ICML 2026 Papers — Page 26
International Conference on Machine Learning · 6554 papers
HaloProbe: Bayesian Detection and Mitigation of Object Hallucinations in Vision-Language Models
Reihaneh Zohrabi (Technical University of Darmstadt & hessian.AI), Marcus Rohrbach (Technical University of Darmstadt & hessian.AI)
Anomaly DetectionExplainability and InterpretabilityTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose the HaloProbe framework, which uses a Bayesian method to jointly utilize internal attention and external description features to detect word-level hallucinations generated by vision-language models.
Hamiltonian Asymmetric Fusion: One-Way Safe Directed Refinement under Modality Imbalance
Bingbing Chen (ShanghaiTech University), Zhuo-Xu Cui (Shenzhen Institutes of Advanced Technology)
Image HarmonizationSegmentationOptimizationTransformerContrastive LearningImageMultimodalityBenchmarkStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Designed a role-asymmetric Hamiltonian Asymmetric Fusion (HAF) module to achieve a secure multi-step iterative fusion under modal imbalance.
Hard Labels In! Rethinking the Role of Hard Labels in Mitigating Local Semantic Drift
Jiacheng Cui (MBZUAI), Zhiqiang Shen (MBZUAI)
ClassificationKnowledge DistillationContrastive LearningImage
🎯 What it does: Reintroduce hard labels as correction signals in dataset distillation to alleviate the problem of local semantic drift caused by limited soft labels.
Hard labels sampled from sparse targets mislead rotation invariant algorithms
Avrajit Ghosh (University of California, Berkeley), Peter Bartlett (Google Deepmind)
ClassificationOptimizationRepresentation LearningContrastive LearningTabularStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper studies the logic regression trained with hard labels under sparse targets for rotation-invariant algorithms, and finds that it cannot reach the Bayes risk, providing a strict lower bound; subsequently, it proposes a non-rotation-invariant algorithm achieved through reparameterization (spindly network), which can achieve the optimal sparse learning rate under over-determined constraints.
Hard-Constrained Graph Generation with Discrete-Projection Diffusion
Xuesong Zhang (Beijing University of Posts and Telecommunications), Jingyu Wang (Beijing University of Posts and Telecommunications)
GenerationData SynthesisGraph Neural NetworkTransformerDiffusion modelScore-based ModelGraph
🎯 What it does: Proposes an NSPSG framework that combines unconstrained diffusion models with discrete projection to enforce hard constraints in graph generation.
HARD-KV: Head-Adaptive Regularization for Decoding-time KV Compression
Yuxuan Yang (Zhejiang University), Huan Li (Zhejiang University)
CompressionComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Proposes the HARD-KV framework, unifying head-adaptive KV cache compression, and achieving compatibility between dynamic selection and static inference engines during decoding through three-level cache, logarithmic calibration, and index regularization.
Harmful Overfitting in Sobolev Spaces
Kedar Karhadkar (University of California, Los Angeles), Guido Montufar (University of California, Los Angeles)
OptimizationExplainability and InterpretabilityComputational EfficiencyRepresentation LearningDiffusion modelScore-based ModelFlow-based ModelRectified FlowContrastive LearningReview/Survey Paper
🎯 What it does: The generalization performance of approximate minimum norm interpolators was studied when they achieve full fitting on training sets with label noise in the Sobolev space W_{k,p}(R^d), proving that even as the sample size tends to infinity, such models still exhibit constant-level overfitting;
Harmonized Dual Policy Improvement for Modelic Reinforcement Learning
Guojian Zhan (Tsinghua University), Shengbo Eben Li (Tsinghua University)
Reinforcement LearningWorld ModelTabularTime Series
🎯 What it does: This paper proposes and verifies Harmonized Dual Policy Improvement (HDPI), which resolves training instability caused by gradient inconsistency by harmonizing the Q-maximization gradient and the planner gradient at the gradient level in model-based reinforcement learning.
Harnessing Non-Adversarial Robustness in Large Language Models
Qinghua Zhou (International Joint Laboratory of AI for Industry), Ivan Y Tyukin (Applied AI Institute)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Improve robustness against semantic neutral random prompt perturbations by bias correction on the outputs of internal modules of LLMs.
Harnessing Reasoning Trajectories for Hallucination Detection via Answer-agreement Representation Shaping
Jianxiong Zhang (Nanyang Technological University), Xuefeng Du (Nanyang Technological University)
Anomaly DetectionExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringDiffusion modelContrastive LearningText
🎯 What it does: This paper proposes the Answer-Agreement Representation Shaping (ARS) method, which utilizes the inference trajectory of large models (LRMs) to enhance hallucination detection capabilities. The method generates counterfactual answers by injecting small perturbations into the hidden states at the end of the inference trajectory, and then performs positive and negative alignment on the generated answer representations based on whether the answers are consistent, thereby learning a lightweight mapping that transforms the original hidden states into more detectable embeddings.
Harnessing Spectrum Video for Subject-Level Few-Shot and Cross-Montage EEG Generalization
Wei Wang (Shanghai Innovation Institute), Yanwei Fu (Shanghai Innovation Institute)
ClassificationDomain AdaptationFederated LearningComputational EfficiencyRepresentation LearningTransformerSupervised Fine-TuningAuto EncoderContrastive LearningVideoTime SeriesBiomedical DataElectrocardiogramReview/Survey Paper
🎯 What it does: This paper transforms electroencephalogram (EEG) signals into structured spectrum videos (Spectrum Video) through time-frequency transformation and spatial mapping, and utilizes video MAE for self-supervised pre-training to achieve generalization across subjects and electrode layouts.
Harnessing Uncertainty: Entropy-Modulated Policy Gradients for Long-Horizon LLM Agents
Jiawei Wang (ByteDance), WANG KE
Autonomous DrivingOptimizationRobotic IntelligenceTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringTextSequentialRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes a framework called Entropy-Modulated Policy Gradients (EMPG), aimed at addressing the credit assignment challenges caused by sparse rewards in large language models (LLMs) during long-horizon tasks.
