ICLR 2026 Papers — Page 33
International Conference on Learning Representations · 5356 papers
OmniCT: Towards a Unified Slice-Volume LVLM for Comprehensive CT Analysis
Tianwei Lin (Zhejiang University), Yingda Xia (DAMO Academy, Alibaba Group)
RecognitionLarge Language ModelMixture of ExpertsVision Language ModelImageTextBiomedical DataComputed TomographyBenchmark
🎯 What it does: Proposed a unified CT slice-volume large vision-language model, OmniCT, which integrates 2D slice and 3D volume information for comprehensive CT image analysis.
OmniCVR: A Benchmark for Omni-Composed Video Retrieval with Vision, Audio, and Text
Junyang Ji (Tsinghua University), Wenming Yang (Tsinghua University)
RetrievalLarge Language ModelVision Language ModelContrastive LearningVideoTextMultimodalityBenchmarkAudio
🎯 What it does: Propose the OmniCVR benchmark, completing combined retrieval of videos using three modalities: visual, audio, and text, and designing an automated generation process and dual verification; propose AudioVLM2Vec, converting audio into textual descriptions and fusing with visual information to enhance retrieval performance.
OmniEVA: Embodied Versatile Planner via Task-Adaptive 3D-Grounded and Embodiment-aware Reasoning
Yuecheng Liu (Huawei Noah's Ark Lab), Xingyue Quan (Huawei Noah's Ark Lab)
OptimizationRobotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelVision-Language-Action ModelImageTextMultimodalityPoint CloudBenchmark
🎯 What it does: Propose OmniEVA, an executable task planning framework that integrates 2D/3D multimodal information, addressing the shortcomings of existing MLLMs in geometric adaptability and robot body constraints.
OmniField: Conditioned Neural Fields for Robust Multimodal Spatiotemporal Learning
Kevin Valencia (UCLA), David Keetae Park (Brookhaven National Laboratory)
Neural Radiance FieldImageMultimodalityTime Series
🎯 What it does: Propose OmniField, a continuous neural field framework capable of performing multi-modal spatiotemporal learning on sparse, noisy, and cross-modal incompletely overlapping experimental data, supporting reconstruction, interpolation, prediction, and cross-modal prediction.
OmniMouse: Scaling properties of multi-modal, multi-task Brain Models on 150B Neural Tokens
Konstantin Friedrich Willeke (Stanford University), Andreas S. Tolias (Stanford University)
TransformerVideoTime SeriesBiomedical DataBenchmark
🎯 What it does: Trained and evaluated a large-scale, multimodal, multitask Transformer model called OmniMouse for predicting mouse visual cortex neuron activity, behavior, and conditional responses to video stimuli.
OmniNav: A Unified Framework for Prospective Exploration and Visual-Language Navigation
Xinda Xue (Amap, Alibaba Group), Zhengzhou Zhu (Peking University)
Robotic IntelligenceTransformerVision Language ModelDiffusion modelFlow-based ModelImageTextPoint CloudChain-of-Thought
🎯 What it does: Proposes a unified dual-system framework called OmniNav, capable of simultaneously handling instruction-based goals, object-based goals, point-based navigation, and frontier exploration.
OmniPortrait: Fine-Grained Personalized Portrait Synthesis via Pivotal Optimization
Dongxu Yue (Tsinghua University), Chun Yuan (Tsinghua University)
GenerationDiffusion modelImageMultimodality
🎯 What it does: Proposed OmniPortrait, a fine-grained identity customization portrait synthesis framework based on diffusion;
OmniSpatial: Towards Comprehensive Spatial Reasoning Benchmark for Vision Language Models
Mengdi Jia, Li Yi
Supervised Fine-TuningVision Language ModelImageVideoTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose the OmniSpatial benchmark to systematically evaluate the comprehensive spatial reasoning capabilities of vision-language models.
OmniSTVG: Toward Spatio-Temporal Omni-Object Video Grounding
Jiali Yao (Hangzhou Institute for Advanced Study, University of Chinese Academy of Sciences), Libo Zhang (Institute of Software Chinese Academy of Sciences)
Object DetectionObject TrackingTransformerVideoTextMultimodality
🎯 What it does: Proposes a novel task called OmniSTVG, aiming to simultaneously locate the spatial and temporal positions of all mentioned targets in videos based on free-text queries;
OmniText: A Training-Free Generalist for Controllable Text-Image Manipulation
Agus Gunawan (KAIST), Munchurl Kim (KAIST)
Image HarmonizationGenerationTransformerVision Language ModelDiffusion modelImageTextMultimodalityBenchmark
🎯 What it does: Proposes a training-free, general-purpose text-image processing framework called OmniText, capable of achieving text insertion, editing, deletion, repositioning, scaling, and style control.
OmniVideoBench: Towards Audio-Visual Understanding Evaluation for Omni MLLMs
Caorui Li (Nanjing University), Jiaheng Liu (Nanjing University)
Large Language ModelVideoMultimodalityBenchmarkChain-of-ThoughtAudio
🎯 What it does: Propose OmniVideoBench, an audio-visual collaborative reasoning benchmark for multimodal large language models, containing 628 real videos, 1000 high-quality question-answer pairs, and their atomic reasoning chains.
OmniVinci: Enhancing Architecture and Data for Omni-Modal Understanding LLM
Hanrong Ye (NVIDIA), Pavlo Molchanov (NVIDIA)
Data-Centric LearningTransformerLarge Language ModelReinforcement LearningContrastive LearningImageVideoTextMultimodalityBenchmarkAudio
🎯 What it does: Developed a multimodal large language model called OmniVinci, capable of simultaneously understanding image, video, audio, and text information.
