ICML 2026 论文列表

Direct Preference Optimization (DPO) has become the dominant approach for aligning large language models with human preferences. However, standard DPO treats all preference pairs uniformly, overlooking the heterogeneous nature of the learning problem: some samples demand sophisticated semantic understanding of the prompt, while others require nuanced discrimination between similar responses. We argue that these two objectives should be disentangled during training. Through gradient analysis, we identify a layer-wise localization phenomenon where semantic complexity predominantly drives lower-layer updates while preference uncertainty modulates upper layers. Building on this insight, we propose Gradient-Guided Disentangled DPO (GDO-DPO), a curriculum framework that independently regulates learning pace along each dimension based on layer-specific gradient stability. Experiments on UltraFeedback and HH-RLHF demonstrate consistent improvements, with GDO-DPO outperforming DPO by 4.1\% on AlpacaEval 2.0 and showing particularly strong gains on reasoning-intensive tasks.

Hierarchical decision-making frameworks are pivotal for addressing complex control tasks, enabling agents to decompose intricate problems into manageable subgoals. Despite their promise, existing hierarchical policies face critical limitations: (i) reinforcement learning (RL)-based methods struggle to guarantee strict constraint satisfaction, and (ii) optimal control (OC)-based approaches often rely on myopic and computationally prohibitive formulations. To reconcile these trade-offs, hierarchical RL-OC architectures have emerged as a promising paradigm. However, the formulation of the lower-level optimization within these frameworks remains underexplored, often relying on heuristic or myopic objectives. In this work, we propose a principled framework that systematically integrates upper-level goal abstraction with structured lower-level decision making. We adopt an inverse optimization approach to inform the structure of the lower-level problem from expert demonstrations, ensuring that the objective of the lower-level policy remains aligned with the overall long-term task goal. To validate the approach, our framework is evaluated on distinct decision making tasks: network-based resource allocation and continuous collision avoidance. Empirical results demonstrate that our method consistently outperforms strong baselines based on end-to-end RL, learning-augmented optimal control, and existing hierarchical RL approaches in both efficiency and decision quality.

We investigate episodic Markov Decision Processes with heavy-tailed losses (HTMDPs). Existing approaches for HTMDPs are conservative in stochastic environments and lack adaptivity in adversarial regimes. In this work, we propose algorithms HT-FTRL-OM and HT-FTRL-UOB for HTMDPs that achieve Best-of-Both-Worlds (BoBW) guarantees: instance-independent regret in adversarial environments and logarithmic instance-dependent regret in self-bounding (including the stochastic case) environments. For the known transition setting, HT-FTRL-OM applies the Follow-The-Regularized-Leader (FTRL) framework over occupancy measures with novel skipping loss estimators, achieving a $\widetilde{\mathcal{O}}(T^{1/\alpha})$ regret bound in adversarial regimes and a $\mathcal{O}(\log T)$ regret in stochastic regimes. Building upon this framework, we develop a novel algorithm HT-FTRL-UOB to tackle the more challenging unknown-transition setting. Under a mild truncative nonnegativity condition on the loss distributions, this algorithm employs a pessimistic skipping loss estimator and achieves a $\widetilde{\mathcal{O}}(T^{1/\alpha} + \sqrt{T})$ regret in adversarial regimes and a $\mathcal{O}(\log^2(T))$ regret in stochastic regimes. Our analysis overcomes key barriers through several technical insights, including a local control mechanism for heavy-tailed shifted losses, a new suboptimal-mass propagation principle, and a novel regret decomposition that isolates transition uncertainty from heavy-tailed estimation errors and skipping bias.

