ICML 2026 论文列表

Reinforcement learning algorithms such as GRPO have driven recent advances in large language model (LLM) reasoning. While scaling the number of rollouts stabilizes training, existing approaches suffer from limited exploration on challenging prompts and leave informative feedback signals underexploited, due to context-independent rollout allocation across prompts (e.g., generating 16 rollouts per prompt) and relying heavily on sparse rewards. This paper presents XRPO (eXplore–eXploit GRPO), a unified framework that recasts policy optimization through the principled lens of rollout exploration–exploitation. To enhance exploration, XRPO introduces a mathematically grounded rollout allocator that adaptively prioritizes prompts with higher potential for uncertainty reduction. It further addresses stagnation on zero-reward prompts through an in-context seeding strategy that injects curated exemplars, steering the model into more difficult reasoning trajectories. To strengthen exploitation, XRPO develops a group-relative, novelty-aware advantage sharpening mechanism that leverages sequence likelihoods to amplify low-probability yet correct responses, thereby extending the policy’s reach beyond sparse rewards. Experiments across diverse math and coding benchmarks on both reasoning and non-reasoning models demonstrate that XRPO outperforms existing advances (e.g., GRPO and GSPO) up to 4% pass@1 and 6% cons@32, while accelerating training convergence by up to 2.7x.

Predictive coding (PC) is a biologically plausible alternative to standard backpropagation (BP) that minimises an energy function with respect to network activities before updating weights. Recent work has improved the training stability of deep PC networks (PCNs) by leveraging some BP-inspired reparameterisations. However, the full scalability and theoretical basis of these methods remain unclear. To address this gap, we study the infinite width and depth limits of PCNs. For linear residual networks, we show that the set of width- and depth-stable feature-learning parameterisations for PC is exactly the same as for BP. Moreover, under any of these parameterisations, the PC energy with equilibrated activities converges to the quadratic BP loss when the model width is much larger than the depth, resulting in PC computing the same gradients as BP. Experiments show that, as long as an activity equilibrium is reached, convergence to BP holds for nonlinear models including convolutional networks and transformers. Overall, this work constrains the types of parameterisation that are scalable with PC, while showing a way in which BP can be effectively implemented with only local updates in much wider than deep networks like the brain.

Personalized Federated Learning enables a collection of agents to collaboratively learn individual models without sharing raw data. We propose a new approach in which each agent optimizes a weighted combination of all agents' empirical risks, with the weights learned from data rather than specified a priori. The novelty of our method lies in formulating the estimation of these collaborative weights as a kernel mean embedding estimation problem with multiple data sources, leveraging tools from multi-task averaging to capture statistical relationships between agents. This perspective yields a fully adaptive procedure that requires no prior knowledge of data heterogeneity and automatically transitions between global and local learning regimes. By recasting the objective as a high-dimensional mean estimation problem, we derive finite-sample guarantees on local excess risks for a broad class of distributions, explicitly quantifying the statistical gains of collaboration. To address communication constraints inherent to federated settings, we also propose a practical implementation based on random Fourier features, which allows one to trade communication cost for statistical efficiency. Numerical experiments validate our theoretical results.

Multimodal unlearning aims to eliminate specific data from pretrained multimodal models, which offers significant advantages in data privacy and model efficiency. Current methods struggle to achieve the desired properties of effectiveness, reliability and locality, due to the complex interdependency of unimodal and multimodal knowledge. By introducing a causal perspective, we propose multimodal unlearning with decoupled knowledge components. To promote fine-grained understanding of multimodal context, we introduce Multimodal Variational Inference (MVI) to infer modal-specific and -consistent factors with incomplete sample observation. With foundation of decoupled knowledge, we propose contrastive semantic editing to regulate multimodal unlearning towards refined forgetting. Experiments on privacy- and copyright-sensitive scenarios validate effectiveness of our method across multiple scenarios, ensuring the unlearned model maintains high reliability and locality.

Continual learning for visual question answering (VQA) is typically implemented by training one expert per task and routing each query using task-ID supervision. Yet continual VQA tasks overlap substantially: on the VQA-v2 task stream, a non-native expert outperforms the task’s own expert on $49.9\%$ of queries, so hard routing both wastes transferable knowledge and can be confidently wrong when mismatched. We propose a calibrated Bayesian mixture-of-experts that trains parameter-efficient per-task adapters, learns routing by directly maximizing expected VQA utility, and marginalizes expert identity at inference via Bayesian aggregation in a unified answer space; an entropy penalty prevents the utility objective from collapsing to one-hot routing, enabling evidence pooling across plausible experts. We reach $64.16$ accuracy with $0.63$ forgetting on VQA-v2 CL-LS ($+5.74\%$ accuracy, $-2.99$ forgetting vs. the strongest prior method), $78.81$ with $0.40$ forgetting on TDIUC CL-LS ($+3.10$, $-1.74$), and $83.41$ with $3.21$ forgetting on TDIUC CL-VS ($+1.58$, $-0.82$). Calibration also improves on VQA-v2, reducing ECE from $0.15$ to $0.07$.

Vision-Language-Action (VLA) models hold promise for generalist robotics but currently struggle with data scarcity, architectural inefficiencies, and the inability to generalize across different hardware platforms. We introduce RDT2, a robotic foundation model built upon a 7B parameter VLM designed to enable zero-shot deployment on novel embodiments for open-vocabulary tasks. To achieve this, we collected one of the largest open-source robotic datasets—over $10,000$ hours of demonstrations in diverse families—using an enhanced, embodiment-agnostic Universal Manipulation Interface (UMI). Our approach employs a novel three-stage training recipe that aligns discrete linguistic knowledge with continuous control via Residual Vector Quantization (RVQ), flow-matching, and distillation for real-time inference. Consequently, RDT2 becomes one of the first models that simultaneously zero-shot generalizes to unseen objects, scenes, instructions, and even robotic platforms. Besides, it outperforms state-of-the-art baselines in dexterous, long-horizon, and dynamic downstream tasks like playing table tennis.

