ICML 2025 Papers — Page 21
International Conference on Machine Learning · 3257 papers
Of Mice and Machines: A Comparison of Learning Between Real World Mice and RL Agents
Shuo Han (Northwestern University), Bradly C. Stadie (Northwestern University)
Robotic IntelligenceReinforcement LearningSequential
🎯 What it does: This paper compares the learning and decision-making behaviors of real mice and reinforcement learning (RL) agents in a predator-avoidance maze environment, and proposes two mechanisms (trauma-inspired safety buffer and variance penalty TD learning) to enhance the risk-averse behavior of RL agents to be more similar to that of mice.
Off-Policy Actor-Critic for Adversarial Observation Robustness: Virtual Alternative Training via Symmetric Policy Evaluation
Kosuke Nakanishi (Honda Research and Development), Shin Ishii (Kyoto University)
Reinforcement Learning
🎯 What it does: A method for achieving adversarial observation robustness in an offline reinforcement learning framework without additional environment interactions is proposed.
Off-Policy Evaluation under Nonignorable Missing Data
Han Wang (North Carolina State University), Rui Song (North Carolina State University)
Reinforcement LearningBiomedical DataElectronic Health Records
🎯 What it does: This paper addresses the bias issue in off-policy evaluation (OPE) in offline reinforcement learning when there is non-random missing (MNAR) data. It proposes an estimator based on inverse probability weighting (IPW) and provides confidence intervals, completing theoretical derivation and empirical validation.
Offline Learning for Combinatorial Multi-armed Bandits
Xutong Liu (Carnegie Mellon University), Wei Chen (Microsoft Research)
OptimizationReinforcement LearningTabular
🎯 What it does: The Off-CMAB framework is proposed to address the learning problem of Combinatorial Multi-Armed Bandit (CMAB) on offline data, introducing the CLCB (Combinatorial Lower Confidence Bound) algorithm.
Offline Model-based Optimization for Real-World Molecular Discovery
Dong-Hee Shin (Korea University), Tae-Eui Kam (Korea University)
OptimizationDrug DiscoveryReinforcement LearningGraph
🎯 What it does: A molecular stitching framework called MolStitch is proposed, which utilizes existing molecules to generate 'stitched molecules' and achieves offline multi-objective molecular optimization through ranking proxies and preference optimization.
Offline Opponent Modeling with Truncated Q-driven Instant Policy Refinement
Yuheng Jing (Chinese Academy of Sciences), Jian Cheng (Chinese Academy of Sciences)
TransformerReinforcement LearningSequential
🎯 What it does: This paper proposes a framework named TIPR, which alleviates performance degradation caused by suboptimal data by learning truncated Q-values in offline opponent modeling (OOM) and instantaneously refining strategies during testing.
Offline-to-Online Reinforcement Learning with Classifier-Free Diffusion Generation
Xiao Huang (Shanghai Jiao Tong University), Shuai Li (Shanghai Jiao Tong University)
Data SynthesisReinforcement LearningDiffusion modelTabular
🎯 What it does: An adaptive data augmentation method based on a classifier-free guidance diffusion model (CFDG) is proposed in the offline-to-online reinforcement learning (O2O RL) framework, which generates both online and offline data, significantly improving the distribution alignment with online policies.
Olica: Efficient Structured Pruning of Large Language Models without Retraining
Jiujun He (Southwestern University of Finance and Economics), Huazhen Lin (Southwestern University of Finance and Economics)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper presents Olica, a method for structured pruning of large language models that does not require retraining.
OmiAD: One-Step Adaptive Masked Diffusion Model for Multi-class Anomaly Detection via Adversarial Distillation
Yaoxuan Feng (Xidian University), Mingyuan Zhou (University of Texas at Austin)
Anomaly DetectionKnowledge DistillationDiffusion modelImage
🎯 What it does: Proposes a first-order adaptive mask diffusion model OmiAD for multi-class unsupervised anomaly detection.
Omni-Angle Assault: An Invisible and Powerful Physical Adversarial Attack on Face Recognition
Shuai Yuan (University of Electronic Science and Technology of China), Guowen Xu (University of Electronic Science and Technology of China)
RecognitionAdversarial AttackReinforcement LearningImage
🎯 What it does: Designed and implemented a hat using a UV light emitter for invisible physical adversarial attacks targeting facial recognition systems.
OmniArch: Building Foundation Model for Scientific Computing
Tianyu Chen (Beihang University), Jianxin Li (Beihang University)
TransformerContrastive LearningTabularPhysics Related
🎯 What it does: Proposes the OmniArch foundational model for unified processing of 1D, 2D, and 3D multi-scale, multi-physics PDE computations, enhancing prediction accuracy through physical alignment.
OmniAudio: Generating Spatial Audio from 360-Degree Video
Huadai Liu (Hong Kong University of Science and Technology), Wei Xue (Hong Kong University of Science and Technology)
GenerationData SynthesisFlow-based ModelVideoBenchmarkAudio
🎯 What it does: A new task called 360V2SA is proposed, which synthesizes three-dimensional space First-order Ambisonics (FOA) audio from 360-degree video.
