ICLR 2026 Papers with AI Summaries
International Conference on Learning Representations · 5356 papers
→ ICLR 2026 papers with code (2207)
Each paper below shows an AI-generated one-line summary. Get the full 6-part summary (innovation, method, data, results, limitations) and search all 5356 ICLR 2026 papers by keyword, author or institution —
free trial on arXivSub.
(U)NFV: (Un)Supervised Neural Finite Volume Methods for Solving Hyperbolic PDEs
Nathan Lichtlé (University of California, Berkeley), Alexandre M Bayen (University of California, Berkeley)
Convolutional Neural NetworkTime SeriesPhysics Related
🎯 What it does: Proposed Neural Finite Volume (NFV) and its unsupervised variant UNFV, which leverage neural networks to learn numerical fluxes within a conservative finite volume framework, enabling efficient approximation of solutions to one-dimensional supersonic PDEs without compromising conservation.
``Noisier'’ Noise Contrastive Estimation is (Almost) Maximum Likelihood
Peiyu Yu (University of California, Los Angeles), Ying Nian Wu (University of California, Los Angeles)
GenerationOptimizationDiffusion modelContrastive LearningImageTabular
🎯 What it does: The paper proposes an improved noise contrastive estimation method called 'Noisier NCE,' which amplifies the scale of the noise distribution to make its gradient approach maximum likelihood estimation, thereby addressing the convergence difficulties of traditional NCE when there is a large distribution gap;
<SO$G_k$>: One LLM Token for Explicit Graph Structural Understanding
Jingyao Wu (Shanghai Jiao Tong University), Chenghu Zhou (IGSNRR, Chinese Academy of Sciences)
ClassificationRepresentation LearningDrug DiscoveryGraph Neural NetworkLarge Language ModelSupervised Fine-TuningAuto EncoderContrastive LearningTextGraph
🎯 What it does: This paper proposes using a special structural marker <SO G k> to directly map graph structures into the LLM's vocabulary, achieving precise and concise representation of graph structures, and aligning markers through hybrid structural QA.
$\alpha$-DPO: Robust Preference Alignment for Diffusion Models via $\alpha$ Divergence
Yang Li (New Laboratory of Pattern Recognition, MAIS, CASIA), Jing Dong (New Laboratory of Pattern Recognition, MAIS, CASIA)
GenerationReinforcement Learning from Human FeedbackReinforcement LearningDiffusion modelImageTextMultimodality
🎯 What it does: This paper proposes α-DPO, a direct preference optimization method based on α-divergence, for robustly aligning diffusion models with human preferences under noisy labels.
$\boldsymbol{\partial^\infty}$-Grid: A Neural Differential Equation Solver with Differentiable Feature Grids
Navami Kairanda (Max Planck Institute for Informatics), Vladislav Golyanik (Max Planck Institute for Informatics)
OptimizationComputational EfficiencyImageMeshPhysics Related
🎯 What it does: Proposed a differentiable feature grid called ∂∞-Grid for efficiently solving partial differential equations.
$\ell_1$ Latent Distance based Continuous-time Graph Representation
Zhao-Rong Lai, Ziliang Chen (Peng Cheng Laboratory)
OptimizationRepresentation LearningGraph
🎯 What it does: Propose a continuous-time graph representation model based on ℓ1 distance, named ℓ1 LD-CTGR, which replaces the squared ℓ2 distance with ℓ1 implicit space distance to address the violation of triangle inequality, and provides a closed-form piecewise exponential integral for the hazard function.
$\mathbf{Li_2}$: A Framework on Dynamics of Feature Emergence and Delayed Generalization
Yuandong Tian (Meta Superintelligence Labs)
OptimizationRepresentation Learning
🎯 What it does: Studied the gradient dynamics of the 'grokking' phenomenon in two-layer networks and proposed the Li² framework, dividing the learning process into three stages: lazy learning, independent feature learning, and interactive feature learning, elucidating how features emerge from gradient signals and lead to sudden generalization.
$\mu$LO: Compute-Efficient Meta-Generalization of Learned Optimizers
Benjamin Thérien (Université de Montréal), Eugene Belilovsky (Mila - Quebec AI Institute)
OptimizationComputational EfficiencyMeta LearningImageText
🎯 What it does: This paper investigates the generalization ability of learning optimizers (LO) when facing widths, depths, and training times that exceed the meta-training range, and proposes migrating LO to maximum update parameterization (θ P), significantly enhancing its meta-generalization performance.
$\nabla$-Reasoner: LLM Reasoning via Test-Time Gradient Descent in Latent Space
Peihao Wang (University of Texas at Austin), Zhangyang Wang (University of Texas at Austin)
OptimizationComputational EfficiencyTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Propose a framework called ∇-Reasoner, which optimizes the generation strategy during inference by performing gradient descent on the token logits of the LLM output, with the core being Differentiable Textual Optimization (DTO);
$\pi^3$: Permutation-Equivariant Visual Geometry Learning
Yifan Wang (Shanghai Jiao Tong University), Tong He (Shanghai AI Laboratory)
Pose EstimationDepth EstimationTransformerContrastive LearningImageVideoPoint Cloud
🎯 What it does: Propose a visual geometry reconstruction model π3 that does not require a reference view, utilizing a fully permutation-equivariant network to predict camera poses and local point clouds.
$\textbf{Re}^{2}$: Unlocking LLM Reasoning via Reinforcement Learning with Re-solving
Pinzheng Wang (Soochow University), Min Zhang (Soochow University)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes a reinforcement learning framework Re², enabling LLMs to self-judge and restart problem-solving during inference, thereby improving reasoning accuracy.
