NeurIPS 2025 Papers — Page 22
Conference on Neural Information Processing Systems · 5275 papers
How Data Mixing Shapes In-Context Learning: Asymptotic Equivalence for Transformers with MLPs
Samet Demir (Koç University), Zafer Dogan (Koç University)
TransformerText
🎯 What it does: This study investigates the contextual learning performance of pre-trained Transformers using a nonlinear MLP head under mixed multi-source data, and proves its equivalence to polynomial predictors in a high-dimensional asymptotic framework.
How Different from the Past? Spatio-Temporal Time Series Forecasting with Self-Supervised Deviation Learning
Haotian Gao (University of Tokyo), Renhe Jiang (University of Tokyo)
Anomaly DetectionRecurrent Neural NetworkGraph Neural NetworkContrastive LearningTime Series
🎯 What it does: This paper proposes a spatiotemporal sequence prediction framework called ST-SSDL based on self-supervised bias learning, using historical averages as anchors and further discretizing the latent space through learnable prototypes, achieving relative distance consistency through contrastive learning and bias loss.
How do Transformers Learn Implicit Reasoning?
Jiaran Ye (Tsinghua University), Juanzi Li (Tsinghua University)
TransformerLarge Language ModelText
🎯 What it does: In a controlled symbolic environment, training a Transformer to learn implicit reasoning from scratch reveals a three-stage developmental trajectory from memory to generalization.
How Does Label Noise Gradient Descent Improve Generalization in the Low SNR Regime?
Wei Huang (RIKEN AIP), Taiji Suzuki (University of Tokyo)
Convolutional Neural NetworkImage
🎯 What it does: This paper studies the introduction of label noise gradient descent to enhance the generalization of deep networks in low signal-to-noise ratio environments.
How Does Sequence Modeling Architecture Influence Base Capabilities of Pre-trained Language Models? Exploring Key Architecture Design Principles to Avoid Base Capabilities Degradation
Xin Lu (Harbin Institute of Technology), Ting Liu (Harbin Institute of Technology)
TransformerLarge Language ModelText
🎯 What it does: This paper proposes and validates the impact of sequence modeling architectures on base capabilities in pre-trained language models. It reveals the degradation of base capabilities in stateful sequence architectures (such as Mamba and RWKV) through a limited pre-training + OOV testing framework. It concludes that 'full sequence arbitrary selection' is the key design principle to avoid base capability degradation and designs a minimal Top-1 Element/Chunk Selection architecture for validation.
How Does Topology Bias Distort Message Passing in Graph Recommender? A Dirichlet Energy Perspective
Yanbiao Ji (Shanghai Jiao Tong University), Hongtao Lu (Shanghai Jiao Tong University)
Recommendation SystemGraph Neural NetworkGraph
🎯 What it does: The paper studies the impact of topological bias on message passing in graph recommendation from the perspective of Dirichlet energy, and proposes a Test-time Simplicial Propagation (TSP) method that uses higher-order Simplicial Complex for message passing during the inference phase to mitigate this bias.
How Ensembles of Distilled Policies Improve Generalisation in Reinforcement Learning
Max Weltevrede (Delft University of Technology), Wendelin Boehmer
Knowledge DistillationReinforcement Learning
🎯 What it does: This study investigates policy distillation after training in reinforcement learning to enhance the generalization ability of zero-shot policy transfer, proposes theoretical generalization bounds, provides practical recommendations, and validates them through experiments.
How Far Are We from Optimal Reasoning Efficiency?
Jiaxuan Gao (Tsinghua University), Yi Wu (Tsinghua University)
Computational EfficiencySupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: This paper studies the reasoning efficiency of large reasoning models (LRM) in chain-of-thought (CoT) and proposes an empirical reasoning efficiency frontier and a unified metric for reasoning efficiency gap (REG), as well as the REO-RL algorithm to optimize reasoning efficiency under a limited token budget.
How Many Domains Suffice for Domain Generalization? A Tight Characterization via the Domain Shattering Dimension
Cynthia Dwork (Harvard University), Han Shao (University of Maryland)
Domain Adaptation
🎯 What it does: In the domain generalization problem, a new metric called Domain Shattering Dimension is proposed, which precisely characterizes the number of domains that need to be sampled (domain sample complexity), thus answering the core question: 'How many random domains need to be observed to maintain good performance across all domains?'
How many measurements are enough? Bayesian recovery in inverse problems with general distributions
Ben Adcock (Simon Fraser University), Nick Huang (Simon Fraser University)
🎯 What it does: This study investigates the sample complexity of Bayesian recovery in inverse problems with general distributions, establishing sufficient conditions for stable and accurate recovery with high probability.
How Many Tokens Do 3D Point Cloud Transformer Architectures Really Need?
Tuan Anh Tran (German Research Centre for Artificial Intelligence), Paul Swoboda (Heinrich Heine University Dusseldorf)
Object DetectionSegmentationComputational EfficiencyTransformerPoint Cloud
🎯 What it does: This study addresses the token redundancy issue in 3D point cloud Transformers and proposes an efficient token merging strategy for 3D data.
How Memory in Optimization Algorithms Implicitly Modifies the Loss
Matias D. Cattaneo (Princeton University), Boris Shigida (Princeton University)
OptimizationTransformerImageText
🎯 What it does: A general method is proposed to approximate optimization algorithms with exponentially decaying memory as memoryless algorithms, and the memory effect is explained as an implicit disturbance to the loss function through a memory correction term, thereby analyzing the implicit regularization effect of memory on optimization dynamics.
