learning
Yuheng Zhang (University of Illinois Urbana-Champaign), Dong Yu (Tencent AI Lab)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningTextBenchmarkOrdinary Differential Equation
🎯 What it does: This paper proposes an online RLHF algorithm based on game theory—Iterative Nash Policy Optimization (INPO), which iteratively solves the Nash strategy under general preferences through no-regret learning.
Iterative Substructure Extraction for Molecular Relational Learning with Interactive Graph Information Bottleneck
Shuai Zhang (University of Science and Technology of China), Yang Wang (University of Science and Technology of China)
Drug DiscoveryGraph Neural NetworkContrastive LearningGraph
🎯 What it does: An iterative substructure extraction framework (ISE) and an interactive graph information bottleneck (IGIB) are proposed, which can accurately extract core interactive substructures from molecular pairs and perform molecular relationship learning.
IterComp: Iterative Composition-Aware Feedback Learning from Model Gallery for Text-to-Image Generation
Xinchen Zhang (Tsinghua University), Bin CUI
GenerationData SynthesisReinforcement LearningDiffusion modelImageTextMultimodality
🎯 What it does: A multi-model ensemble preference dataset is constructed, and an IterComp iterative feedback learning framework is proposed, utilizing a reward model and a benchmark diffusion model for closed-loop co-evolution to enhance the text-to-image combination generation capability.
IterGen: Iterative Semantic-aware Structured LLM Generation with Backtracking
Shubham Ugare (University of Illinois Urbana-Champaign), Sasa Misailovic (University of Illinois Urbana-Champaign)
GenerationSafty and PrivacyTransformerLarge Language ModelText
🎯 What it does: Designed the ITERGEN framework, achieving forward and backward LLM generation control based on grammatical symbols, supporting semantic constraints and backtracking.
IV-mixed Sampler: Leveraging Image Diffusion Models for Enhanced Video Synthesis
Shitong Shao (Hong Kong University of Science and Technology), Zeke Xie (Baidu Inc.)
GenerationData SynthesisDiffusion modelVideoOrdinary Differential Equation
🎯 What it does: A training-free sampling algorithm called IV-Mixed Sampler is designed, which enhances the visual quality of video diffusion models (VDM) by utilizing high-quality sampling from image diffusion models (IDM), and ensures temporal consistency through the alternating use of DDIM and DDIM-Inversion.
Jailbreak Antidote: Runtime Safety-Utility Balance via Sparse Representation Adjustment in Large Language Models
Guobin Shen (Beijing Institute of AI Safety and Governance), Yi Zeng (Beijing Institute of AI Safety and Governance)
OptimizationSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: A mechanism for real-time adjustable LLM security has been designed—Jailbreak Antidote, which utilizes sparse internal state adjustments to suppress jailbreak attacks while maintaining model utility.
Jailbreaking as a Reward Misspecification Problem
Zhihui Xie (University of Hong Kong), Lingpeng Kong (University of Hong Kong)
Adversarial AttackReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Proposes viewing LLM vulnerabilities as a reward mis-specification problem, defines ReGap to measure reward mis-specification, and designs the ReMiss automated red team attack system based on this;
Jailbreaking Leading Safety-Aligned LLMs with Simple Adaptive Attacks
Maksym Andriushchenko (École Polytechnique Fédérale de Lausanne), Nicolas Flammarion (École Polytechnique Fédérale de Lausanne)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: The study demonstrates that even the latest safety-aligned large language models (LLMs) remain highly vulnerable to simple adaptive jailbreak attacks, and proposes a set of adaptive attack frameworks utilizing techniques such as prompt templates, random search, self-transfer, and pre-filling.
Jamba: Hybrid Transformer-Mamba Language Models
Barak Lenz (AI21 Labs), Yoav Shoham (AI21 Labs)
TransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsText
🎯 What it does: A new hybrid Transformer-Mamba structure is proposed and implemented, on which two large-scale MoE language models (Jamba-1.5-Mini and Jamba-1.5-Large) are built.