$\textit{MADFormer}$: Mixed Autoregressive and Diffusion Transformers for Continuous Image Generation
Junhao Chen (Tsinghua University), Xiaochuang Han (University of Washington)
GenerationTransformerDiffusion modelAuto EncoderImageMultimodalityBenchmark
🎯 What it does: Propose a hybrid autoregressive and diffusion Transformer called MADFormer for generating high-resolution images in a continuous latent space, and systematically explore the impact of AR and diffusion allocation at hierarchical and block levels on generation quality and efficiency through experiments.
$AutoDrive\text{-}P^3$: Unified Chain of Perception–Prediction–Planning Thought via Reinforcement Fine-Tuning
Yuqi Ye (Peking University), Wei Gao (Peking University)
Autonomous DrivingSupervised Fine-TuningReinforcement LearningVision Language ModelChain-of-Thought
🎯 What it does: Proposing the AutoDrive P-3 framework to achieve end-to-end driving decision-making through chain-of-thought reasoning that integrates perception, prediction, and planning.
$p\textrm{-less}$ Sampling: A Robust Hyperparameter-Free Approach for LLM Decoding
Runyan Tan (Thoughtworks), Phillip Howard (Thoughtworks)
GenerationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose a parameter-free truncation sampling method called p-less, which dynamically determines the truncation threshold through information theory to generate high-quality text.
$PhyWorldBench$: A Comprehensive Evaluation of Physical Realism in Text-to-Video Models
Jing Gu (University of California, Santa Cruz), Xin Eric Wang (University of California, Santa Cruz)
GenerationLarge Language ModelPrompt EngineeringVideoTextBenchmarkPhysics RelatedChain-of-Thought
🎯 What it does: Proposed PhyWorldBench, a physics realism evaluation benchmark for text-to-video models, containing 10 major physics categories, 7 sub-scenes, and 3 prompt types, with basic and key criteria established through human-machine review; simultaneously developed a context-aware prompting-based zero-shot evaluation method for MLLMs.
3D Aware Region Prompted Vision Language Model
An-Chieh Cheng (UC San Diego), Sifei Liu (NVIDIA)
Depth EstimationTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageVideoTextPoint Cloud
🎯 What it does: Proposes SR-3D, a unified 3D-perception vision-language model that can be jointly trained with single-view and multi-view data to achieve efficient spatial reasoning and region-level interaction.
3D Scene Prompting for Scene-Consistent Camera-Controllable Video Generation
JoungBin Lee (KAIST AI), Seungryong Kim (KAIST AI)
GenerationPrompt EngineeringDiffusion modelSimultaneous Localization and MappingOptical FlowVideoPoint Cloud
🎯 What it does: Proposes the 3DScenePrompt framework, which can generate subsequent video segments from input videos of arbitrary length, while maintaining scene consistency and achieving precise camera control.
3D-aware Disentangled Representation for Compositional Reinforcement Learning
Sungbin Mun (Carnegie Mellon University), Young Min Kim (Carnegie Mellon University)
Representation LearningTransformerReinforcement LearningImage
🎯 What it does: Proposed a 3D block-slot representation method, combining multi-view Transformer with attribute factorization, and designed a block Transformer strategy for goal-conditioned reinforcement learning;
3DCS: Datasets and Benchmark for Evaluating Conformational Sensitivity in Molecular Representations
Xi Wang (New York University), Shenji Wan
Drug DiscoveryProtein Structure PredictionGraphBenchmark
🎯 What it does: This work proposes the 3DCS benchmark, systematically evaluating the sensitivity of molecular representations to conformational changes within molecules, including geometry, chirality, and energy.
3DGEER: 3D Gaussian Rendering Made Exact and Efficient for Generic Cameras
Zixun Huang (Bosch Research North America and Bosch Center for AI), Liu Ren (Bosch Research North America and Bosch Center for AI)
Computational EfficiencyGaussian SplattingPoint CloudMesh
🎯 What it does: Propose 3DGEER, a geometrically accurate and efficient 3D Gaussian rendering framework that can achieve real-time rendering under any field of view (including wide field-of-view and fisheye cameras).
3DSMT: A Hybrid Spiking Mamba-Transformer for Point Cloud Analysis
Zhouzhiming, Ajmal Saeed Mian
ClassificationSegmentationSpiking Neural NetworkTransformerPoint Cloud
🎯 What it does: This paper proposes a hybrid pulsed Mamba-Transformer (3DSMT) network for point cloud analysis, combining pulsed local offset attention and pulsed Mamba blocks to efficiently extract local geometric details and global context.
A Balanced Neuro-Symbolic Approach for Commonsense Abductive Logic
Joseph Cotnareanu (McGill University), Mark Coates (Huawei Noah's Ark Lab)
Explainability and InterpretabilityTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Proposes the ARGOS framework, which leverages interaction between LLM and SAT solvers to iteratively complete missing common sense, enabling logical reasoning for logical problems
A Bayesian Nonparametric Framework For Learning Disentangled Representations
Vaishnavi S Patil (University of Maryland), Joseph JaJa (University of Maryland)
Representation LearningAuto EncoderImage
🎯 What it does: Propose an autoencoder framework based on Bayesian nonparametric hierarchical mixed prior to learn interpretable and disentangled representations
A Bayesian Nonparametric Framework for Private, Fair, and Balanced Tabular Data Synthesis
Forough Fazeli-Asl (University of Alberta), Bei Jiang (University of Alberta)
Data SynthesisSafty and PrivacyGenerative Adversarial NetworkTabular
🎯 What it does: This paper proposes a conditional generative model based on Bayesian nonparametric learning (BNPL) to synthesize tabular data while ensuring privacy, fairness, and class balance.