How Particle System Theory Enhances Hypergraph Message Passing
Yixuan Ma (Shanghai Jiao Tong University), Yu Guang Wang (Shanghai Jiao Tong University)
ClassificationGraph Neural NetworkGraphStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A novel hypergraph information propagation framework HAMP based on particle dynamics is proposed, utilizing hyperedges as field forces and introducing attraction-repulsion and Allen-Cahn forces to model high-order interactions in hypergraphs and suppress excessive smoothing.
How Patterns Dictate Learnability in Sequential Data
Mario Morawski, Remi Rehm
Contrastive LearningSequential
🎯 What it does: Evaluate the learnability and minimum achievable risk of sequential data through an information-theoretic framework.
How to Auto-optimize Prompts for Domain Tasks? Adaptive Prompting and Reasoning through Evolutionary Domain Knowledge Adaptation
Yang Zhao (Johns Hopkins University), Hao Frank Yang (Johns Hopkins University)
Domain AdaptationOptimizationPrompt EngineeringText
🎯 What it does: The automated framework EGO-Prompt improves domain task performance by iteratively refining prompts and reasoning processes through text gradients, in conjunction with a semantically constructed causal graph (SCG) without fine-tuning.
How to build a consistency model: Learning flow maps via self-distillation
Nicholas Matthew Boffi, Eric Vanden-Eijnden (Courant Institute of Mathematical Sciences)
Data SynthesisKnowledge DistillationFlow-based ModelGenerative Adversarial NetworkImage
🎯 What it does: A direct learning flow-map framework based on self-distillation is proposed, unifying and extending various accelerated sampling methods such as consistency models and progressive distillation, avoiding the need for a pre-trained teacher.
How to Learn a Star: Binary Classification with Starshaped Polyhedral Sets
Marie-Charlotte Brandenburg (Ruhr University Bochum), Katharina Jochemko (KTH Royal Institute of Technology)
ClassificationOptimizationTabular
🎯 What it does: This paper studies a binary classifier with star-shaped polytope sets as decision boundaries, analyzing the geometric and combinatorial structure of its parameter space, and provides a description of the VC dimension and sub-level sets;
How to Scale Second-Order Optimization
Zixi Chen, Andrew Gordon Wilson (New York University)
OptimizationTransformerLarge Language ModelText
🎯 What it does: The study developed and formulated hyperparameter scaling rules for second-order optimizers (such as Shampoo, SOAP, Muon) that are width and depth scalable, and validated their effectiveness in large-scale Transformer training.
How to Train Your LLM Web Agent: A Statistical Diagnosis
Dheeraj Vattikonda (ServiceNow Research), Massimo Caccia (ServiceNow Research)
Computational EfficiencyHyperparameter SearchTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIText
🎯 What it does: This study investigates how to train LLM web agents through supervised fine-tuning (SFT) and reinforcement learning (RL) under constrained computational costs, systematically evaluating the impact of different computational allocation strategies on performance.
How Well Can Differential Privacy Be Audited in One Run?
Amit Keinan (Hebrew University of Jerusalem), Katrina Ligett (Hebrew University of Jerusalem)
Safty and PrivacyTabular
🎯 What it does: This study investigates and quantifies the effectiveness of one-run auditing on differential privacy algorithms, revealing its inherent limitations.
HPSERec: A Hierarchical Partitioning and Stepwise Enhancement Framework for Long-tailed Sequential Recommendation
Xiaolong Xu (Nanjing University of Information Science and Technology), Lianyong Qi (China University of Petroleum East China)
Recommendation SystemKnowledge DistillationTransformerContrastive LearningSequential
🎯 What it does: The HPSERec framework is proposed, utilizing hierarchical partitioning, expert networks, knowledge distillation, and the Sinkhorn OT feedback mechanism to model and enhance long-tail users and long-tail items in sequential recommendation.
HQA-VLAttack: Towards High Quality Adversarial Attack on Vision-Language Pre-Trained Models
Han Liu (Dalian University of Technology), Hong Yu (Dalian University of Technology)
RetrievalAdversarial AttackContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes a query-free high-quality cross-modal adversarial attack method called HQA-VLAttack, which achieves a higher attack success rate against vision-language pre-trained models.
HubGT: Fast Graph Transformer with Decoupled Hierarchy Labeling
Ningyi Liao (Nanyang Technological University), Gao Cong (Nanyang Technological University)
Computational EfficiencyGraph Neural NetworkTransformerGraph
🎯 What it does: Proposes HubGT, an efficient graph transformer that utilizes pivot labels and supports batch training of large-scale graphs;
Human Texts Are Outliers: Detecting LLM-generated Texts via Out-of-distribution Detection
Cong Zeng, zhiqiang xu
Anomaly DetectionLarge Language ModelContrastive LearningText
🎯 What it does: Proposes transforming the detection of LLM-generated text from a binary classification problem into an outlier detection problem, considering human text as out-of-distribution samples.
Human-assisted Robotic Policy Refinement via Action Preference Optimization
Wenke Xia (Renmin University of China), Di Hu (Renmin University of China)
OptimizationRobotic IntelligenceReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningVision Language ModelVision-Language-Action ModelMultimodalitySequential
🎯 What it does: Collect interaction trajectories through human-machine collaboration in real environments and simulations, and iteratively refine the Vision-Language-Action (VLA) model using Action Preference Optimization (APO) to enable learning from failure examples and achieve continuous improvement.