OmniWorld: A Multi-Domain and Multi-Modal Dataset for 4D World Modeling
Yang Zhou (Fudan University), Tong He (Shanghai AI Laboratory)
SegmentationDepth EstimationRobotic IntelligenceOptical FlowImageVideoTextMultimodalityBenchmark
🎯 What it does: Built and released OmniWorld, a cross-domain, multi-modal large-scale 4D world modeling dataset, and established benchmarks for 3D geometry prediction and camera-controlled video generation based on this dataset;
On Code-Induced Reasoning in LLMs
Abdul Waheed, Daphne Ippolito
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Constructed a parallel instruction dataset containing 10 programming languages, applying normalization and generative perturbations to the code, followed by fine-tuning on multiple LLMs and evaluating their reasoning performance on natural language, mathematical, and code tasks.
On Coreset for LASSO Regression Problem with Sensitivity Sampling
YuanBin Zou, Qilong Feng (Central South University)
OptimizationImageTabularComputed Tomography
🎯 What it does: Constructed a coreset for the LASSO regression problem, achieving data subsampling through sensitivity sampling.
On Discovering Algorithms for Adversarial Imitation Learning
Shashank Reddy Chirra (University of Oxford), Pradeep Varakantham (Singapore Management University)
OptimizationLarge Language ModelReinforcement Learning
🎯 What it does: The study discovers and automatically designs a reward allocation function for adversarial imitation learning (AIL), proposing the Discovered Adversarial Imitation Learning (DAIL) algorithm;
On Discriminative vs. Generative classifiers: Rethinking MLLMs for Action Understanding
Zhanzhong Pang (National University of Singapore), Angela Yao (National University of Singapore)
ClassificationRecognitionLarge Language ModelSupervised Fine-TuningVision Language ModelVision-Language-Action ModelVideoTextMultimodalityBenchmark
🎯 What it does: This paper compares two learning methods for multi-modal large language models (MLLM) in closed-set action understanding tasks—generative classifiers and discriminative classifiers—and proposes a Generative-Aided Discriminative (GAD) framework to integrate their advantages.
On Entropy Control in LLM-RL Algorithms
Han Shen (Ant Group)
TransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: This paper investigates the dilemma of entropy control in large language model reinforcement learning (LLM-RL). To address the poor performance of traditional entropy regularization in scenarios with large vocabularies and sparse optimal outputs, we propose an adaptive entropy regularization method called AEnt;
On Fairness of Task Arithmetic: The Role of Task Vectors
Laura Gomezjurado Gonzalez (Stanford University), Ryotaro Shimizu (ZOZO Research)
ClassificationTransformerSupervised Fine-TuningImageText
🎯 What it does: Investigates the fairness of task vectors (task arithmetic) in text and image binary classification, compares their performance with full fine-tuning (FFT) and LoRA, and explores the impact of scaling coefficients and subgroup task vectors on group fairness metrics (DPD/EOD).
On learning linear dynamical systems in context with attention layers
Maria-Luiza Vladarean (Technical University of Munich), Suvrit Sra (Technical University of Munich)
Meta LearningTransformerTime Series
🎯 What it does: This paper investigates the expressive power of linear attention layers in in-context learning (ICL) linear dynamic systems (LDS), particularly when trained on sequences generated by noisy LDS.
On Measuring Influence in Avoiding Undesired Future
Lue Tao (Nanjing University), Zhi-Hua Zhou (Nanjing University)
OptimizationReinforcement LearningTabular
🎯 What it does: Propose a new metric called 'Influence Power' to evaluate the impact of manipulable variables on preventing undesired future events, and provide a recursive formula based on maximum expected utility along with its Monte-Carlo tree search approximation implementation.
On Optimal Hyperparameters for Differentially Private Deep Transfer Learning
Aki Rehn (University of Helsinki), Antti Honkela (University of Helsinki)
Safty and PrivacyHyperparameter SearchTransformerImageText
🎯 What it does: Systematically study the optimal selection of clipping threshold C and batch size B in differential privacy transfer learning, and provide parameter tuning rules based on problem difficulty.
On Powerful Ways to Generate: Autoregression, Diffusion, and Beyond
Chenxiao Yang (Toyota Technological Institute at Chicago), Zhiyuan Li (Toyota Technological Institute at Chicago)
GenerationData SynthesisTransformerSupervised Fine-TuningDiffusion modelTextGraphSequential
🎯 What it does: This paper systematically studies the computational power and generation effects of autoregressive models (ARM), masked diffusion models (MDM), and the proposed arbitrary process generation (AP-MDM) from both theoretical and experimental perspectives, demonstrating the significant advantages of AP-MDM in parallelizability, editability, and cross-domain generation.
On Predictability of Reinforcement Learning Dynamics for Large Language Models
Cai Yuchen (University of Science and Technology of China), Junfeng Fang (National University of Singapore)
TransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Conduct a systematic analysis of parameter updates in large language models (LLMs) during reinforcement learning (RL) training, revealing that the main performance gains are attributable to a rank-1 subspace of the update matrix, which evolves linearly with training; based on this, propose the AlphaRL framework, which predicts the final updates using an early training window, thereby significantly accelerating RL training.
