ICLR 2025 Papers — Page 2
International Conference on Learning Representations · 3704 papers
Accelerating Neural ODEs: A Variational Formulation-based Approach
Hongjue Zhao (University of Illinois Urbana-Champaign), Huajie Shao (William and Mary)
OptimizationComputational EfficiencyTime SeriesOrdinary Differential Equation
🎯 What it does: A NODE training method based on the variational formula (VF) is proposed—VF-NODE, which uses global integration to replace numerical ODE solvers, significantly reducing the number of function evaluations and suppressing autoregressive errors.
Accelerating Task Generalisation with Multi-Level Skill Hierarchies
Thomas P Cannon, Özgür Şimşek (University of Bath)
Robotic IntelligenceReinforcement Learning
🎯 What it does: This paper proposes and implements Fracture Cluster Options (FraCOs) — a multi-level hierarchical reinforcement learning framework that automatically generates transferable skills from 'fractures' in agent behavior and enhances learning speed and generalization ability in new tasks through hierarchical options.
Accelerating Training with Neuron Interaction and Nowcasting Networks
Boris Knyazev (Samsung), Simon Lacoste-Julien (University of Montreal)
OptimizationComputational EfficiencyGraph Neural NetworkTransformerImageText
🎯 What it does: This paper proposes the NiNo (Neuron Interaction and Nowcasting Networks) method, which accelerates training by periodically predicting network parameters, significantly reducing the training steps of Adam.
Accessing Vision Foundation Models via ImageNet-1K
Yitian Zhang (Northeastern University), Yun Fu (Northeastern University)
ClassificationSegmentationDepth EstimationKnowledge DistillationRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: This paper proposes the Proteus framework, which extracts general visual representations from large vision foundation models (such as DINOv2, CLIP, SynCLR) through knowledge distillation using only the ImageNet-1K dataset, achieving task-agnostic model compression and retraining.
Accurate and Scalable Graph Neural Networks via Message Invariance
Zhihao Shi (University of Science and Technology of China), Feng Wu (University of Science and Technology of China)
Graph Neural NetworkGraph
🎯 What it does: A new subgraph sampling method called TOP is proposed, which is based on message invariance. It can maintain the message passing results of the entire graph while only using message passing among the nodes within the batch, significantly alleviating the neighborhood explosion problem.
ACE: All-round Creator and Editor Following Instructions via Diffusion Transformer
Zhen Han (Alibaba Group), Jingren Zhou (Alibaba Group)
GenerationData SynthesisTransformerDiffusion modelImageTextMultimodalityBenchmark
🎯 What it does: We propose ACE, a versatile generation and editing model based on a diffusion Transformer, which supports various visual tasks such as text-guided, controllable generation, semantic editing, element editing, and local/global editing, as well as multi-turn and long-context interactions.
ACES: Automatic Cohort Extraction System for Event-Stream Datasets
Justin Xu (University of Oxford), Matthew B.A. McDermott
Data-Centric LearningDrug DiscoveryTabularBiomedical DataElectronic Health Records
🎯 What it does: An automated queue extraction system called ACES has been developed, simplifying the definition and reproduction of machine learning tasks and queues in electronic health records (EHR).
Achieving Dimension-Free Communication in Federated Learning via Zeroth-Order Optimization
Zhe Li (Rochester Institute of Technology), Haibo Yang (ComboCurve)
OptimizationFederated LearningComputational EfficiencyText
🎯 What it does: The DeComFL algorithm is proposed, achieving dimension-independent communication in federated learning through zero-order optimization, reducing the communication volume per round from O(d) to O(1);
Action abstractions for amortized sampling
Oussama Boussif (Mila - Quebec AI Institute), Yoshua Bengio (Mila - Quebec AI Institute)
Reinforcement LearningSequential
🎯 What it does: This paper proposes an online action abstraction method called ACTIONPIECE, which can automatically extract and incorporate high-level actions ('chunks') during the training process, thereby reducing the sampling path length and enhancing the sample efficiency and pattern discovery capabilities of reinforcement learning and GFlowNet.
Action Sequence Augmentation for Action Anticipation
Yihui Qiu (Nanyang Technological University), Deepu Rajan (Nanyang Technological University)
RecognitionTransformerAgentic AIVideoSequential
🎯 What it does: Proposes the ActSeq sequence augmentation framework, which augments the observed action sequence and the next action to be predicted separately, and obtains context-free grammar through grammar induction, then uses a cross-tree Earley parser to generate reasonable candidates for the next action; at the same time, it randomly deletes, modifies, and replaces the observed sequence to enhance model robustness.
ActionReasoningBench: Reasoning about Actions with and without Ramification Constraints
Divij Handa (Arizona State University), Chitta Baral (New Mexico State University)
TransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: A new diagnostic benchmark called ACTIONREASONINGBENCH is proposed to comprehensively evaluate the performance of large language models in action and change reasoning (RAC).
Actions Speak Louder Than Words: Rate-Reward Trade-off in Markov Decision Processes
Haotian Wu (Imperial College London), Deniz Gunduz (Imperial College London)
OptimizationTransformerReinforcement LearningAgentic AISequential
🎯 What it does: This paper proposes implicit communication through the actions of agents in Markov Decision Processes (MDP), studies the capacity of the action-state channel and its trade-off with rewards, and designs the Act2Comm framework based on Transformer for collaborative optimization of control and communication.