Understanding complex interactions between brain regions is critical for early neurodegenerative disease classification such as Alzheimer’s Disease (AD) and Parkinson’s Disease (PD). While graph-based models are widely used to analyze brain networks, most existing approaches primarily focus on pairwise interactions between directly connected nodes, limiting their ability to capture higher-order dependencies across multiple regions. Although hypergraph-based methods have been proposed to model higher-order relations, many rely on predefined hyperedges or restrict learning to hyperedge weights, reducing flexibility and limiting their capacity to capture multi-resolution structural patterns. In this regard, we introduce an adaptive multi-scale hyperedge learning framework, i.e., MuHL, which constructs hierarchical node features and dynamically learns high-order interaction through continuous hyper-edge construction over multi-resolution graph signals. Extensive experiments on multiple brain network benchmarks demonstrate that MuHL consistently improves disease classification performance across different stages, and further identifies key regions of interest (ROIs) and their group-wise interactions from the learned hyperedges that are associated with disease progression, highlighting its potential as a powerful tool for brain network analysis with neurodegenerative disorders.

Large language model (LLM)-powered multi-agent systems (MAS) enable agents to communicate and share information, achieving strong performance on complex tasks. However, this communication also creates an attack surface where malicious agents can propagate misinformation and manipulate group decisions, undermining MAS safety. Existing embedding-based defenses aim to detect and prune suspicious agents, but their effectiveness depends on a clear separation between the text embeddings of malicious and benign messages. Attackers can circumvent such defenses by crafting messages whose embeddings lie close to benign ones. We analyze this failure mode theoretically and validate it empirically with three attacks, Slow Drift, Benign Wrapper, and Chaos Seeding. Our analysis further reveals a fundamental limitation of embedding-based defenses: because they rely solely on the text embeddings, they ignore token-level confidence signals such as logits, which can remain informative when embeddings are not distinguishable under attack. We propose using confidence scores to prune or down-weight messages during MAS communication. Experiments show improved robustness across models, datasets, and communication topologies. Moreover, we find that the effectiveness of confidence signals decays over communication rounds, highlighting the importance of early intervention.

Masked diffusion language models decode by iteratively unmasking tokens, where the unmasking order defines an ``order of thought'' that strongly influences generation quality yet is typically chosen heuristically. We derive a tractable upper bound on the sequential decoding mismatch, measured by the Kullback–Leibler divergence and expressed in terms of the model’s pathwise log-likelihood, with tightness under sufficient model expressivity. This bound induces a dense self-aware reward for a target sequence $x$ and unmasking order $\sigma$, over ordered paths, casting order selection as a principled policy optimization problem with a frozen denoiser. We instantiate this idea as **Self-Aware Scheduling (SAS)**, which learns a lightweight order policy using Group Relative Policy Optimization and applies seamlessly to both sequential and semi-autoregressive decoding. On Sudoku with 1B MDM, SAS improves puzzle accuracy from $82.0\%$ (best heuristic schedule) to $91.8\%$, and reaches $97.9\%$ with second-stage fine-tuning along learned trajectories. On LLaDA-8B, SAS improves pass@1 on GSM8K from $64\%$ to $76\%$ (full diffusion) and on MBPP from $39.5\%$ to $41\%$, while consistently matching or exceeding heuristic schedules across generation lengths and block sizes.

Large language models (LLMs) demonstrate strong capabilities in high-level semantic reasoning and strategic planning, making them appealing for complex decision-making tasks; however, their quantitative reasoning remains unreliable despite recent progress in tool-augmented and structured inference. To address this limitation, we decouple reasoning from computation by assigning LLMs to planning, analysis, and result interpretation, while delegating numerical computation and statistical inference to specialized external tools. Rather than being hard-coded, these tools are constructed in a constrained and structured manner during planning as explicit intermediate reasoning artifacts, enabling adaptive and scenario-dependent quantitative reasoning. LLMs iteratively analyze tool outputs under diverse market conditions and leverage performance-based feedback to refine subsequent tool selection and construction, thereby forming a bounded self-evolving loop. We instantiate this process through self-play in a controllable digital twin market, DecoupledMarket, where LLM agents continuously test, compare, and adapt their strategies. By coupling high-level planning with robust quantitative execution, the proposed framework enhances the quantitative reliability of LLM-driven decision-making. All code and data are available at https://github.com/MTMQuantAI/Agent-Trading-Arena.git.

