ICML 2026 Papers — Page 27
International Conference on Machine Learning · 6554 papers
How Does the Lagrangian Guide Safe Reinforcement Learning through Diffusion Models?
Xiaoyuan Cheng (University College London), Yukun Hu (University College London)
OptimizationRecurrent Neural NetworkTransformerReinforcement LearningDiffusion modelScore-based ModelTabularTime SeriesSequentialStochastic Differential Equation
🎯 What it does: Propose a safety reinforcement learning framework based on Augmented Lagrangian Guided Diffusion (ALGD), which uses diffusion models to generate multi-modal policies in offline and online offline safe RL, and stabilizes training through the Augmented Lagrangian local convexification of the energy landscape.
How Does the Pretraining Distribution Shape In-Context Learning? A Fundamental Trade-Off
Waïss Azizian (University of Grenoble Alpes), Ali Hasan (Morgan Stanley)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningMeta LearningTransformerLarge Language ModelContrastive LearningTextSequentialReview/Survey PaperStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper constructs a theoretical framework to study the impact of pre-training distributions (especially tail behaviors and temporal dependencies) on in-context learning (ICL) in large language models.
How Far Ahead Do LLMs Plan? Uncovering the Latent Horizon in Chain-of-Thought Reasoning
Liyan Xu (WeChat AI, Tencent Inc), Jie Zhou (WeChat AI, Tencent Inc)
Explainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringContrastive LearningTextChain-of-Thought
🎯 What it does: This paper designs a low-rank adapter method called Tele‑Lens to probe the internal hidden states of large language models (LLMs) layer by layer across 12 different types of tasks, revealing their planning horizon during the chain-of-thought (CoT) process. Based on this finding, we propose an adaptive uncertainty estimation that focuses only on key 'pivot' positions and a CoT skipping mechanism that leverages early answer clues.
How Far Can LLM Agents Reason with Tables? Benchmarking Multi-Turn Agentic Table Question Answering in the Wild
Jingwang Huang (Chongqing University), Jian Yang (Beihang University)
Autonomous DrivingOptimizationFederated LearningMeta LearningDrug DiscoveryAI Code AssistantReinforcement Learning from Human FeedbackNeural Architecture SearchProtein Structure PredictionTransformerLarge Language ModelAgentic AIPrompt EngineeringTextTabularBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed TableAgent-Bench, an industrial-level benchmark for multi-turn table-based question answering;
How Few-Shot Examples Add Up: A Causal Decomposition of Function Vectors in In-Context Learning
Entang Wang (Saarland Univesity), Michael Hahn (Saarland Univesity)
Explainability and InterpretabilityRepresentation LearningMeta LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper investigates the formation mechanism of functional vectors (FV) in large language models under few-shot prompting, demonstrating that FV can be approximated as a linear superposition of individual example embeddings, and revealing that the contextualization process enhances task representation through attention reweighting;
How Good is Post-Hoc Watermarking With Language Model Rephrasing?
Pierre Fernandez (FAIR, Meta Superintelligence Labs), Alexandre Mourachko (FAIR, Meta Superintelligence Labs)
GenerationExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: Conduct post-watermarking experiments on large language models by restating the original text to embed detectable watermarks, evaluating the trade-off between text quality and detection strength.
How Hard Can It Be? Hardness-Aware Multi-Objective Unlearning
Jiangwei Chen (National University of Singapore), Bryan Kian Hsiang Low (National University of Singapore)
OptimizationFederated LearningSafty and PrivacyConvolutional Neural NetworkTransformerSupervised Fine-TuningRectified FlowContrastive LearningImageText
🎯 What it does: This paper proposes a hardness-aware multi-objective machine learning algorithm called HAMU without learning, which determines the update direction at each iteration based on the gradient dot product, ensuring both the improvement of forgotten data quality and the minimization of performance degradation in retained data.
How Hard Is Science?
Adil Soubki (University of Cambridge), Miles Cranmer (University of Cambridge)
OptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyRepresentation LearningMeta LearningDrug DiscoveryAI Code AssistantReinforcement Learning from Human FeedbackNeural Architecture SearchProtein Structure PredictionReview/Survey PaperBenchmarkPhysics Related
🎯 What it does: This paper studies the computational difficulty of symbolic regression through parameterized complexity theory, proving that it is FPT when the expression depth or tree size is limited, and W[1]-hard when the number of variables or primitive choices is limited. Furthermore, it provides proofs of ETH lower bounds, approximation difficulty, and the absence of a polynomial kernel.
How High is ‘High’? Rethinking the Roles of Dimensionality in Topological Data Analysis and Manifold Learning
Hannah Sansford (University of Bristol), Patrick Rubin-Delanchy (University of Edinburgh)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningContrastive LearningPoint CloudTime SeriesBiomedical DataBenchmark
🎯 What it does: Studied the impact of effective dimension, correlation rank, and latent intrinsic dimension on topological data analysis and manifold learning in high-dimensional data, clarifying when high-dimensional data is beneficial, and validated the theory on grid cell activity data.
How Language Models Process Negation
Zhejian Zhou (University of Southern California), Jonathan May (University of Southern California)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringAuto EncoderContrastive LearningText
🎯 What it does: This paper systematically analyzes the internal mechanisms of large language models in processing sentences with negation words (such as 'not', 'cannot', etc.) using various interpretability techniques. It finds that models have both mechanisms that correctly understand negation, namely 'construction' and 'inhibition', as well as mechanisms caused by 'shortcut' attention heads that lead to incorrect answers. The paper proposes the Attention Sink ablation method, which effectively eliminates the impact of shortcuts and improves the model's accuracy on negated sentences.
How Many Different Outputs Can a Transformer Generate?
Maxime Meyer (National University of Singapore), Vincent Y. F. Tan (National University of Singapore)
GenerationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: This paper studies the number of different sequences that Transformers can generate, defines reachable sequences, and proves that their maximum length grows linearly with the prompt length. Beyond a threshold, the proportion of reachable sequences decays exponentially, explaining the failure of Transformers on simple tasks such as copying. Through theoretical derivation and experimental validation, upper and lower bounds on the length of reachable sequences are established.