OmniBal: Towards Fast Instruction-Tuning for Vision-Language Models via Omniverse Computation Balance
Yongqiang Yao (Shanghai Jiao Tong University), Ningyi Xu (Shanghai Jiao Tong University)
OptimizationComputational EfficiencyTransformerSupervised Fine-TuningVision Language ModelMultimodality
🎯 What it does: Designed and implemented the OmniBal framework, which systematically balances the computational load of data, models, and memory, significantly accelerating the instruction fine-tuning training of VLM.
On Differential Privacy for Adaptively Solving Search Problems via Sketching
Shiyuan Feng (Peking University), Lichen Zhang (Massachusetts Institute of Technology)
OptimizationSafty and PrivacyTabular
🎯 What it does: This paper studies the use of differential privacy to solve search problems under adaptive queries, with a particular focus on nearest neighbor queries and regression problems, and proposes corresponding algorithms.
On Efficient Estimation of Distributional Treatment Effects under Covariate-Adaptive Randomization
Undral Byambadalai (CyberAgent), Shota Yasui (CyberAgent)
Supervised Fine-TuningReinforcement LearningTabular
🎯 What it does: In a randomized trial based on Covariate Adaptive Randomization (CAR), a regression adjustment method using distribution regression combined with machine learning is proposed to estimate Distribution Treatment Effects (DTE) and Probability Treatment Effects (PTE).
On Exact Bit-level Reversible Transformers Without Changing Architecture
Guoqiang Zhang (University of Exeter), W. Bastiaan Kleijn (Victoria University of Wellington)
TransformerSupervised Fine-TuningImageTextOrdinary Differential Equation
🎯 What it does: Proposes BDIA-transformer, a model that achieves precise bit-level reversible training while maintaining the original Transformer architecture.
On Explaining Equivariant Graph Networks via Improved Relevance Propagation
Hongyi Ling (Texas A&M University), Shuiwang Ji (Texas A&M University)
Explainability and InterpretabilityGraph Neural NetworkGraph
🎯 What it does: The EquiGX method is proposed, which implements layer-wise relevance propagation for spherical equivariant GNNs using deep Taylor decomposition, thereby providing an interpretable decomposition of the prediction results.
On Expressive Power of Looped Transformers: Theoretical Analysis and Enhancement via Timestep Encoding
Kevin Xu (University of Tokyo), Issei Sato (University of Tokyo)
TransformerText
🎯 What it does: This paper theoretically analyzes the expressive power of the Looped Transformer, providing its approximation rate in relation to the number of loops and three types of continuous moduli, and proposes improvements to the model's construction and training through Timestep Encoding.
On Fine-Grained Distinct Element Estimation
Ilias Diakonikolas (University of Wisconsin), Samson Zhou (Texas A&M University)
Tabular
🎯 What it does: This paper studies the distinct element estimation problem and proposes a parameterized method based on the number of collisions between servers, providing a protocol with lower communication costs.
On Learning Parallel Pancakes with Mostly Uniform Weights
Ilias Diakonikolas (University of Wisconsin Madison), Thanasis Pittas (University of Wisconsin Madison)
🎯 What it does: The paper studies the learning complexity of Gaussian Mixture Models (GMM) in high-dimensional spaces, constrained by the parallel pancake structure with component weight distributions that are primarily uniform or nearly uniform.
On Linear Convergence in Smooth Convex-Concave Bilinearly-Coupled Saddle-Point Optimization: Lower Bounds and Optimal Algorithms
Ekaterina Borodich (Higher School of Economics), Dmitry Kovalev (Yandex Research)
Optimization
🎯 What it does: This paper studies the smooth convex-concave bilinear coupling saddle point optimization problem and presents an optimal linear convergence algorithm.
On Measuring Long-Range Interactions in Graph Neural Networks
Jacob Bamberger (University of Oxford), Xiaowen Dong (University of Oxford)
Graph Neural NetworkGraph
🎯 What it does: A formal definition of long-range graph tasks is proposed, and a long-distance metric based on Jacobian and Hessian is designed. Various GNN structures are evaluated in synthetic and LRGB experiments.
On Mitigating Affinity Bias through Bandits with Evolving Biased Feedback
Matthew Faw (Georgia Institute of Technology), Jessica Hoffmann (Google DeepMind)
Recommendation SystemOptimizationReinforcement LearningTabular
🎯 What it does: This paper proposes a new non-stationary multi-armed bandit model - the 'Affinity Bandit', which characterizes the feedback loop caused by affinity bias during the recruitment process; and presents an elimination-based algorithm that minimizes the true rewards without knowing the actual rewards, only observing biased feedback.
On Path to Multimodal Generalist: General-Level and General-Bench
Hao Fei (National University of Singapore), Hanwang Zhang (Nanyang Technological University)
Large Language ModelPrompt EngineeringImageVideoTextMultimodalityBenchmarkAudio
🎯 What it does: A General-Level evaluation framework and General-Bench large-scale multimodal benchmark are proposed, conducting zero-shot evaluations on over 100 multimodal LLMs.
On Teacher Hacking in Language Model Distillation
Daniil Tiapkin (École Polytechnique), Mathieu Blondel (Google DeepMind)
Knowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper systematically studies the phenomenon of 'Teacher Hacking' that occurs during the distillation process of language models by constructing a semi-synthetic controlled experimental framework, and proposes detection and mitigation strategies.