Multimodal image fusion aims to integrate complementary information from different modalities, but cross-modal discrepancies and local conflicts often make modality allocation uncertain, causing information loss or artifact propagation. We address this problem by formulating fusion as an observation-conditioned probabilistic decision-learning problem, where local modality contribution is explicitly modeled as a decision variable. Based on this view, we propose a dIfference-aware Decision-lEArning muLtimodal image fusion paradigm (IDEAL). IDEAL uses cross-modal differences as decision triggers and constructs spatial and spectral decision conditions from multi-scale difference attention, power-spectrum energy, complementary spectra, and spectral-entropy reliability. These conditions are mapped to interpretable contribution policies through a symmetric Beta prior, while uncertainty modulation pulls unreliable decisions toward conservative mixing when evidence is insufficient. Extensive experiments on multiple fusion tasks demonstrate stable and competitive performance against state-of-the-art methods. Code is available at: https://github.com/Pon915/IDEAL-main.

Vision-language models (VLMs) have progressed rapidly with large-scale high-quality data and adaptation strategies, yet remain brittle under real-world corruptions, where both visual recognition and language-grounded reasoning degrade. Beyond cascaded image restoration, a natural alternative is parameter-efficient adaptation that aligns corrupted features with clean references; however, Euclidean alignment alone is not semantics-preserving and can even harm downstream reasoning. We attribute this to a semantic misalignment gap, where features become geometrically closer while drifting off the in-distribution support on which multimodal reasoning is calibrated. To address this, we propose Manifold-Adversarial Adapters (MAA), parameter-efficient layer-wise modules for a frozen vision encoder that explicitly steer corrupted features back onto the clean in-distribution manifold rather than merely shrinking feature-space distance. MAA combines paired feature self-distillation with a token-level adversarial manifold constraint to prevent off-manifold semantic shortcuts. At inference, only the adapters are retained, enabling single-stage robustness with negligible overhead and avoiding the latency and semantic drift of restoration pipelines. Across benchmarks and corruption settings, MAA consistently improves performance over strong baselines.

Learning motion latents for robotic manipulation heavily relies on extracting motion patterns from visual sequences, yet effective action abstractions require understanding three-dimensional geometric transformations. Here, we introduce GeoMoLa (Geometry-Aware Motion Latents), which learns discrete motion latent codes by predicting how point clouds evolve during manipulation rather than reconstructing visual observations. This four-dimensional objective – spatial geometry changing through time – forces latent representations to encode actual physical motion rather than appearance patterns. GeoMoLa achieves state-of-the-art performance using only single-view RGB-D input, while existing methods require multi-view reconstruction, succeeding across diverse manipulation benchmarks. Our ablations reveal that geometric prediction is the key to driving performance, quantitatively validating that manipulation depends on spatial understanding. Furthermore, the learned codes exhibit effective motion abstraction: applying them to novel scenes produces physically consistent transformations regardless of visual context. Our real-world experiments also confirm this robustness capability, achieving robust manipulation with minimal demonstrations in cluttered environments where geometric reasoning determines success. Thus, we demonstrate that effective motion latents for robot control can better emerge from understanding motion through its three-dimensional effects rather than pixel-level patterns.

Autoregressive (AR) image generation has recently gained momentum as a scalable alternative to diffusion models, benefiting from unified next-token prediction paradigm and strong instruction following ability. However, AR visual generation must decode excessively long sequences of visual tokens, making inference heavily bottlenecked by the memory footprint and latency of the self-attention KV cache. While KV cache compression is well studied in Large Language Model, its counterparts in AR image generation remain underexplored. The reason is fundamental: visual tokens are highly redundant, and their spatial information density is highly non-uniform. In this work, we introduce SparseAR, a training-free, entropy-aware sparse attention method that is specifically designed for AR image generation and editing. Our key insight is that information-rich regions exhibit higher entropy and require broader attention, while redundant regions show lower entropy and allow aggressive sparsification. Based on this insight, we dynamically identify information-rich regions during decoding and adaptively adjust attention sparsity to reduce KV-cache overhead. SparseAR is plug-and-play and can be readily applied to mainstream AR models. Extensive experiments on four representative AR models across multiple benchmarks demonstrate that SparseAR significantly improves inference efficiency while maintaining, and often even improving, generation and editing quality.