How much can language models memorize?
John Xavier Morris (Cornell University), Saeed Mahloujifar (FAIR at Meta)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Proposed a formal framework to distinguish between accidental memory and generalization, and measured the capacity of GPT-style transformers by using compression rate and Kolmogorov information estimation, finding that each parameter can store approximately 3.6 bits of information.
How Out-of-Distribution Detection Learning Theory Enhances Transformer: Learnability and Reliability
Yijin Zhou (Shanghai Jiao Tong University), Yu Guang Wang
Anomaly DetectionRepresentation LearningTransformerSupervised Fine-TuningContrastive LearningImageText
🎯 What it does: This paper constructs a learnability framework for Transformer in OOD detection based on PAC learning theory, and provides necessary and sufficient conditions as well as error bounds. Subsequently, it designs and implements the GROD algorithm to enhance model reliability.
How Powerful are LLMs in Generating Formal Program Specifications?
Fanpeng Yang (Institute of Software, Chinese Academy of Sciences), Fanjiang Xu (Institute of Software, Chinese Academy of Sciences)
OptimizationExplainability and InterpretabilityComputational EfficiencyAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Propose the COINS evaluation framework, using test cases based on Rocq to verify formal program specifications generated by LLMs.
How Reasoning Evolves from Post-Training Data: An Empirical Study Using Chess
Lucas Dionisopoulos (University of California San Diego), Prithviraj Ammanabrolu (University of California San Diego)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringDiffusion modelTextSequentialRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper studies how to enhance the reasoning ability of language models in the verifiable sequential decision-making task of chess through supervised fine-tuning (SFT) and reinforcement learning (RL).
How RL Unlocks the Aha Moment in Geometric Interleaved Reasoning
Xiangxiang Zhang (ByteDance), Jingxuan Wei (Shenyang Institute of Computing Technology, Chinese Academy of Sciences)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper studies interactive reasoning in geometric problems, proposing an approach that alternates between drawing diagrams and performing logical deductions, and achieving functional alignment between drawing and reasoning through reinforcement learning;
How RLHF Amplifies Sycophancy
Itai Shapira (Harvard University), Ariel D. Procaccia (Harvard University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper reveals through formal analysis how RLHF amplifies the sycophancy of LLMs during alignment, and proposes a method to minimize reward correction to prevent this amplification.
How Should Transformers Encode Numeric Values in Electronic Health Records?
Maria Elkjær Montgomery (University of Copenhagen), Mads Nielsen (University of Copenhagen)
TransformerSupervised Fine-TuningTabularBiomedical DataElectronic Health Records
🎯 What it does: In Transformer-based electronic medical record models, systematically evaluate and compare discrete, continuous, and hybrid numerical encoding strategies, propose a unified experimental framework, and verify their effectiveness using synthetic arithmetic tasks and real clinical prediction tasks.
How to Avoid Debate: Scalable AI Safety via Doubly-Efficient Interactive Proofs
Liyan Chen (Massachusetts Institute of Technology), Zoe Xi (Massachusetts Institute of Technology)
Safty and PrivacyComputational EfficiencyReview/Survey Paper
🎯 What it does: Proposes a scalable AI security interactive proof framework with a single prover, overcoming the limitations of traditional two-prover debate models;
How to Correctly Report LLM-as-a-Judge Evaluations
Chungpa Lee (Yonsei University), Kangwook Lee (University of WisconsinMadison)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose an unbiased estimation framework based on misclassification adjustment, using LLM as a judge to correct results and provide statistical confidence intervals, while designing an adaptive calibration sample allocation strategy to shorten the interval length.
How to Fine-Tune a Reasoning Model? A Teacher–Student Cooperation Framework to Synthesize Student-Consistent SFT Data
Zixian Huang (Shanghai AI Laboratory), Qipeng Guo (Shanghai AI Laboratory)
Data SynthesisComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the TESSY framework based on teacher-student collaboration, alternately generating capability and style texts for supervised fine-tuning of reasoning models.
How to Guide Your Flow: Few-Step Alignment via Flow Map Reward Guidance
Jerry Yuyang Huang (Carnegie Mellon University), Nicholas Matthew Boffi (Carnegie Mellon University)
GenerationOptimizationComputational EfficiencyReinforcement Learning from Human FeedbackTransformerReinforcement LearningFlow-based ModelImageTextMultimodality
🎯 What it does: Propose a flow-map-based deterministic optimal control framework for efficient reward-guided inference in flow/flow-map generation models.
How to Price Data: A Market Equilibrium Based Approach
Pooja Ravi Kulkarni (University of Chicago), Ruta Mehta (University of Illinois)
OptimizationFederated LearningData-Centric LearningTabularBiomedical DataReview/Survey PaperBenchmarkFinance Related
🎯 What it does: This paper proposes a novel data pricing mechanism based on Nash equilibrium, specifically tailored for non-rivalrous data goods (replicable and non-rivalrous), and analyzes the existence of equilibrium, computational efficiency, welfare, and revenue properties under two types of buyer preferences: perfect complementarity (Leontief) and perfect substitutability (linear).
How to Train Your Advisor: Steering Black-Box LLMs with Advisor Models
Parth Asawa (University of California, Berkeley), Joseph E. Gonzalez
OptimizationKnowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper proposes a lightweight open-weight model (Advisor) trained to generate instance-specific natural language suggestions, dynamically guiding and enhancing the performance of black-box large language models (LLMs).
How Transformers Represent Hierarchies: A Local-to-Global Mechanism
Zhiling Zhou (Yale University), Zhuoran Yang (Yale University)
Explainability and InterpretabilityRepresentation LearningTransformerTextSequential
🎯 What it does: This paper trains an autoregressive Transformer on sequences generated by restricted PCFGs and analyzes its hidden layer representations through probing, revealing how the Transformer captures hierarchical structures.