On Temperature Scaling and Conformal Prediction of Deep Classifiers
Lahav Dabah (Bar-Ilan University), Tom Tirer (Bar-Ilan University)
ClassificationConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper studies the impact of Temperature Scaling (TS) on the adaptive non-conformity prediction (Conformal Prediction, CP) methods of deep classifiers, providing theoretical explanations and practical guidance.
On the Adversarial Robustness of Multi-Kernel Clustering
Hao Yu (National University of Defence Technology), Xinwang Liu (National University of Defence Technology)
Adversarial AttackReinforcement LearningTabular
🎯 What it does: A black-box adversarial attack framework for multi-kernel clustering (MKC) called AdvMKC is proposed.
On the Alignment between Fairness and Accuracy: from the Perspective of Adversarial Robustness
Junyi Chai (Purdue University), Xiaoqian Wang (Purdue University)
Adversarial AttackTabular
🎯 What it does: This study investigates the relationship between fairness and accuracy under adversarial robustness, proposing a unified fairness adversarial attack framework and enhancing the model's fairness robustness through adversarial training.
On the Benefits of Active Data Collection in Operator Learning
Unique Subedi (University of Michigan), Ambuj Tewari (University of Michigan)
Time SeriesStochastic Differential Equation
🎯 What it does: This study investigates active data collection strategies in linear operator learning, proving that when the input functions come from a random process with zero mean and continuous covariance kernel, active sampling can achieve a faster error convergence rate than passive (i.i.d.) sampling.
On the Clean Generalization and Robust Overfitting in Adversarial Training from Two Theoretical Views: Representation Complexity and Training Dynamics
Binghui Li (Peking University), Yuanzhi Li (Carnegie Mellon University)
Representation LearningAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This study investigates the phenomenon of clean generalization and robust overfitting in adversarial training, providing theoretical explanations from the perspectives of representation complexity and training dynamics.
On The Concurrence of Layer-wise Preconditioning Methods and Provable Feature Learning
Thomas TCK Zhang, Nikolai Matni (University of Pennsylvania)
OptimizationRepresentation Learning
🎯 What it does: In two types of simplified feature learning models (linear representation learning and single exponential learning), it is proven that conventional SGD learns features slowly and poorly when the input distribution has non-homogeneous variance, while hierarchical preprocessing (KFAC) can eliminate these issues and achieve faster and more accurate feature learning;
On the Convergence of Continuous Single-timescale Actor-critic
Xuyang Chen (National University of Singapore), Lin Zhao (National University of Singapore)
Reinforcement Learning
🎯 What it does: Under continuous state-action space, a finite-time convergence analysis of the single time-scale single-sample Actor-Critic algorithm is provided, proving that with appropriate step size selection, both the critic's error and the actor's gradient error can converge at a rate of O(1/√T).
On the Diversity of Adversarial Ensemble Learning
Jun-Qi Guo (Nanjing University), Zhi-Hua Zhou (Nanjing University)
Adversarial AttackConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: This paper studies the concept of diversity in adversarial ensemble learning, proves that calculating exact diversity is NP-Hard, provides a first-order approximation-based diversity decomposition, and designs the AdvEOAP ensemble method to enhance gradient and cross diversity through orthogonal adversarial prediction.
On the Duality between Gradient Transformations and Adapters
Lucas Torroba Hennigen, Yoon Kim (Massachusetts Institute of Technology)
TransformerLarge Language ModelText
🎯 What it does: This study investigates the equivalence between gradient linear transformations and adapter reparameterization, improving memory-efficient training of large language models.
On the Dynamic Regret of Following the Regularized Leader: Optimism with History Pruning
Naram Mhaisen (Delft University of Technology), George Iosifidis (Delft University of Technology)
Optimization
🎯 What it does: Re-evaluating the FTRL framework in online convex optimization, this paper proposes an Optimistic Follow the Pruned Leader (OptFPRL) algorithm with prediction and historical pruning, and provides an upper bound dynamic regret analysis in dynamic environments.
On the Emergence of Position Bias in Transformers
Xinyi Wu (Massachusetts Institute of Technology), Ali Jadbabaie (Massachusetts Institute of Technology)
RetrievalTransformerSequential
🎯 What it does: This paper constructs a graph theory framework to theoretically analyze the combined effects of attention masks and position encodings in multi-layer Transformers, and verifies their impact on positional bias through numerical experiments.
On the Generalization Ability of Next-Token-Prediction Pretraining
Zhihao Li (Huazhong Agricultural University), Feng Zheng (Southern University of Science and Technology)
TransformerLarge Language ModelText
🎯 What it does: This paper establishes a detailed generalization analysis of next token prediction (NTP) pre-training based on Rademacher complexity, exploring how NTP pre-training affects the model's generalization ability.
On the Guidance of Flow Matching
Ruiqi Feng (Westlake University), Tailin Wu (Westlake University)
RestorationGenerationReinforcement LearningDiffusion modelFlow-based ModelImageTabular
🎯 What it does: A general guiding framework for flow matching models is proposed, deriving various guiding methods based on MC, approximation, and training, covering the classic guidance of diffusion models.