A Benchmark for Deep Information Synthesis
Debjit Paul (Huawei Noah's Ark Lab), Gerasimos Lampouras (Huawei Noah's Ark Lab)
Data SynthesisLarge Language ModelAgentic AITextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed the DEEPSYNTH benchmark, comprising 120 cross-border, multi-source, verifiable complex information synthesis tasks, evaluating the planning, retrieval, reasoning, and integration capabilities of LLMs and intelligent agents.
A Biologically Plausible Dense Associative Memory with Exponential Capacity
Mohadeseh Shafiei Kafraj (University College London), Peter E. Latham (University College London)
RetrievalRepresentation LearningImage
🎯 What it does: Propose a biologically plausible two-layer dense associative memory network that uses threshold nonlinearity to achieve distributed hidden representations, obtaining exponential storage capacity; meanwhile, the corresponding learning rules are provided.
A Block Coordinate Descent Method for Nonsmooth Composite Optimization under Orthogonality Constraints
Ganzhao Yuan (Shenzhen University of Advanced Technology)
Optimization
🎯 What it does: This paper proposes an algorithm called OBCD based on block coordinate descent for non-smooth composite optimization problems with orthogonality constraints, updating only k rows of variables in each iteration and globally solving subproblems in low-dimensional subspaces.
A Brain Graph Foundation Model: Pre-Training and Prompt-Tuning across Broad Atlases and Disorders
Xinxu Wei (Lehigh University), Yu Zhang (Stanford University)
ClassificationRepresentation LearningMeta LearningGraph Neural NetworkTransformerPrompt EngineeringAuto EncoderContrastive LearningMultimodalityGraphBiomedical DataMagnetic Resonance ImagingAlzheimer's Disease
🎯 What it does: Propose a brain benchmark model called BrainGFM based on graph neural networks, which is pre-trained on a large-scale multi-modal dataset consisting of 27 fMRI datasets, 25 common brain diseases, and 8 brain parcellation templates. Subsequently, few-shot and zero-shot transfer is achieved through graph prompts and language prompts.
A Brain-Inspired Gating Mechanism Unlocks Robust Computation in Spiking Neural Networks
Qianyi Bai (Tianjin University), Qiang Yu (Tianjin University)
RecognitionAdversarial AttackSpiking Neural NetworkStochastic Differential EquationAudio
🎯 What it does: Propose a spiking neuron model called Dynamic Gated Neuron (DGN) based on dynamic conductance, and verify its advantages in speech recognition and robustness tasks within multi-layer spiking neural networks.
A Cognitive Process-Inspired Architecture for Subject-Agnostic Brain Visual Decoding
Jingyu Lu (Hong Kong University of Science and Technology), Xiaomeng Li (Shenzhen Loop Area Institute)
ClassificationImage TranslationSegmentationGenerationTransformerDiffusion modelContrastive LearningBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Proposed a novel fMRI-to-video decoding framework called VCFLOW, which does not require additional training for new subjects. It utilizes a dual-stream structure of the brain's visual cortex (early, dorsal, and ventral streams) to extract multi-level cognitive features and align and reconstruct them.
A Comprehensive Information-Decomposition Analysis of Large Vision-Language Models
Lixin Xiu (University of Tokyo), Hideki Nakayama (University of Tokyo)
Explainability and InterpretabilityRepresentation LearningMultimodalityBenchmark
🎯 What it does: This paper constructs an unsupervised evaluation framework based on Partial Information Decomposition (PID), performing information spectrum decomposition (redundant, visual-unique, language-unique, synergistic information) on the decision-making processes of 26 large vision-language models (LVLM) across four multiple-choice VQA benchmarks, and systematically analyzes the information flow from three perspectives: model hierarchy, task dimensions, and training stages.
A Convergence Analysis of Adaptive Optimizers under Floating-point Quantization
Xuan Tang (University of Hong Kong), Difan Zou (University of Hong Kong)
OptimizationImageText
🎯 What it does: Propose a convergence theory framework for adaptive optimizers (Adam, Muon) under floating-point quantization, providing convergence rates when gradients, weights, momentum, and second moments are all quantized.
A cross-species neural foundation model for end-to-end speech decoding
Yizi Zhang (Columbia University), Liam Paninski (Stanford University)
Representation LearningTransformerLarge Language ModelContrastive LearningTime SeriesBiomedical Data
🎯 What it does: Built an end-to-end brain-to-text framework called BIT, directly translating human and monkey EEG signals into complete sentences, supporting cross-species and cross-task transfer learning.
A Dense Subset Index for Collective Query Coverage
Kartik Nair (Carnegie Mellon University), Abir De (IIT Bombay)
RetrievalTextTabularBenchmark
🎯 What it does: This paper proposes a new 'Dense Subset Index' (DISCO), aiming to achieve multi-content collaborative retrieval through set cover approaches, solving the problem that traditional single-document retrieval cannot cover scenarios such as multi-step reasoning and table retrieval.
A Derandomization Framework for Structure Discovery: Applications in Neural Networks and Beyond
Nikos Tsikouras (National and Kapodistrian University of Athens), Christos Tzamos (National and Kapodistrian University of Athens)
OptimizationRepresentation LearningGraph
🎯 What it does: Proposed a structure discovery framework based on derandomization, proving that under weak assumptions neural networks can automatically learn low-rank first-layer weights via optimization methods such as gradient descent, thereby achieving structural discovery.
A Fair Bayesian Inference through Matched Gibbs Posterior
Jihu Lee (Seoul National University), Yongdai Kim (Seoul National University)
ClassificationSafty and PrivacyComputational EfficiencyImageTextTabular
🎯 What it does: Propose matching Gibbs posterior as a proxy distribution to achieve group fair Bayesian inference, addressing the challenge of posterior computation under demographic balance constraints in deep models.