HumanCrafter: Synergizing Generalizable Human Reconstruction and Semantic 3D Segmentation
Panwang Pan (ByteDance), Yixuan Yuan (Xiamen University)
SegmentationGenerationPose EstimationTransformerDiffusion modelGaussian SplattingImagePoint Cloud
🎯 What it does: A unified framework called HUMANCRAFTER is proposed, which can achieve high-quality 3D human reconstruction and semantic segmentation of body parts using explicit 3D Gaussian splats with a single image input.
HumanoidGen: Data Generation for Bimanual Dexterous Manipulation via LLM Reasoning
Zhi Jing (Fudan University), Chenjia Bai (China Telecom)
Data SynthesisOptimizationRobotic IntelligenceTransformerLarge Language ModelReinforcement LearningDiffusion modelMultimodality
🎯 What it does: Proposes the HumanoidGen framework, which utilizes large language models and Monte Carlo tree search to automatically generate scenes, planning, and demonstration data for dual-arm humanoid robot grasping tasks.
Hybrid Autoencoders for Tabular Data: Leveraging Model-Based Augmentation in Low-Label Settings
Erel Naor (Bar Ilan University), Ofir Lindenbaum (Bar Ilan University)
ClassificationData-Centric LearningAuto EncoderTabular
🎯 What it does: A hybrid self-supervised autoencoder called TANDEM has been developed, which combines a neural network encoder with an Oblivious Soft Decision Tree encoder, and implements data transformation through sample-specific random gating; during training, only the tree encoder is used for model base augmentation, while only the neural network is retained during inference.
Hybrid Boundary Physics-Informed Neural Networks for Solving Navier-Stokes Equations with Complex Boundary
Chuyu Zhou (Northwest University), Hangzhou Yang (Northwest University)
Neural Architecture SearchTime SeriesPhysics Related
🎯 What it does: This paper proposes a Hybrid Boundary Physics-Informed Neural Network (HB-PINN) to solve the Navier-Stokes equations under complex boundaries using a pre-trained special solution network and a distance metric network.
Hybrid Latent Reasoning via Reinforcement Learning
Zhenrui Yue (University of Illinois Urbana-Champaign), Dong Wang
OptimizationTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: A hybrid implicit reasoning framework HRPO based on reinforcement learning is proposed, allowing LLM to use both discrete words and continuous hidden states for reasoning during the generation process.
Hybrid Latent Representations for PDE Emulation
Ali Can Bekar (Helmholtz Centre Hereon Geesthacht), David S. Greenberg (Helmholtz Centre Hereon Geesthacht)
Time SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: A hybrid latent variable representation is proposed, combining locally averaged coarsened PDE fields with spatial structure latent variables extracted from high-resolution inputs, achieving efficient and accurate PDE simulations.
Hybrid Re-matching for Continual Learning with Parameter-Efficient Tuning
Weicheng Wang (Nankai University), Jufeng Yang (Nankai University)
OptimizationKnowledge DistillationPrompt EngineeringImage
🎯 What it does: Proposes HRM-PET, a parameter-efficient fine-tuning method for replay-free continual learning;
Hybrid-Balance GFlowNet for Solving Vehicle Routing Problems
Ni Zhang (Singapore Management University), Zhiguang Cao (Singapore Management University)
OptimizationGraph Neural NetworkGenerative Adversarial NetworkGraph
🎯 What it does: Proposes the Hybrid-Balance GFlowNet (HBG), which integrates Trajectory Balance and Detailed Balance to enhance both global and local optimization in vehicle routing planning;
Hybrid-Collaborative Augmentation and Contrastive Sample Adaptive-Differential Awareness for Robust Attributed Graph Clustering
TianxiangZhao, Jipeng Guo (University of Sydney)
Representation LearningGraph Neural NetworkContrastive LearningGraph
🎯 What it does: A robust attribute graph clustering framework RAGC is proposed, which integrates node-level and edge-level embeddings with hybrid collaborative enhancement and adaptive differentiated contrastive sample perception.
HybridNorm: Towards Stable and Efficient Transformer Training via Hybrid Normalization
Zhijian Zhuo (Peking University), Jinwen Ma
TransformerMixture of ExpertsText
🎯 What it does: This paper proposes the HybridNorm structure, which uses QKV normalization in the attention layer of the Transformer and Post-Norm in the feedforward network, balancing the gradient stability of Pre-Norm with the performance advantages of Post-Norm.
HyGen: Efficient LLM Serving via Elastic Online-Offline Request Co-location
Ting Sun (University of Illinois Urbana-Champaign), Fan Lai (University of Illinois Urbana-Champaign)
TransformerLarge Language ModelText
🎯 What it does: The HyGen system is proposed to achieve elastic co-location for online and offline LLM tasks, maintaining the SLO of online requests while significantly improving overall throughput.
Hyper-Modality Enhancement for Multimodal Sentiment Analysis with Missing Modalities
Yan Zhuang (University of Electronic Science and Technology of China), Fuji Ren (University of Electronic Science and Technology of China)
ClassificationRecognitionData-Centric LearningTransformerPrompt EngineeringMultimodality
🎯 What it does: A Hyper-Modality Enhancement (HME) framework is proposed to enhance the observed modalities with cross-sample semantic information, and to improve the robustness of multimodal sentiment analysis in the absence of modalities through an uncertainty-aware fusion mechanism.