On the Impact of Performative Risk Minimization for Binary Random Variables
Nikita Tsoy (INSAIT Sofia University St Kliment Ohridski), Nikola Konstantinov (INSAIT Sofia University St Kliment Ohridski)
OptimizationReinforcement Learning
🎯 What it does: This paper analyzes the long-term impact of performative risk minimization (PRM) on model predictions and data distribution for binary random variables, and proposes two impact metrics: prediction bias and mean shift.
On the Importance of Embedding Norms in Self-Supervised Learning
Andrew Draganov (Aarhus University), Erik J Bekkers
ClassificationRepresentation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This study investigates the impact of embedding norms on convergence speed and model confidence in self-supervised learning, and proposes three strategies to regulate the norms.
On the Importance of Gaussianizing Representations
Daniel Eftekhari (University of Toronto), Vardan Papyan (University of Toronto)
ClassificationRepresentation LearningConvolutional Neural NetworkTransformerImage
🎯 What it does: A novel normalization layer (Normality Normalization) is proposed, which approximates activations to a normal distribution through power transformation and adds scaled additive Gaussian noise during training to enhance the model's generalization and robustness.
On the Interplay between Graph Structure and Learning Algorithms in Graph Neural Networks
Junwei Su (University of Hong Kong), Chuan Wu (University of Hong Kong)
Graph Neural NetworkGraph
🎯 What it does: This study investigates the interaction between learning algorithms (SGD, Ridge) and graph structures in Graph Neural Networks (GNN), providing a theoretical analysis of generalization (overfitting risk) in the presence of noise.
On the Learnability of Distribution Classes with Adaptive Adversaries
Tosca Lechner (Vector Institute), Gautam Kamath (University of Waterloo)
Adversarial Attack
🎯 What it does: It is proposed that when there is an adaptive additive adversary, distributional learnability is not equivalent to robust learnability, and a distribution family is constructed that is learnable but not learnable under adaptive additive/subtractive attacks;
On the Local Complexity of Linear Regions in Deep ReLU Networks
Niket Nikul Patel (University of California Los Angeles), Guido Montufar (University of California Los Angeles)
OptimizationRepresentation LearningAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This paper proposes and quantifies the density of linear regions of deep ReLU networks around the input distribution—essentially a measure of local complexity—and establishes theoretical connections with feature learning, total variation, robustness, and implicit regularization in optimization.
On the Out-of-Distribution Generalization of Self-Supervised Learning
Wenwen Qiang (Institute of Software Chinese Academy of Sciences), Changwen Zheng (Institute of Software Chinese Academy of Sciences)
Domain AdaptationRepresentation LearningContrastive LearningImage
🎯 What it does: This paper conducts an in-depth analysis of the out-of-distribution (OOD) generalization problem in self-supervised learning (SSL) from the perspectives of batch construction and causality. It introduces the concept of post-intervention distribution (PID) and designs a batch sampling strategy based on latent variable models and propensity scores to eliminate spurious correlations among factors during the self-supervised training process, thereby enhancing SSL performance on OOD tasks.
On the Power of Context-Enhanced Learning in LLMs
Xingyu Zhu (Princeton University), Sanjeev Arora (Princeton University)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper studies Context-Enhanced Learning (CEL) and verifies its advantages in sample efficiency by constructing a Multi-level Translation (MLT) task.
On the Power of Learning-Augmented Search Trees
Jingbang Chen (University of Waterloo), Li Chen (Georgia Institute of Technology)
Tabular
🎯 What it does: Proposed a learning-enhanced Treap based on composite priority (and its B-Tree variant), achieving static optimality under arbitrary access distributions while supporting dynamic updates and being robust to prediction errors;
On the Private Estimation of Smooth Transport Maps
Clément Lalanne (University of Toulouse), Julien Chhor (Toulouse School of Economics)
OptimizationSafty and PrivacyTabular
🎯 What it does: A private algorithm for estimating smooth optimal transport mappings under differential privacy constraints is proposed, along with its theoretical upper and lower bounds on error.
On the Provable Separation of Scales in Maximal Update Parameterization
Letong Hong, Zhangyang Wang
OptimizationHyperparameter SearchTransformerLarge Language ModelTextStochastic Differential Equation
🎯 What it does: Under the Maximal Update Parameterization (µP), it is proven that the convergence speed of macro variables (such as loss curves) with respect to network width is O(1/n), while the convergence speed of micro variables (such as individual weights) is Θ(1/√n), thereby explaining the theoretical basis for the effective transfer of hyperparameter tuning in the early stages to larger models.
On the Query Complexity of Verifier-Assisted Language Generation
Edoardo Botta (Carnegie Mellon University), Andrej Risteski (Carnegie Mellon University)
GenerationData SynthesisTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Proposes the use of a process validator to reduce the query complexity of constrained language generation and designs a backtracking Tokenwise rejection sampling algorithm.
On the Resilience of LLM-Based Multi-Agent Collaboration with Faulty Agents
Jen-tse Huang (Chinese University of Hong Kong), Maarten Sap (Carnegie Mellon University)
AI Code AssistantTransformerLarge Language ModelText
🎯 What it does: This paper studies the robustness of multi-agent systems with large language models in the presence of erroneous agents, evaluating the performance of three structures—linear, flat, and hierarchical—across four tasks: code generation, mathematical reasoning, translation, and text evaluation. It also proposes methods for automatic error introduction, AUTOTRANSFORM and AUTOINJECT, along with improvement strategies involving challengers and inspectors to address errors.