A Fano-Style Accuracy Upper Bound for LLM Single-Pass Reasoning in Multi-Hop QA
Kaiyang Wan (INSAIT), Xiuying Chen (MBZUAI)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: An information-theoretic analysis of the information bottleneck in single-inference LLMs for multi-hop question answering, proposing a Fano-style accuracy upper bound and discovering the 'accuracy cliff,' leading to the design of the multi-call framework InfoQA.
A Faster Parameter-Free Regret Matching Algorithm
Linjian Meng (Shanghai Artificial Intelligence Laboratory), Yang Gao (Nanjing University)
OptimizationReinforcement Learning
🎯 What it does: Propose a new parameter-free smooth Regret Matching+ variant called MI-SPRM+, which guarantees a theoretical convergence rate of O(1/T) while achieving parameter-free updates;
A Federated Generalized Expectation-Maximization Algorithm for Mixture Models with an Unknown Number of Components
Michael Ibrahim (Georgia Institute of Technology), Weijun Xie (Georgia Institute of Technology)
OptimizationFederated LearningImageTabular
🎯 What it does: Propose FedGEM: a federated generalized expectation maximization (GEM) algorithm that can train a mixture model without knowing the global number of clusters and when the local cluster sets across clients are inconsistent, utilizing uncertainty sets to detect cross-client cluster overlaps and aggregating them on the server side;
A foundation model with multi-variate parallel attention to generate neuronal activity
Francesco S. Carzaniga (IBM Research), Abbas Rahimi (IBM Research)
GenerationData SynthesisTransformerContrastive LearningTime SeriesBiomedical Data
🎯 What it does: Proposed a new multi-variate parallel attention mechanism (MVPA) and built MVPFormer, a foundational model for generative pre-training on large-scale iEEG data across various channel configurations.
A Framework for Studying AI Agent Behavior: Evidence from Consumer Choice Experiments
Manuel Cherep (MIT), Nikhil Singh (Dartmouth College)
TransformerLarge Language ModelAgentic AIPrompt EngineeringTextTabularFinance RelatedChain-of-Thought
🎯 What it does: Propose the ABXLAB framework, systematically evaluating LLM agents' decision-making behaviors in real-world e-commerce environments by controlling attributes such as price, ratings, sequence, and implications, and quantifying their biases through over 80k experiments;
A Function-Centric Graph Neural Network Approach for Predicting Electron Densities
Manuel V. Klockow (Heidelberg University), Fred A. Hamprecht (Heidelberg University)
Graph Neural NetworkGraphPhysics Related
🎯 What it does: Built an equivariant graph neural network based on the quadratic expansion of the ground-state electron density to predict the electron density under molecular geometry.
A General Framework for Black-Box Attacks Under Cost Asymmetry
Mahdi Salmani (University of Southern California), Seyed-Mohsen Moosavi-Dezfooli (Apple)
OptimizationAdversarial AttackConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes a general framework for performing decision-based black-box adversarial attacks under scenarios of asymmetric query costs;
A General Spatio-Temporal Backbone with Scalable Contextual Pattern Bank for Urban Continual Forecasting
Aoyu Liu (Tongji University), Yaying Zhang (Tongji University)
Graph Neural NetworkPrompt EngineeringTime Series
🎯 What it does: Proposed a continuous spatiotemporal prediction framework named STBP, capable of performing efficient prediction in real-world streaming data environments where nodes continuously expand and graph structures and distributions experience persistent drift;
A Generalized Geometric Theoretical Framework of Centroid Discriminant Analysis for Linear Classification of Multi-dimensional Data
Yue Wu (Tsinghua University), Carlo Vittorio Cannistraci (Tsinghua University)
ClassificationHyperparameter SearchImageBiomedical Data
🎯 What it does: Proposed the Geometric Discriminant Analysis (GDA) framework, and designed an efficient linear classifier called Center Point Discriminant Analysis (CDA), as well as its kernelized nonlinear extension.
A Genetic Algorithm for Navigating Synthesizable Molecular Spaces
Alston Lo (Massachusetts Institute of Technology), Wojciech Matusik (Massachusetts Institute of Technology)
Drug DiscoveryBiomedical Data
🎯 What it does: Propose SynGA, a genetic algorithm that directly operates on synthetic routes (synthesis trees), further enhancing search efficiency with a lightweight machine learning filter (MLP or neural additive model), and finally embedding it into Bayesian optimization to obtain SynGBO, achieving the design of synthetically feasible molecules and property optimization.
A Graph Meta-Network for Learning on Kolmogorov–Arnold Networks
Guy Bar-Shalom (Technion), Haggai Maron (Technion)
ClassificationGraph Neural NetworkImage
🎯 What it does: Proposes WS-KAN, a weight-space graph primitive network for Kolmogorov-Arnold networks (KAN), designed to directly learn and infer KAN parameters.
A Guardrail for Safety Preservation: When Safety-Sensitive Subspace Meets Harmful-Resistant Null-Space
Bingjie Zhang (Jilin University), Bernard Ghanem (Jilin University)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose the GuardSpace framework, which protects safety alignment during LLM fine-tuning through safe-sensitive subspace initialization and harm-resistant zero-space projection.
A Hidden Semantic Bottleneck in Conditional Embeddings of Diffusion Transformers
Trung X. Pham (Korea Advanced Institute of Science and Technology), Chang D. Yoo (Korea Advanced Institute of Science and Technology)
GenerationTransformerDiffusion modelImageVideoMultimodalityAudio
🎯 What it does: This paper systematically analyzes conditional embeddings in diffusion Transformers, revealing their extremely high cosine similarity and sparse features, and demonstrates that up to 66% of dimensions can be pruned without affecting generation quality while preserving key information dimensions.