Hyperbolic Dataset Distillation
Wenyuan Li (Hokkaido University), Miki Haseyama (Hokkaido University)
Data SynthesisKnowledge DistillationImage
🎯 What it does: This paper introduces hyperbolic geometry into dataset distillation, utilizing the negative curvature of hyperbolic space to hierarchically embed the original data, and constructs a smaller synthetic dataset that maintains model performance by matching centroids in hyperbolic space.
Hyperbolic Fine-Tuning for Large Language Models
Menglin Yang (Hong Kong University of Science and Technology), Rex Ying (Chinese University of Hong Kong)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This study investigates the non-Euclidean characteristics of LLM word embeddings and proposes the HypLoRA method for low-rank fine-tuning in hyperbolic space.
HyperET: Efficient Training in Hyperbolic Space for Multi-modal Large Language Models
Zelin Peng (Shanghai Jiao Tong University), Wei Shen (Shanghai Jiao Tong University)
OptimizationComputational EfficiencyRepresentation LearningTransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: This paper proposes a method called HyperET that dynamically adjusts the hyperbolic radius of visual representations through learnable matrices and Möbius multiplication in hyperbolic space, enabling efficient training of multimodal large language models.
Hypergraph-Enhanced Contrastive Learning for Multi-View Clustering with Hyper-Laplacian Regularization
Zhibin Gu (Hebei Normal University), Weili Wang
Representation LearningGraph Neural NetworkAuto EncoderContrastive LearningMultimodality
🎯 What it does: A multi-view clustering framework called HOPER is proposed, which is based on hypergraph enhanced contrastive learning and hypergraph Laplacian regularization, capable of capturing both high-order relationships and local geometric structures.
HyperGraphRAG: Retrieval-Augmented Generation via Hypergraph-Structured Knowledge Representation
Haoran Luo (Beijing University of Posts and Telecommunications), Anh Tuan Luu
GenerationRetrievalGraph Neural NetworkLarge Language ModelTextBiomedical DataAgriculture RelatedRetrieval-Augmented Generation
🎯 What it does: A new hypergraph-based retrieval-augmented generation method, HyperGraphRAG, is proposed to improve knowledge retrieval and generation by representing n-ary relational facts through hyperedges.
HYPERION: Fine-Grained Hypersphere Alignment for Robust Federated Graph Learning
Guancheng Wan (Wuhan University), Bo Du (Wuhan University)
Federated LearningKnowledge DistillationGraph Neural NetworkGraph
🎯 What it does: The HYPERION framework is proposed to achieve robust federated graph learning based on hyperspherical fine-grained alignment.
HyperMARL: Adaptive Hypernetworks for Multi-Agent RL
Kale-ab Tessera (University of Edinburgh), Stefano V. Albrecht (DeepFlow)
Reinforcement LearningAgentic AIBenchmark
🎯 What it does: This paper proposes HyperMARL, a shared policy framework using agent-conditioned hypernetworks that can adaptively achieve specialization, homogenization, or mixed behaviors in multi-agent reinforcement learning.
HyperMixup: Hypergraph-Augmented with Higher-order Information Mixup
Kaixuan Yao (Shanxi University), Feilong Cao (Zhejiang Normal University)
Graph Neural NetworkGraphBiomedical Data
🎯 What it does: Proposes HyperMixup, an adaptive data augmentation framework for hypergraphs;
HyPINO: Multi-Physics Neural Operators via HyperPINNs and the Method of Manufactured Solutions
Rafael Bischof (ETH Zurich), Bernd Bickel (ETH Zurich)
TransformerPhysics Related
🎯 What it does: A multi-physics neural operator HyPINO based on hypernetworks is proposed, which can zero-shot generalize to various two-dimensional linear PDEs without task-specific fine-tuning, and further improve accuracy through iterative residual-driven refinement.
HyPlaneHead: Rethinking Tri-plane-like Representations in Full-Head Image Synthesis
Heyuan Li (Chinese University of Hong Kong), Xiaoguang Han (Chinese University of Hong Kong)
GenerationData SynthesisGenerative Adversarial NetworkImage
🎯 What it does: Proposes the Hy-Plane mixed plane representation and a unified splitting strategy, constructing the HyPlaneHead 3D-aware GAN for high-quality full-head image synthesis.
HypoBootstrap: A Bootstrapping Framework for Inductive Reasoning
Si Chen (Beihang University), Richong Zhang (Beihang University)
TransformerLarge Language ModelText
🎯 What it does: Developed the HypoBootstrap framework for inductive reasoning in large language models, employing a bootstrapping hypothesis generation and bootstrapping confirmation method to reduce model hallucinations.
HYPRL: Reinforcement Learning of Control Policies for Hyperproperties
Tzu-Han Hsu (Michigan State University), Borzoo Bonakdarpour (Michigan State University)
Reinforcement Learning
🎯 What it does: Proposes the HYPRL framework, which uses HyperLTL specifications to guide multi-agent reinforcement learning to maximize the probability of satisfying hyperproperties.