HASTE: Hardware-Aware Dynamic Sparse Training for Large Output Spaces
Nasib Ullah (Aalto University), Rohit Babbar (University of Bath)
ClassificationTextTabular
🎯 What it does: The HASTE framework is proposed for extreme multi-label classification, achieving efficient sparse training by grouping labels and sharing fixed fan-in connections, eliminating the need for auxiliary objectives.
HDFlow: Hierarchical Diffusion-Flow Planning for Long-horizon Tasks
Gireesh Nandiraju (Peking University), He Wang (Peking University)
Robotic IntelligenceTransformerReinforcement LearningDiffusion modelFlow-based ModelRectified FlowContrastive LearningWorld ModelImageVideoTabularSequentialBenchmark
🎯 What it does: Propose a new hierarchical planning framework called HDFlow, which uses diffusion models to generate high-level subgoal sequences, and then employs a regularized flow model to quickly synthesize low-level smooth trajectories, achieving long-horizon robot task planning;
HDTree: Generative Modeling of Cellular Hierarchies for Robust Lineage Inference
Zelin Zang (Hong Kong Institute of Science and Innovation), Stan Z. Li (Westlake University)
GenerationData SynthesisRepresentation LearningTransformerDiffusion modelAuto EncoderContrastive LearningImageTextBiomedical Data
🎯 What it does: Proposed the HDTree generative model, which uses a unified hierarchical vector quantization codebook and quantized diffusion process to model, reconstruct, and generate single-cell differentiation trees, thus achieving cell lineage inference and generative validation.
HE-SNR: Uncovering Latent Logic via Entropy for Guiding Mid-Training on SWE-bench
Yueyang Wang (Peking University), Xiaoqing Liu (Meituan)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsTextSequentialBenchmark
🎯 What it does: Propose the HE-SNR metric and data filtering strategy for evaluating the potential reasoning ability of LLMs on SWE-Bench tasks during the mid-training phase, and explain the significance of high-entropy decision points through the Entropy Compression Hypothesis.
Head-in-Head in Linear Attention
Shijie Mei (Institute of Automation Chinese Academy of Sciences), Guoqi Li (Institute of Automation Chinese Academy of Sciences)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelContrastive LearningText
🎯 What it does: This paper proposes the Head-in-Head structure, which partitions memory states within a single linear attention head, thereby enhancing the expressive power of linear attention models while maintaining a fixed memory size.
Hearing Without Noticing? Attention-Aware Stealthy Black-Box Adversarial Audio Attacks
Tianyi Xu (Chinese Academy of Sciences), Kai Chen (Chinese Academy of Sciences)
Adversarial AttackDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningAudio
🎯 What it does: Propose a black-box ASR attention-aware stealthy attack method called HWN, which generates adversarial audio that is 'inaudible but recognizable' in the spectral domain through music carrier selection and attention dilution loss.
HEARTS: Benchmarking LLM Reasoning on Health Time Series
Sirui Li (Ucla), Yuzhe Yang (Ucla)
TransformerLarge Language ModelPrompt EngineeringMultimodalityTime SeriesBiomedical DataElectronic Health RecordsBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed the HEARTS benchmark to evaluate the reasoning capabilities of large language models (LLMs) on health time series, covering 12 health domains, 20 signal modalities, 110 tasks, and constructing a diverse evaluation through four hierarchical levels (perception, reasoning, generation, deduction);
Heavy-tailed Physics-Informed Neural Networks
Jephte Abijuru (RPTU University Kaiserslautern-Landau), Sophie Fellenz (RPTU University Kaiserslautern-Landau)
OptimizationComputational EfficiencyConvolutional Neural NetworkRecurrent Neural NetworkTransformerContrastive LearningTabularTime SeriesSequentialBenchmarkPhysics RelatedStochastic Differential Equation
🎯 What it does: The study treats PINN training as likelihood inference of residuals, and proposes using a Student-t residual model with EM training method to reduce the impact of extreme residuals on gradients, thereby improving the accuracy and robustness of solving nonlinear and scale-diverse PDEs.
HECTOR: Hybrid Editable Compositional Object References for Video Generation
Guofeng Zhang (ByteDance), Chongyang Ma
GenerationData SynthesisTransformerPrompt EngineeringDiffusion modelFlow-based ModelAuto EncoderGaussian SplattingOptical FlowImageVideo
🎯 What it does: This paper proposes the HECTOR framework, which enables editable and compositional generation of multiple objects in videos, supporting both static and dynamic image references, and precisely specifies the spatial position, scale, and velocity of each object through trajectory control.
Hedging on the Frontier: Learning New Tasks with Few Samples
Tobias Wegel (ETH Zurich), Fanny Yang (ETH Zurich)
Meta LearningTextMultimodalityBenchmark
🎯 What it does: Studied new task learning under the assumption of weak monotonicity, and proposed a model selection and transfer method based on the Pareto frontier.
HEDP: A Hybrid Energy-Distance Prompt-based Framework for Domain Incremental Learning
Yu Feng (China Mobile Research Institute), Yifan Zhu (Beijing University of Posts and Telecommunications)
ClassificationDomain AdaptationRepresentation LearningTransformerPrompt EngineeringContrastive LearningImageBenchmark
🎯 What it does: Proposed the HEDP framework, achieving domain incremental learning without forgetting and generalization through energy regularization and energy-distance weighted reasoning.
HelioX: A GPU-Native Framework for Simulation and Training of Biophysically Detailed Networks
JUNFENG LU, Tiejun Huang (Peking University)
Computational EfficiencyDrug DiscoveryNeural Architecture SearchProtein Structure PredictionSpiking Neural NetworkAuto EncoderContrastive LearningBiomedical DataBenchmarkStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Propose the HelioX framework, achieving high-performance simulation on GPU and scalable training of neurophysiological detail neural networks
HELIX: Hybrid Encoding with Learnable Identity and Cross-dimensional Synthesis for Time Series Imputation
Fengming Zhang (Beijing Institute of Technology), Shen Qu (Beijing Institute of Technology)
RestorationAnomaly DetectionRepresentation LearningTransformerMixture of ExpertsContrastive LearningTabularTime SeriesElectronic Health Records
🎯 What it does: A hybrid encoding framework called HELIX is proposed for missing value imputation in multivariate time series, with the core idea of introducing learnable feature identity embeddings to provide persistent semantic anchors for each feature.