Activation Gradient based Poisoned Sample Detection Against Backdoor Attacks
Danni Yuan (Chinese University of Hong Kong), Baoyuan Wu (Chinese University of Hong Kong)
Anomaly DetectionImage
🎯 What it does: This paper proposes a method for detecting contaminated samples in backdoor attacks based on the Gradient Circular Distribution (GCD), called AGPD.
Active Learning for Continual Learning: Keeping the Past Alive in the Present
Jaehyun Park (Korea Advanced Institute of Science and Technology), Jae-Gil Lee (Korea Advanced Institute of Science and Technology)
ClassificationImage
🎯 What it does: AccuACL is proposed, a query strategy for Active Continual Learning (ACL) aimed at reducing catastrophic forgetting while quickly learning new tasks under a limited annotation budget.
Active Learning for Neural PDE Solvers
Daniel Musekamp (University of Stuttgart), Mathias Niepert (University of Stuttgart)
Convolutional Neural NetworkTime SeriesBenchmarkPhysics Related
🎯 What it does: This paper proposes AL4PDE, the first active learning benchmark framework for neural PDE solvers, aimed at automating the collection of high-quality training data and evaluating active learning methods.
Active Task Disambiguation with LLMs
Kasia Kobalczyk, Mihaela van der Schaar (University of Cambridge)
TransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: This paper proposes an active task disambiguation method based on Bayesian experimental design, allowing large language models to actively ask questions when faced with ambiguous tasks to obtain additional information.
ACTIVE: Offline Reinforcement Learning via Adaptive Imitation and In-sample $V$-Ensemble
Tianyuan Chen (Beihang University), Xiao Zhang (Beihang University)
Reinforcement LearningTabularBenchmark
🎯 What it does: An offline reinforcement learning algorithm named ACTIVE is proposed, which incorporates V-ensemble and adaptive replication temperature based on implicit TD backups to alleviate the issues of V function overfitting and excessive regularization.
ActSafe: Active Exploration with Safety Constraints for Reinforcement Learning
Yarden As (ETH Zurich), Andreas Krause (ETH Zurich)
Safty and PrivacyReinforcement Learning
🎯 What it does: This paper proposes ACTSAFE, a model-based reinforcement learning algorithm that utilizes model uncertainty as intrinsic reward to achieve safe and efficient exploration in continuous state-action spaces.
Ada-K Routing: Boosting the Efficiency of MoE-based LLMs
Tongtian Yue (Institute of Automation Chinese Academy of Sciences), Jing Liu (Institute of Automation Chinese Academy of Sciences)
OptimizationComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningMixture of ExpertsText
🎯 What it does: This paper proposes the Ada-K routing strategy, allowing each token to dynamically decide how many experts to activate, thereby achieving more efficient MoE inference.
AdaFisher: Adaptive Second Order Optimization via Fisher Information
Damien MARTINS GOMES, Mahdi S. Hosseini (Concordia University)
OptimizationConvolutional Neural NetworkTransformerImageText
🎯 What it does: A new adaptive second-order optimizer, AdaFisher, is proposed, which preprocesses gradients using a diagonal block Kronecker approximation of the Fisher information matrix, significantly improving the convergence speed and generalization performance of deep network training.
AdaGrad under Anisotropic Smoothness
Yuxing Liu (University of Illinois Urbana-Champaign), Tong Zhang (University of Illinois Urbana-Champaign)
OptimizationTransformerSupervised Fine-TuningText
🎯 What it does: This paper studies the convergence of AdaGrad under anisotropic smoothness and noise assumptions, introduces a new anisotropic (L0, L1) smoothness assumption, and provides theoretical convergence upper bounds in large batch settings; it then validates the theory through experiments and compares it with SGD and AdaGrad-Norm.
AdaIR: Adaptive All-in-One Image Restoration via Frequency Mining and Modulation
Yuning Cui (Technical University of Munich), Fahad Shahbaz Khan (Mohammed Bin Zayed University of AI)
RestorationTransformerImage
🎯 What it does: A single model named AdaIR for full-scene image denoising/dehazing/rain removal/deblurring/low-light enhancement is proposed, capable of adaptively recognizing and processing different types of image distortions.
Adam Exploits $\ell_\infty$-geometry of Loss Landscape via Coordinate-wise Adaptivity
Shuo Xie (Toyota Technological Institute at Chicago), Zhiyuan Li (Toyota Technological Institute at Chicago)
OptimizationTransformerLarge Language ModelText
🎯 What it does: This paper proposes a new convergence analysis method, exploring the advantages of the Adam optimization algorithm over SGD when training language models, particularly its performance under ℓ ∞ geometry.
ADAM Optimization with Adaptive Batch Selection
Gyu Yeol Kim (Seoul National University), Min-hwan Oh (Seoul National University)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: Designed and implemented AdamCB, an adaptive mini-batch selection method that combines the Adam optimizer with combinatorial bandit sampling;
Adam-mini: Use Fewer Learning Rates To Gain More
Yushun Zhang (Chinese University of Hong Kong), Ruoyu Sun (Shenzhen Research Institute of Big Data)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes Adam-mini, an optimizer that reduces memory usage of the optimizer state by approximately 50% while maintaining performance close to AdamW. It achieves a significant reduction in the number of learning rates by dividing model parameters into several blocks based on the Hessian structure and using a single learning rate for each block (taking the mean of Adam's v).