On the Robustness of Reward Models for Language Model Alignment
Jiwoo Hong (Korea Advanced Institute of Science and Technology), James Thorne
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningText
🎯 What it does: This study addresses the issue of over-optimization in the Reward Model (RM) when trained using Bradley-Terry (BT) loss and proposes a simple Batch Sum Zero Regularization (BSR) to alleviate this problem, further validating its positive impact on RLHF training.
On the Role of Label Noise in the Feature Learning Process
Andi Han (University of Sydney), Taiji Suzuki (University of Tokyo)
Representation LearningConvolutional Neural NetworkImage
🎯 What it does: A rigorous theoretical analysis of the feature learning process of neural networks in the presence of label noise is conducted, revealing a two-phase behavior during training (first learning clean signals and then memorizing noise).
On the Similarities of Embeddings in Contrastive Learning
Chungpa Lee (Yonsei University), Jy-yong Sohn (Yonsei University)
ClassificationRepresentation LearningContrastive LearningImage
🎯 What it does: Theoretical analysis of embedding similarity in contrastive learning, and the proposal of an auxiliary loss to reduce the variance of negative sample similarity.
On the Statistical Mechanisms of Distributional Compositional Generalization
Jingwen Fu (Xi'an Jiaotong University), Nanning Zheng (Xi'an Jiaotong University)
Large Language ModelText
🎯 What it does: This paper explores the mechanism of Distributed Combination Generalization (DCG) from the perspective of statistical learning, proposing an invariant measure and a general upper bound on generalization error, explaining the transferability and adaptability between methods.
On the Tension between Byzantine Robustness and No-Attack Accuracy in Distributed Learning
Yi-Rui Yang (Nanjing University), Wu-Jun Li (Nanjing University)
OptimizationFederated LearningConvolutional Neural NetworkImage
🎯 What it does: This study investigates the aggregation error and convergence performance of robust aggregators in distributed learning without Byzantine attacks, revealing the tension between robustness and accuracy in the absence of attacks.
On the Training Convergence of Transformers for In-Context Classification of Gaussian Mixtures
Wei Shen (University of Virginia), Cong Shen (University of Virginia)
ClassificationOptimizationTransformer
🎯 What it does: This study investigates the training dynamics, convergence, and upper bounds of inference error of a single-layer Transformer in the context of Gaussian mixture models.
On the Vulnerability of Applying Retrieval-Augmented Generation within Knowledge-Intensive Application Domains
Xun Xian (University of Minnesota), Jie Ding (University of Minnesota)
RetrievalAdversarial AttackTextBiomedical DataRetrieval-Augmented Generation
🎯 What it does: Conduct a security assessment of the retrieval module in retrieval-augmented generation (RAG) systems, demonstrating the existence of common poisoning attacks in medical and legal question-answering scenarios.
On Understanding Attention-Based In-Context Learning for Categorical Data
Aaron T Wang, Lawrence Carin (Duke University)
ClassificationGenerationOptimizationTransformerLarge Language ModelImageTextTabular
🎯 What it does: This paper proposes a Transformer architecture based on attention blocks, utilizing self-attention and cross-attention to achieve multi-step function gradient descent, thereby accurately inferring contextual learning problems, treating the next word generation of language models as such tasks.
On Volume Minimization in Conformal Regression
Batiste Le bars, Pierre Humbert (Sorbonne Université et Univ. Paris Cité, CNRS, LPSM)
OptimizationTabular
🎯 What it does: This paper studies the volume minimization problem in split conformal regression and proposes new efficiency-oriented methods EffOrt and Ad-EffOrt.
On Zero-Initialized Attention: Optimal Prompt and Gating Factor Estimation
Nghiem Tuong Diep, Nhat Ho
TransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextBenchmark
🎯 What it does: The theoretical foundation of zero-initialized attention mechanisms is studied, and a nonlinear prompt structure is proposed.
On-Device Collaborative Language Modeling via a Mixture of Generalists and Specialists
Dongyang Fan (Ecole Polytechnique Federale de Lausanne), Martin Jaggi (Ecole Polytechnique Federale de Lausanne)
Federated LearningSafty and PrivacyComputational EfficiencyLarge Language ModelMixture of ExpertsText
🎯 What it does: A federated learning framework CoMiGS based on a mixture of general experts and dedicated experts is proposed for personalized fine-tuning of language models on edge devices.
On-the-Fly Adaptive Distillation of Transformer to Dual-State Linear Attention for Long-Context LLM Serving
Yeonju Ro (University of Texas at Austin), Aditya Akella (University of Texas at Austin)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: This paper proposes a Dual-State Linear Attention (DSLA) module and designs an adaptive distillation framework called DSLAServe, which can convert self-attention layers to linear attention on demand during inference.