Helpful to a Fault: Measuring Illicit Assistance in Multi-Turn, Multilingual LLM Agents
Nivya Talokar (Independent Researcher), Antoine Bosselut (EPFL)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Proposes the STING framework for multi-round, multi-lingual LLM agent red team evaluation, aiming to detect the extent to which agents assist in complex malicious plans.
HeraSys: Collaborative Serving of Multiple LLM Workflows via Fine-Grained End-to-End Optimization
Size Li (Beijing Normal University), Weijia Jia (Beijing Normal University)
OptimizationComputational EfficiencyTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Proposes HeraSys, a system for multi-workflow LLM services, employing fine-grained graph fusion and load-aware joint scheduling to achieve cross-workflow computation redundancy elimination and end-to-end performance optimization.
Hermes: An Evidence-Driven Agentic Framework for Trustworthy and Explainable AI-Generated Video Detection
Shuaibo Li (Hong Kong University of Science and Technology), Lei Zhu (Hong Kong University of Science and Technology)
Anomaly DetectionExplainability and InterpretabilityGraph Neural NetworkTransformerAgentic AIVision Language ModelVideoMultimodalityRetrieval-Augmented Generation
🎯 What it does: This paper proposes and implements the Hermes framework, specifically designed for detecting AI-generated videos and providing credible and interpretable judgments;
HERMES: Towards Efficient and Verifiable Mathematical Reasoning in LLMs
Azim Ospanov (Huawei), Farzan Farnia (Chinese University of Hong Kong)
Explainability and InterpretabilityComputational EfficiencyAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes Hermes, a tool-enhanced agent that interacts large language model (LLM) reasoning with Lean4 theorem proving, capable of formalizing, verifying, and feeding back key reasoning steps during the reasoning process;
Hermite-NGP: Gradient-Augmented Hash Encoding for Learning PDEs
Jinjin He (Georgia Institute of Technology), Bo Zhu (Georgia Institute of Technology)
OptimizationDiffusion modelNeural Radiance FieldContrastive LearningPoint CloudMeshTabularBenchmarkPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Propose Hermite-NGP, a gradient-enhanced multi-resolution hash encoding that directly stores function values and mixed partial derivatives on hash grid points, enabling fast and accurate gradient computation for neural PDE solvers.
Heterogeneity-Aware Knowledge Sharing for Graph Federated Learning
Wentao Yu (Nanjing University of Science and Technology), Chen Gong (Shanghai Jiao Tong University)
ClassificationFederated LearningRepresentation LearningGraph Neural NetworkAuto EncoderContrastive LearningGraph
🎯 What it does: Propose a framework called FedSSA that simultaneously addresses node feature heterogeneity and structural heterogeneity in graph federated learning.
Heterogeneous Customizable Personalized Federated Fine-Tuning Approach for Large Language Models
xin tong, Baojiang cui
Federated LearningComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: This paper proposes a heterogeneous customizable personalized federated LoRA fine-tuning framework called Het-CPFLoRA, which allows each client to simultaneously learn shared and personalized knowledge through a single adapter, and dynamically adjust weights during inference.
HexGen-3: A Fully Disaggregated LLM Serving Framework with Fine-Grained Heterogeneous Resource Autoscaling
Youhe Jiang (Hong Kong University of Science and Technology), Binhang Yuan (Hong Kong University of Science and Technology)
OptimizationComputational EfficiencyTransformerLarge Language ModelAuto EncoderTextRetrieval-Augmented Generation
🎯 What it does: Propose HEXGEN-3, a cost-efficient service framework that combines fully discretized LLM inference with fine-grained heterogeneous resource adaptive scaling;
HEXST: Hexagonal Shifted-Window Transformer for Spatial Transcriptomics Gene Expression Prediction
Keunho Byeon (Korea University), Jin Tae Kwak (Korea University)
Image TranslationRepresentation LearningTransformerSupervised Fine-TuningContrastive LearningImageBiomedical DataReview/Survey Paper
🎯 What it does: Propose the HEXST model, which directly predicts spatial gene expression from H&E slices using a Transformer based on a hexagonal window, addressing the geometric mismatch and over-smoothing issues of traditional methods.
HGMem: Hypergraph-based Working Memory to Improve Multi-step RAG for Long-Context Complex Relational Modeling
Chulun Zhou (Chinese University of Hong Kong), Mo Yu (Tencent)
RetrievalExplainability and InterpretabilityComputational EfficiencyRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningTextGraphRetrieval-Augmented Generation
🎯 What it does: Propose a hypergraph-based working memory called HGMem, aimed at enhancing the ability of multi-step retrieval-augmented generation (RAG) to model complex relationships and construct global meaning in long contexts.
Hi-Time: Hierarchical Latent Prediction for Multivariate Time Series Classification
Kun Zeng (South China University of Technology), Qianli Ma (South China University of Technology)
ClassificationAnomaly DetectionRepresentation LearningTransformerLarge Language ModelPrompt EngineeringAuto EncoderContrastive LearningTabularTime SeriesChain-of-Thought
🎯 What it does: Studied the application of large language models to multivariate time series classification, and proposed the Hi-Time framework, which achieves classification through self-learning hierarchical latent predictions (coarse-to-fine semantic codes).
HIAL: Towards Semantics-Aware Hypergraph Active Learning via Dual-Perspective Information Maximization
Yanheng Hou (Beijing Institute of Technology), Guoren Wang (Beijing Institute of Technology)
ClassificationOptimizationComputational EfficiencyRepresentation LearningGraph Neural NetworkDiffusion modelContrastive LearningGraph
🎯 What it does: Propose a training-agnostic hypergraph active learning framework, HIAL, which constructs a node selection strategy that maximizes information by utilizing high-order interaction weighted projection and linear diffusion.