ADAM: An Embodied Causal Agent in Open-World Environments
Shu Yu (Shanghai Artificial Intelligence Laboratory), Chaochao Lu (Shanghai Artificial Intelligence Laboratory)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIMultimodality
🎯 What it does: In the Minecraft 1.19 environment, a physical causal intelligent agent named ADAM was built, capable of autonomous interaction, perceiving multimodal information, utilizing causal discovery to construct a technology tree, and executing tasks through planning;
AdaManip: Adaptive Articulated Object Manipulation Environments and Policy Learning
Yuanfei Wang (Peking University), Hao Dong (Peking University)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningDiffusion modelMultimodalityPoint Cloud
🎯 What it does: This paper studies adaptive manipulation strategies for robotic arms under various joint mechanisms, proposing a new experimental environment and dataset. It learns multimodal manipulation strategies based on a 3D visual diffusion policy that can dynamically adjust according to historical actions and feedback.
Adapt-$\infty$: Scalable Continual Multimodal Instruction Tuning via Dynamic Data Selection
Adyasha Maharana (University of North Carolina), Mohit Bansal (University of North Carolina)
OptimizationTransformerSupervised Fine-TuningVision Language ModelMultimodality
🎯 What it does: Proposes Adapt‑∞, which achieves lifelong multimodal instruction fine-tuning through dynamic data selection, significantly reducing forgetting and enhancing forward transfer.
ADAPT: Attentive Self-Distillation and Dual-Decoder Prediction Fusion for Continual Panoptic Segmentation
Ze Yang (Nanyang Technological University), Guosheng Lin (Nanyang Technological University)
SegmentationKnowledge DistillationTransformerImage
🎯 What it does: The ADAPT framework is proposed for the task of continuous learning in panoramic segmentation, where only the decoder is updated at each step while the other modules remain frozen, achieving efficient and sustainable knowledge retention and learning of new tasks.
Adapters for Altering LLM Vocabularies: What Languages Benefit the Most?
HyoJung Han (University of Maryland), Huda Khayrallah (Amazon)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A vocabulary adaptation method based on adapters, VocADT, is proposed, which can integrate new vocabulary sets into LLMs without modifying the weights of the pre-trained model.
Adapting Multi-modal Large Language Model to Concept Drift From Pre-training Onwards
Xiaoyu Yang (Australian Artificial Intelligence Institute), En Yu (Australian Artificial Intelligence Institute)
ClassificationAnomaly DetectionTransformerLarge Language ModelMixture of ExpertsContrastive LearningImageTextMultimodality
🎯 What it does: This paper studies the robustness of multimodal large language models in long-tail and anomalous distribution drift environments, proposing a unified drift theory framework and designing a T-distribution adapter to mitigate the negative effects of tail drift and OOD drift.
Adaptive $Q$-Network: On-the-fly Target Selection for Deep Reinforcement Learning
Théo Vincent (DFKI GmbH), Carlo D'Eramo (Center for AI and Data Science, University of Wrzburg)
Hyperparameter SearchReinforcement LearningSequential
🎯 What it does: An adaptive Q-network, AdaQN, is proposed, which maintains multiple sets of Q-networks with different hyperparameters during training and selects the network with the smallest error as the shared target during each target network update, achieving online hyperparameter tuning without increasing additional environment interactions.
Adaptive backtracking line search
Joao V. Cavalcanti (Massachusetts Institute of Technology), Ashia C. Wilson (Massachusetts Institute of Technology)
OptimizationTabular
🎯 What it does: This paper studies Adaptive Backtracking Line Search (ABLS), which improves the traditional fixed-ratio backtracking method by setting the step size adjustment factor as an online adaptive function related to the degree of violation of conditions.
Adaptive Batch Size for Privately Finding Second-Order Stationary Points
Daogao Liu (University of Washington), Kunal Talwar (Apple)
OptimizationSafty and Privacy
🎯 What it does: An algorithm for finding second-order stationary points (SOSP) under differential privacy constraints is proposed, along with a theoretical convergence upper bound.
Adaptive Camera Sensor for Vision Models
Eunsu Baek (Seoul National University), Hyung-Sin Kim (Seoul National University)
ClassificationRecognitionDomain AdaptationConvolutional Neural NetworkTransformerSupervised Fine-TuningImageBenchmark
🎯 What it does: This paper proposes an adaptive camera sensor control method called Lens, based on the perspective of deep learning models. It utilizes a real-time quality evaluator, VisiT, to significantly improve the model's recognition accuracy in domain shift environments by adjusting parameters such as ISO, shutter speed, and aperture without modifying the model.
Adaptive Data Optimization: Dynamic Sample Selection with Scaling Laws
Yiding Jiang (Carnegie Mellon University), J Zico Kolter
OptimizationData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: An online data selection algorithm called Adaptive Data Optimization (ADO) is proposed, which dynamically adjusts the data ratio from different domains during the model training process. It utilizes domain-level learning curves to predict the learning potential of the model across various domains, thereby achieving adaptive optimization of data allocation.
Adaptive Deployment of Untrusted LLMs Reduces Distributed Threats
Jiaxin Wen (Tsinghua University), Akbir Khan (George Washington University)
AI Code AssistantTransformerLarge Language ModelText
🎯 What it does: In a distributed threat scenario, a dual-layer deployment framework composed of micro-protocols and macro-protocols is proposed to securely utilize large language models (LLMs) that may have biases.