HyRF: Hybrid Radiance Fields for Memory-efficient and High-quality Novel View Synthesis
Zipeng Wang (Hong Kong University of Science and Technology), Dan Xu (Hong Kong University of Science and Technology)
GenerationData SynthesisComputational EfficiencyNeural Radiance FieldGaussian SplattingPoint Cloud
🎯 What it does: Proposes Hybrid Radiance Fields (HyRF), which achieves high-quality real-time novel view synthesis by combining sparse explicit 3D Gaussians with grid-based neural fields.
I2-NeRF: Learning Neural Radiance Fields Under Physically-Grounded Media Interactions
Shuhong Liu (University of Tokyo), Tatsuya Harada (University of Tokyo)
RestorationDepth EstimationNeural Radiance FieldImagePhysics Related
🎯 What it does: This paper proposes I²-NeRF, which models light propagation in participating media by combining physical laws, and implements unified rendering of objects and media within the NeRF framework, supporting 3D reconstruction and enhancement in underwater, foggy, and low-light scenes.
IA-GGAD: Zero-shot Generalist Graph Anomaly Detection via Invariant and Affinity Learning
Xiong Zhang (Yunnan University), Cheng Xie (Yunnan University)
Anomaly DetectionGraph Neural NetworkAuto EncoderGraphFinance Related
🎯 What it does: A zero-shot general graph anomaly detection framework IA-GGAD is proposed, which can detect anomalous nodes in graphs from different domains without any fine-tuning of the target graph.
IBGS: Image-Based Gaussian Splatting
Hoang Chuong Nguyen (Australian National University), Miaomiao Liu (Australian National University)
GenerationData SynthesisGaussian SplattingImage
🎯 What it does: This paper proposes an image-based Gaussian splatting method that utilizes high-resolution source images to supplement details and view-dependent colors, significantly enhancing the rendering quality of novel view synthesis (NVS).
ICLScan: Detecting Backdoors in Black-Box Large Language Models via Targeted In-context Illumination
Xiaoyi Pang (Hong Kong University of Science and Technology), Zhibo Wang (Hong Kong University of Science and Technology)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes a method for detecting backdoors in black-box large language models (LLMs) using In-context Learning (ICL).
Identifiability of Deep Polynomial Neural Networks
Konstantin Usevich (Universite Lorraine), Marianne Clausel (Universite Lorraine)
🎯 What it does: This study investigates the identifiability of deep polynomial neural networks (PNN) and proposes a localization theorem that reduces the identifiability of deep networks to that of 2-layer sub-networks. It provides conditions for identifiability for pyramid and bottleneck structures, as well as upper bounds on activation thresholds.
Identifying interactions across brain areas while accounting for individual-neuron dynamics with a Transformer-based variational autoencoder
Qi Xin (Carnegie Mellon University), Robert Kass
TransformerAuto EncoderTime Series
🎯 What it does: This study proposes a hybrid GLM-Transformer model that utilizes a Transformer-based variational autoencoder to capture trial-to-trial dynamics at the individual neuron level, while estimating directed interactions across regions through a low-rank GLM coupling term.
Identifying Macro Causal Effects in C-DMGs over DMGs
Simon Ferreira (Sorbonne Université), Charles K. Assaad (Sorbonne Université)
Graph Neural Network
🎯 What it does: This paper conducts a theoretical study on the identification of macro causal effects through the construction of C-DMG using clustering methods on directed mixed graphs (DMG) with cycles. It proves that σ-separation maintains consistency and completeness on C-DMG, and the do-calculus rules derived from this are also applicable to the identification of macro causal effects. Additionally, it introduces SC-hedge as a new graphical criterion for detecting non-identifiability, and further shows that if each cyclic node in the clustering has at least two dimensions, SC-hedge can ensure that the macro effect is non-identifiable.
Identifying multi-compartment Hodgkin-Huxley models with high-density extracellular voltage recordings
Ian Christopher Tanoh (Stanford University), Scott Linderman
Biomedical DataStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Using high-density extracellular voltage recordings combined with a multi-compartment Hodgkin-Huxley model and extended Kalman filtering to estimate single-cell membrane voltage, ion channel parameters, and relative probe positions.
IDOL: Meeting Diverse Distribution Shifts with Prior Physics for Tropical Cyclone Multi-Task Estimation
HantingYan, Cong Bai (Zhejiang University of Technology)
Graph Neural NetworkTransformerTime SeriesPhysics Related
🎯 What it does: Proposes the IDOL framework, which utilizes prior physical knowledge to impose identity distribution constraints for multi-task estimation of tropical cyclones.
IF-Guide: Influence Function-Guided Detoxification of LLMs
Zachary Coalson (Oregon State University), Sanghyun Hong (Anthropic)
OptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a method that uses Influence Functions to actively identify and suppress content in training data that leads large language models (LLMs) to exhibit toxic behavior, thereby reducing toxicity during the training phase.
iFinder: Structured Zero-Shot Vision-Based LLM Grounding for Dash-Cam Video Reasoning
Manyi Yao (NEC Laboratories America), Abhishek Aich (NEC Laboratories America)
Object DetectionObject TrackingAutonomous DrivingTransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoMultimodality
🎯 What it does: A structured semantic grounding framework called Finder is proposed, which extracts hierarchical interpretable features from driving videos using a pre-trained visual module and guides the LLM for post-test reasoning through three prompts.