JetFormer: An autoregressive generative model of raw images and text
Michael Tschannen (Google DeepMind), Alexander Kolesnikov (Google DeepMind)
GenerationData SynthesisTransformerFlow-based ModelImageTextMultimodality
🎯 What it does: This paper proposes JetFormer, an end-to-end multimodal generation model that combines regularized flow with autoregressive Transformers;
Joint Fine-tuning and Conversion of Pretrained Speech and Language Models towards Linear Complexity
Mutian He (Idiap Research Institute), Philip N. Garner (Idiap Research Institute)
Computational EfficiencyKnowledge DistillationTransformerSupervised Fine-TuningTextAudio
🎯 What it does: A cross-architecture layer-wise distillation method (CALD) has been developed, which directly converts the pre-trained Transformer into a model with linear complexity during the fine-tuning phase;
Joint Gradient Balancing for Data Ordering in Finite-Sum Multi-Objective Optimization
Hansi Yang (Hong Kong University of Science and Technology), James Kwok (Hong Kong University of Science and Technology)
SegmentationDepth EstimationOptimizationImage
🎯 What it does: This paper proposes a Joint Gradient Balancing Sample Ranking method (JoGBa) for multi-objective finite sum optimization problems, which determines the sample order of different objectives through online vector balancing to accelerate gradient descent.
Joint Graph Rewiring and Feature Denoising via Spectral Resonance
Jonas Linkerhägner (University of Basel), Ivan Dokmanić (University of Basel)
ClassificationOptimizationGraph Neural NetworkGraph
🎯 What it does: This paper proposes a joint algorithm for graph structure rearrangement and feature denoising (JDR), which enhances the performance of downstream graph neural networks in node classification tasks by aligning the principal spectral spaces of the graph adjacency matrix and the node feature matrix.
Joint Reward and Policy Learning with Demonstrations and Human Feedback Improves Alignment
Chenliang Li (Texas A&M University), Mingyi Hong (University of Minnesota)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningText
🎯 What it does: A joint learning framework named AIHF is proposed, which simultaneously trains the reward model and policy using demonstration data and preference data to achieve better alignment.
JPEG Inspired Deep Learning
Ahmed H. Salamah (University of Waterloo), EN-HUI YANG
ClassificationCompressionAdversarial AttackImage
🎯 What it does: A new deep learning framework called JPEG-DL is proposed, which adds a trainable JPEG compression layer in front of any base deep neural network (DNN) architecture to enhance deep learning performance.
Judge Decoding: Faster Speculative Sampling Requires Going Beyond Model Alignment
Gregor Bachmann (Meta GenAI), Jonas K Kohler
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Improved the inference acceleration technique speculative decoding by proposing a new validation strategy (Judge Decoding) to increase the acceptance rate of correct tokens generated by draft models, thereby achieving higher inference speed.
JudgeBench: A Benchmark for Evaluating LLM-Based Judges
Sijun Tan (University of California Berkeley), Ion Stoica (University of California Berkeley)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: A hierarchical evaluation framework aimed at LLM judges is proposed, and based on this, the JudgeBench benchmark is created to objectively measure the factual and logical reasoning abilities of LLM judges.
JudgeLM: Fine-tuned Large Language Models are Scalable Judges
Lianghui Zhu (Beijing Academy of Artificial Intelligence), Xinlong Wang (Beijing Academy of Artificial Intelligence)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A scalable language model evaluator, JudgeLM, is proposed for efficiently and accurately assessing the responses of large language models in open-ended tasks.
Jump Your Steps: Optimizing Sampling Schedule of Discrete Diffusion Models
Yong-Hyun Park, Yuki Mitsufuji (Sony Group Corporation)
Data SynthesisOptimizationComputational EfficiencyDiffusion modelImageText
🎯 What it does: A sampling schedule optimization method called Jump Your Steps (JYS) is proposed to improve the sampling quality and speed of the Discrete Diffusion Model (DDM) without incurring additional computational costs.
Justice or Prejudice? Quantifying Biases in LLM-as-a-Judge
Jiayi Ye (University of Notre Dame), Xiangliang Zhang (University of Notre Dame)
TransformerLarge Language ModelText
🎯 What it does: This paper proposes and implements the CALM framework for the automatic quantification of 12 types of biases when using LLMs as judges.
K-HALU: Multiple Answer Korean Hallucination Benchmark for Large Language Models
Jaehyung Seo (Korea University), Heuiseok Lim (Korea University)
Large Language ModelTextBenchmark
🎯 What it does: A large language model hallucination detection benchmark for Korean, K-HALU, is proposed, which includes multiple answer types and temporal consistency checks.
KAA: Kolmogorov-Arnold Attention for Enhancing Attentive Graph Neural Networks
Taoran Fang (Zhejiang University), Yang Yang (Zhejiang University)
Graph Neural NetworkGraph
🎯 What it does: Proposed the Kolmogorov-Arnold Attention (KAA), introducing the Kolmogorov-Arnold Network (KAN) into the scoring function of attention graph neural networks, unifying and enhancing their expressive capability.