On Robustness of Vision-Language-Action Model against Multi-Modal Perturbations
Jianing Guo (Beihang University), Simin Li (Beihang University)
Adversarial AttackRobotic IntelligenceVision-Language-Action ModelFlow-based ModelRectified FlowMultimodality
🎯 What it does: Systematic evaluation of Vision-Language-Action (VLA) models under 17 perturbations across four modalities (action, observation, environment, language), and propose RobustVLA which enhances multimodal robustness by imposing semantic consistency constraints between worst-case action noise in outputs and inputs.
On Smoothness Bounds for Non-Clairvoyant Scheduling with Predictions
Tianming Zhao (University of Sydney), Albert Zomaya (University of Sydney)
Optimization
🎯 What it does: This paper introduces an improved smoothness measure in non-omniscient scheduling and provides lower and upper bounds for the total completion time on a single machine, the latest completion time on parallel identical machines, and the latest completion time on uniformly related machines. It further proves the actual improvement of predictive information on algorithm performance.
On the $O(1/T)$ Convergence of Alternating Gradient Descent–Ascent in Bilinear Games
Tianlong Nan (Columbia University), Christian Kroer (Columbia University)
Optimization
🎯 What it does: Studied the convergence of the alternating gradient descent-ascent (AltGDA) algorithm in two-player zero-sum games, proving that AltGDA converges at a rate of O(1/T) when an internal Nash equilibrium exists.
On the Alignment Between Supervised and Self-Supervised Contrastive Learning
Achleshwar Luthra (Texas A&M University), Tomer Galanti (Texas A&M University)
ClassificationRepresentation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: Investigated the representation layer alignment between self-supervised contrastive learning (CL) and its supervised variant (NSCL) during training, proposed an analysis framework based on similarity space, and provided lower bounds for CKA and RSA alignment under high probability;
On the Bayes Inconsistency of Disagreement Discrepancy Surrogates
Neil G Marchant, Sarah Monazam Erfani
Domain AdaptationOptimizationAdversarial AttackImageTabularBenchmark
🎯 What it does: This paper studies the surrogate loss for measuring distribution drift in deep learning, proving that existing surrogates have defects in Bayes consistency, and proposes a new provably consistent surrogate loss, verifying its superiority on multiple benchmarks.
On the Benefits of Weight Normalization for Overparameterized Matrix Sensing
Yudong Wei (ETH Zurich), Niao He (ETH Zurich)
OptimizationImage
🎯 What it does: This paper investigates the theoretical convergence properties of weight normalization (WN) in over-parameterized matrix sensing problems, and shows that Riemannian gradient descent with random initialization can achieve linear convergence.
On the Computational Limits of AI4S-RL : A Unified $\varepsilon$-$N$ Analysis
Qili Shen (Tongji University), Dake Zhang (Shanghai Jiao Tong University)
Computational EfficiencyReinforcement LearningPhysics Related
🎯 What it does: Propose a unified ε-N framework to quantitatively analyze the balance between computational cost and accuracy in AI4S simulations and RL sampling, and derive the optimal grid and time step ratio through theoretical analysis;
On the Convergence Behavior of Preconditioned Gradient Descent Toward the Rich Learning Regime
Shuai Jiang (Sandia National Laboratories), Alexey Voronin (Sandia National Laboratories)
OptimizationImagePhysics Related
🎯 What it does: Investigate the impact of preconditioned gradient descent (Gauss-Newton/Levenberg-Marquardt) on spectral bias and grokking, demonstrating its ability to achieve uniform convergence in the NTK/lazy phase and significantly reduce grokking delay.
On the Convergence Direction of Gradient Descent
Shuo Chen (Beijing Jiaotong University), Yao Zhao (Beijing Jiaotong University)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: This paper demonstrates through theoretical proof and experimental validation that gradient descent (GD) converges either along a fixed direction or oscillates along a straight line with large step sizes if it converges; this phenomenon is also observable in stochastic gradient descent (SGD) and Adam.
On the Convergence of Two-Layer Kolmogorov-Arnold Networks with First-Layer Training
Seyed Mohammad Eshtehardian (Sharif University of Technology), Babak Khalaj (Sharif University of Technology)
OptimizationRepresentation Learning
🎯 What it does: This paper proves that gradient descent can converge to a global optimum with zero training error when only the first layer coefficients of a two-layer Kolmogorov-Arnold network (KAN) are trained in an overparameterized scenario.
On the Design of KL-Regularized Policy Gradient Algorithms for LLM Reasoning
Yifan Zhang (Princeton University), Andrew C Yao (Tsinghua University)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposed a unified KL regularization policy gradient framework (RPG), achieving stable offline reinforcement learning on LLM inference tasks.
On the Design of One-step Diffusion via Shortcutting Flow Paths
Haitao Lin (Westlake University), Stan Z. Li (Westlake University)
GenerationDiffusion modelFlow-based ModelRectified FlowImage
🎯 What it does: Propose a general 'one-step shortcut diffusion' design framework, decomposing components such as discrete/continuous, path, and time sampling, and improving training methods within this framework to ultimately achieve high-quality single-step generation.
On the Eligibility of LLMs for Counterfactual Reasoning: A Decompositional Study
Shuai Yang (Binghamton University), Zhaohan Xi (Binghamton University)
Large Language ModelPrompt EngineeringMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposes a decomposed evaluation framework based on causal structure models, splitting causal reasoning for large language models (LLMs) into four stages: causal variable identification, causal graph construction, causal intervention identification, and result inference, and builds 11 benchmark datasets covering multimodal tasks on this basis.