How2Everything: Mining the Web for How-to Procedures to Evaluate and Improve LLMs
Yapei Chang (University of Maryland), Luca Soldaini (Allen Institute for AI)
Knowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes an scalable framework called HOW2EVERYTHING for large-scale mining of real step-by-step 'how-to' procedures from the web, and builds an evaluation benchmark HOW2BENCH and evaluation protocol HOW2SCORE based on these procedures; meanwhile, the evaluation signal is used for RL training to enhance the model's program generation capability.
HPS: Hyperspherical Parameter Sharing for Efficient Multi-Agent Reinforcement Learning
Hu Fu (Huazhong University of Science and Technology), Yihua Tan (Huazhong University of Science and Technology)
Reinforcement LearningBenchmark
🎯 What it does: Propose Hyperspherical Parameter Sharing (HPS) for multi-agent reinforcement learning to address the insufficient behavioral diversity caused by parameter sharing.
HSGG: Training-Free Hierarchical Scene Graph Generation with Geometry-Guided Relation Reasoning
yunzhe Liu, Shichao Kan (Central South University)
RecognitionObject DetectionGenerationData SynthesisGraph Neural NetworkTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Proposes a fully training-free hierarchical scene graph generation framework called HSGG, which first constructs a multi-layer entity tree through bidirectional hierarchical entity perception, and then generates structured and geometrically consistent open-world scene graphs by using geometry-guided contrastive relation reasoning for filtering and decoding.
HSMAD: Heterophily-Driven Spectral and Manifold Learning for Graph Anomaly Detection
Chen Zhu (Tongji University), Yaying Zhang (Tongji University)
Anomaly DetectionGraph Neural NetworkScore-based ModelFlow-based ModelRectified FlowNeural Radiance FieldAuto EncoderContrastive LearningGaussian SplattingGraph
🎯 What it does: This paper proposes a graph anomaly detection framework called HSMAD that simultaneously models heterogeneity in the spectral domain and the manifold domain, enhancing the ability to identify abnormal nodes by utilizing heterogeneity-weighted spectral filtering and manifold routing message updates.
HTAC: Hierarchical Task-Aware Composition for Continual Offline Reinforcement Learning
Qiyang Zhou (Shenzhen Campus of Sun Yat-sen University), Li Shen (Shenzhen Campus of Sun Yat-sen University)
TransformerReinforcement LearningMixture of ExpertsTextBenchmark
🎯 What it does: Propose the Hierarchical Task-Aware Composition (HTAC) framework for continual offline reinforcement learning, addressing issues of heterogeneity in dynamics, rewards, and behavior policies across tasks, as well as distribution drift;
Hugging Carbon: Quantifying the Training Carbon Emissions of AI Models at Scale
Xinlei Wang (Sofia University St. Kliment Ohridski), Jinjin Gu (Sofia University St. Kliment Ohridski)
Computational EfficiencyBenchmark
🎯 What it does: By hierarchically processing metadata from Hugging Face open-source models, we construct a FLOPs-based carbon emission estimation framework to quantify the training carbon emissions of thousands of models.
Human-AI Collaborative Uncertainty Quantification
Sima Noorani (University of Pennsylvania), Hamed Hassani (University of Pennsylvania)
ClassificationAnomaly DetectionExplainability and InterpretabilityData-Centric LearningImageTextTabular
🎯 What it does: Proposed a human-AI collaborative uncertainty quantification framework, which utilizes a two-threshold rule to prune and augment based on a given initial human prediction set, and provides distribution-free calibration algorithms for both offline and online settings;
Human-in-the-Loop Policy Optimization for Preference-Based Multi-Objective Reinforcement Learning
Tianmeng Hu (University of Exeter), Ke Li (University of Exeter)
OptimizationRobotic IntelligenceReinforcement LearningTabularTime SeriesSequential
🎯 What it does: Propose a human-in-the-loop multi-objective reinforcement learning framework called PBMORL, which guides and accelerates multi-policy optimization by actively learning the implicit preferences of the decision-maker;
HumanLM: Simulating Users with State Alignment Beats Response Imitation
Shirley Wu (Stanford), James Zou (Stanford)
Data SynthesisReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: Propose HUMANLM, constructing a user simulator based on a large language model. By first generating potential states consistent with the user's true psychological state and then synthesizing the final response, it generates replies that are more in line with real users.
Hunt Instead of Wait: Evaluating Deep Data Research on Large Language Models
Wei Liu (King's College London), Yulan He (King's College London)
Data-Centric LearningTransformerLarge Language ModelAgentic AITextTabularBiomedical DataElectronic Health RecordsBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes the Deep Data Research (DDR) task and the DDR-Bench benchmark, aiming to evaluate the investigatory intelligence of large language models in scenarios where they are given only structured databases and a set of tools, enabling them to autonomously explore data, generate insights, and decide when to terminate.
HVAE: Hyperbolic Variational Autoencoder For Flexible Knowledge Transfer Across Multiple Domains
Xiaolei Liu (MYbank, Ant Group), Shaoshuai Li (MYbank, Ant Group)
Domain AdaptationRecommendation SystemGraph Neural NetworkFlow-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningGraphTabularFinance Related
🎯 What it does: Propose a hybrid curvature hyperbolic variational autoencoder (HVAE) to achieve knowledge extraction and transfer in cross-domain recommendation
HVR-Met: A Hypothesis-Verification-Replanning Agentic System for Extreme Weather Diagnosis
Shuo Tang (Chinese Academy of Sciences), Cheng-Lin Liu (Chinese Academy of Sciences)
Autonomous DrivingOptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningDrug DiscoveryAI Code AssistantReinforcement Learning from Human FeedbackNeural Architecture SearchTransformerLarge Language ModelAgentic AIPrompt EngineeringVision Language ModelImageTextGraphTabularTime SeriesReview/Survey PaperBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed HVR-Met — a multi-agent extreme weather diagnostic system that utilizes a Hypothesis-Verification-Replanning closed-loop to achieve automated, professional-level extreme weather diagnosis and report generation.
Hybrid Policy Distillation for LLMs
Wenhong Zhu (Shanghai Jiao Tong University), Pengfei Liu (Shanghai Jiao Tong University)
Computational EfficiencyKnowledge DistillationRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningContrastive LearningText
🎯 What it does: Propose a knowledge distillation method called Hybrid Policy Distillation (HPD), which simultaneously leverages the advantages of forward KL (covering modes) and reverse KL (focusing on high-probability modes) by using the teacher model distribution and student self-sampled token-level rewards.