IGD: Token Decisiveness Modeling via Information Gain in LLMs for Personalized Recommendation
Zijie Lin (National University of Singapore), Tat-Seng Chua (National University of Singapore)
Recommendation SystemTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper measures the decisiveness of each token when LLM generates recommendations using Information Gain (IG) and proposes the IGD (IG-based Decisiveness-aware Token Handling) strategy, which re-weights low decisiveness tokens during the fine-tuning and inference stages to correct the token bias of existing models and improve recommendation effectiveness.
IllumiCraft: Unified Geometry and Illumination Diffusion for Controllable Video Generation
Yuanze Lin, Ming-Hsuan Yang
GenerationTransformerDiffusion modelVideo
🎯 What it does: IllumiCraft integrates HDR environment maps, synthetic lighting videos, and 3D point tracking into a Diffusion model, achieving controllable video lighting under text or background instructions.
Image as a World: Generating Interactive World from Single Image via Panoramic Video Generation
Dongnan Gui (University of Science and Technology of China), Yan Lu (Microsoft Research Asia)
GenerationVision Language ModelDiffusion modelImageVideo
🎯 What it does: This paper proposes a framework called IaaW for generating interactive panoramic video worlds from a single image.
Image Editing As Programs with Diffusion Models
Yujia Hu (National University of Singapore), Xinchao Wang (National University of Singapore)
GenerationTransformerLarge Language ModelDiffusion modelImageChain-of-Thought
🎯 What it does: Proposes the IEAP framework, which breaks down instruction-driven image editing into five atomic operations, and achieves multi-step complex editing through chain reasoning and a neural program interpreter.
Image Stitching in Adverse Condition: A Bidirectional-Consistency Learning Framework and Benchmark
Zengxi Zhang (University of Tokyo), Jinyuan Liu (Dalian University of Technology)
RestorationGaussian SplattingImageBenchmark
🎯 What it does: This study investigates the problem of image stitching in harsh environments such as low light, haze, and underwater, proposing a bidirectional consistency learning framework and a motion-tolerant seamless fusion network.
Image Super-Resolution with Guarantees via Conformalized Generative Models
Eduardo Adame (Getulio Vargas Foundation), Guilherme Tegoni Goedert (Getulio Vargas Foundation)
RestorationSuper ResolutionImage
🎯 What it does: This work proposes a generalizable confidence mask method applicable to any black-box generative image super-resolution model, providing reliable and intuitive uncertainty estimates for generated images using conformal prediction techniques.
Image Token Matters: Mitigating Hallucination in Discrete Tokenizer-based Large Vision-Language Models via Latent Editing
Weixing Wang (Hasso Plattner Institute University of Potsdam), Haojin Yang (Hasso Plattner Institute University of Potsdam)
RecognitionGenerationGraph Neural NetworkVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: This paper proposes to identify and suppress visual priors that may lead to hallucinations by clustering the co-occurrence relationships of image tokens in discrete image tokenizers, thereby reducing the hallucination phenomenon in large visual language models.
ImageNet-trained CNNs are not biased towards texture: Revisiting feature reliance through controlled suppression
Tom Burgert (Technische Universität Berlin), Begüm Demir (Technische Universität Berlin)
ClassificationRecognitionConvolutional Neural NetworkTransformerImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper proposes a framework based on targeted feature suppression to quantify the dependence of CNNs on shape, texture, and color features, and re-evaluates the texture bias hypothesis of ImageNet pre-trained CNNs through this framework.
ImageSentinel: Protecting Visual Datasets from Unauthorized Retrieval-Augmented Image Generation
Ziyuan Luo (Hong Kong Baptist University), Renjie Wan (Hong Kong Baptist University)
GenerationRetrievalSafty and PrivacyTransformerVision Language ModelImageRetrieval-Augmented Generation
🎯 What it does: The ImageSentinel framework is proposed, utilizing random character keys and visually consistent sentinel images to protect visual datasets from unauthorized use in retrieval-augmented image generation (RAIG) systems and to achieve detection.
Imagine Beyond ! Distributionally Robust Autoencoding for State Space Coverage in Online Reinforcement Learning
Nicolas Castanet (Sorbonne Université), Sylvain Lamprier
Robotic IntelligenceReinforcement LearningAuto EncoderImage
🎯 What it does: This paper studies a distributed robust optimization-based autoencoder DRAG, which learns a latent space representation that covers the complete state space from pixel inputs in online reinforcement learning, thereby enhancing the exploration and control performance of goal-oriented RL.
Imagine360: Immersive 360 Video Generation from Perspective Anchor
Jing Tan (Chinese University of Hong Kong), Dahua Lin (Shanghai Artificial Intelligence Laboratory)
GenerationData SynthesisPose EstimationDiffusion modelVideo
🎯 What it does: Convert ordinary perspective videos into complete 360° panoramic videos, providing an immersive panoramic viewing experience.
Imagined Autocurricula
Ahmet H. Güzel (University College London), Jack Parker-Holder (University College London)
Reinforcement LearningDiffusion modelWorld ModelTabularBenchmark
🎯 What it does: Train a diffusion-based world model using offline data to generate an 'imagined' environment, and build an adaptive Imagined Self-Curriculum (IMAC) using Priority Level Replay (PLR) to train reinforcement learning agents that can generalize to unseen levels.