A Hierarchical Circuit Symbolic Discovery Framework for Efficient Logic Optimization
Yinqi Bai (University of Science and Technology of China), Jianye HAO
OptimizationExplainability and InterpretabilityComputational EfficiencyGraph Neural NetworkTransformerReinforcement LearningGraph
🎯 What it does: This paper proposes a hierarchical circuit symbolic discovery framework (HIS), which identifies invalid subgraphs in logic optimization by learning lightweight, interpretable symbolic functions, thereby achieving efficient pruning.
A High Quality Dataset and Reliable Evaluation for Interleaved Image-Text Generation
Yukang Feng (Nankai University), Kaipeng Zhang (Shanghai Innovation Institute)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelDiffusion modelMultimodalityBenchmark
🎯 What it does: This paper constructs a large-scale, high-quality interleaved image-text question answering dataset called InterSyn, and proposes an automatic iterative evaluation-refinement (SEIR) process for generating and optimizing samples; meanwhile, it designs a four-dimensional evaluator called SynJudge to systematically assess the completeness of generated text, completeness of images, image quality, and image-text synergy.
A Joint Diffusion Model with Pre-Trained Priors for RNA Sequence-Structure Co-Design
Xiner Li (Texas A&M University), Shuiwang Ji (Texas A&M University)
Protein Structure PredictionReinforcement LearningDiffusion modelBiomedical Data
🎯 What it does: Developed RiboDiff, a joint discrete/continuous diffusion model based on the pre-trained RoseTTAFold2NA, for co-designing RNA sequences and 3D structures, supporting single RNA, RNA-protein complexes, and RNA binding under protein conditions.
A Law of Data Reconstruction for Random Features (And Beyond)
Leonardo Iurada (Politecnico di Torino), Marco Mondelli (Institute of Science and Technology Austria)
Convolutional Neural NetworkImage
🎯 What it does: Proposed the data reconstruction law: when the number of parameters p in the random feature model satisfies p ≫ d n, the complete training set can be recovered from the model parameters.
A Memory-Efficient Hierarchical Algorithm for Large-scale Optimal Transport Problems
Wenzhou Xia, Xiaoqun Zhang (Shanghai Jiao Tong University)
OptimizationImagePoint Cloud
🎯 What it does: Propose a memory-efficient large-scale optimal transport algorithm HALO based on multi-level hierarchical structure, proactive support updates, and GPU-parallel LP solving.
A Minimum Variance Path Principle for Accurate and Stable Score-Based Density Ratio Estimation
Wei Chen (South China University of Technology), Delu Zeng (South China University of Technology)
Score-based ModelMultimodalityTabularBenchmark
🎯 What it does: Propose the Minimum Variance Path (MVP) principle, derive a closed-form expression for path variance, and parameterize paths using the Kumaraswamy Mixture Model (KMM) to learn data-adaptive, low-variance probabilistic paths, thereby achieving more accurate and stable score-based density ratio estimation.
A Near-Optimal Best-of-Both-Worlds Algorithm for Federated Bandits
Zicheng Hu (East China Normal University), Cheng Chen (East China Normal University)
OptimizationFederated LearningTabular
🎯 What it does: Studied the federated multi-armed bandit problem and proposed the FEDFTRL algorithm, achieving near-optimal performance in heterogeneous feedback and distributed communication environments.
A New Approach to Controlling Linear Dynamical Systems
Anand Paresh Brahmbhatt, Elad Hazan
OptimizationTime SeriesSequential
🎯 What it does: Propose an online control algorithm (OSC) based on spectral filtering, which transforms the control problem of linear dynamic systems under adversarial perturbations and convex costs into low-dimensional feature regression, achieving efficient learning through convex relaxation and online gradient descent.
A New Initialization to Control Gradients in Sinusoidal Neural Networks
Andrea Combette (ENS de Lyon), Antoine Venaille (ENS de Lyon)
OptimizationComputational EfficiencyRepresentation LearningImageVideoPhysics RelatedAudio
🎯 What it does: A new weight and bias initialization method for networks with sine activation functions (SIREN) was designed and verified, which can maintain gradient stability as network depth increases, avoid gradient explosion or vanishing, and control the network's spectrum during early training.
A New Paradigm for Genome-wide DNA Methylation Prediction Without Methylation Input
Xiaoke Huang (University Of California Santa Cruz), Wenpin Hou (Columbia University)
TransformerTextTabularBiomedical Data
🎯 What it does: Introduces MethylProphet, a Transformer-based model that can predict genome-wide DNA methylation levels using only gene expression and DNA sequences.
A Noise is Worth Diffusion Guidance
Donghoon Ahn (UC Berkeley), Seungryong Kim (KAIST)
GenerationDiffusion modelScore-based ModelImageText
🎯 What it does: Propose the NoiseRefine method, achieving high-quality image generation in diffusion models with a single noise refinement step without guidance.
A Physics-Inspired Optimizer: Velocity Regularized Adam
Pranav Vaidhyanathan (University of Oxford), Michael A Osborne
OptimizationConvolutional Neural NetworkTransformerImageTextPhysics Related
🎯 What it does: Proposed VRAdam, a physics-inspired optimizer based on a quartic kinetic term, achieving faster convergence across various tasks.