HiCI: Hierarchical Construction–Integration for Long-Context Attention
Xiangyu Zeng (University of Sydney), Chang Xu (University of Sydney)
RetrievalComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose the HiCI (Hierarchical Construction-Integration) layer, introducing a hierarchical attention mechanism to the Transformer, enabling the model to better handle long contexts.
Hidden in Plain Sight -- Class Competition Focuses Attribution Maps
Nils Philipp Walter (CISPA Helmholtz Center for Information Security), Jonas Fischer (Max Planck Institute for Informatics)
ClassificationObject DetectionSegmentationExplainability and InterpretabilityConvolutional Neural NetworkTransformerImageBiomedical Data
🎯 What it does: Proposed a lightweight post-processing method called Attribution Lens (AL), which transforms attribution results from a single logit into an attribution distribution across multiple classes, significantly improving the target specificity and interpretability of attribution while keeping the original method unchanged.
Hidden in Plain Tokens: Simply Robust, Gradient-Free Watermark for Synthetic Audio
Georgios Milis (University of Maryland), Heng Huang (University of Maryland)
GenerationData SynthesisSafty and PrivacyTransformerPrompt EngineeringDiffusion modelContrastive LearningAudio
🎯 What it does: Propose a gradient-free audio generation watermarking method, which reduces re-embedding error by performing community detection on the discrete audio vocabulary, achieving high detectability and robustness.
Hide and Seek in Embedding Space: Geometry-based Steganography and Detection in Large Language Models
Charles Westphal (University College London), Fernando Rosas (University of Sussex)
Safty and PrivacyExplainability and InterpretabilityAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Propose a low recoverability LLM steganography attack based on random hyperplane projection in embedding space, and demonstrate its implementation on multiple large language models; simultaneously propose a method for mechanism explanation detection by utilizing linear probes on internal activations.
HiDe: Rethinking The Zoom-IN method in High Resolution MLLMs via Hierarchical Decoupling
Xianjie Liu (Alimama Tech, Taobao & Tmail Group of Alibaba), Bo Zheng (Alimama Tech, Taobao & Tmail Group of Alibaba)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose the HiDe framework, rethinking the scaling methods for high-resolution multi-modal large language models;
Hide&Seek: Learning to Explain in an End-to-End Differentiable Network
Tal Ellinson (University of Technology Sydney), Sally Cripps (University of Technology Sydney)
Explainability and InterpretabilityRepresentation LearningContrastive LearningImageTabularBiomedical DataBenchmark
🎯 What it does: Proposed an end-to-end differentiable instance-level feature selection framework called Hide&Seek, which jointly trains the selector and predictor.
HieraMAS: Optimizing Intra-Node LLM Mixtures and Inter-Node Topology for Multi-Agent Systems
Tianjun Yao (Shenzhen Loop Area Institute), Zhiqiang Shen (Mohamed bin Zayed University of Artificial Intelligence)
Autonomous DrivingOptimizationFederated LearningComputational EfficiencyAI Code AssistantReinforcement Learning from Human FeedbackGraph Neural NetworkTransformerLarge Language ModelReinforcement LearningPrompt EngineeringMixture of ExpertsContrastive LearningTextGraphBenchmark
🎯 What it does: Built a super node containing internal LLMs, and jointly optimized the selection of LLMs inside the super node and the communication topology between nodes, forming a complete multi-agent collaboration framework.
Hierarchical Abstract Tree for Cross-Document Retrieval Augmented Generation
Ziwen Zhao (Hong Kong University of Science and Technology (Guangzhou)), Menglin Yang (Hong Kong University of Science and Technology (Guangzhou))
GenerationData SynthesisRetrievalTransformerAgentic AIPrompt EngineeringDiffusion modelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Construct a hierarchical abstract tree index combined with multi-granularity agent retrieval, named Ψ-RAG, to enhance cross-document multi-hop retrieval and generation tasks.
Hierarchical Anchor Graph Learning for Multi-View Clustering
Xingchen Hu (National University of Defence Technology), Wenjing Yang (National University of Defence Technology)
Representation LearningGraph Neural NetworkMultimodalityGraph
🎯 What it does: Propose a hierarchical anchor graph learning framework called HAG-MVC, which uses multi-layer anchor co-evolution to capture the fine-grained to coarse-grained semantic structure of multi-view data;
Hierarchical Causal Abduction: A Foundation Framework for Explainable Model Predictive Control
Ramesh Arvind Naagarajan (Chemnitz University of Technology), Stefan Streif (Chemnitz University of Technology)
OptimizationExplainability and InterpretabilityReinforcement Learning from Human FeedbackGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringGraphTabularTime SeriesAgriculture RelatedPhysics RelatedRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed and implemented Hierarchical Causal Abduction (HCA), an interpretable model predictive control (MPC) explanation framework that integrates physical knowledge graphs, KKT multipliers, and PCMCI temporal causal discovery;
Hierarchical Decision Making with Structured Policies: A Principled Design via Inverse Optimization
Yuexuan Wang (National University of Singapore), Kaidi Yang (National University of Singapore)
OptimizationRobotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningTabularTime SeriesSequential
🎯 What it does: This paper proposes a hierarchical RL-OC framework, where the upper-level RL generates sub-goals, and the lower-level constructs structured decision policies through inverse optimization.
Hierarchical Goal Abstractions via Learned Subset Relations
Fabian Wurzberger (Ulm University), Daniel Alexander Braun (Ulm University)
Recurrent Neural NetworkReinforcement LearningContrastive LearningSequential
🎯 What it does: Under a self-supervised goal-conditioned reinforcement learning framework, a traversable hierarchical latent goal space is constructed, and subset relationships are learned through an energy function to map specific observations to goals at different abstraction levels.