Imbalances in Neurosymbolic Learning: Characterization and Mitigating Strategies
Efthymia Tsamoura (Huawei Labs), Dan Roth (University of Pennsylvania)
Image
🎯 What it does: This paper studies the learning imbalance problem under hidden labels in neural symbolic learning (NSL) and proposes theoretical analysis along with practical mitigation methods for training and testing.
Imitation Beyond Expectation Using Pluralistic Stochastic Dominance
Ali Farajzadeh (University of Illinois Chicago), Brian D Ziebart
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningTabular
🎯 What it does: A multi-modal imitation learning method is proposed—Multi-dimensional Stochastic Dominance (PSD), which ensures that the reward distribution of the imitation policy is probabilistically superior to the demonstration distribution under all possible convex combinations of reward functions.
Imitation Learning with Temporal Logic Constraints
Zining Fan (Rutgers University), He Zhu (Rutgers University)
Robotic IntelligenceReinforcement LearningGenerative Adversarial NetworkSequential
🎯 What it does: A demonstration-based LTL constraint reinforcement learning framework TiLoIL is proposed, which guides the agent to frequently visit LDBA accepting states in infinite-horizon tasks through piecewise imitation learning, enhancing the probability of task satisfaction.
Impact of Dataset Properties on Membership Inference Vulnerability of Deep Transfer Learning
Marlon Tobaben (University of Helsinki), Antti Honkela (University of Helsinki)
Adversarial AttackTransformerSupervised Fine-TuningImage
🎯 What it does: This study investigates the impact of dataset attributes (number of samples per class, number of classes) on the susceptibility of membership inference attacks (MIA) in deep transfer learning. It theoretically derives and experimentally verifies that the susceptibility to MIA decreases according to a power law as the number of samples per class increases.
Impact of Layer Norm on Memorization and Generalization in Transformers
Rishi Singhal (North Carolina State University), Jung-Eun Kim (North Carolina State University)
ClassificationTransformerImageText
🎯 What it does: This study investigates the impact of LayerNorm (LN) on memorization and learning in Pre-LN and Post-LN Transformers, explores the effect of removing LN parameters on model performance, and explains the different roles of LN in the two architectures through gradient analysis.
IMPACT: Irregular Multi-Patch Adversarial Composition Based on Two‑Phase Optimization
Zenghui Yang (Beijing University of Posts and Telecommunications), Tianle Zhang (Beijing University of Posts and Telecommunications)
OptimizationAdversarial AttackConvolutional Neural NetworkTransformerGenerative Adversarial NetworkImage
🎯 What it does: Proposes the IMPACT framework, which utilizes gradient-free evolutionary algorithms to jointly optimize the shape, position, number, and content of multiple patches, generating irregular and printable adversarial patches.
Impartial Selection with Predictions
Javier Cembrano (Max Planck Institute for Computer Science), Max Klimm (Technische Universität Berlin)
🎯 What it does: This paper studies how to improve the election effectiveness of top nominees by incorporating predictive information into an 'impartial' election mechanism while ensuring fairness.
Implicit Bias of Spectral Descent and Muon on Multiclass Separable Data
Chen Fan (University of British Columbia), Christos Thrampoulidis (University of British Columbia)
ClassificationOptimizationTabular
🎯 What it does: This paper studies the implicit optimization bias of spectral gradient descent (Spectral-GD) and Muon under cross-entropy loss in multi-class linear classification, and provides the corresponding marginal convergence rates.
Implicit Generative Property Enhancer
Pedro O. Pinheiro (Prescient Design), Natasa Tagasovska
GenerationOptimizationDrug DiscoveryAuto EncoderBiomedical Data
🎯 What it does: Unsupervised generation of attribute enhancement using matched data
Implicit Modeling for Transferability Estimation of Vision Foundation Models
Yaoyan Zheng (Beihang University), Di Huang (Beihang University)
ClassificationDomain AdaptationOptimizationImage
🎯 What it does: An Implicit Transferability Modeling (ITM) framework is proposed, which uses implicit variational approximation methods to quickly estimate the transfer performance of visual foundation models.
Implicit Reward as the Bridge: A Unified View of SFT and DPO Connections
Bo Wang (Fudan University), Xipeng Qiu (Shanghai Artificial Intelligence Laboratory)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This study explores a unified theoretical perspective of SFT and DPO in post-training of large language models, proving that SFT is actually a special case of implicit reward learning, and proposes to enhance the performance of post DPO by reducing the learning rate and using different f-divergence SFT objectives.
Implicit-ARAP: Efficient Handle-Guided Neural Field Deformation via Local Patch Meshing
Daniele Baieri (University of Milano-Bicocca), Zorah Lähner (University of Bonn)
OptimizationComputational EfficiencyNeural Radiance FieldPoint CloudMesh
🎯 What it does: This paper proposes a method for performing 'As-Rigid-As-Possible' (ARAP) deformation on neural fields using local patch grids, allowing users to control the geometric deformation of the neural implicit model via handles.
Improve Temporal Reasoning in Multimodal Large Language Models via Video Contrastive Decoding
Daiqing Qi (University of Virginia), Sheng Li (University of Virginia)
Large Language ModelVision Language ModelContrastive LearningVideoMultimodality
🎯 What it does: A video LLM temporal contrastive decoding method is proposed that does not require additional training. By inducing temporal distortion on key frames, the model generates temporally irrelevant errors, which are suppressed during contrastive decoding, thereby enhancing the video temporal reasoning ability.