Hybrid Reinforcement Learning in Adversarial Markov Decision Processes
Duo Cheng (Virginia Tech), Bo Ji (Virginia Tech)
Reinforcement LearningContrastive Learning
🎯 What it does: In adversarial Markov decision processes, a hybrid reinforcement learning framework that combines on-policy and off-policy feedback is studied, which can achieve coverage-related convergence under any comparison policy while maintaining worst-case guarantees.
Hybrid-Gym: Training Coding Agents to Generalize Across Tasks
Yiqing Xie (Carnegie Mellon University), Daniel Fried (Carnegie Mellon University)
AI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningAgentic AIPrompt EngineeringTextSequentialBenchmarkRetrieval-Augmented Generation
🎯 What it does: Designed and constructed HYBRID-GYM, a scalable multi-task training dataset to enhance the generalization ability of coding agents across different software engineering tasks.
HybridFlow: Resource-Adaptive Subtask Routing for Efficient Edge-Cloud LLM Inference
Jiangwen Dong (Hong Kong Polytechnic University), Wanyu Lin (Hong Kong Polytechnic University)
Federated LearningComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Developed a resource-adaptive edge-cloud collaborative inference framework called HybridFlow, which utilizes task decomposition into DAGs to enable subtask-level parallel execution and dynamically routes subtasks to the edge or cloud based on real-time budgets.
HybridOM: Hybrid Physics-Based and Data-Driven Global Ocean Modeling with Efficient Regional Downscaling
Ruiqi Shu (Tsinghua University), Xiaomeng Huang (Tsinghua University)
TransformerDiffusion modelFlow-based ModelNeural Radiance FieldAuto EncoderContrastive LearningTabularTime SeriesPhysics Related
🎯 What it does: Developed a HybridOM framework that combines differentiable physical cores with neural network residual corrections, achieving efficient and physically consistent global ocean simulations and predictions.
Hydra-Nav: Object Navigation via Adaptive Dual-Process Reasoning
Zixuan Wang (ByteDance Seed), Yiming Gan (Institute of Computing Technology, Chinese Academy of Sciences)
Autonomous DrivingComputational EfficiencyRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVision-Language-Action ModelDiffusion modelScore-based ModelContrastive LearningImageTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose HYDRA-NAV, a unified vision-language model (VLM) architecture that can adaptively switch between slow inference (global planning) and fast reaction (low-level actions) in navigation tasks.
HyMTRL: A Hybrid Multi-Task Reinforcement Learning Framework via Phased Policy Evolution
Jinmin He (Institute of Automation Chinese Academy of Sciences), Jian Cheng
Meta LearningReinforcement LearningMixture of ExpertsTabularTime SeriesSequentialBenchmark
🎯 What it does: Proposes the HyMTRL framework, which divides multi-task reinforcement learning tasks into an exploration phase and a refinement phase of imitation learning, using data-driven mastery assessment to dynamically transfer tasks, reducing the representational burden on shared critics;
HypCL: Adapting CLIP in Hyperbolic Space for Continual Learning
Quan Cheng (Nanjing University), Lijun Zhang (Nanjing University)
ClassificationRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: Propose the HypCL framework, which freezes the CLIP visual encoder while using a lightweight task-specific adapter and sequentially combines them via Mobius operations in hyperbolic space to achieve continual learning.
Hyper-ICL: Attention Calibration with Hyperbolic Anchor Distillation for Multimodal In-Context Learning
Niloufar Alipour Talemi (Clemson University), Fatemeh Afghah (Clemson University)
Computational EfficiencyKnowledge DistillationRepresentation LearningMeta LearningTransformerPrompt EngineeringVision Language ModelContrastive LearningMultimodality
🎯 What it does: Propose Hyper-ICL, a training-free multi-modal ICL approach that directly corrects the distribution in attention to achieve ICD-free reasoning.
Hyper-LLaVA: Hyperbolic Uncertainty-aware Modality-Balanced Routing for Multimodal Continual Instruction Tuning
Kunlun Xu (Wangxuan Institute of Computer Technology, Peking University), Jiahuan Zhou (Wangxuan Institute of Computer Technology, Peking University)
OptimizationFederated LearningComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: Propose the Hyper-LLaVA framework for multi-modal continual instruction tuning, improving parameter routing
HyPER: Bridging Exploration and Exploitation for Scalable LLM Reasoning with Hypothesis Path Expansion and Reduction
Shengxuan Qiu (Peking University), Meng Li (Peking University)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes HyPER, an untrained, online-controlled multi-path reasoning strategy that dynamically switches between exploration (path expansion) and exploitation (fine-grained rethinking/voting).
Hyperbolic Associative Memory Networks
Boliang Hao (Zhejiang Sci-Tech University), Fangyu Wu (Xi'an JiaotongLiverpool University)
ClassificationRetrievalOptimizationRepresentation LearningContrastive LearningImageTextGraphTabular
🎯 What it does: Propose Hyperbolic Associative Memory Networks (HAMNs), which migrate modern Hopfield networks to negative curvature hyperbolic spaces, achieving hierarchical memory retrieval based on arc-length energy.
Hyperbolic Hierarchical Alignment for Video-Based Visible-Infrared Person Re-Identification
Shuang Li (Chongqing University of Posts and Telecommunications), Xinbo Gao (Chongqing University of Posts and Telecommunications)
RetrievalTransformerVision Language ModelDiffusion modelContrastive LearningImageVideoMultimodality
🎯 What it does: Propose Hyperbolic Hierarchical Alignment (HHA), unifying spatiotemporal modeling and cross-modal alignment of videos on the Poincaré ball for visible-infrared person re-identification.
Hyperbolic Multimodal Continual Learning
Jiahong Liu (Chinese University of Hong Kong), Irwin King (Chinese University of Hong Kong)
ClassificationRetrievalOptimizationRepresentation LearningTransformerContrastive LearningImageTextMultimodality
🎯 What it does: Propose a multi-modal continual learning framework under hyperbolic geometry, and design geometry-preserving parameter update rules to prevent forgetting in multi-task learning.