On The Expressive Power of GNN Derivatives
Yam Eitan, Haggai Maron (Technion Israel Institute of Technology)
Representation LearningDrug DiscoveryProtein Structure PredictionGraph Neural NetworkGraphBiomedical DataBenchmark
🎯 What it does: Designed and implemented a new framework, HOD-GNN, which enhances the expressiveness of message passing neural networks (MPNN) by leveraging higher-order derivatives;
On the Expressive Power of GNNs for Boolean Satisfiability
Saku Peltonen (ETH Zürich), Roger Wattenhofer (ETH Zürich)
OptimizationRepresentation LearningGraph Neural NetworkGraphBenchmark
🎯 What it does: This paper studies the expressive power of graph neural networks (GNNs) in Boolean satisfiability problems, proving that the full Weisfeiler–Leman hierarchy cannot distinguish between satisfiable and unsatisfiable instances, and analyzing the distinguishability of different instance families.
On the Expressiveness of State Space Models via Temporal Logics
Eric Alsmann (University of Kassel), Martin Lange (University of Kassel)
Representation LearningTransformer
🎯 What it does: Analyze and establish lower bounds on the expressive power of different types of State Space Models (SSM), using linear temporal logic and its extensions to characterize recognizable languages.
On The Fragility of Benchmark Contamination Detection in Reasoning Models
Han Wang (University of Illinois Urbana-Champaign), Huan Zhang (University of Illinois Urbana-Champaign)
Explainability and InterpretabilityTransformerSupervised Fine-TuningReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: This paper systematically studies the problem of benchmark pollution detection in large reasoning models (LRM), exploring the stealthiness and detection effectiveness of pollution in two scenarios: during the SFT and RL stages when models evolve from base models to LRM, and CoT pollution in mature LRM.
On the Generalization Capacities of MLLMs for Spatial Intelligence
Gongjie Zhang (DAMO Academy, Alibaba Group), Ran Xu (DAMO Academy, Alibaba Group)
Object DetectionPose EstimationDepth EstimationLarge Language ModelImageVideoBenchmark
🎯 What it does: Built a camera-aware multimodal large language model (Camera-Aware MLLM) specifically for spatial intelligence tasks, addressing the geometric ambiguity issue in traditional RGB-only MLLMs during cross-camera generalization.
On the Generalization of SFT: A Reinforcement Learning Perspective with Reward Rectification
Yongliang Wu (Southeast University), Xu Yang (Southeast University)
Supervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: Propose Dynamic Fine-Tuning (DFT), which improves the generalization performance of large language models by dynamically weighting tokens based on their probabilities in the supervised fine-tuning (SFT) loss.
On The Geometry and Topology of Representations: the Manifolds of Modular Addition
Gabriela Moisescu-Pareja (McGill University), Jonathan Love (Leiden University)
Representation LearningTransformer
🎯 What it does: This study investigates the performance of different neural network architectures on the modular addition task, proposing a unified geometric and topological perspective to understand the learning representations of these architectures.
On the identifiability of causal graphs with multiple environments
Francesco Montagna (Institute of Science and Technology Austria)
Explainability and InterpretabilityRepresentation LearningGraphTabular
🎯 What it does: Proposed a method to uniquely identify any nonlinear structural causal graph using observational data from only two different environments under the Gaussian noise assumption.
On the Impact of the Utility in Semivalue-based Data Valuation
Mélissa Tamine (Criteo AI lab), Patrick Loiseau (Inria)
Explainability and InterpretabilityComputational EfficiencyTabular
🎯 What it does: Investigate the robustness of semivalue data evaluation methods under different utility choices.
On the Impossibility of Separating Intelligence from Judgment: The Computational Intractability of Filtering for AI Alignment
Sarah Ball (LMU Munich), Guy N. Rothblum (Apple)
Computational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Studied the feasibility of achieving alignment in large language models (LLMs) through filters, demonstrating fundamental computational barriers;
On the Interaction of Compressibility and Adversarial Robustness
Melih Barsbey (Imperial College London), Tolga Birdal (Imperial College London)
CompressionAdversarial AttackTransformerImage
🎯 What it does: This paper investigates the impact of structural compression (neuron-level compression and spectral compression) on the adversarial robustness of neural networks, establishing theoretical upper bounds and verifying them through various experiments.
On the Interpolation Effect of Score Smoothing in Diffusion Models
Zhengdao Chen (Google)
GenerationDiffusion modelScore-based ModelOrdinary Differential Equation
🎯 What it does: Investigated the interpolation effect of fractional smoothing on diffusion model generation, and proved that regularized neural networks can learn smooth empirical score functions, thereby avoiding memorization and achieving interpolation generation within subspaces.
On the Limits of Sparse Autoencoders: A Theoretical Framework and Reweighted Remedy
Jingyi Cui (Peking University), Yisen Wang (Peking University)
Explainability and InterpretabilityRepresentation LearningAuto EncoderImageText
🎯 What it does: Theoretical analysis of feature recovery in sparse autoencoders (SAE) and derivation of a closed-form solution
On the Lipschitz Continuity of Set Aggregation Functions and Neural Networks for Sets
Giannis Nikolentzos (University of Peloponnese), Konstantinos Skianis (University of Ioannina)
OptimizationRepresentation LearningTextPoint Cloud
🎯 What it does: Studied the Lipschitz continuity of multiset aggregation functions (SUM, MEAN, MAX) and attention-based aggregation functions for three unordered set distances (EMD, Hausdorff, matching distance), deriving corresponding Lipschitz constants; further extended these results to set neural networks composed of multi-layer perceptrons (MLP) and aggregation functions, providing upper bounds for their Lipschitz constants; analyzed network stability against input perturbations and generalization error under distribution drift, validating theoretical and empirical consistency on ModelNet40 and Polarity datasets.