One Arrow, Two Hawks: Sharpness-aware Minimization for Federated Learning via Global Model Trajectory
Yuhang Li (Shanghai University), Xiaoqiang Li (Singapore Management University)
OptimizationFederated LearningConvolutional Neural NetworkTransformerImageText
🎯 What it does: This paper studies the introduction of global model trajectories in federated learning, directly measuring and minimizing the sharpness of global loss, and proposes the FedGMT algorithm to alleviate the client drift problem caused by data heterogeneity, while reducing the computational cost of SAM through a single backward propagation.
One Diffusion Step to Real-World Super-Resolution via Flow Trajectory Distillation
Jianze Li (Shanghai Jiao Tong University), Yulun Zhang (Shanghai Jiao Tong University)
RestorationSuper ResolutionKnowledge DistillationDiffusion modelFlow-based ModelAuto EncoderImage
🎯 What it does: A single-step image super-resolution model called FluxSR based on Flux is proposed, utilizing a flow matching framework to achieve knowledge distillation from multi-step to single-step.
One Example Shown, Many Concepts Known! Counterexample-Driven Conceptual Reasoning in Mathematical LLMs
Yinghui Li (Tsinghua University), Philip S. Yu (University of Illinois Chicago)
TransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: A university-level counterexample-driven conceptual reasoning benchmark, COUNTERMATH, has been constructed to evaluate the mathematical reasoning capabilities of LLMs.
One Image is Worth a Thousand Words: A Usability Preservable Text-Image Collaborative Erasing Framework
Feiran Li (University of Chinese Academy of Sciences), Qingming Huang (University of Chinese Academy of Sciences)
GenerationData SynthesisDiffusion modelImageText
🎯 What it does: A text-image collaborative concept elimination framework (Co-Erasing) is proposed, which jointly guides the elimination of undesirable concepts in diffusion models through text prompts and corresponding image prompts.
One Leaf Reveals the Season: Occlusion-Based Contrastive Learning with Semantic-Aware Views for Efficient Visual Representation
Xiaoyu Yang (Shenzhen University of Advanced Technology), Shaoting Zhang (Centre for Perceptual and Interactive Intelligence)
Computational EfficiencyRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: Proposes Occlusion-based Contrastive Learning (OCL), which generates views with semantic differences by randomly occluding images, and performs unsupervised pre-training using only a ViT encoder and contrastive learning.
One Stone, Two Birds: Enhancing Adversarial Defense Through the Lens of Distributional Discrepancy
Jiacheng Zhang (University of Melbourne), Feng Liu (University of Melbourne)
OptimizationAdversarial AttackGaussian SplattingImage
🎯 What it does: A two-stage adversarial defense framework DAD based on distribution difference is proposed, which first optimizes to obtain MMD-OPT as a common signal for discrimination and denoising training through Maximum Mean Discrepancy (MMD); during inference, MMD-OPT is used to determine whether a sample is an adversarial sample, and then either directly classify or classify after denoising.
One Wave To Explain Them All: A Unifying Perspective On Feature Attribution
Gabriel Kasmi (Mines Paris PSL University), Jayneel Parekh (ISIR Sorbonne Universite)
Explainability and InterpretabilityImageMultimodalityAudio
🎯 What it does: A unified method for feature attribution in the wavelet domain (WAM) is proposed, which can be used for post-hoc interpretability in three different modalities: audio, image, and volumetric data.
One-dimensional Path Convolution
Xuanshu Luo (Technical University of Munich), Martin Werner (Technical University of Munich)
ClassificationComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: A fully 1D convolutional network called PathConv is proposed, which maintains image locality by utilizing Hilbert/Z-order curve paths and introducing a path displacement method, achieving a reduction in parameter count to 1/3 while maintaining performance comparable to ResNet.
One-Pass Feature Evolvable Learning with Theoretical Guarantees
Cun-Yuan Xing (Nanjing University), Zhi-Hua Zhou (Nanjing University)
OptimizationTabular
🎯 What it does: A one-pass learning feature evolution framework is proposed, utilizing Kernel Ortho-Mapping (KOM) mismatch to measure the relationship in feature space, and providing complete theoretical error and convergence guarantees.
One-Shot Heterogeneous Federated Learning with Local Model-Guided Diffusion Models
Mingzhao Yang (Fudan University), Xiangyang Xue (Fudan University)
GenerationData SynthesisFederated LearningKnowledge DistillationDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: We propose FedLMG, a heterogeneous federated learning framework that requires only a single round of communication and does not need to deploy a base model on the client side. Clients first locally train classifiers and upload the models; the server uses these models and their BN statistics to guide the diffusion model in generating synthetic images that conform to each client's distribution; finally, the global model is aggregated through distillation or fine-tuning.
One-Step Diffusion Policy: Fast Visuomotor Policies via Diffusion Distillation
Zhendong Wang (University of Texas at Austin), Yu Zeng (NVIDIA)
Knowledge DistillationRobotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningDiffusion modelScore-based Model
🎯 What it does: This paper proposes a One-Step Diffusion Policy (OneDP), which achieves real-time inference for robot control by distilling a pre-trained diffusion policy into a single-step action generator.
One-Step Generalization Ratio Guided Optimization for Domain Generalization
Sumin Cho (Sungkyunkwan University), Kwangsu Kim (Sungkyunkwan University)
Domain AdaptationOptimizationImage
🎯 What it does: A new optimizer called GENIE is proposed, which uses the first-order generalization ratio (OSGR) to pre-tune parameter updates, balancing parameter contributions to enhance domain generalization performance.