Hyperbolic Neural Operator
Jieyuan Pei (Zhejiang University of Technology), Jianwei Zheng (Zhejiang University of Technology)
OptimizationComputational EfficiencyTransformerDiffusion modelContrastive LearningPoint CloudMeshGraphTime SeriesBenchmarkPhysics Related
🎯 What it does: Proposed the Hyperbolic Neural Operator (HNO), which constructs a Gibbs kernel based on geodesic distance on the Lorentz hyper-sphere to replace the traditional Euclidean dot product attention, learning a solver for parameterized PDEs, while considering multi-scale coupling and irregular grids.
Hyperbolic neural population geometry benefits computation
Dennis Wu (Northwestern University), Han Liu (Northwestern University)
ClassificationOptimizationRepresentation LearningSupervised Fine-TuningContrastive LearningImageMultimodalityTabularTime SeriesReview/Survey PaperBenchmarkPhysics Related
🎯 What it does: Proposed a theoretical framework that proves that when the width of Gaussian tuning curves of hippocampal cells follows an exponential distribution, it induces statistically hyperbolic geometry in neural population activities; established a connection between Bayes optimal decoding and the update rules of modern Hopfield networks (MHN), and defined a new associative memory model (Karcher-flow) on hyperbolic space, theoretically proving that its memory capacity can achieve doubly exponential growth.
Hyperbolic RQ-VAE enhanced Generative Recommendation with Differential-Length Codebook Strategy
Aoran Zhang (Nanjing University), Yonghong Yu (Nanjing University)
Recommendation SystemTransformerLarge Language ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningSequential
🎯 What it does: This paper proposes HG-Rec, a generative recommendation framework based on residual quantization and differential length codebook in hyperbolic space, aiming to improve codebook utilization and recommendation performance.
HyperMLP: An Integrated Perspective for Sequence Modeling
Jiecheng Lu (Georgia Institute of Technology), Shihao Yang (Georgia Institute of Technology)
Computational EfficiencyRepresentation LearningData-Centric LearningTransformerLarge Language ModelMixture of ExpertsTextSequentialBenchmark
🎯 What it does: Reinterpret autoregressive attention as a dynamic two-layer MLP and propose HyperMLP/HyperGLU by learning low-order sequence mixing and inverse offset layouts to enhance model expressiveness.
Hyperparameter Transfer Laws for Non-Recurrent Multi-Path Neural Networks
Haosong Zhang (Fudan University), Wei Lin (Fudan University)
OptimizationHyperparameter SearchConvolutional Neural NetworkRecurrent Neural NetworkTransformerSupervised Fine-TuningContrastive LearningImage
🎯 What it does: Proposed and verified a unified deep learning rate scaling rule applicable to non-recursive multi-path networks (CNN, ResNet, Transformer), and derived the -3/2 power law relationship between learning rate and depth through the concept of effective depth and the arithmetic mean maximum update (AMµP) constraint.
Hyperparameter Transfer with Mixture-of-Expert Layers
Tianze Jiang (Princeton University), Boris Hanin (Princeton University)
OptimizationComputational EfficiencyHyperparameter SearchTransformerLarge Language ModelMixture of ExpertsContrastive LearningText
🎯 What it does: A set of hyperparameter transfer parameterization applicable to Mixture-of-Experts (MoE) layers is proposed, and the parameterization is theoretically proven to maintain consistent training dynamics across different model scales using Dynamic Mean-Field Theory (DMFT). Subsequently, experiments are conducted on models with various widths, depths, numbers of experts, and expert sizes.
HyperPotter: Spell the Charm of High-Order Interactions in Audio Deepfake Detection
Qing Wen (Zhejiang University), Kui Ren (Zhejiang University)
Anomaly DetectionGraph Neural NetworkTransformerSupervised Fine-TuningContrastive LearningAudio
🎯 What it does: Propose the HyperPotter framework, which models high-order interactions using hypergraphs to enhance cross-scenario generalization in audio deepfake detection.
Hyperspectral Image Fusion with Spectral-Band and Fusion-Scale Agnosticism
Yujie Liang, Malu Zhang (University of Electronic Science and Technology of China)
RestorationSuper ResolutionConvolutional Neural NetworkTransformerDiffusion modelAuto EncoderContrastive LearningImage
🎯 What it does: Propose a unified MS/HS fusion framework named SSA, which can simultaneously accommodate the number of spectral bands from different sensors and arbitrary spatial magnification scales, achieving multi-sensor joint training with a single model.
HyPOLE: Hyperproperty-Guided Multi-Agent Reinforcement Learning under Partial Observation
Arshia Rafieioskouei (Michigan State University), Borzoo Bonakdarpour (Michigan State University)
Reinforcement LearningAuto Encoder
🎯 What it does: Propose the HYPOLE framework, which leverages HyperLTL hyper-properties to guide learning in partially observable multi-agent reinforcement learning, combining Skolemization, robustness measurement, and CTDE for policy synthesis;
HypoSpace: A Diagnostic Benchmark for Set-Valued Hypothesis Generation under Underdetermination and Sublinear Coverage Bounds
Tingting Chen (National University of Singapore), Dianbo Liu (National University of Singapore)
TransformerLarge Language ModelPrompt EngineeringTextGraphTabularBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the HypoSpace benchmark to evaluate the generation capabilities of large language models in multiple feasible hypothesis spaces.
HypRAG: Hyperbolic Dense Retrieval for Retrieval Augmented Generation
Hiren Madhu (Yale University), Rex Ying (Yale University)
RetrievalRepresentation LearningTransformerContrastive LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: In the retrieval-augmented generation (RAG) task, the authors propose a framework that utilizes hyperbolic space for dense retrieval, introducing two models: HyTE-FH (Fully Hyperbolic Transformer) and HyTE-H (Hybrid architecture, projecting pre-trained Euclidean embeddings into hyperbolic space).