A Primer on SO(3) Action Representations in Deep Reinforcement Learning
Martin Schuck (Technical University of Munich), Angela P. Schoellig (Technical University of Munich)
Representation LearningRobotic IntelligenceReinforcement Learning
🎯 What it does: Studied the action representation of SO(3) in deep reinforcement learning, systematically evaluated the performance of multiple parameterizations under algorithms such as PPO, SAC, and TD3, and provided practical usage guidelines.
A Probabilistic Hard Concept Bottleneck for Steerable Generative Models
María Martínez-García (Saarland University), Isabel Valera (Saarland University)
GenerationExplainability and InterpretabilityDiffusion modelAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: Propose the Variational Hard Concept Bottleneck (VHCB) layer, which maps internal representations to hard concepts using a binary VAE, achieving interpretable and controllable generation.
A Problem-Oriented Perspective and Anchor Verification for Code Optimization
Tong Ye (Zhejiang University), Wenhai Wang (Zhejiang University)
OptimizationAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Propose a problem-oriented code optimization perspective and an anchor verification framework to optimize program performance.
A Recovery Guarantee for Sparse Neural Networks
Sara Fridovich-Keil (Georgia Institute of Technology), Mert Pilanci (Stanford University)
OptimizationImage
🎯 What it does: This paper reformulates the training problem of a sparse two-layer ReLU MLP as a convexified linear measurement problem, proving that under random Gaussian inputs, the Iterative Hard Thresholding (IHT) algorithm can exactly recover sparse network weights in polynomial time;
A Relative Error-Based Evaluation Framework of Heterogeneous Treatment Effect Estimators
Jiayi Guo (Peking University), Peng Wu (Beijing Technology and Business University)
Tabular
🎯 What it does: Proposed a HTE estimator evaluation framework based on relative error, and constructed a neural network implementation that maintains robustness under deviations in the potential outcome regression model;
A Representer Theorem for Hawkes Processes via Penalized Least Squares Minimization
Hideaki Kim (NTT, Inc.), Tomoharu Iwata (NTT, Inc.)
OptimizationTime SeriesSequentialFinance Related
🎯 What it does: A penalized least squares estimation method is proposed in RKHS for the triggering kernel function of multivariate Hawkes processes, along with a corresponding representer theorem. It is proven that the optimal solution can be expressed as a linear combination of equivalent kernel functions, with all dual coefficients equal to 1, leading to a non-iterative closed-form solution.
A Resolution-Agnostic Geometric Transformer for Chromosome Modeling Using Inertial Frame
Yize Zhou (Wave Intelligence Lab), Shengchao Liu (Wave Intelligence Lab)
TransformerBiomedical Data
🎯 What it does: Propose an InertialGenome framework based on Transformer for 3D chromosome reconstruction, which utilizes an inertial frame to normalize the initial coordinates' pose and incorporates geometric-aware position encoding in the Transformer to achieve high-precision 3D chromosomal structure reconstruction across resolutions.
A Reward-Free Viewpoint on Multi-Objective Reinforcement Learning
Ying-Tu Chen (National Yang Ming Chiao Tung University), Ping-Chun Hsieh (National Yang Ming Chiao Tung University)
Reinforcement LearningBenchmark
🎯 What it does: Proposes an algorithm called MORL-FB that combines reward-free RL (RFRL) with multi-objective RL (MORL), improving sample efficiency and generalization by using RFRL's training objectives as auxiliary tasks within MORL.
A Rich Knowledge Space for Scalable Deepfake Detection
Inho Jung (Sungkyunkwan University), Simon S. Woo (Sungkyunkwan University)
Anomaly DetectionTransformerSupervised Fine-TuningVision Language ModelContrastive LearningMultimodality
🎯 What it does: By constructing the MMI-DD 3.6M multimodal deepfake dataset and training a CLIP-based SD 2 framework on it, scalable deepfake detection is achieved.
A Scalable Constant-Factor Approximation Algorithm for $W_p$ Optimal Transport
Pankaj K Agarwal, Keegan Yao (North Carolina State University)
OptimizationImage
🎯 What it does: Propose a scalable constant factor approximation algorithm that achieves a (4+ε) approximation for Wp optimal transport (including p=∞) in any metric space, along with a corresponding matching problem solution.
A Scalable Distributed Framework for Multimodal GigaVoxel Image Registration
Rohit Jena (University of Pennsylvania), James Gee (University of Pennsylvania)
OptimizationComputational EfficiencyBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Propose an expandable distributed framework named FFDP, leveraging IO-aware fused kernel and multi-GPU parallelism technologies to achieve large-scale multimodal image registration.
A Scalable Inter-edge Correlation Modeling in CopulaGNN for Link Sign Prediction
Jinkyu Sung (Seoul National University), Joonseok Lee (Seoul National University)
Computational EfficiencyRepresentation LearningGraph Neural NetworkGraphFinance Related
🎯 What it does: Propose a scalable edge correlation model named CopulaLSP for link sign prediction on signed graphs.
A Scene is Worth a Thousand Features: Feed-Forward Camera Localization from a Collection of Image Features
Axel Barroso-Laguna (Niantic Spatial), Eric Brachmann (Niantic Spatial)
Pose EstimationComputational EfficiencyTransformerSimultaneous Localization and MappingImage
🎯 What it does: Propose the FastForward method, which enables rapid map construction and image localization with only a single forward inference.
A Schrödinger Eigenfunction Method for Long-Horizon Stochastic Optimal Control
Louis Claeys (ETH Zürich), Niao He (ETH Zürich)
OptimizationGraphTabularPhysics RelatedStochastic Differential Equation
🎯 What it does: Propose mapping long-term stochastic optimal control problems to the eigenvalue problem of the Schrödinger operator, solving optimal control using its discrete spectrum;
A Sharp KL Convergence Analysis for Diffusion Models under Minimal Assumptions
Nishant Jain (University of Illinois Urbana-Champaign), Tong Zhang (University of Illinois Urbana-Champaign)
GenerationDiffusion modelScore-based ModelOrdinary Differential Equation
🎯 What it does: Perform KL convergence analysis of the inverse process of diffusion models under minimal assumptions, propose an improved Exponential Integrator discretization scheme, and prove the convergence rate in linear dimensions.