OneForecast: A Universal Framework for Global and Regional Weather Forecasting
Yuan Gao (Tsinghua University), Xiaomeng Huang (Tsinghua University)
Graph Neural NetworkTime Series
🎯 What it does: A unified global-regional nested weather forecasting framework called OneForecast is proposed, capable of performing both global and high-resolution regional forecasts.
Online Clustering of Dueling Bandits
Zhiyong Wang (Chinese University of Hong Kong), Zhongxiang Dai (Chinese University of Hong Kong)
Recommendation SystemOptimizationGraph Neural NetworkReinforcement LearningTabular
🎯 What it does: A new online adversarial multi-armed bandit clustering algorithm is proposed, aimed at improving the preference feedback-based decision-making process through user collaboration.
Online Conformal Prediction via Online Optimization
Felipe Areces (Stanford University), John Duchi (Stanford University)
OptimizationTime SeriesFinance Related
🎯 What it does: A parameterized online adaptive confidence set prediction framework based on online gradient descent is proposed, which can simultaneously satisfy long-term coverage and temporal conditional coverage under adversarial and random data.
Online Curvature-Aware Replay: Leveraging $\mathbf{2^{nd}}$ Order Information for Online Continual Learning
Edoardo Urettini (University of Pisa), Antonio Carta (University of Pisa)
OptimizationReinforcement LearningImage
🎯 What it does: This paper proposes a second-order optimization algorithm called OCAR, based on the K-FAC approximate Fisher information matrix, to address the stability and plasticity balance problem in online continual learning.
Online Detection of LLM-Generated Texts via Sequential Hypothesis Testing by Betting
Can Chen (University of California San Diego), Jun-Kun Wang (University of California San Diego)
Anomaly DetectionLarge Language ModelText
🎯 What it does: A betting-based sequential hypothesis testing algorithm is proposed for the online detection of text generated by large language models (LLMs), aimed at quickly and accurately distinguishing between machine-generated text and human-written text.
Online Differentially Private Conformal Prediction for Uncertainty Quantification
Qiangqiang Zhang (Zhongtai Securities Institute for Financial Studies Shandong University), Jinhan Xie (Yunnan University)
OptimizationSafty and PrivacyTabularTime Series
🎯 What it does: An online differential privacy confidence prediction framework (ODPCP) is proposed, which can generate privacy-preserving prediction intervals in real-time from data streams, and is further extended to confidence quantile regression (ODPCQR).
Online Episodic Convex Reinforcement Learning
Bianca Marin Moreno (Univ. Grenoble Alpes), Nadia Oudjane (EDF)
OptimizationReinforcement Learning
🎯 What it does: This paper proposes an online convex reinforcement learning framework (CURL) and presents an algorithm that achieves near-optimal asymptotic unbounded rewards under unknown transition dynamics and variable opponent losses.
Online Laplacian-Based Representation Learning in Reinforcement Learning
Maheed H. Ahmed (Purdue University), Mahsa Ghasemi (Purdue University)
Representation LearningReinforcement LearningGraph
🎯 What it does: An online learning framework is proposed for synchronously updating policies and Laplacian representations in reinforcement learning, utilizing Asymmetric Graph Drawing Objectives (AGDO) for representation learning;
Online Learning in Risk Sensitive constrained MDP
Arnob Ghosh (New Jersey Institute of Technology), Mehrdad Moharrami (University of Iowa)
OptimizationReinforcement Learning
🎯 What it does: An algorithm for online learning risk-sensitive constrained Markov decision processes (RSCMDP) is proposed, and theoretical upper bounds for sublinear regret and constraint violation are provided.
Online Learning in the Random-Order Model
Martino Bernasconi (Bocconi University), Matteo Russo (Sapienza University)
OptimizationReinforcement Learning from Human FeedbackTabularTime Series
🎯 What it does: This paper proposes a general template for adapting stochastic learning algorithms to stochastic order models without significantly affecting their regret guarantees. This allows us to recover improved regret bounds for delayed predictions, constrained online learning, and bandit problems with switching costs.
Online Learning with Unknown Constraints
Karthik Sridharan (Cornell University), Seung Won Wilson Yoo (Cornell University)
Optimization
🎯 What it does: A general meta-algorithm is proposed for the online learning problem with unknown safety constraints, enabling the safe execution of unconstrained actions while minimizing the cumulative loss relative to the best safe policy.
Online Linear Classification with Massart Noise
Ilias Diakonikolas (University of Wisconsin Madison), Nikos Zarifis (University of Wisconsin Madison)
ClassificationOptimizationReinforcement Learning
🎯 What it does: An online linear classifier under the Massart noise model is proposed, which can achieve effective learning with an error of ηT + o(T) when the γ-margin assumption is satisfied; this method is also extended to k-armed context-free games (monotone rewards), obtaining an expected reward improvement of approximately (1-1/k)ΔT compared to random selection.