On the Mechanisms of Collaborative Learning in VAE Recommenders
Long Tung Vuong (Amazon), Vu Nguyen (Amazon)
Recommendation SystemAuto EncoderTabular
🎯 What it does: This paper systematically analyzes the collaborative mechanisms in Variational Autoencoder (VAE) collaborative filtering from both theoretical and experimental perspectives, revealing that the core of user update sharing is determined by the distance in the latent space;
On the Predictive Power of Representation Dispersion in Language Models
Yanhong Li (Allen Institute for AI), Jiawei Zhou (Stony Brook University)
Representation LearningTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Investigate and demonstrate that the representational diversity (average cosine distance) of language models is negatively correlated with predictive performance (perplexity), and apply it to unlabeled diagnosis, model and layer selection, and enhance training effectiveness through an auxiliary 'push-away' loss.
On the Reasoning Abilities of Masked Diffusion Language Models
Anej Svete (ETH Zürich), Ashish Sabharwal (Allen Institute for AI)
Explainability and InterpretabilityComputational EfficiencyTransformerDiffusion modelChain-of-Thought
🎯 What it does: This paper theoretically explains the inference capability of masked diffusion language models (MDM), proving its equivalence to filling loop transformers (PLT) and its ability to solve all NC₁-level problems within logarithmic steps, further revealing the sequential bottleneck of chain-of-thought (CoT);
On the Shelf Life of Fine-Tuned LLM-Judges: Future-Proofing, Backward-Compatibility, and Question Generalization
Janvijay Singh (Salesforce AI Research), Shafiq Joty (Salesforce AI Research)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Investigated the long-term usability of fine-tuned LLM discriminators (LLM-as-a-judge), systematically evaluating their performance when facing stronger model-generated text, old model text, and new unseen questions.
On the Spectral Differences Between NTK and CNTK and Their Implications for Point Cloud Recognition
Yuanqu Mou (Nanjing University), Jia Liu (Nanjing University)
RecognitionConvolutional Neural NetworkImagePoint Cloud
🎯 What it does: This paper investigates the spectral differences between Neural Tangent Kernel (NTK) and Convolutional Neural Tangent Kernel (CNTK) under infinitely wide networks, derives theoretical bounds on their mean and spectral spread, and proposes a convolution applicability index β based on this. Subsequently, CNTK and its hybrid with NTK (CNTK-NTK) kernel regression are applied for the first time in point cloud recognition tasks, demonstrating advantages in low-sample environments.
On The Surprising Effectiveness of a Single Global Merging in Decentralized Learning
Tongtian Zhu (Zhejiang University), Can Wang (Zhejiang University)
Federated LearningConvolutional Neural NetworkTransformerImage
🎯 What it does: Studied and empirically analyzed time allocation strategies for communication budgets in decentralized learning, finding that concentrating communication in the later stages of training with a single global merge significantly improves model performance.
On the Tension Between Optimality and Adversarial Robustness in Policy Optimization
Haoran Li (University of Chinese Academy of Sciences), Nan Jiang (University of Illinois Urbana-Champaign)
OptimizationAdversarial AttackReinforcement Learning
🎯 What it does: Studied the trade-off between standard policy optimization and adversarial robust policy optimization in reinforcement learning, and proposed the BARPO two-layer framework to balance optimality and robustness.
On the Theoretical Limitations of Embedding-Based Retrieval
Orion Weller (Google DeepMind), Jinhyuk Lee (Google DeepMind)
RetrievalRepresentation LearningTransformerContrastive LearningTextBenchmark
🎯 What it does: This paper investigates the theoretical limits of unidirectional vector embedding models in retrieval tasks and proves that, given a certain dimensionality, there exist certain document combinations that cannot appear as retrieval results. Subsequently, the theory is validated through optimal vector optimization experiments, leading to the construction of the LIMIT dataset based on this theory. The dataset was tested on various state-of-the-art embedding models, revealing that even with high dimensions, achieving ideal recall remains challenging. Finally, the paper discusses alternative approaches such as multi-vector, sparse, and cross-encoders.
On the Thinking-Language Modeling Gap in Large Language Models
Chenxi Liu (Hong Kong Baptist University), Kun Zhang (MBZUAI)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper studies the bias in large language models during the reasoning process, finding that implicit expressions in training data lead models to ignore key information, and proposes structural causal models and a LoT method based on prompting to alleviate this bias.
On the trade-off between expressivity and privacy in graph representation learning
Patrick Indri (TU Wien), Thomas Gärtner (TU Wien)
Safty and PrivacyRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: This paper studies the trade-off between expressiveness and privacy in graph representation learning, proposing a private graph embedding method based on noised homomorphic density;
On the Universality and Complexity of GNN for Solving Second-order Cone Programs
Ruizhe Li (Southern University of Science and Technology), Minghua Chen (City University of Hong Kong)
OptimizationGraph Neural NetworkGraph
🎯 What it does: Proposed a new graph representation and corresponding Graph Neural Network (SOCP-GNN) for second-order cone programming (SOCP), and proved its generality and generalizability in predicting key attributes such as feasibility and optimal solutions.