Adaptive Energy Alignment for Accelerating Test-Time Adaptation
Wonjeong Choi (Korea Advanced Institute of Science and Technology), Jaekyun Moon (Korea Advanced Institute of Science and Technology)
Domain AdaptationImage
🎯 What it does: Proposes the Adaptive Energy Alignment (AEA) method, which utilizes energy alignment to achieve online target domain adaptation (TTA), significantly accelerating the model's adaptation process to the target domain.
Adaptive Gradient Clipping for Robust Federated Learning
Youssef Allouah (École Polytechnique Fédérale de Lausanne), John Stephan (École Polytechnique Fédérale de Lausanne)
Federated LearningImage
🎯 What it does: This paper studies an adaptive gradient clipping method ARC suitable for robust federated learning.
Adaptive Length Image Tokenization via Recurrent Allocation
Shivam Duggal (Massachusetts Institute of Technology), William T. Freeman (Massachusetts Institute of Technology)
GenerationCompressionRepresentation LearningTransformerAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: A variable-length image tokenizer ALIT is proposed, which recursively compresses two-dimensional image tokens into one-dimensional latent tokens and adaptively adds new tokens in each iteration, achieving a variable representation from 32 to 256 tokens.
Adaptive Methods through the Lens of SDEs: Theoretical Insights on the Role of Noise
Enea Monzio Compagnoni (University of Basel), Aurelien Lucchi (University of Basel)
OptimizationConvolutional Neural NetworkTransformerImageTabularStochastic Differential Equation
🎯 What it does: This paper first derives the first formal stochastic differential equation (SDE) for SignSGD, and based on this, provides improved SDEs for AdamW and RMSpropW. It then reveals the roles of noise, curvature, and adaptive mechanisms in the dynamics of optimizers through SDE analysis, particularly pointing out the convergence properties of SignSGD under three stages (high signal-to-noise ratio, moderate, low signal-to-noise ratio), and proves its robustness against noise with infinite variance. Additionally, the author proposes a batch size scaling rule for AdamW and verifies the stabilizing effect of decoupled weight decay under high noise conditions.
Adaptive Pruning of Pretrained Transformer via Differential Inclusions
Yizhuo Ding (Fudan University), Yanwei Fu (Fudan University)
RetrievalCompressionOptimizationTransformerLarge Language ModelImageText
🎯 What it does: This paper proposes a dynamic sparsification method based on differential inclusion—Solution Path Pruning (SPP), which can generate a family of Transformer compressed models with different sparsity levels within a single search phase.
Adaptive Rank Allocation: Speeding Up Modern Transformers with RaNA Adapters
Roberto Garcia (Stanford University), Sabri Eyuboglu (Stanford University)
CompressionOptimizationComputational EfficiencyTransformerSupervised Fine-TuningText
🎯 What it does: Proposed the Adaptive Rank Allocation framework and RaNA adapter, which compresses the MLP and QKV linear layers of the Transformer using low-rank decomposition and adaptive routers to improve inference speed;
Adaptive Retention & Correction: Test-Time Training for Continual Learning
Haoran Chen (Fudan University), Yu-Gang Jiang (Fudan University)
ClassificationImage
🎯 What it does: This paper proposes a framework for Adaptive Retention and Correction (ARC) during the testing phase, utilizing Out-of-Task Detection (OTD) to dynamically balance the classifier and correct misclassifications with only one gradient update, thereby alleviating classifier bias in continual learning.
Adaptive Shrinkage Estimation for Personalized Deep Kernel Regression in Modeling Brain Trajectories
Vasiliki Tassopoulou (University of Pennsylvania), Christos Davatzikos (University of Pennsylvania)
Time SeriesBiomedical DataAlzheimer's Disease
🎯 What it does: This paper proposes an adaptive shrinkage deep kernel regression framework that can predict the long-term trajectories of brain biomarkers in real-time based on a small amount of individual follow-up data.
Adaptive teachers for amortized samplers
Minsu Kim (Mila), Yoshua Bengio (Mila)
OptimizationData-Centric LearningReinforcement LearningMultimodalityBiomedical Data
🎯 What it does: This paper proposes a 'teacher-student' framework, where the teacher model generates sampling trajectories in high-loss areas to guide the student (GFlowNet) for incremental sampling, significantly improving exploration efficiency and mode coverage in multimodal distributions.
Adaptive Transformer Programs: Bridging the Gap Between Performance and Interpretability in Transformers
Quoc-Vinh Lai-Dang (Korea Advanced Institute of Science and Technology), Seungah Son (Korea Advanced Institute of Science and Technology)
Explainability and InterpretabilityComputational EfficiencyTransformerText
🎯 What it does: An Adaptive Transformer Programs framework is proposed to enhance the interpretability and performance of Transformers.
AdaRankGrad: Adaptive Gradient Rank and Moments for Memory-Efficient LLMs Training and Fine-Tuning
Yehonathan Refael (Tel Aviv University), Ofir Lindenbaum (Bar-Ilan University)
OptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes AdaRankGrad, which achieves efficient training and fine-tuning of all parameters by adaptively projecting low-rank gradients in the Adam optimization steps, significantly reducing memory usage;
AdaWM: Adaptive World Model based Planning for Autonomous Driving
Hang Wang (Bosch Research North America), Junshan Zhang (University of California)
Autonomous DrivingRecurrent Neural NetworkReinforcement LearningWorld ModelSequential
🎯 What it does: An adaptive world model planning framework, AdaWM, is proposed to address the performance degradation issue that arises in new driving tasks after pre-training.