Automated AI research holds great potential to accelerate scientific discovery. However, current LLMs often generate plausible-looking but ineffective ideas. Execution grounding may help, but it is unclear whether automated execution is feasible and whether LLMs can learn from the execution feedback. To investigate these, we first build an automated executor to implement ideas and launch large-scale parallel GPU experiments to verify their effectiveness. We then convert two realistic research problems -- LLM pre-training and post-training -- into execution environments and demonstrate that our automated executor can implement a large fraction of the ideas sampled from frontier LLMs. We analyze two methods to learn from the execution feedback: evolutionary search and reinforcement learning. Execution-guided evolutionary search is sample-efficient: it finds a method that significantly outperforms the GRPO baseline on post-training, and finds a pre-training recipe that outperforms the nanoGPT baseline on pre-training, all within just ten search epochs. Frontier LLMs often generate meaningful algorithmic ideas during search, but they tend to saturate early and only occasionally exhibit scaling trends. Reinforcement learning from execution reward, on the other hand, suffers from mode collapse. It successfully improves the average reward of the ideator model but not the upper-bound, due to models converging on simple ideas. We thoroughly analyze the executed ideas and training dynamics to facilitate future efforts.

Federated learning (FL) enables distributed model training, yet in heterogeneous deployments, Bandwidth-Constrained Clients (BCCs) often contribute inefficiently due to limited uplink bandwidth. In model-heterogeneous FL with fixed small sub-models, BCCs with sub-models may improve quickly in early rounds but become under-parameterized later, resulting in slow convergence and poor generalization. To address this challenge, we propose **FedGMR**, a federated learning framework centered around **Gradual Model Restoration (GMR)**, where GMR progressively increases each client’s sub-model density during training, allowing BCCs to remain effective contributors throughout optimization. To make GMR practical under real-world heterogeneity, FedGMR is realized as an end-to-end workflow with asynchronous coordination and stable, mask-aware aggregation. We further establish convergence guarantees, showing that the aggregation error scales with the **average** sub-model density across clients and rounds, and that GMR provably narrows the gap toward full-model FL. Extensive experiments on FEMNIST, CIFAR-10, ImageNet-100, and StackOverflow demonstrate that FedGMR improves both convergence speed and final accuracy, especially under severe heterogeneity and non-IID data distributions. Code is available at https://github.com/machengjie321/ICML2026-FedGMR.

Existing video object removal methods predominantly rely on diffusion models following a noise-to-data paradigm, where generation starts from uninformative Gaussian noise. This approach discards the rich structural and contextual priors present in the original input video. Consequently, such methods often lack sufficient guidance, leading to incomplete object erasure or the synthesis of implausible content that conflicts with the scene's physical logic. In this paper, we reformulate video object removal as a video-to-video translation task via a stochastic bridge model. Unlike noise-initialized methods, our framework establishes a direct stochastic path from the source video (with objects) to the target video (objects removed). This bridge formulation effectively leverages the input video as a strong structural prior, guiding the model to perform precise removal while ensuring that the filled regions are logically consistent with the surrounding environment. To address the trade-off where strong bridge priors hinder the removal of large objects, we propose a novel adaptive mask modulation strategy. This mechanism dynamically modulates input embeddings based on mask characteristics, balancing background fidelity with generative flexibility. Extensive experiments demonstrate that our approach significantly outperforms existing methods in both visual quality and temporal consistency. The project page is https://bridgeremoval.github.io/.