Hierarchical Image Tokenization for Multi-Scale Image Super Resolution
Isma Hadji (Samsung AI Center), Georgios Tzimiropoulos (Samsung AI Center)
Super ResolutionConvolutional Neural NetworkTransformerVision Language ModelDiffusion modelAuto EncoderContrastive LearningImage
🎯 What it does: Propose a multi-scale image super-resolution method, achieving step-by-step reasoning from low-resolution to high-resolution through Hierarchical Image Segmentation (HIT);
Hierarchical Multi Scale Graph Neural Networks: Scalable Heterophilous Learning with Oversmoothing and Oversquashing Mitigation
MD SAZZAD Hossen, Avimanyu Sahoo (University of Alabama in Huntsville)
ClassificationRepresentation LearningGraph Neural NetworkAuto EncoderContrastive LearningGraph
🎯 What it does: Proposed a hierarchical multi-scale Haar graph neural network (HMH), which constructs signed affinity through adaptive heterogeneous encoders, performs soft hierarchical aggregation, sparse orthogonal Haar basis and frequency domain filtering, and finally fuses multi-scale information through skip connection transposed convolution.
Hierarchical ODE: Learning Continuous-Time Physical Prototypes for Early Link Failure Detection
Jiaen Lv (Nanjing University), Shaowei Wang (Nanjing University)
Anomaly DetectionRecurrent Neural NetworkAuto EncoderContrastive LearningTabularTime SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: This paper proposes a continuous-time prototype learning framework based on hierarchical neural ordinary differential equations for early failure detection of irregularly sampled link signals.
Hierarchical Policy Learning via Spectral Decomposition
Shuxin Cao (Georgia Institute of Technology), Animesh Garg (Georgia Institute of Technology)
Robotic IntelligenceTransformerReinforcement LearningDiffusion modelContrastive LearningTime SeriesSequential
🎯 What it does: Propose a causal cascading spectral policy (CSP) based on discrete cosine transform (DCT), decomposing robot actions into low-frequency coarse-grained motion and high-frequency fine-grained correction, forming a coarse-to-fine generation process;
Hierarchical Procedural Meta-Reasoning for Generalizable Multimodal Agents
Yao Fu (University of Michigan), Manchen Wang (Meta)
Autonomous DrivingRobotic IntelligenceMeta LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision-Language-Action ModelTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose a multi-modal agent called PAMAR based on natural language programs, which can explicitly evaluate whether the current program is sufficient during execution and proactively retrieve external information when necessary;
Hierarchical Reinforcement Learning for Sparse-Reward Search in Commutative Algebra
Giorgi Butbaia (California Institute of Technology), Sergei Gukov (California Institute of Technology)
Graph Neural NetworkSupervised Fine-TuningReinforcement LearningGraphReview/Survey PaperChain-of-Thought
🎯 What it does: This paper proposes a hierarchical reinforcement learning (HRL) framework for searching non-Hirsch ideals in conformal algebra environments with sparse rewards, thereby constructing counterexamples to the Hirsch conjecture in Kalai algebra.
Hierarchical Representations for Cross-task Automated Heuristic Design using LLMs
Fei Liu (City University of Hong Kong), Qingfu Zhang (City University of Hong Kong)
OptimizationMeta LearningReinforcement Learning from Human FeedbackNeural Architecture SearchTransformerLarge Language ModelPrompt EngineeringGraphTabularBenchmark
🎯 What it does: Designed a multi-task hierarchical search framework, MTHS, which leverages large language models to evolve general metaheuristics and task-specific procedures, enabling automatic heuristic algorithm design across tasks.
Hierarchical Retrieval at Scale: Bridging Interpretability and Efficiency
Shubham Gupta (ServiceNow Research), Valentina Zantedeschi (ServiceNow Research)
RetrievalExplainability and InterpretabilityComputational EfficiencyTransformerContrastive LearningImageTextMultimodalityRetrieval-Augmented GenerationAudio
🎯 What it does: Proposes a binary tree-based retrieval framework called RETREEVER, which can achieve high accuracy, low latency, and provide interpretable hierarchical organization in large-scale retrieval tasks.
Hierarchical Successor Representation for Robust Transfer
Changmin Yu (University of Cambridge), Máté Lengyel (Central European University)
Representation LearningReinforcement LearningTabular
🎯 What it does: Propose Hierarchical Successor Representation (HSR), integrating temporal abstraction (options) into traditional successor representation to obtain more robust, interpretable, and transferable state features.
HieraScaffold: Learning Compact Hierarchical Representations for Scalable 4D LiDAR Generation
Zijie Wu (Singapore University of Technology and Design), Na Zhao (Singapore University of Technology and Design)
GenerationData SynthesisAutonomous DrivingTransformerDiffusion modelNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkContrastive LearningPoint Cloud
🎯 What it does: Propose the HieraScaffold framework, achieving compact representation and generation of 4D LiDAR scenes through multi-resolution unsigned distance scaffold and neural contourlet reuse;
HieRD: Hierarchical Relational Distillation for Vision-Language Embedding Models
Vinh Le (Hanoi University of Science and Technology), Trung Le (Monash University)
CompressionKnowledge DistillationRepresentation LearningTransformerVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: Propose the HieRD framework, which uses a hierarchical relationship distillation method to compress large vision-language models into efficient embedding models, maintaining object-level clustering of visual tokens and multi-granularity correspondence with text phrases.
High-accuracy and dimension-free sampling with diffusions
Khashayar Gatmiry (University of California Berkeley), Adil Salim (Stealth)
GenerationData SynthesisDiffusion modelScore-based ModelMultimodalityStochastic Differential Equation
🎯 What it does: Proposed a high-precision sampler based on diffusion models, which can generate samples on multi-modal distributions with extremely low error;
High-Accuracy Sampling for Diffusion Models and Log-Concave Distributions
Fan Chen (Massachusetts Institute of Technology), Alexander Rakhlin (Massachusetts Institute of Technology)
GenerationData SynthesisDiffusion modelScore-based ModelTabularTime SeriesSequentialStochastic Differential Equation
🎯 What it does: Proposed a high-precision diffusion model sampling algorithm that can achieve δ-error within multiple logarithmic steps, given access to accurate score estimates in L2.