On the Wasserstein Geodesic Principal Component Analysis of probability measures
Nina Vesseron (CREST-ENSAE, IP Paris), Klein
OptimizationRepresentation LearningImagePoint Cloud
🎯 What it does: Investigated Geodesic PCA (GPCA) in the Wasserstein space, and proposed two exact algorithms for Gaussian distributions and general absolutely continuous distributions;
On the Wings of Imagination: Conflicting Script-based Multi-role Framework for Humor Caption Generation
Wenbo Shang (Hong Kong Baptist University), Xin Huang (Hong Kong Baptist University)
GenerationTransformerLarge Language ModelPrompt EngineeringImageTextRetrieval-Augmented Generation
🎯 What it does: Proposed a multi-role LLM framework called HOMER based on the GTVH theory to generate humorous captions for images.
On Universality of Deep Equivariant Networks
Marco Pacini (University of Trento), Robin Walters (Northeastern University)
Representation Learning
🎯 What it does: Proposes a general theory of separability constraint universality for depth-invariant networks and equivariant networks, and identifies depth and readout layers as key mechanisms for achieving universality.
On-Policy RL Meets Off-Policy Experts: Harmonizing Supervised Fine-Tuning and Reinforcement Learning via Dynamic Weighting
Wenhao Zhang (Alibaba Group), Jingren Zhou (Alibaba Group)
Supervised Fine-TuningReinforcement LearningTextBiomedical Data
🎯 What it does: Proposes the CHORD framework, unifying supervised fine-tuning (SFT) with reinforcement learning (RL) into a dynamically weighted hybrid objective, avoiding mode conflicts and overfitting caused by the traditional SFT→RL approach.
On-the-Fly Adaptation to Quantization: Configuration-Aware LoRA for Efficient Fine-Tuning of Quantized LLMs
Rongguang Ye (Southern University of Science and Technology), Edith C. H. Ngai (University of Hong Kong)
Computational EfficiencyHyperparameter SearchTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose CoA-LoRA, a configuration-aware LoRA tuning method that can dynamically generate low-rank adapters according to arbitrary quantization configurations without repeated fine-tuning, improving deployment efficiency of quantized LLMs.
Once-More: Continuous Self-Correction for Large Language Models via Perplexity-Guided Intervention
Jiaxun Gao (University of British Columbia), Z. Jane Wang (University of British Columbia)
Large Language ModelAgentic AIText
🎯 What it does: Proposes the Once-More framework, achieving a no-training, model-agnostic continuous self-correction method;
One Demo Is All It Takes: Planning Domain Derivation with LLMs from A Single Demonstration
Jinbang Huang (Huawei Noah's Ark Lab), Yingxue Zhang (Huawei Noah's Ark Lab)
Robotic IntelligenceTransformerLarge Language ModelPrompt Engineering
🎯 What it does: Proposes the PDDLLM framework, which utilizes a single demonstration and automatically generates PDDL planning domains through LLM+physical simulation, seamlessly integrating with low-level motion planners to achieve end-to-end robotic planning;
One for Two: A Unified Framework for Imbalanced Graph Classification via Dynamic Balanced Prototype
Guanjun Wang (University of Science and Technology of China), Yang Wang (University of Science and Technology of China)
ClassificationDrug DiscoveryGraph Neural NetworkGraphBiomedical Data
🎯 What it does: Propose UniImb, a unified framework that employs Dynamic Balanced Prototype (DBP) and personalized graph perturbation techniques to address class imbalance and topological imbalance issues in graph classification.
One Life to Learn: Inferring Symbolic World Models for Stochastic Environments from Unguided Exploration
Zaid Khan (UNC Chapel Hill), Mohit Bansal (UNC Chapel Hill)
Explainability and InterpretabilityLarge Language ModelWorld ModelTabularBenchmark
🎯 What it does: Learning an interpretable symbolic world model in complex random environments during a single episode of goal-free, random exploration.
One Model for All Tasks: Leveraging Efficient World Models in Multi-Task Planning
Yuan Pu (Shanghai Artificial Intelligence Laboratory), Hongsheng Li (Chinese University of Hong Kong)
Computational EfficiencyTransformerReinforcement LearningMixture of ExpertsWorld ModelImageTextBenchmark
🎯 What it does: Proposed the ScaleZero unified world model and combined it with Dynamic Parameter Scaling (DPS) to achieve efficient learning for multi-task planning
One Patch Doesn’t Fit All: Adaptive Patching for Native-Resolution Multimodal Large Language Models
Wenzhuo Liu (University of Chinese Academy of Sciences), Cheng-Lin Liu (University of Chinese Academy of Sciences)
Computational EfficiencyTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: Propose the Adaptive Patching method, which dynamically adjusts patch sizes in multi-modal large language models based on image resolution and information density, and realizes a scheme to convert fixed-patch models into arbitrary-patch models without training.
One protein is all you need
Anton Bushuiev (Czech Technical University), Josef Sivic (Massachusetts Institute of Technology)
Protein Structure PredictionTransformerLarge Language ModelSupervised Fine-TuningSequentialBiomedical Data
🎯 What it does: Propose a self-supervised on-the-fly customization method called ProteinTTT, which can instantly fine-tune existing protein language models with only a single protein sequence, thereby improving the accuracy of structure, compatibility, and function prediction.