IACW: Intent-Aware Controllable Watermarking for Scalable Authorial Intent Attribution
Hao Huang (Chinese Academy of Sciences), Yuling Liu (Chinese Academy of Sciences)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes an intent-aware controllable watermark (IACW) for large language models (LLMs), which can embed the author's 'edit' or 'generate' intent into a recoverable watermark signal during text generation, and achieve semantically lossless and robust intent attribution through entropy-gated injection and confidence-based erasure decoding.
IAPO: Information-Aware Policy Optimization for Token-Efficient Reasoning
Yinhan He (University of Virginia), Jundong Li (University of Virginia)
OptimizationComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: To address the issue of large language models generating lengthy reasoning chains during inference, this paper proposes the IAPO framework. During the RL fine-tuning phase, the framework explicitly guides the model to generate information-rich and redundant-free reasoning steps by connecting the advantage of each token with its conditional mutual information (MI) with respect to the final answer, and combines an exploration adjustment term to maintain model accuracy.
IBMA: Information Bottleneck-Based Multimodal Alignment
Yancheng Wang (Arizona State University), Yingzhen Yang (Arizona State University)
ClassificationComputational EfficiencyRepresentation LearningData-Centric LearningTransformerAuto EncoderContrastive LearningTextMultimodalityBiomedical DataBenchmarkAudio
🎯 What it does: Propose the Information Bottleneck based Multimodal Alignment framework (IBMA), which applies information bottleneck constraints on both the fused and unimodal representations to achieve cross-modal alignment.
ICR-RL: Deep Reinforcement Learning via In-Context-Regression
David Schiff (Bar Ilan University), Yonathan Efroni (Tel Aviv University)
TransformerReinforcement LearningPrompt EngineeringTabular
🎯 What it does: Propose a gradient-agnostic deep reinforcement learning framework called ICR-RL based on In-Context Regression (ICR), which directly uses the pre-trained TabPFN regression model to approximate the Q-value function;
Ideal Attribution and Faithful Watermarks for Language Models
Min Jae Song (University of Chicago), Kameron Shahabi (University of Washington)
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Proposes an abstract model of the ideal attribution mechanism, viewing it as the ideal goal for watermarking schemes;
Identifiable Equivariant Networks are Layerwise Equivariant
Vahid Shahverdi (Kth Royal Institute Of Technology), Kathlén Kohn (Kth Royal Institute Of Technology)
ClassificationConvolutional Neural NetworkTransformerSupervised Fine-TuningContrastive LearningImage
🎯 What it does: This paper proves that in parameter-identifiable deep networks, end-to-end equivariance necessarily leads to hierarchical equivariance;
Identifiable Markov Switching Models with Instantaneous Effects and Exponential Families
Roel Hulsman (University of Amsterdam), Sara Magliacane (University of Amsterdam)
Flow-based ModelTabularTime SeriesFinance Related
🎯 What it does: Constructing identifiable Markov switching models in non-stationary time series, allowing for instantaneous effects, nonlinear lagged effects, and restricting noise to the exponential family, while proposing a switching detection and causal structure discovery framework based on conditional regularized flows (FlowMSM).
Identifiable Nonlinear Differentiable Causal Discovery via Independence and Adaptive Group Sparsity
Ruicong Yao (Ghent University), Jakob Raymaekers (University of Antwerp)
OptimizationExplainability and InterpretabilityComputational EfficiencyRepresentation LearningGraph Neural NetworkDiffusion modelScore-based ModelFlow-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningGraphTabular
🎯 What it does: Propose a differentiable causal structure learning framework called NOTIME‑CAM, which utilizes the independence measure of residuals and adaptive group Lasso regularization to achieve identifiable DAG learning under arbitrary noise distributions for the nonlinear additive noise model (ANM).
Identifiable Smooth Conjugacy Learning via Adversarial Orthogonality
In Huh (Purdue University), Muhammad Alam
OptimizationExplainability and InterpretabilityComputational EfficiencyRepresentation LearningFlow-based ModelGenerative Adversarial NetworkTabularTime SeriesSequentialPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes a identifiable smooth conjugate learning framework that can be applied when the topologies of the source and target systems are mismatched. It utilizes low-rank context modulation to make controllable corrections to the source dynamics, and enforces orthogonality between the correction space and the orbit tangent space through adversarial regularization, thereby achieving a minimal and identifiable conjugate recovery of the target dynamics;
Identifiable Token Correspondence for World Models
Youngin Kim (Seoul National University), Hyun Oh Song (Seoul National University)
TransformerReinforcement LearningPrompt EngineeringWorld ModelImageVideoBenchmark
🎯 What it does: Propose a method of inserting an identifiable token correspondence (ITC) decoding step into a token-based transformer world model, utilizing Optimal Transport to compute token correspondence relationships, copying tokens from the previous frame or generating the next frame tokens from the transformer, significantly improving object consistency and overall performance in long-term prediction.
Identifying and Correcting Label Noise for Robust GNNs via Influence Contradiction
Wei Ju (Sichuan University), Ming Zhang (Peking University)
ClassificationAnomaly DetectionExplainability and InterpretabilityComputational EfficiencyRepresentation LearningGraph Neural NetworkContrastive LearningGraphBenchmark
🎯 What it does: Propose an influence contradiction score-based graph neural network (ICGNN) for identifying and correcting label noise in graph data, and enhancing robustness in semi-supervised scenarios.
Identifying and Mitigating Errors in Gradient Aggregation of Distributed Data Parallel Training
Zhenheng Tang (Hong Kong University of Science and Technology), Bo Li (Hong Kong University of Science and Technology)
OptimizationFederated LearningComputational EfficiencyTransformerImageText
🎯 What it does: This paper addresses silent data corruption (SDC) errors that occur during gradient aggregation in distributed data parallel training. It theoretically derives gradient inconsistency and analyzes its impact on model divergence. Subsequently, the paper proposes the PAFT system, which includes two major modules: periodic synchronization (PAFT-Sync) and dynamic adaptive synchronization (PAFT-Dyn), and implements asynchronous communication overlap and error measurement adaptive scheduling.