High-Dimensional Learning Dynamics of Quantized Models with Straight-Through Estimator
Yuma Ichikawa (Fujitsu Limited), Ayaka Sakata (Riken Center for AIP)
OptimizationComputational EfficiencyRepresentation LearningContrastive LearningTabularStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes a dynamic analysis of learning in a linear regression model with simultaneous quantization of weights and inputs using the straight-through estimator (STE) under the high-dimensional limit, and derives the corresponding deterministic ordinary differential equations and fixed points, revealing the impact of quantization parameters on the learning phase, convergence speed, and generalization error;
High-Dimensional Sensitivity Analysis for Genomic Studies: An Adversarial Framework for Learning Worst-Case Latent Confounders
Yifan Lin (University of California), Kevin Z. Lin (University of Washington)
Explainability and InterpretabilityDrug DiscoverySupervised Fine-TuningGenerative Adversarial NetworkContrastive LearningTabularBiomedical DataAlzheimer's DiseaseElectronic Health Records
🎯 What it does: A framework called sensGAN based on adversarial learning is proposed to systematically evaluate the impact of potential unobserved confounding factors on the discovery of differentially expressed genes in high-dimensional genomic data.
High-Fidelity ANN-to-SNN Conversion via Closed-Loop CKA Distillation
Bozhou Li (Hunan University), Kenli Li (Hunan University)
ClassificationComputational EfficiencyKnowledge DistillationConvolutional Neural NetworkSpiking Neural NetworkSupervised Fine-TuningContrastive LearningImage
🎯 What it does: Propose a closed-loop fine-tuning framework that significantly improves the accuracy and low-latency performance of ANN-to-SNN conversion without modifying the source ANN, through a dual alignment strategy.
High-Probability Convergence Guarantees of Decentralized SGD
Aleksandar Armacki (EPFL), Ali H. Sayed (EPFL)
OptimizationFederated LearningContrastive LearningTabularStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Studied the high probability convergence of decentralized stochastic gradient descent (DSGD) under light-tailed noise, and provided optimal convergence rates and linear acceleration for non-convex and strongly convex problems.
Higher-Order Certified Robustness for Regression
Claire Jie Zhang (University of Washington), Natalie Frank (University of Washington)
Anomaly DetectionImageTabular
🎯 What it does: Propose a high-order regression robustness proof framework based on random smoothing, providing a computable robust radius with prediction centering;
HilbertA: Hilbert-Curve–Aligned Sparse Attention for 2D Structured Data
Shaoyi Zheng (New York University), Shenji Wan (New York University)
Image TranslationRestorationGenerationComputational EfficiencyTransformerSupervised Fine-TuningVision Language ModelDiffusion modelContrastive LearningImageMultimodality
🎯 What it does: This paper proposes HilbertA, a 2D structural sparse attention mechanism based on Hilbert curve reordering, which significantly accelerates inference in diffusion models and vision-language models.
HiMAP-Travel: Hierarchical Multi-Agent Planning for Long-Horizon Constrained Travel
The Viet Bui (Singapore Management University), Yong Liu (Independent Researcher)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes HiMAP-Travel, a hierarchical multi-agent framework for long-horizon travel itinerary generation. It first performs global resource allocation by the Coordinator, and then multiple Day Executors work in parallel to complete daily itineraries. Global constraints are enforced in real time and dynamically replanned through synchronized global state, cooperative bargaining, and single-role conditional strategies.
HiMe: Hierarchical Embodied Memory for Long-Horizon Vision-Language-Action Control
Li Ji (Fudan University), Xipeng Qiu (Fudan University)
Robotic IntelligenceTransformerVision Language ModelVision-Language-Action ModelMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose a hierarchical embodied memory framework called HiMe, which decomposes robotic intelligence into high-speed actuators, monitors, and planners, supporting cross-modal semantic memory and active management (Add/Update/Delete), addressing the frequency-capability paradox in long-sequence tasks.
HInT: Hypergraph Infusion at the Structural Layers Improves Table Understanding
Wonjin Lee (POSTECH), Kwang In Kim (POSTECH)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextMultimodalityTabularChain-of-Thought
🎯 What it does: Without altering the architecture of language models, this paper analyzes and utilizes the encoding of table structures (row-column relationships) within the decoder of unidirectional LLMs. Subsequently, structural features based on hypergraphs are injected into these key layers, achieving an improvement in table reasoning capabilities through lightweight message passing and gated fusion.
HiPER: Hierarchical Plan–Execute RL for Multi-Turn LLM Agents
Jiangweizhi Peng (University of Minnesota), Mingyi Hong (University of Minnesota)
TransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringText
🎯 What it does: Proposed a Hierarchical Plan-Execute RL framework called HiPER, designed to train large language models (LLMs) to complete multi-turn interactive tasks by explicitly separating high-level planning from low-level execution.
HiPhO: How Far Are (M)LLMs from Humans in the Latest High School Physics Olympiad Benchmark?
Fangchen Yu (Shanghai AI Laboratory), Peng Ye (Shanghai AI Laboratory)
TransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkPhysics RelatedRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the HiPhO benchmark to evaluate the reasoning capabilities of large language models (LLMs) and multimodal LLMs (MLLMs) in the 2024-2025 high school physics Olympiad.
HiPPO Zoo: Explicit Memory Mechanisms for Interpretable State Space Models
Jack Goffinet (Duke University), David Carlson (Duke University)
Explainability and InterpretabilityComputational EfficiencyRecurrent Neural NetworkTransformerDiffusion modelScore-based ModelFlow-based ModelRectified FlowAuto EncoderContrastive LearningTextTime SeriesSequentialStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes the HiPPO Zoo framework, which integrates HiPPO's explicit memory mechanism with modern SSMs through five extensions (Volterra, Salience, Associative, Multiscale, Forecasting), achieving interpretable sequence models.
HiST: A Hierarchical Sparse Transformer for Cross-Modal Spatial Transcriptomics Modeling
Weiyi Wu (Dartmouth College), Jiang Gui (Dartmouth College)
Computational EfficiencyData-Centric LearningDrug DiscoveryTransformerVision Language ModelAuto EncoderContrastive LearningImageMultimodalityBiomedical DataBenchmark
🎯 What it does: Designed and implemented a HiST (Hierarchical Sparse Transformer) model for predicting gene expression in spatial transcriptomics (ST) from conventional H&E tissue sections.