Online Pre-Training for Offline-to-Online Reinforcement Learning
Yongjae Shin (Korea Advanced Institute of Science and Technology), Woohyung Lim (LG AI Research)
Reinforcement LearningTabularBenchmark
🎯 What it does: This paper proposes a new online pre-training framework called OPT, aimed at addressing the performance degradation issue caused by inaccurate value estimation during the online fine-tuning phase of offline pre-trained agents.
Online Robust Reinforcement Learning Through Monte-Carlo Planning
Tuan Quang Dam (Hanoi University of Science and Technology), Adam Wierman (California Institute of Technology)
OptimizationReinforcement LearningTabularFinance Related
🎯 What it does: A robust version of Monte-Carlo Tree Search (Robust-Power-UCT) is proposed for reinforcement learning problems with uncertainty in transition and reward models.
Online Sparsification of Bipartite-Like Clusters in Graphs
Joyentanuj Das (University of Edinburgh), He Sun (University of Edinburgh)
OptimizationComputational EfficiencyGraph Neural NetworkGraph
🎯 What it does: This paper proposes an online sparsification algorithm that can construct sparse subgraphs in nearly linear time while preserving the structure and metrics of k bipartite-like clusters in undirected and directed graphs.
OOD-Chameleon: Is Algorithm Selection for OOD Generalization Learnable?
Liangze Jiang (École Polytechnique Fédérale de Lausanne), Damien Teney (Idiap Research Institute)
Domain AdaptationMeta LearningText
🎯 What it does: This study investigates how to select the most suitable training algorithm for OOD generalization, proposing OOD-Chameleon, which predicts the best algorithm by learning dataset features.
Open Materials Generation with Stochastic Interpolants
Philipp Höllmer (New York University), Stefano Martiniani (New York University)
GenerationGraph Neural NetworkDiffusion modelGraphStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposed and implemented the OMatG framework, which uses Stochastic Interpolants for the generation and prediction of solid materials.
Open Your Eyes: Vision Enhances Message Passing Neural Networks in Link Prediction
Yanbin Wei (Southern University of Science and Technology), Yu Zhang (Hong Kong University of Science and Technology)
Graph Neural NetworkGraph
🎯 What it does: Proposed GVN (Graph Vision Network) and its efficient variant E-GVN for the link prediction task, utilizing graph visualization to generate images and extracting structural features through a visual encoder, which is then combined with MPNN.
Open-Det: An Efficient Learning Framework for Open-Ended Detection
Guiping Cao (Southern University of Science and Technology), Dongmei Jiang (Pengcheng Laboratory)
Object DetectionComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelVision Language ModelImageMultimodality
🎯 What it does: This paper proposes Open-Det, an efficient end-to-end open detection framework that does not rely on a predefined vocabulary during inference, while achieving both bounding box prediction and free-text target name generation.
OpenworldAUC: Towards Unified Evaluation and Optimization for Open-world Prompt Tuning
Cong Hua (Institute of Computing Technology, Chinese Academy of Sciences), Qingming Huang (University of Chinese Academy of Sciences)
ClassificationOptimizationTransformerPrompt EngineeringVision Language ModelImage
🎯 What it does: Proposed the OpenworldAUC metric and the Gated Mixture-of-Prompts (GMoP) framework, achieving a unified evaluation and optimization of Open-world Prompt Tuning.
Optimal Algorithm for Max-Min Fair Bandit
Zilong Wang (Shanghai Jiao Tong University), Shuai Li (Shanghai Jiao Tong University)
OptimizationReinforcement LearningTabular
🎯 What it does: A Distributed Fair Elimination (DFE) algorithm is proposed for maximum-minimum fair matching in multi-player multi-armed bandits (MP-MAB);
Optimal and Practical Batched Linear Bandit Algorithm
Sanghoon Yu (Seoul National University), Min-hwan Oh (Seoul National University)
OptimizationReinforcement Learning
🎯 What it does: A new batch linear Bandit algorithm BLAE is proposed, which can achieve approximately optimal cumulative returns with a limited number of updates.
Optimal Auction Design in the Joint Advertising
Yang Li (Renmin University of China), Qi Qi (Renmin University of China)
OptimizationGraph Neural NetworkReinforcement LearningTabular
🎯 What it does: This paper proposes the BundleNet framework for single-slot optimal mechanisms and multi-slot automated mechanism learning in the context of joint advertising, aiming to maximize platform revenue while ensuring approximate incentive compatibility and individual rationality.
Optimal Decision Tree Pruning Revisited: Algorithms and Complexity
Juha Harviainen (University of Helsinki), Stefan Szeider (TU Wien)
ClassificationOptimizationTabularBenchmark
🎯 What it does: This paper systematically studies two common pruning operations of decision trees—subtree replacement and subtree lifting—and provides the corresponding optimal pruning algorithms and complexity classifications. It also conducts an in-depth analysis of the parameterized complexity of these problems and compares the effects of heuristic pruning and optimal pruning on multiple public datasets.
Optimal Fair Learning Robust to Adversarial Distribution Shift
Sushant Agarwal (Northeastern University), Ravi Sundaram (Northeastern University)
ClassificationOptimizationAdversarial Attack
🎯 What it does: This paper studies the robustness of the Fair Bayes Optimal Classifier (BOC) under adversarial distribution shift and proves that the randomized fair BOC can maintain Lipschitz robustness against distribution perturbations while satisfying fairness constraints.