A Simple "Motivation" Can Enhance Reinforcement Finetuning of Large Reasoning Models
Junjie Zhang (Nanyang Technological University), Dacheng Tao (Alibaba Group)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: In reinforcement learning fine-tuning, injecting task reward specifications into prompts enables large language models to 'understand the rules' during training, thereby enhancing reasoning ability.
A Single Architecture for Representing Invariance Under Any Space Group
Cindy Zhang, Ryan P Adams
TransformerGraphPhysics Related
🎯 What it does: Propose a single adaptive crystal Fourier Transformer that enforces strict invariance for any of the 230 space groups.
A Spectral-Grassmann Wasserstein metric for operator representations of dynamical systems
Thibaut Germain (Ecole Polytechnique), Karim Lounici (Ecole Polytechnique)
OptimizationRepresentation LearningTime SeriesPhysics Related
🎯 What it does: Proposed a Spectral-Grassmann Wasserstein metric (SGOT) based on optimal transport, representing the Koopman/transfer operator of dynamical systems through the joint distribution of its eigenvalues and eigensubspaces, and utilizing this metric to achieve system comparison, clustering, dimensionality reduction, and system interpolation and averaging (Basel).
A State-Transition Framework for Efficient LLM Reasoning
Liang Zhang (Xiamen University), Jinsong Su (Alibaba International Digital Commerce)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Propose an efficient framework that models the reasoning process of large language models as state transitions, replacing the original attention mechanism with a hybrid attention (Softmax + linear attention). By utilizing a state matrix to store historical reasoning information, the framework significantly improves reasoning speed and efficiency while maintaining the full length of Chain-of-Thought (CoT).
A Statistical Benchmark for Diffusion-Posterior-Sampling Algorithms
Martin Zach (École Polytechnique Fédérale de Lausanne), Michael Unser (École Polytechnique Fédérale de Lausanne)
Diffusion modelBenchmarkStochastic Differential Equation
🎯 What it does: Proposes a statistical benchmark for diffusion posterior sampling algorithms in linear inverse problems, using discrete Lévy processes to generate posterior samples that can be exactly solved by Gibbs sampling as a gold standard.
A Statistical Learning Perspective on Semi-dual Adversarial Neural Optimal Transport Solvers
Roman Tarasov, Alexander Korotin (Applied AI Institute)
OptimizationGenerative Adversarial NetworkTabularBenchmark
🎯 What it does: This paper conducts theoretical research on a semi-dual adversarial neural optimal transport solver from a statistical learning perspective, bridging the gap in theoretical analysis of existing methods.
A Statistical Theory of Overfitting for Imbalanced Classification
Jingyang Lyu (University of Wisconsin-Madison), Yiqiao Zhong (University of Wisconsin-Madison)
ClassificationImageTextBiomedical Data
🎯 What it does: Studied the problem of high-dimensional imbalanced classification, constructed the statistical theory of linear classifiers, and revealed the truncation effect of the training set on the numerical (logit) distribution;
A Step to Decouple Optimization in 3DGS
Renjie Ding (Hunan University), Xiang Chen (Hunan University)
OptimizationComputational EfficiencyGaussian SplattingImage
🎯 What it does: This paper proposes decoupling and restructuring the optimization process of 3D Gaussian Splatting (3DGS), designing Sparse Adam, Re-State Regularization (RSR), and Decoupled Attribute Regularization (DAR), and integrating them into a new optimizer called AdamW-GS to improve optimization efficiency, controllable regularization, and automatic redundancy removal.
A Stitch in Time Saves Nine: Proactive Self-Refinement for Language Models
Jinyi Han (East China Normal University), Yanghua Xiao (East China Normal University)
OptimizationComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Developed a proactive self-refinement method (PASR) that enables large language models to actively decide whether, when, and how to improve their own outputs during the generation process;
A Structured, Tagged, and Localized Visual Question Answering Dataset with Full Sentence Answers and Scene Graphs for Chest X-ray Images
Philip Müller (Technical University of Munich), Daniel Rueckert (Technical University of Munich)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityBiomedical DataElectronic Health RecordsBenchmark
🎯 What it does: Automatically constructed a chest X-ray visual question answering dataset named CXR‑QBA containing 42 million QA pairs, featuring complete sentence answers, precise bounding boxes, and multi-level labels.
A Study of Posterior Stability in Time-Series Latent Diffusion
Yangming Li, Mihaela van der Schaar (University of Cambridge)
Diffusion modelAuto EncoderTime Series
🎯 What it does: This paper systematically analyzes the problem of posterior collapse in latent diffusion models for time series and proposes a new posterior-stable latent diffusion framework to address this issue.