Identifying Common Hubs in Multiple Gaussian Graphical Models
José Á Sánchez Gómez (University of California Riverside), Yufeng Liu (University of Michigan)
OptimizationFederated LearningComputational EfficiencyRepresentation LearningDrug DiscoveryGraph Neural NetworkContrastive LearningGaussian SplattingGraphBiomedical DataReview/Survey Paper
🎯 What it does: This paper proposes a method called JIC-HD for identifying common hub nodes from multiple Gaussian graphical models.
Identifying Connectivity Distributions from Neural Dynamics Using Flows
Timothy Doyeon Kim (Allen Institute), Uygar Sümbül
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningNeural Architecture SearchRecurrent Neural NetworkFlow-based ModelContrastive LearningTime SeriesBiomedical DataStochastic Differential Equation
🎯 What it does: By constructing a framework called Connector, which combines low-rank recurrent neural networks (lrRNN) and continuous normalizing flows (CNF) to infer variable connectivity distributions that match neural dynamics, the effectiveness is evaluated on multiple synthetic and real neural datasets.
Identifying dependent components from multi-domain linear mixtures
Danru Xu (University of Amsterdam), Aapo Hyvarinen (University of Helsinki)
OptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyRepresentation LearningData-Centric LearningAuto EncoderContrastive LearningBiomedical DataDiffusion Tensor ImagingReview/Survey Paper
🎯 What it does: Propose a multi-domain linear mixture model, where different domains are distinguished only by scale changes, and allow arbitrary dependencies between latent variables, proving that under sufficient domain variations, latent variables and mixture matrices can be uniquely identified.
Identifying Latent Concepts and Structures for Generalized Category Discovery
Boyang Dai (University of Hong Kong), Yizhou Yu (University of Hong Kong)
ClassificationRecognitionRepresentation LearningTransformerAuto EncoderContrastive LearningImage
🎯 What it does: Propose a module called CPF that rewrites visual features through low-rank learnable primitives to enhance the performance of generalized class discovery (GCD).
Identifying Learnwares via Reduced Neural Conditional Mean Embedding
Zi-Yu Mao (Nanjing University), Ming Li (Nanjing University)
Federated LearningSafty and PrivacyComputational EfficiencyRepresentation LearningAuto EncoderContrastive LearningImageTabular
🎯 What it does: Proposed the RNCME (Reduced Neural Conditional Mean Embedding) standard generation method, which directly models the input-output distribution using conditional mean embedding, combining learnable neural feature mapping and low-rank approximation to achieve privacy-friendly learnware specifications;
Identifying Partially Observed Causal Models from Heterogeneous/Nonstationary Data
Xinshuai Dong (Carnegie Mellon University), Kun Zhang (Carnegie Mellon University)
Domain AdaptationAnomaly DetectionGraphTabularTime SeriesBenchmark
🎯 What it does: This paper proposes the LCD-NOD method, which identifies structure and change points in linear causal models with potential variables and non-stationary/heterogeneous data.
IdEst: Assessing Self-Supervised Learning Representations via Intrinsic Dimension
Julie Mordacq (Inria), Steve Oudot (Inria)
Representation LearningHyperparameter SearchData-Centric LearningTransformerSupervised Fine-TuningAuto EncoderContrastive LearningImage
🎯 What it does: Propose IDEST, an unsupervised evaluation method that estimates the intrinsic dimension (ID) of self-supervised learning (SSL) representations using a minimum spanning tree (MST), aiming to replace traditional linear probes;
IDLM: Inverse-distilled Diffusion Language Models
David Li (Mohamed Bin Zayed University of AI), Alexander Korotin (Applied AI Institute)
GenerationComputational EfficiencyKnowledge DistillationTransformerDiffusion modelScore-based ModelText
🎯 What it does: Propose the inverse distillation method IDLM in discrete diffusion language models, significantly accelerating sampling
IDRBench: Understanding the Capability of Large Language Models on Interdisciplinary Research
Yuanhao Shen (Queen's University), Xiaodan Zhu (Queen's University)
ClassificationRecognitionRetrievalRecommendation SystemTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Construct the IDRBench benchmark framework to evaluate the ability of large language models in interdisciplinary research (IDR) to identify papers, integrate viewpoints, and recommend relevant works.
IEC: When Information-Driven Exploration Meets Spectral Consensus via Primal–Dual Reward Regularization in Decentralized Multi-Agent RL
Xuefeng Du (University of Dalian Jiaotong), Fengqi Li (University of Dalian Jiaotong)
Reinforcement Learning
🎯 What it does: Propose a framework named IEC, aiming to dynamically balance exploration and coordination in decentralized multi-agent reinforcement learning by combining information-driven exploration with spectral consistency constraints.
iGRPO: Fast Online RL for Flow Matching Model with Instant Reward
Sucheng Ren (Johns Hopkins University), Jiasen Lu (Apple)
GenerationOptimizationTransformerReinforcement LearningScore-based ModelFlow-based ModelImageMultimodalityStochastic Differential Equation
🎯 What it does: Propose an online reinforcement learning framework called iGRPO based on immediate rewards, which replaces complete denoising trajectories with single-step mappings, significantly accelerating the alignment training of flow matching models;
iLoRA: Bayesian Low-Rank Adaptation with Latent Interaction Graphs for Microbiome Diagnosis
Yang Song (University of Copenhagen), Hengguan Huang (University of Copenhagen)
ClassificationExplainability and InterpretabilityDrug DiscoveryGraph Neural NetworkTransformerContrastive LearningGraphTabularBiomedical Data
🎯 What it does: Propose a Bayesian Graph Conditional LoRA framework (iLoRA), which simultaneously learns a prediction model and a sample-level microbial interaction network in the microbiome diagnosis task, and generates LoRA updates conditioned on this network.
Image Restoration via Diffusion Models with Dynamic Resolution
Yang Zheng (University of Electronic Science and Technology of China), Zhaoqiang Liu (University of Electronic Science and Technology of China)
RestorationSuper ResolutionSupervised Fine-TuningDiffusion modelScore-based ModelImage
🎯 What it does: Modeling and inference for general image restoration tasks using dynamic resolution diffusion models (dynamic resolution DMs), with fine-tuning on pre-trained pixel space diffusion models, leading to the development of SubDPS, SubDAPS, and the improved version SubDAPS++.