One step further with Monte-Carlo sampler to guide diffusion better
Minsi Ren (Westlake University), Tailin Wu (Westlake University)
GenerationData SynthesisDiffusion modelScore-based ModelImageTextBiomedical DataStochastic Differential Equation
🎯 What it does: Proposed an ABMS method that improves conditional guidance in training-agnostic diffusion models by incorporating additional backward denoising steps and Monte-Carlo sampling, thereby reducing estimation errors and enhancing generation quality.
One-Prompt Strikes Back: Sparse Mixture of Experts for Prompt-based Continual Learning
Minh Le (Trivita AI), Nhat Ho (Northeastern University)
TransformerPrompt EngineeringMixture of ExpertsImage
🎯 What it does: Propose SMoPE, a prompt-based continual learning framework that combines sparse Mixture of Experts (MoE) with Prefix Tuning;
One-Shot Exemplars for Class Grounding in Self-Supervised Learning
Haowen Cui (Nanjing University of Science and Technology), Jian Yang (Nanjing University)
Representation LearningContrastive LearningImage
🎯 What it does: Propose a One-Shot Exemplar SSL framework that requires only one labeled image per class for self-supervised pre-training
One-Step Flow for Image Super-Resolution with Tunable Fidelity-Realism Trade-offs
Yuanzhi Zhu (Nanjing University), Kai Zhang (Nanjing University)
Super ResolutionFlow-based ModelRectified FlowImageOrdinary Differential Equation
🎯 What it does: Proposed OFTSR, a flow-based single-step image super-resolution framework that can generate high-resolution images with adjustable realism and fidelity in a single forward pass.
One-Step Flow Q-Learning: Addressing the Diffusion Policy Bottleneck in Offline Reinforcement Learning
Thanh Xuan Nguyen (Korea Advanced Institute of Science and Technology), Chang D. Yoo (Korea Advanced Institute of Science and Technology)
Reinforcement LearningFlow-based ModelBenchmarkOrdinary Differential Equation
🎯 What it does: Proposed a One-Step Flow Q-Learning (OFQL) to achieve single-step action generation in offline reinforcement learning, avoiding the multi-step diffusion inference and backpropagation of traditional Diffusion Q-Learning.
One2Scene: Geometric Consistent Explorable 3D Scene Generation from a Single Image
Pengfei Wang (Hong Kong Polytechnic University), Lei Zhang (Hong Kong Polytechnic University)
GenerationDepth EstimationTransformerDiffusion modelGaussian SplattingImageBenchmark
🎯 What it does: Propose a three-stage method to generate explorable 3D scenes from a single image.
OneTwoVLA: A Unified Vision-Language-Action Model with Adaptive Reasoning
Fanqi Lin (Tsinghua University), Yang Gao (Tsinghua University)
Robotic IntelligenceLarge Language ModelVision-Language-Action ModelFlow-based ModelImageTextMultimodalityChain-of-Thought
🎯 What it does: Proposed a unified vision-language-action model called OneTwoVLA, which can adaptively switch between reasoning and execution modes during task execution;
Online Black-Box Prompt Optimization with Regret Guarantees under Noisy Feedback
Jinjie Fang (Jilin University), Bin Gu (Jilin University)
OptimizationTransformerLarge Language ModelPrompt EngineeringDiffusion modelImageText
🎯 What it does: Propose AOZPT, combining online black-box prompt tuning with zeroth-order gradient estimation, and introduce an adaptive uncertainty scaling mechanism to address noise and high variance in generative AI.
Online Conformal Prediction with Adversarial Semi-bandit Feedback via Regret Minimization
Junyoung Yang (POSTECH), Sangdon Park (POSTECH)
ClassificationOptimizationReinforcement LearningImageTabular
🎯 What it does: Propose an online conformal prediction algorithm under adversarial semi-bandit feedback. The problem is first discretized into a multi-armed adversarial bandit problem, and a loss function combining coverage error and prediction set size is designed. The adversarial EXP3.P algorithm is utilized with estimators incorporating full unlocking and partial unlocking, ultimately leading to the OCP-Unlock+ algorithm. Long-term coverage guarantees and experimental validation are provided.
Online Decision Making with Generative Action Sets
Jianyu Xu (Carnegie Mellon University), Aarti Singh (Carnegie Mellon University)
OptimizationReinforcement LearningTextBiomedical Data
🎯 What it does: Studied an online learning problem where an agent can generate new actions at any time step by paying a one-time cost, and the generated actions can be permanently used.
Online Decision-Focused Learning
Aymeric Capitaine (École Polytechnique), Alain Oliviero Durmus
OptimizationTabular
🎯 What it does: Propose a framework called Decision-Focused Learning in dynamic, online decision environments, and design two feasible algorithms, DF-FTPL and DF-OGD;
Online Inventory Optimization in Non-Stationary Environment
Koji Ichikawa (NEC Corporation), Tatsuya Matsuoka (NEC Corporation)
Optimization
🎯 What it does: Proposed an online inventory optimization (OIO) algorithm that achieves a dynamic regret upper bound in non-stationary environments
Online Learning and Equilibrium Computation with Ranking Feedback
Mingyang Liu (Massachusetts Institute of Technology), Kaiqing Zhang (University of Maryland)
OptimizationText
🎯 What it does: This paper studies the problem of online learning and game equilibrium computation when only action ranking feedback is available, analyzes immediate and time-averaged ranking mechanisms, provides difficulty proofs, and proposes a sublinear regret algorithm under total variation linear; further proving convergence to coarse equilibrium in multi-player repeated games.