Improved Algorithms for Fair Matroid Submodular Maximization
Sepideh Mahabadi (Microsoft Research), Jakub Tarnawski (Microsoft Research)
Recommendation SystemOptimizationGraphTabular
🎯 What it does: This paper proposes a randomized algorithm and a deterministic algorithm to maximize a monotone submodular function under the constraints of base matroid and group fairness, with only a loss of a tunable parameter ε on the fairness lower bound constraint; it also provides a deterministic approximation method under two matroid intersection constraints.
Improved Algorithms for Overlapping and Robust Clustering of Edge-Colored Hypergraphs: An LP-Based Combinatorial Approach
Changyeol Lee (Yonsei University), Hyung-Chan An (Yonsei University)
OptimizationComputational EfficiencyGraph
🎯 What it does: A combinatorial algorithm based on linear programming is proposed for overlapping and robust clustering edge coloring hypergraphs, addressing the limitations of traditional edge coloring clustering.
Improved Approximation Algorithms for Chromatic and Pseudometric-Weighted Correlation Clustering
Chenglin Fan (Seoul National University), Euiwoong Lee (University of Michigan)
OptimizationGraph
🎯 What it does: Two new approximation algorithms are proposed for the pseudometric-weighted correlation clustering (pseudometric-weighted CC) and chromatic correlation clustering (CCC) problems, along with improved approximation ratios.
Improved Balanced Classification with Theoretically Grounded Loss Functions
Corinna Cortes (Google Research), Yutao Zhong (Google Research)
ClassificationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes and studies two new proxy loss functions—Generalized Logit-Adjusted (GLA) and Generalized Class-Aware (GCA)—to achieve balanced loss in multi-class classification and overcome the class imbalance problem.
Improved Best-of-Both-Worlds Regret for Bandits with Delayed Feedback
Ofir Schlisselberg (Tel Aviv University), Yishay Mansour (Google Research)
OptimizationReinforcement Learning
🎯 What it does: A new algorithm is proposed that simultaneously achieves approximately optimal stochastic and adversarial performance in a multi-armed bandit environment with arbitrary delayed feedback.
Improved Bounds for Swap Multicalibration and Swap Omniprediction
Haipeng Luo (University of Southern California), Vatsal Sharan (University of Southern California)
Optimization
🎯 What it does: This paper proposes an online algorithm that achieves better error bounds on the swap multicalibration and swap omniprediction problems, providing an ℓ₂-swap multicalibration error of O(T^{1/3}) for linear predictors, which leads to faster convergence rates for ℓ₁-swap multicalibration and swap omniprediction.
Improved Confidence Regions and Optimal Algorithms for Online and Offline Linear MNL Bandits
Yuxuan Han (New York University), Zhengyuan Zhou (New York University)
Optimization
🎯 What it does: This paper addresses the data-driven combinatorial optimization problem in the Linear Multinomial Logit (MNL) model, proposing improved confidence intervals, offline and online algorithms, and providing approximate optimal sample complexity and return guarantees.
Improved Regret and Contextual Linear Extension for Pandora's Box and Prophet Inequality
Junyan Liu (University of Washington), Lillian J. Ratliff (University of Washington)
Optimization
🎯 What it does: This paper proposes an online learning algorithm for the Pandora's Box problem and Prophet Inequality under unknown reward distributions, primarily achieving approximately optimal cumulative utility and providing corresponding theoretical upper bounds.
Improved Regret Bounds for Gaussian Process Upper Confidence Bound in Bayesian Optimization
Shogo Iwazaki (LY Corporation)
Optimization
🎯 What it does: This paper studies the Bayesian optimization problem and proposes an improved method for minimizing cumulative regret under known Gaussian processes, particularly analyzing the cumulative regret of the Gaussian Process Upper Confidence Bound (GP-UCB) algorithm.
Improved Regret Bounds for Linear Bandits with Heavy-Tailed Rewards
Artin Tajdini (University of Washington), Kevin Jamieson (University of Washington)
🎯 What it does: Proposed and analyzed the lower and upper bounds of linear bandit problems with a finite (1+ϵ)-th moment in the reward distribution, providing improved lower and upper bounds;
Improved Representation Steering for Language Models
Zhengxuan Wu (Stanford University), Christopher Potts (Stanford University)
OptimizationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: A reference-independent bidirectional preference optimization objective, RePS, is proposed for fine control of language model representations, which can both guide concept generation and suppress concepts.
Improved Robust Estimation for Erdős-Rényi Graphs: The Sparse Regime and Optimal Breakdown Point
Hongjie Chen (ETH Zurich), Stefan Tiegel (ETH Zurich)
Graph Neural NetworkGraph
🎯 What it does: Developed the first polynomial-time algorithm that can achieve near-optimal edge density estimation in node-corrupted Erdős‑Rényi graphs;
Improved Scaling Laws in Linear Regression via Data Reuse
Licong Lin (University of California Berkeley), Peter Bartlett
OptimizationTabular
🎯 What it does: The study investigates the use of multiple data reuse (multi-round SGD) in linear regression and derives the scaling law of its test error;
Improved Training Technique for Shortcut Models
Anh Nguyen (Qualcomm AI Research), Anh Tuan Tran (Qualcomm AI Research)
GenerationData SynthesisDiffusion modelFlow-based ModelImage
🎯 What it does: Overall improvements to the Shortcut Model enable high-quality sampling from one step to multiple steps on a single network.