Hista and Numca: Estimate State Value Effectively for Large Language Model Reinforcement Learning
Zizhe Chen (Chinese University of Hong Kong), James Cheng (Chinese University of Hong Kong)
TransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: This paper proposes the SVEB benchmark and investigates the defects of state value estimation in LLM RL. Subsequently, two methods, Numca and Hista, are proposed to improve state value estimation.
History-Bootstrapped Flow Matching for Inverse Boiling Reconstruction
Xianwei Zou (University of California Irvine), Aparna Chandramowlishwaran (University of California Irvine)
Convolutional Neural NetworkRecurrent Neural NetworkTransformerDiffusion modelFlow-based ModelAuto EncoderOptical FlowVideoTabularSequentialPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes a history-bootstrap assisted flow matching (HB-ARFM) framework for reconstructing complete temperature and velocity fields in reverse, under the condition that only the geometric and motion of the boiling interface are observed;
HIVE-3D: Hierarchical Voxel Enhancement for High-Quality 3D Scene Generation
Bin Zang (Zhejiang University), Rengan Xie (Zhejiang University)
GenerationData SynthesisTransformerVision Language ModelDiffusion modelContrastive LearningGaussian SplattingImagePoint CloudMesh
🎯 What it does: Propose HIVE-3D, a hierarchical voxel-enhanced framework that can generate high-resolution, high-quality 3D scenes from a single RGB image.
HO-SFL: Hybrid-Order Split Federated Learning with Backprop-Free Clients and Dimension-Free Aggregation
Qiyuan Chen (University of Hong Kong), Xianhao Chen (University of Hong Kong)
OptimizationFederated LearningComputational EfficiencyRepresentation LearningContrastive LearningImageText
🎯 What it does: Propose Hybrid-Order Split Federated Learning (HO-SFL), a federated learning framework that achieves client-side training without backpropagation and dimension-agnostic aggregation.
HOBIT: Hardness Optimized Batch Sampling for InfoNCE Training
Himanshu Dutta (Microsoft Research, India), Yashoteja Prabhu (Microsoft Research, India)
RetrievalOptimizationRepresentation LearningContrastive LearningText
🎯 What it does: Proposes a hardness-optimized batch sampling method called HOBIT based on InfoNCE contrastive learning to improve the quality of negative samples during training.
HodgeFlow Policy Search by Topologically Dissecting Temporal-Difference Signals in Non-Markovian Environments
Zuyuan Zhang (George Washington University), Tian Lan (George Washington University)
Reinforcement LearningContrastive LearningTabularTime SeriesSequential
🎯 What it does: A new topological perspective is proposed to handle temporal difference learning in non-Markovian environments, demonstrating that temporal difference errors can be viewed as 1-cochains in the topological space of state transitions, and obtaining a Hodge-type decomposition of temporal difference errors through the Bellman-Darmstadt projection.
HOI-PAGE: Zero-Shot Human-Object Interaction Generation with Part Affordance Guidance
Lei Li (University of Virginia), Angela Dai (Technical University of Munich)
GenerationData SynthesisPose EstimationOptimizationGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringVision-Language-Action ModelDiffusion modelScore-based ModelVideoTextPoint CloudMeshGraphChain-of-Thought
🎯 What it does: Propose HOI-PAGE, a zero-shot 4D human-object interaction generation framework, which uses part affinity graphs (PAG) inferred by LLM to guide the video diffusion and optimization process, generating realistic multi-object, multi-human interactions.
Hölder++: Improving Quality-Coherence Trade-off in Multimodal VAEs
Huyen Thuc Khanh Vo (Saarland University), Isabel Valera (Saarland University)
GenerationRepresentation LearningMixture of ExpertsDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningImageMultimodality
🎯 What it does: This paper proposes Holder++, a multi-modal variational autoencoder, which improves the balance between generation quality and consistency by implementing symmetric Holder pooling, introducing shared and private latent subspaces (Holder+), and hierarchical inference (Holder++).
Holi-Spatial: Evolving Video Streams into Holistic 3D Spatial Intelligence
Yuanyuan Gao (Shanghai AI Lab), Zhihang Zhong (Shanghai Jiao Tong University)
SegmentationData SynthesisDepth EstimationOptimizationTransformerPrompt EngineeringVision Language ModelGaussian SplattingVideoTextPoint CloudMeshRetrieval-Augmented Generation
🎯 What it does: Proposed a fully automated pipeline called Holi-Spatial, which converts raw video streams into high-quality 3D geometric models and multi-layer spatial annotations, and based on this pipeline, constructed a large-scale Holi-Spatial-4M dataset.
HoloFair: Unified T2I Fairness Evaluation and Fair-GRPO Debiasing
Ruyi Chen (Nanjing University of Aeronautics and Astronautics), Liming Fang (Nanjing University of Aeronautics and Astronautics)
GenerationExplainability and InterpretabilityData-Centric LearningTransformerSupervised Fine-TuningReinforcement LearningPrompt EngineeringDiffusion modelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: Research and propose a unified text-to-image model fairness evaluation framework called HoloFair, as well as a multi-attribute debiasing method based on reinforcement learning called Fair-GRPO.
Holonomy Grid Codes for Generalisation Under Directed Actions
Pradeep Singh (Indian Institute of Information Technology Surat), Balasubramanian Raman (Indian Institute of Technology Roorkee)
OptimizationExplainability and InterpretabilityComputational EfficiencyRepresentation LearningMeta LearningReinforcement Learning from Human FeedbackDiffusion modelScore-based ModelFlow-based ModelRectified FlowNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkContrastive LearningOptical FlowPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: In discrete lattice dynamics with non-trivial holonomy phase, block Fourier benchmarks that can be reused across all actions are constructed, providing a novel 'Holonomy Grid Codes' framework for achieving long-term period prediction.
Hom-PGD$^+$: Fast Reparameterized Optimization over Non-convex Ball-Homeomorphic Set
Chenghao Liu (City University of Hong Kong), Minghua Chen (City University of Hong Kong)
OptimizationGraph Neural NetworkTransformerFlow-based ModelImagePoint CloudGraphTabular
🎯 What it does: Propose the Hom-PGD+ method, which uses invertible neural networks to learn mappings under non-convex spherical homeomorphic constraints, transforming the original problem into a spherical constraint problem, and then efficiently solves it in the transformed space using projected gradient descent and binary projection.