Online Minimization of Polarization and Disagreement via Low-Rank Matrix Bandits
Federico Cinus (Intesa Sanpaolo AI Research), Francesco Bonchi (Intesa Sanpaolo AI Research)
OptimizationReinforcement LearningGraph
🎯 What it does: Proposes an online low-rank matrix bandit method to minimize polarization and disagreement under the Friedkin-Johnsen model in the absence of prior knowledge about individual's innate opinions.
Online Navigation Refinement: Achieving Lane-Level Guidance by Associating Standard-Definition and Online Perception Maps
Jiaxu Wan (BUAA), Yifan Yang (BUAA)
Autonomous DrivingOptimizationComputational EfficiencyTransformerPoint CloudGraphBenchmark
🎯 What it does: Proposed the Online Navigation Refinement (ONR) task, achieving the transition from road-level navigation to lane-level navigation by associating standard definition (SD) maps with real-time perception (OP) maps.
Online Prediction of Stochastic Sequences with High Probability Regret Bounds
Matthias Frey (University of Melbourne), Jingge Zhu (University of Melbourne)
OptimizationTime SeriesSequential
🎯 What it does: This paper proposes a high-probability regret upper bound for generic prediction of random sequences under a finite time window, extending the previous theoretical framework that only provided expected regret upper bounds.
Online Pseudo-Zeroth-Order Training of Neuromorphic Spiking Neural Networks
Mingqing Xiao (Microsoft Research Asia), Zhouchen Lin
OptimizationComputational EfficiencySpiking Neural NetworkImageVideo
🎯 What it does: Proposed an online pseudo-zeroth-order (OPZO) training method, achieving spatial and temporal credit assignment in spiking neural networks through a single forward pass with noise injection and direct top feedback.
Online Rounding and Learning Augmented Algorithms for Facility Location
Silvio Lattanzi (Google Research), Ola Svensson (EPFL)
Optimization
🎯 What it does: Proposed two online rounding algorithms to convert fractional facility location solutions into integer solutions, and applied them to the online facility location problem with machine learning recommendations;
Online time series prediction using feature adjustment
Xiannan Huang (Tongji University), Chao Yang (Tongji University)
TransformerTime SeriesSequential
🎯 What it does: Propose an online time series prediction feature space adaptation framework ADAPT-Z, which directly updates the feature representations corresponding to latent factors;
Only Brains Align with Brains: Cross-Region Alignment Patterns Expose Limits of Normative Models
Larissa Höfling (University Clinics Tübingen), Katrin Franke (University Clinics Tübingen)
Representation LearningVideoBiomedical DataDiffusion Tensor ImagingBenchmark
🎯 What it does: This paper benchmarks 47 visual models on the BOLD Moments video fMRI dataset, finding that many models achieve similar scores on traditional point-to-point alignment metrics (RSA/LP); to address this limitation, we propose Cross-Region Alignment Pattern Analysis (APA), which distinguishes models that are truly brain-similar by comparing alignment pattern similarity between model layers and brain regions.
Open Data Synthesis for Deep Research
Ziyi Xia (Beijing Academy of Artificial Intelligence), Zheng Liu (Beijing Academy of Artificial Intelligence)
Data SynthesisRetrievalSupervised Fine-TuningReinforcement LearningDiffusion modelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Construct the InfoSeek framework, using HCSP to define deep search problems and generate over 50k QA pairs.
Open-Set Semantic Gaussian Splatting SLAM with Expandable Representation
Yucheng Yan (Zhejiang University), Yi Yang (Zhejiang University)
Contrastive LearningGaussian SplattingSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: Proposes an open semantic Gaussian Splatting SLAM system implementable on daily devices, capable of real-time capturing and reconstructing 3D scenes with scalable semantic information.
OpenAgentSafety: A Comprehensive Framework For Evaluating Real-World AI Agent Safety
Sanidhya Vijayvargiya (Carnegie Mellon University), Maarten Sap (Carnegie Mellon University)
Safty and PrivacyLarge Language ModelAgentic AITextBenchmark
🎯 What it does: Proposed the OPENAGENTSAFETY framework, constructed over 350 executable tasks covering 8 risk categories, and assessed the safety of AI agents through real-world tools and multi-user, multi-round interactions.
OpenApps: Simulating Environment Variations to Measure UI Agent Reliability
Karen Ullrich (FAIR at Meta), Mark Ibrahim (FAIR at Meta)
Agentic AITextBenchmark
🎯 What it does: Developed a lightweight configurable multi-application simulator, OPENAPPS, for evaluating the reliability of UI agents across different application variants.
OpenEstimate: Evaluating LLMs on Reasoning Under Uncertainty with Real-World Data
Alana Marzoev (MIT), Jacob Andreas (MIT)
TransformerLarge Language ModelTabularBiomedical DataBenchmark
🎯 What it does: Designed and implemented the OPENESTIMATE benchmark to evaluate probability estimation and uncertainty reasoning of language models under real-world conditional statistics.
OpenFly: A COMPREHENSIVE PLATFORM FOR AERIAL VISION-LANGUAGE NAVIGATION
Yunpeng Gao (The University of Hong Kong), Xuelong Li (Shanghai AI Laboratory)
Autonomous DrivingLarge Language ModelVision-Language-Action ModelGaussian SplattingImageTextMultimodalityPoint CloudBenchmark
🎯 What it does: This paper proposes the OpenFly platform, which provides multi-rendering engines, an automated data generation pipeline, and the first large-scale aerial visual language navigation (VLN) dataset with 100K trajectories, along with the OpenFly-Agent model based on keyframe selection and visual token compression;