ADBM: Adversarial Diffusion Bridge Model for Reliable Adversarial Purification
Xiao Li (Tsinghua University), Xiaolin Hu (Tsinghua University)
RestorationAdversarial AttackDiffusion modelImage
🎯 What it does: Proposed and implemented the Adversarial Diffusion Bridge Model (ADBM), which constructs an inverse bridge in the diffusion model to perform denoising defense against adversarial samples.
Add-it: Training-Free Object Insertion in Images With Pretrained Diffusion Models
Yoad Tewel (NVIDIA), Gal Chechik (NVIDIA)
GenerationData SynthesisTransformerDiffusion modelImageTextBenchmark
🎯 What it does: A no-training object insertion method called Add-it, based on a pre-trained diffusion model, has been developed, allowing users to naturally add objects to images through brief text prompts.
Addax: Utilizing Zeroth-Order Gradients to Improve Memory Efficiency and Performance of SGD for Fine-Tuning Language Models
Zeman Li (University of Southern California), Vahab Mirrokni (Google Research)
OptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A memory-efficient fine-tuning method called Addax, which integrates zero-order and first-order gradients, has been developed for fine-tuning large-scale language models.
Adding Conditional Control to Diffusion Models with Reinforcement Learning
Yulai Zhao (Princeton University), Ehsan Hajiramezanali (Genentech)
GenerationOptimizationReinforcement LearningDiffusion modelImageText
🎯 What it does: Fine-tuning a pre-trained diffusion model through reinforcement learning, adding new conditional controls to enable the model to generate samples based on additional labels.
Addressing Label Shift in Distributed Learning via Entropy Regularization
Zhiyuan Wu (University of Oslo), Ali Ramezani-Kebrya (University of Oslo)
OptimizationFederated LearningImage
🎯 What it does: This paper proposes the VRLS method, which improves the maximum likelihood estimation of label transfer through entropy regularization in distributed learning, allowing for a more accurate estimation of the importance ratio of test/training labels; it also combines IW-ERM in a multi-node environment to robustly handle label shifts for internal/external nodes.
ADePT: Adaptive Decomposed Prompt Tuning for Parameter-Efficient Fine-tuning
Pengwei Tang (Renmin University of China), Yong Liu (Renmin University of China)
TransformerSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: A parameter-efficient fine-tuning method called ADePT is proposed, which utilizes short soft prompts and a shared shallow feedforward network to generate input embedding offsets.
ADIFF: Explaining audio difference using natural language
Soham Deshmukh (Carnegie Mellon University), Bhiksha Raj (Carnegie Mellon University)
GenerationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringAudio
🎯 What it does: This paper studies and proposes the audio difference explanation task, constructing two new datasets based on AudioCaps and Clotho.
Adjoint Matching: Fine-tuning Flow and Diffusion Generative Models with Memoryless Stochastic Optimal Control
Carles Domingo-Enrich (Meta AI), Ricky T. Q. Chen (Meta AI)
GenerationOptimizationReinforcement Learning from Human FeedbackReinforcement LearningDiffusion modelFlow-based ModelImageText
🎯 What it does: This paper proposes a stochastic optimal control (SOC) framework based on a memoryless noise schedule for reward fine-tuning of flow matching and diffusion models, and introduces a novel Adjoint Matching algorithm to efficiently solve the SOC problem.
ADMM for Nonconvex Optimization under Minimal Continuity Assumption
Ganzhao Yuan (Peng Cheng Laboratory)
Optimization
🎯 What it does: An IPDS-ADMM algorithm is proposed for solving multi-block non-convex composite optimization problems, utilizing adaptive penalty terms and smooth parameter updates.
ADMM for Structured Fractional Minimization
Ganzhao Yuan (Peng Cheng Laboratory)
OptimizationFinance Related
🎯 What it does: This paper proposes a novel ADMM method for structured fractional minimization problems—FADMM—and presents two variants: FADMM-D (based on the Dinkelbach parameter method) and FADMM-Q (based on the quadratic transformation method).
Advancing Graph Generation through Beta Diffusion
Xinyang Liu (University of Texas at Austin), Mingyuan Zhou (University of Texas at Austin)
GenerationData SynthesisGraph Neural NetworkTransformerDiffusion modelGraph
🎯 What it does: A graph generation framework based on Beta diffusion, Graph Beta Diffusion (GBD), is proposed to address the challenge of generating graph data with mixed discrete and continuous features.
Advancing LLM Reasoning Generalists with Preference Trees
Lifan Yuan (University of Illinois Urbana Champaign), Maosong Sun (Tsinghua University)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: The EURUS series LLM and the ULTRAINTERACT dataset are proposed, constructing a multi-level preference tree and enhancing reasoning capabilities through preference learning;
Advancing Mathematical Reasoning in Language Models: The Impact of Problem-Solving Data, Data Synthesis Methods, and Training Stages
Zui Chen (Jinan University), Zitao Liu (Jinan University)
Data SynthesisOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper explores the introduction of problem-solving data during the Continuous Pre-Training (CPT) phase, investigates different data synthesis methods, and compares the differences in enhancing mathematical reasoning abilities between CPT and Supervised Fine-Tuning (SFT), ultimately training the MathGPT-8B model.