A Study on PAVE Specification for Learnware
Hao-Yu Shi (Nanjing University), Zhi-Hua Zhou (Nanjing University)
RetrievalTransformerLarge Language ModelSupervised Fine-TuningImageTextBiomedical Data
🎯 What it does: Proposed a learnware specification based on parameter vectors (PAVE), which leverages parameter variations of pre-trained models to characterize model capabilities and task requirements, and implemented efficient learnware retrieval and reuse in the Learnware Dock System;
A Tale of Two Geometries: Adaptive Optimizers and Non-Euclidean Descent
Shuo Xie (Toyota Technological Institute at Chicago), Zhiyuan Li (Toyota Technological Institute at Chicago)
Optimization
🎯 What it does: This paper investigates the convergence properties of adaptive optimizers (such as Adam, AdaGrad, Shampoo) and their corresponding non-Euclidean descent methods (such as Lion, Muon) in both non-convex and convex scenarios, introducing the concepts of 'adaptive smoothness' and 'adaptive variance' and providing a unified theoretical analysis;
A tale of two tails: Preferred and anti-preferred natural stimuli in visual cortex
Rabia Gondur (Cold Spring Harbor Laboratory), Benjamin R. Cowley (Cold Spring Harbor Laboratory)
Explainability and InterpretabilityRepresentation LearningData-Centric LearningConvolutional Neural NetworkVision Language ModelImageBiomedical Data
🎯 What it does: Through experimental recordings, data-driven DNN modeling, psychological tasks, and image statistics, the study systematically validated and quantified the bimodal response distribution of macaque V4 neurons to anti-preferred stimuli, and proposed the ImageBeagle tool to efficiently search for these stimuli.
A Theoretical Analysis of Mamba’s Training Dynamics: Filtering Relevant Features for Generalization in State Space Models
Mugunthan Shandirasegaran (New Jersey Institute of Technology), Shuai Zhang (New Jersey Institute of Technology)
Explainability and InterpretabilityRepresentation LearningTransformerSequential
🎯 What it does: This paper theoretically analyzes the training dynamics of the Mamba model, proving that gradient descent can achieve convergence under two structured data scenarios (majority voting and localization), providing non-asymptotic upper bounds on sample complexity and iteration counts, and further demonstrating that gating vectors automatically focus on class-related features while suppressing irrelevant ones.
A Training-Free Framework for Long Video Understanding via Video-Query-Options Similarity
Zhirong Wu (Peking University), Peixi Peng (Peking University)
RetrievalTransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes an untrained long video understanding framework that integrates adaptive frame sampling, dynamic resolution allocation, and video-query-option similarity calculation.
A Two-Phase Deep Learning Framework for Adaptive Time-Stepping in High-Speed Flow Modeling
Jacob Helwig (Texas A&M University), Shuiwang Ji (Texas A&M University)
Convolutional Neural NetworkGraph Neural NetworkMixture of ExpertsMeshPhysics Related
🎯 What it does: Propose the ShockCast two-stage framework: the first stage uses a neural network to predict adaptive time steps, and the second stage drives the flow field evolution with these steps.
A Unification of Discrete, Gaussian, and Simplicial Diffusion
Nuria Alina Chandra (New York University), Andrew Gordon Wilson (New York University)
GenerationData SynthesisDiffusion modelImageTextBiomedical Data
🎯 What it does: Unify discrete, Gaussian, and simplex diffusion models under the Wright-Fisher evolutionary model framework, and propose a sufficient statistic parameterization, enabling the same network to be simultaneously trained and inferred across three domains.
A Unified Federated Framework for Trajectory Data Preparation via LLMs
Zhihao Zeng (Zhejiang University), Yunjun Gao (Zhejiang University)
Federated LearningSafty and PrivacyComputational EfficiencyKnowledge DistillationLarge Language ModelSupervised Fine-TuningPrompt EngineeringAuto EncoderTime Series
🎯 What it does: Propose FedTDP, a unified federated trajectory data preprocessing framework that enables multi-task preprocessing on vertically partitioned data across different regions;
A Unified Total Variation Framework for Membrane Potential Perturbation Dynamic
Zhao-Rong Lai (Jinan University), Yongsen Zheng (Nanyang Technological University)
ClassificationAdversarial AttackSpiking Neural NetworkImage
🎯 What it does: Prove that membrane potential perturbation dynamics (MPPD) is equivalent to total variation (TV), and propose a new TVℓ1 regularization framework to enhance the robustness of spiking neural networks (SNN)
A Unifying View of Coverage in Linear Off-policy Evaluation
Philip Amortila (University of California, Berkeley), Nan Jiang (University of Illinois Urbana-Champaign)
Reinforcement Learning
🎯 What it does: Investigate the statistical error bounds of linear offline policy evaluation (LSTDQ) under minimal assumptions, and propose a new coverage metric called feature dynamics coverage;
A universal compression theory for lottery ticket hypothesis and neural scaling laws
Hong-Yi Wang (Princeton University), Liu Ziyin (MIT)
CompressionComputational EfficiencyTabular
🎯 What it does: This paper proposes and proves a general compression theorem, demonstrating that any smooth function invariant to object permutations can be compressed to a polynomial logarithmic number of objects with negligible error; based on this, it proves the dynamic lottery ticket hypothesis and shows that compression significantly enhances the scaling laws of neural networks and datasets.
A-TPT: Angular Diversity Calibration Properties for Test-Time Prompt Tuning of Vision-Language Models
Shihab Aaqil Ahamed (University of Moratuwa), Muhammad Haris Khan (Mohamed bin Zayed University of Artificial Intelligence)
ClassificationRepresentation LearningTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityBiomedical Data
🎯 What it does: Propose A-TPT in test-time prompt tuning by maximizing the minimal angular distance of text features to enhance the calibration performance of vision-language models (VLM);
A.I.R.: Enabling Adaptive, Iterative, and Reasoning-based Frame Selection For Video Question Answering
Yuanhao Zou (University of Central Florida), Chen Chen (University of Central Florida)
Computational EfficiencyTransformerVision Language ModelContrastive LearningVideo
🎯 What it does: Propose an untrained framework A.I.R. that selects the most relevant frames in video question answering through adaptive initial sampling and iterative reasoning.