Image-to-Brain Signal Generation for Visual Prosthesis with CLIP Guided Multimodal Diffusion Models
Ganxi Xu (Jinan University), Jinyi Long (Jinan University)
GenerationData SynthesisTransformerVision Language ModelDiffusion modelContrastive LearningMultimodalityBiomedical Data
🎯 What it does: Built an image-to-M/EEG brain signal generation framework for the brain encoding stage of visual prosthetics.
Imagination Helps Visual Reasoning, But Not Yet in Latent Space
You Li (Beijing Jiaotong University), Maosong Sun (Tsinghua University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: This paper systematically analyzes latent visual reasoning (LVR) in multimodal large language models, and through causal mediation analysis, it finds that there is almost no causal association between latent tokens and inputs or answers; subsequently, it proposes CapImagine, a method based on the text space, which directly transforms intermediate visual operations into text descriptions, allowing the model to complete visual imagination through pure text reasoning;
ImgCoT: Compressing Long Chain of Thought into Compact Visual Tokens for Efficient Reasoning of Large Language Model
Xiaoshu Chen (National University of Defense Technology), Xinwang Liu (Nanjing University)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringAuto EncoderImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose the ImgCoT framework, which compresses long-chain thinking (CoT) into visual latent tokens, allowing LLMs to complete reasoning without generating complete text chains.
Imitation Learning for Multi-turn LM Agents via On-policy Expert Corrections
Niklas Lauffer (University of California Berkeley), Jeff Da (Inflection AI)
Data-Centric LearningAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIDiffusion modelContrastive LearningTextSequential
🎯 What it does: Propose a new data generation method called On-Policy Expert Corrections (OEC), used to combat covariate shift in multi-round language model agent training.
ImmersePro: End-to-End Stereo Video Synthesis Via Implicit Disparity Learning
Jian Shi (KAUST), Peter Wonka (KAUST)
GenerationData SynthesisConvolutional Neural NetworkTransformerDiffusion modelAuto EncoderGenerative Adversarial NetworkOptical FlowVideo
🎯 What it does: Developed an end-to-end video stereo conversion framework called ImmersePro, which can generate stereo views from monocular videos.
Immuno-VLM: Immunizing Large Vision-Language Models via Generative Semantic Antibodies for Open-World Trustworthiness
Xiang Fang (Huazhong University of Science and Technology), Wei Ji (Nanjing University)
ClassificationRecognitionAnomaly DetectionAdversarial AttackTransformerLarge Language ModelVision Language ModelGenerative Adversarial NetworkContrastive LearningImageTextMultimodality
🎯 What it does: By proactively generating 'semantic antibodies' in the latent space of visual-language models to counter unknown samples, enhancing the reliability of open-world recognition.
Impact of Connectivity on Laplacian Representations in Reinforcement Learning
Tommaso Giorgi (Baker Hughes), Matteo Papini (Universit` a degli Studi di Milano)
Graph Neural NetworkReinforcement LearningContrastive LearningGraph
🎯 What it does: This paper analyzes the approximation error of state representation based on the graph Laplacian spectrum in reinforcement learning, and proves that the error varies with the algebraic connectivity (λ₂) of the state graph.
IMPACT: Influence Modeling for Open-Set Time Series Anomaly Detection
Xiaohui Zhou (National Key Laboratory of Parallel and Distributed Computing), Guansong Pang (Singapore Management University)
Anomaly DetectionRecurrent Neural NetworkTransformerAuto EncoderGenerative Adversarial NetworkContrastive LearningTime Series
🎯 What it does: For open-ended time series anomaly detection, the IMPACT framework is proposed, which evaluates the impact of training samples using influence functions, performs anomaly decontamination and pseudo-anomaly generation, and trains a dual-head model to achieve joint detection of known and unknown anomalies.
Implicit Action Chunking for Smooth Continuous Control
Bosun Liang (University of Hong Kong), Yong Wang (University of Hong Kong)
Autonomous DrivingRobotic IntelligenceRecurrent Neural NetworkTransformerReinforcement LearningDiffusion modelScore-based ModelContrastive LearningTime SeriesSequentialBenchmarkStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes the Dual-Window Smoothing (DWS) framework, which achieves implicit action chunking through the execution window and value window, thereby improving the smoothness of continuous control while maintaining the standard action dimension.
Implicit Actor Critic Coupling via a Supervised Learning Framework for RLVR
Jiaming Li (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences), Min Yang (Shenzhen University of Advanced Technology)
OptimizationAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: Propose the PACS framework, which transforms the RLVR problem into supervised learning by training LLM policies through predicting reward labels
Implicit Intelligence - Evaluating Agents on What Users Don’t Say
Ved Sirdeshmukh (Labelbox), Marc Wetter (Labelbox)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringWorld ModelTextBenchmark
🎯 What it does: Propose the Implicit Intelligence evaluation framework to assess whether AI agents can identify and satisfy users' implicit needs, accompanied by the Agent-as-a-World (AaW) mechanism, which uses LLMs to simulate an interactive world for large-scale evaluation.
Implicit Preference Alignment for Human Image Animation
Yuanzhi Wang (Beijing Normal University), Zhen Cui (Beijing Normal University)
GenerationPose EstimationTransformerSupervised Fine-TuningDiffusion modelFlow-based ModelImageVideo
🎯 What it does: Propose a post-training framework named Implicit Preference Alignment (IPA), aimed at enhancing the high fidelity of hand movements in portrait animation, eliminating the need for strictly labeled good and bad contrast samples;
Implicit Safety Alignment from Crowd Preferences
Qian Lin (University of Utah), Daniel S. Brown (University of Utah)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerReinforcement LearningAuto EncoderContrastive LearningTextSequential
🎯 What it does: By learning implicit safety rules from crowd preferences, constructing transferable low-level skills, and using hierarchical policies to combine these skills in downstream reinforcement learning tasks, achieving safe behaviors;