Advancing Out-of-Distribution Detection via Local Neuroplasticity
Alessandro Canevaro (Mercedes-Benz AG), Julian Jordan (Mercedes-Benz AG)
Anomaly DetectionImageBiomedical Data
🎯 What it does: Utilizing the principle of local neural plasticity of the Kolmogorov-Arnold Network (KAN), we achieve OOD detection by comparing the activation patterns of KAN after training and before training;
Advancing Prompt-Based Methods for Replay-Independent General Continual Learning
Zhiqi KANG, Karteek Alahari (Inria)
ClassificationRecognitionTransformerPrompt EngineeringImage
🎯 What it does: A method called MISA is proposed to enhance prompt-based models in general continuous learning (GCL) scenarios, primarily including forgetting-aware adaptation of prompt parameters (ISA-FAM) during the initial session and the use of a parameter-free logit mask at the output layer to mitigate catastrophic forgetting.
Advantage Alignment Algorithms
Juan Agustin Duque (University of Montreal and Mila), Aaron Courville (University of Montreal and Mila)
Recurrent Neural NetworkTransformerReinforcement LearningSequential
🎯 What it does: Proposed and implemented Advantage Alignment and its Proximal version, a counterfactual shaping algorithm based on advantage alignment that can induce reciprocal cooperative behavior in general and social games.
Advantage-Guided Distillation for Preference Alignment in Small Language Models
Shiping Gao (Sun Yat-sen University), Qifan Wang (Meta AI)
Knowledge DistillationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Two methods for preference alignment of small language models using aligned teacher models are proposed: Dual-Constrained Knowledge Distillation (DCKD) and Advantage-Guided Distillation for Preference Alignment (ADPA);
Adversarial Attacks on Data Attribution
Xinhe Wang (University of Michigan), Jiaqi W. Ma (University of Illinois Urbana-Champaign)
Adversarial AttackImageText
🎯 What it does: This study investigates the vulnerability of data attribution methods in value assessment and compensation, proposing two attack strategies—Shadow Attack (based on data distribution and shadow training) and Outlier Attack (based on black-box queries and outlier bias)—to make subtle perturbations to training data in order to enhance their own compensation share.
Adversarial Generative Flow Network for Solving Vehicle Routing Problems
Ni Zhang (Singapore Management University), Xu Chi (Singapore Institute of Manufacturing Technology)
OptimizationGraph Neural NetworkFlow-based ModelGenerative Adversarial NetworkGraph
🎯 What it does: A framework for Vehicle Routing Problem (VRP) using Adversarial Generative Flow Networks (AGFN) is proposed, which directly constructs high-quality paths through the generator;
Adversarial Latent Feature Augmentation for Fairness
Hoin Jung (Purdue University), Xiaoqian Wang (Purdue University)
ClassificationAdversarial AttackGenerative Adversarial NetworkImageTextTabular
🎯 What it does: A method based on Adversarial Latent Feature Augmentation (ALFA) is proposed to mitigate group bias in classification models by performing fair attacks on the latent space and fine-tuning with perturbed samples.
Adversarial Machine Unlearning
Zonglin Di (University of California), Yang Liu (Washington University in St. Louis)
Safty and PrivacyAdversarial AttackConvolutional Neural NetworkTransformerImageText
🎯 What it does: A machine unlearning framework based on Stackelberg game is proposed, actively embedding Membership Inference Attacks (MIA) into the design of unlearning algorithms to eliminate the influence of specified training data.
Adversarial Mixup Unlearning
Zhuoyi PENG, Yi Yang (Hong Kong University of Science and Technology)
ClassificationSafty and PrivacyData-Centric LearningTransformerGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: A machine forgetting framework called MixUnlearn based on adversarial mixup is proposed, which utilizes a generator to synthesize challenging samples by mixing forgotten samples with remaining samples, driving the model to achieve complete forgetting of sensitive data while retaining other knowledge without complete retraining, thus addressing the catastrophic forgetting problem.
Adversarial Perturbations Cannot Reliably Protect Artists From Generative AI
Robert Hönig (ETH Zurich), Florian Tramèr (ETH Zurich)
GenerationAdversarial AttackDiffusion modelImage
🎯 What it does: Evaluated and demonstrated that adversarial perturbation-based art style protection tools cannot prevent style imitation by generative models, proposing and validating several low-cost robust imitation methods.
Adversarial Policy Optimization for Offline Preference-based Reinforcement Learning
Hyungkyu Kang (Seoul National University), Min-hwan Oh (Seoul National University)
Reinforcement LearningTabular
🎯 What it does: An algorithm called APPO for offline preference reinforcement learning is proposed, which implements conservative learning through an adversarial game framework and provides an upper bound on sample complexity.
Adversarial Score identity Distillation: Rapidly Surpassing the Teacher in One Step
Mingyuan Zhou (Google DeepMind), Hai Huang (University of Texas at Austin)
GenerationKnowledge DistillationDiffusion modelScore-based ModelGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes the SiDA framework, which incorporates adversarial loss into single-step diffusion model distillation, combining the Fisher divergence of SiD with the discriminator of Diffusion GAN.