The generation of immersive and navigable 3D environments is increasingly prevalent with the growing adoption of virtual reality and 3D content. However, recent methods face a fundamental limitation: they cannot produce 3D worlds that simultaneously (i) are navigable over long-range spatial extents and (ii) cover the complete omnidirectional field of view ($360^\circ$ horizontally and $180^\circ$ vertically). To address this challenge, we introduce SphericalDreamer, a method for generating fully immersive and long-range 3D outdoor environments from textual prompts. Our approach is built on the generation of multiple panoramic images, which are subsequently lifted into 3D and fused together while maintaining visual and geometric consistency. SphericalDreamer produces highly detailed, fully immersive 3D environments, while substantially improving scale and navigability compared to prior approaches.

Supervised fine-tuning is the predominant approach for adapting autoregressive vision–language models to downstream tasks. Recent work has shown that this paradigm is highly vulnerable to backdoor attacks, and that existing defenses are ineffective in open-ended generation settings. In response, we propose BYORn, a backdoor-robust fine-tuning framework motivated by the observation that poisoned target responses are often semantically implausible given the corresponding image–text inputs and a pretrained model. BYORn identifies such misaligned responses and dynamically replaces them with alternative responses generated by the model, thereby breaking the correlation between triggers and target outputs. The resulting objective gradient corresponds to the gradient of the empirical estimate of the population risk upper bound over the clean data distribution. Empirically, BYORn consistently improves robustness to backdoor attacks while preserving clean-task performance, establishing a new trade-off frontier between generalization and attack success rate. Finally, we demonstrate that BYORn remains effective against adaptive attacks specifically designed to circumvent the proposed defense.

Recent advances in generative models, such as diffusion models, have raised concerns related to privacy, copyright infringement, and data curation. Prior work has shown that training data can be reconstructed from such models, but existing attacks typically rely on substantial computational resources, access to the training set, or carefully engineered prompts. In this work, we present a low-resource reconstruction attack that operates through seemingly benign prompts and requires little to no access to the training data. Our attack targets **template-memorized images (TMI)**, where recurring layouts and visual structures are memorized during training. We show that such memorization manifests under potentially realistic usage. This raises a possibility of unintentional reconstruction by naive users that don't carry explicit adversarial intent. For example, we observe that a simple prompt such as "blue Unisex T-Shirt" can reproduce visual content depicting a real individual. Beyond extraction, we observe novel phenomena occurring in TMI (e.g., interpolation), raising questions about the novelty of generated content and the effectiveness of established methods for detecting memorized content. Our code is available at \url{https://github.com/TheSolY/lr-tmi}.

Spiking Neural Networks (SNNs) emulate the spiking behavior of biological neurons and are promising for energy-efficient neuromorphic computing. A widely used strategy to train SNNs is to convert pretrained Artificial Neural Networks (ANNs), where the accuracy and efficiency are determined by the spike encoding scheme. Traditional methods based on spike count or timing severely underutilize the available encoding space, leading to large accuracy degradation under low-timestep constraints. More expressive alternatives involve complex dynamics, which hinder scalability and practical deployment. To address these challenges, we propose Temporal Weighted Encoding (TWE). Spikes are implicitly assigned exponentially decaying weights through a recursive integration, drawing an analogy to a temporal bit sequence. We systematically analyze the temporal mismatch caused by this weight pattern and propose temporal relaxation and threshold relaxation to resolve this issue, enabling fast and accurate activation encoding. Extensive experiments demonstrate that TWE achieves negligible conversion loss with significantly fewer timesteps, offering a scalable and efficient solution for SNN deployment.

Catastrophic forgetting remains a fundamental challenge for neural networks when tasks are trained sequentially. In this work, we reformulate continual learning as a control problem where learning and preservation signals compete within neural activity dynamics. We convert regularization penalties into preservation signals that protect prior-task representations. Learning then proceeds by minimizing the control effort required to integrate new tasks while competing with the preservation of prior tasks. At equilibrium, the neural activities produce weight updates that implicitly encode the full prior-task curvature, a property we term the *continual-natural gradient*, requiring no explicit curvature storage. Experiments confirm that our learning framework recovers true prior-task curvature and enables task discrimination, outperforming existing methods on standard benchmarks without replay.