Homophily-Heterogeneity Gradient Surgery for Federated Graph Learning
Sujia Huang (Nanjing University of Science and Technology), Zhen Cui (Beijing Normal University)
OptimizationFederated LearningRepresentation LearningGraph Neural NetworkSupervised Fine-TuningContrastive LearningGraph
🎯 What it does: In federated graph learning, FedGCM is proposed, which groups clients based on their homogeneity levels and uses residual projection gradient surgery (RPGrad) on the server side to resolve optimization conflicts between different subgraphs, thereby improving the global model's generalization.
HONet: Data-Efficient Learning for Exact Cover Tasks via Hypergraph Optimization
Pengyang Huang (Beijing University of Posts and Telecommunications), Jianxin Liao (Beijing University of Posts and Telecommunications)
OptimizationData-Centric LearningGraph Neural NetworkSupervised Fine-TuningContrastive LearningImageGraphTabular
🎯 What it does: Proposed an end-to-end neuro-symbolic framework called HONet for Exact Cover tasks with known constraints, combining a hypergraph residual encoder and a differentiable fixed QP parser to achieve a complete pipeline from perception to constraint solving.
How (Not) to Hybridize Neural and Mechanistic Models for Epidemiological Forecasting
Yiqi Su (Virginia Tech), Naren Ramakrishnan (Virginia Tech)
OptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyDrug DiscoveryRecurrent Neural NetworkTransformerAuto EncoderContrastive LearningTabularTime SeriesBiomedical DataPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposed and implemented a hybrid neural-physical framework called EpiNode, combining multi-scale decomposition, controlled neural ODEs, and SIRS dynamics for infectious disease prediction and inference of time-varying parameters.
How can embedding models bind concepts?
Arnas Uselis (University of Tübingen), Seong Joon Oh (University of Tübingen)
Explainability and InterpretabilityRepresentation LearningData-Centric LearningConvolutional Neural NetworkTransformerMixture of ExpertsVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper analyzes the geometric structure of multimodal embedding models such as CLIP to study the implementation of concept binding. It finds that scene embeddings can be approximately decomposed into additive components at the object level, while the binding function between objects and concepts has high complexity, leading to failure in cross-modal binding; subsequently, a Transformer-based dual encoder model is trained on controlled synthetic data, demonstrating that low-complexity multiplicative binding functions can achieve systematic generalization to unseen concept combinations.
How Can Mamba Learn In Context with Outliers and Generalize Provably?
Hongkang Li (University of Pennsylvania), Meng Wang (Rensselaer Polytechnic Institute)
ClassificationTransformerSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper provides the first theoretical analysis of the training dynamics and generalization ability of a single-layer Mamba model in in-context learning (ICL) tasks with outliers, proving that it can maintain robust ICL performance on unseen tasks.
How can we assess human-agent interactions? Case studies in software agent design
Valerie Chen (Carnegie Mellon University), Graham Neubig (Carnegie Mellon University)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextTabularBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes and verifies the PULSE framework for evaluating the interaction effects between LLM-driven agents and humans in real-world scenarios.
How do Humans Process AI-generated Hallucination Contents: a Neuroimaging Study
Shuqi Zhu (Tsinghua University), Yiqun LIU
Anomaly DetectionExplainability and InterpretabilityImageTextMultimodality
🎯 What it does: In a laboratory setting, 64-channel EEG was used to record the brain activity of 27 participants while they viewed text-image descriptions generated by AI (Qwen2.5-VL-3B-Instruct), which either contained or did not contain hallucinations. Participants were asked to judge the accuracy of the descriptions, followed by ERP analysis and machine learning prediction of the EEG signals.
How Do Language Models Speak Languages? A Case Study on Unintended Code-Switching
Yuxin Xiao (Zhejiang University), Jieping Ye (Alibaba Cloud)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: This study investigates the phenomenon of unintentional code switching in large language models during multilingual generation and constructs an expandable circuit discovery framework to perform causal analysis and explainable interpretation of the internal mechanisms causing this phenomenon.
How do LLMs Compute Verbal Confidence?
Dharshan Kumaran (Google DeepMind), Petar Veličković (Google DeepMind)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Investigate how large language models internally generate verbal confidence reports, verify whether confidence is computed on-the-fly or automatically cached during answer generation, and explore the type of signal it represents.
How does Bayesian Sampling help Membership Inference Attacks?
Zhenlong Liu (Southern University of Science and Technology), Hongxin Wei (Southern University of Science and Technology)
Safty and PrivacyAdversarial AttackContrastive LearningImageTextMultimodalityTabular
🎯 What it does: Propose a membership inference attack based on Bayesian sampling (BMIA), which directly estimates the conditional score distribution for each sample by using Laplace approximation on a single reference model to obtain the parameter posterior, and further proposes a multi-reference model version (MR-BMIA).
How does Chain of Thought decompose complex tasks?
Amrut Nadgir (University of Pennsylvania), Pratik Chaudhari (University of Pennsylvania)
ClassificationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringImageTextTabularChain-of-Thought
🎯 What it does: Analyze how Chain of Thought (CoT) decomposes complex tasks into multi-level classification problems, and theoretically derive the effectiveness and limitations of CoT.
How does information access affect LLM monitors' ability to detect sabotage?
Rauno Arike (Aether Research), Francis Rhys Ward (Independent)
Anomaly DetectionFederated LearningSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextSequentialBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper studies the impact of information access levels on the performance of large language model (LLM) monitors in detecting attackers' 'sabotage' behaviors, and proposes a hierarchical information filtering 'Extract-and-Evaluate (EaE)' monitoring scheme.
How Does Reasoning Flow? Tracing Attention-Induced Information Flow for Targeted RL in LLMs
Zhichen Dong (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
Explainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningFlow-based ModelText
🎯 What it does: Propose the FlowTracer framework, which utilizes the global information flow attention of LLMs to perform fine-grained token-level credit allocation, thereby improving reward shaping in reinforcement learning;