Adversarial Search Engine Optimization for Large Language Models
Fredrik Nestaas (ETH Zurich), Florian Tramèr (ETH Zurich)
Recommendation SystemAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposed Preference Manipulation Attacks, which can induce large language models (LLMs) to prefer the attacker's content by embedding covert instructions or misleading information in web pages or plugin documentation, thereby increasing its recommendation rate;
Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data
Binghui Li (Peking University), Yuanzhi Li (Carnegie Mellon University)
Representation LearningAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This paper studies the impact of adversarial training on feature learning in deep networks, theoretically proving that standard training tends to learn non-robust features that lead to adversarial examples, while adversarial training can explicitly enhance robust feature learning.
Adversarial Training for Defense Against Label Poisoning Attacks
Melis Ilayda Bal (Max Planck Institute for Intelligent Systems), Michael Muehlebach (Max Planck Institute for Intelligent Systems)
OptimizationAdversarial AttackTabularSequential
🎯 What it does: This paper proposes a defense method based on Support Vector Machines (SVM) called FLORAL, designed to resist label poisoning attacks.
Adversarially Robust Anomaly Detection through Spurious Negative Pair Mitigation
Hossein Mirzaei (École Polytechnique Fédérale de Lausanne), Mohammad Hossein Rohban (Sharif University of Technology)
Anomaly DetectionContrastive LearningImage
🎯 What it does: The COBRA method is proposed, which enhances the robustness of anomaly detection under adversarial attacks by self-supervising the generation of pseudo-anomalous samples and combining adversarial training with contrastive learning loss.
Adversarially Robust Out-of-Distribution Detection Using Lyapunov-Stabilized Embeddings
Hossein Mirzaei (Ecole Polytechnique Federale de Lausanne), Mackenzie W Mathis
Anomaly DetectionAdversarial AttackImageAlzheimer's DiseaseOrdinary Differential Equation
🎯 What it does: Proposes the AROS framework, which utilizes Lyapunov stability theory to improve the robustness of OOD detection under adversarial attacks.
Adversaries With Incentives: A Strategic Alternative to Adversarial Robustness
Maayan Ehrenberg (Technion Israel Institute of Technology), Nir Rosenfeld (Technion Israel Institute of Technology)
Autonomous DrivingAdversarial AttackConvolutional Neural NetworkTransformerGenerative Adversarial NetworkImage
🎯 What it does: This paper researches and implements a strategy robust learning method based on opponent incentives, aiming to replace traditional adversarial training.
AdvPaint: Protecting Images from Inpainting Manipulation via Adversarial Attention Disruption
Joonsung Jeon (Korea Advanced Institute of Science and Technology), Sung-eui Yoon
RestorationAdversarial AttackDiffusion modelImage
🎯 What it does: A defense framework named ADVPAINT has been designed and implemented, which directly interferes with the self-attention and cross-attention mechanisms of the Stable Diffusion Inpainting model in image inpainting tasks using adversarial perturbations, thereby preventing malicious actors from replacing or inserting content in published images.
AdvWave: Stealthy Adversarial Jailbreak Attack against Large Audio-Language Models
Mintong Kang (University of Illinois at Urbana Champaign), Bo Li (University of Illinois at Urbana Champaign)
Adversarial AttackLarge Language ModelAudio
🎯 What it does: This paper presents AdvWave, a white-box covert adversarial jailbreak attack framework for large audio-language models (ALM), aimed at generating adversarial audio that can induce the model to output unsafe content while being difficult to detect by humans or systems.
Affine Steerable Equivariant Layer for Canonicalization of Neural Networks
Yikang Li (Peking University), Zhouchen Lin (Northeastern University)
ClassificationConvolutional Neural NetworkImage
🎯 What it does: A steerable EquivarLayer is proposed to achieve equivariance for arbitrary input-output feature types on affine groups and their continuous subgroups, and it is applied as a canonicalization function in image classification tasks.
AFlow: Automating Agentic Workflow Generation
Jiayi Zhang (The Hong Kong University of Science and Technology), Chenglin Wu (DeepWisdom)
OptimizationAI Code AssistantLarge Language ModelAgentic AITextBenchmark
🎯 What it does: By combining the code-represented LLM invocation nodes and edges that form a workflow with predefined operators, and utilizing MCTS with LLM as the optimizer, an agentic workflow is automatically generated and optimized, achieving fully automated workflow generation.
Agent S: An Open Agentic Framework that Uses Computers Like a Human
Saaket Agashe (Simular Research), Xin Eric Wang (Simular Research)
TransformerLarge Language ModelAgentic AITextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: This study investigates a GUI agent, Agent S, based on a multimodal large language model, capable of autonomously completing multi-step computer tasks.
Agent Security Bench (ASB): Formalizing and Benchmarking Attacks and Defenses in LLM-based Agents
Hanrong Zhang (Zhejiang University), Yongfeng Zhang (Rutgers University)
Adversarial AttackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper constructs the Agent Security Bench (ASB), a security assessment benchmark for LLM-based intelligent agents, covering 10 business scenarios, 10 agents, and 400 tools/tasks. It systematically defines and evaluates 10 types of attacks (such as direct/indirect injection, memory poisoning, Plan-of-Thought backdoors, and hybrid attacks) and 11 defense strategies.
Agent Skill Acquisition for Large Language Models via CycleQD
So Kuroki (Sakana AI), Yujin Tang (Sakana AI)
TransformerLarge Language ModelReinforcement LearningAgentic AIImageText
🎯 What it does: A CycleQD framework based on Quality Diversity (QD) and evolutionary algorithms is proposed to achieve multi-skill acquisition in LLMs through expert model fusion and SVD mutation.
Agent-Oriented Planning in Multi-Agent Systems
Ao Li (Hong Kong University of Science and Technology), Yaliang Li (Alibaba Group)
TransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringText
🎯 What it does: Designed and implemented an AOP framework that uses meta-proxies to split, allocate, and evaluate tasks in a multi-agent system, achieving more efficient agent-oriented planning.
Agent-to-Sim: Learning Interactive Behavior Models from Casual Longitudinal Videos
Gengshan Yang (Meta), Angjoo Kanazawa (University of California Berkeley)
Object DetectionGenerationPose EstimationDiffusion modelNeural Radiance FieldVideo
🎯 What it does: Constructing 4D spatiotemporal reconstructions (including animals, scenes, and observers) from long-term casual videos taken with a mobile phone, and using this reconstruction to learn interactive behavior models;
AgentHarm: A Benchmark for Measuring Harmfulness of LLM Agents
Maksym Andriushchenko (Gray Swan AI), Xander Davies (UK AI Security Institute)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: Proposes the AgentHarm benchmark to evaluate the ability and safety of large language model (LLM) agents in completing malicious tasks in multi-step tool invocation scenarios.
AgentOccam: A Simple Yet Strong Baseline for LLM-Based Web Agents
Ke Yang (University of Illinois Urbana-Champaign), Huzefa Rangwala (Amazon)
TransformerLarge Language ModelAgentic AIText
🎯 What it does: A web task execution agent AGENTOCCAM based on a large language model has been designed, achieving strong performance in zero-shot scenarios by optimizing only the observation space and action space.
AgentRefine: Enhancing Agent Generalization through Refinement Tuning
Dayuan Fu (Beijing University of Posts and Telecommunications), Weiran Xu (Beijing University of Posts and Telecommunications)
Robotic IntelligenceTransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: The AgentRefine framework is proposed, which enhances the generalization ability of open-source LLMs by allowing them to self-correct erroneous actions through environmental feedback in generated diverse environments.
Agents' Room: Narrative Generation through Multi-step Collaboration
Fantine Huot (Google DeepMind), Mirella Lapata (Google DeepMind)
GenerationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: A multi-agent collaborative writing framework called AGENTS' ROOM has been developed, which gradually generates long stories using specialized planning and writing agents.
AgentSquare: Automatic LLM Agent Search in Modular Design Space
Yu Shang (Tsinghua University), Yong Li (Tsinghua University)
TransformerLarge Language ModelAgentic AIPrompt EngineeringText
🎯 What it does: Designed and implemented an automated LLM agent search framework called AgentSquare based on a modular design space, which automatically discovers and optimizes LLM agents across six different benchmarks using module evolution, module recombination, and performance predictors.
AgentStudio: A Toolkit for Building General Virtual Agents
Longtao Zheng (Nanyang Technological University), Shuicheng YAN
Large Language ModelAgentic AIVision Language ModelImageVideoTextMultimodalityBenchmark
🎯 What it does: AgentStudio is proposed, integrating a general observation/action space, an interactive online environment, a toolchain, and three major datasets for building and evaluating general virtual agents.
AgentTrek: Agent Trajectory Synthesis via Guiding Replay with Web Tutorials
Yiheng Xu (University of Hong Kong), Tao Yu (University of Hong Kong)
Data SynthesisRobotic IntelligenceTransformerLarge Language ModelVision Language ModelVision-Language-Action ModelTextMultimodality
🎯 What it does: Proposes a fully automated AgentTrek solution that generates multimodal GUI trajectory data from online tutorials and trains visual/text agents;
Agree to Disagree: Demystifying Homogeneous Deep Ensembles through Distributional Equivalence
Yipei Wang (Purdue University), Xiaoqian Wang (Purdue University)
Convolutional Neural NetworkRecurrent Neural NetworkImage
🎯 What it does: Theoretical and experimental research on deep ensemble methods reveals the true mechanisms behind their performance improvement.
AHA: A Vision-Language-Model for Detecting and Reasoning Over Failures in Robotic Manipulation
Jiafei Duan (NVIDIA), Yijie Guo (NVIDIA)
Anomaly DetectionRobotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelImageTextMultimodality
🎯 What it does: This paper presents AHA, a vision-language model aimed at failure detection and reasoning for robotic manipulation tasks, and constructs a large-scale training set through the automated failure generation pipeline FailGen, completing the model's instruction tuning and downstream integration;
AI as Humanity’s Salieri: Quantifying Linguistic Creativity of Language Models via Systematic Attribution of Machine Text against Web Text
Ximing Lu (University of Washington), Yejin Choi (University of Washington)
GenerationRetrievalTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This study investigates a method to quantify the linguistic creativity of texts through the reconstruction of retrieved web text segments, and proposes the CREATIVITY INDEX and the DJ SEARCH algorithm.
AI Sandbagging: Language Models can Strategically Underperform on Evaluations
Teun van der Weij, Francis Rhys Ward (Imperial College London)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper studies the possibility of AI systems strategically underestimating their own performance (i.e., bagging behavior) in capability assessments, and demonstrates the ability of modern language models to achieve selective underestimation in dangerous capability assessments through prompting and fine-tuning; it also explores the feasibility of hiding dangerous capabilities through password locking.