Recent progress has expanded the use of large language models (LLMs) in drug discovery, including synthesis planning. However, objective evaluation of retrosynthesis performance remains limited. Existing benchmarks and metrics typically rely on published synthetic procedures and Top-K accuracy based on single ground-truth, which does not capture the open-ended nature of real-world synthesis planning. We propose a new benchmarking framework for single-step retrosynthesis that evaluates both general-purpose and chemistry-specialized LLMs using ChemCensor, a novel metric for chemical plausibility. By emphasizing plausibility over exact match, this approach better aligns with human synthesis planning practices. We also introduce CREED, a novel dataset comprising millions of ChemCensor-validated reaction records for LLM training, and use it to train a model that improves over the LLM baselines under this benchmark.

With diffusion and flow matching models achieving state-of-the-art generating performance, the interest of the community now turned to reducing the inference time without sacrificing sample quality. Consistency Models (CMs), which are trained to be consistent on diffusion or probability flow ordinary differential equation (PF-ODE) trajectories, enable one or two-step flow or diffusion sampling. However, CMs typically require prolonged training with large batch sizes to obtain competitive sample quality. In this paper, we examine the training dynamics of CMs near convergence and discover that CM trajectory tangents -- CM output update directions -- are quite oscillatory, in the sense that they move parallel to the data manifold, not towards the manifold. To mitigate oscillatory trajectory tangents, we propose a new loss function, called the {\em manifold feature distance (MFD)}, which provides manifold-aligned trajectory tangents that point toward the data manifold. Consequently, our method -- dubbed {\em Align Your Trajectory Tangent (AYT)} -- can accelerate CM training by orders of magnitude and even out-perform the learned perceptual image patch similarity metric (LPIPS). Furthermore, we find that our loss enables training with extremely small batch sizes without compromising sample quality. Code is available at \url{https://github.com/1202kbs/AYT}

Self-supervised learning (SSL) has emerged as a powerful paradigm for learning meaningful representations from unlabeled data. However, the standard protocol for evaluating these representations, linear probing, is computationally expensive, sensitive to hyperparameters, and provides limited insight into the geometric structure of the representation space. In this work, motivated by connections between neural network generalization and intrinsic dimension (ID) we propose IdEst, a method for estimating the ID of SSL representations via the Minimum Spanning Tree dimension estimator ($\mathrm{dim}_\mathrm{MST}$). Across diverse datasets, architectures, and SSL pretraining objectives, we show that IdEst strongly correlates with downstream linear probe performances. Furthermore, we demonstrate that IdEst enables efficient hyperparameter selection, significantly reducing the computational cost compared to supervised alternatives. Our results highlight intrinsic dimensionality as a principled geometric proxy for assessing SSL representations, complementing standard supervised probing protocols.

While existing methods for reconstructing hand–object interactions have made impressive progress, they either focus on rigid or part-wise rigid objects—limiting their ability to model real-world objects (e.g., cloth, stuffed animals) that exhibit highly non-rigid deformations—or model deformable objects without full 3D hand reconstruction. To bridge this gap, we present PhysHanDI (Physics-based Reconstruction of Hand and Deformable Object Interactions), a framework that enables full 3D reconstruction of both interacting hands and non-rigid objects. Our key idea is to physically simulate object deformations driven by forces induced from densely reconstructed 3D hand motions, ensuring that the reconstructed object dynamics are both physically plausible and coherent with the interacting hand movements. Furthermore, we demonstrate that such simulation of object deformations can, in turn, refine and improve hand reconstruction via inverse physics. In experiments, PhysHanDI outperforms the state-of-the-art baseline across reconstruction and future prediction.

Discrete flow matching, a recent framework for modeling categorical data, has shown competitive performance with autoregressive models. However, unlike continuous flow matching, the rectification strategy cannot be applied due to the stochasticity of discrete paths, necessitating alternative methods to minimize state transitions. We propose a dynamic-optimal-transport-like minimization objective and derive its Kantorovich formulation for discrete flows with convex interpolants, where transport cost depends solely on inter-state dissimilarity and can be optimized via minibatch strategies. We show that such methods can reduce the number of transitions up to 32 times (1024 to 32) to reach the same generative perplexity without compromising diversity. Additionally, path nondeterminism in discrete flows precludes an instantaneous change-of-variables analogue, preventing precise probability estimation available to continuous flows. We therefore propose two upper bounds on perplexity, enabling principled training, evaluation and model comparison. Finally, we introduce Multimask Flows which outperform masked flows in generative perplexity without compromising diversity, particularly when utilizing minibatch Optimal Transport.

Statistical evaluation aims to estimate the generalization performance of a model using held-out i.i.d. test data sampled from the ground-truth distribution. In supervised learning settings such as classification, performance metrics such as error rate are well-defined, and test error reliably approximates population error given sufficiently large datasets. In contrast, evaluation is more challenging for generative models due to their open-ended nature: it is unclear which metrics are appropriate and whether such metrics can be reliably evaluated from finite samples. In this work, we introduce a theoretical framework for evaluating language models and establish evaluability results for commonly used metrics. We study two categories of metrics: test-based metrics, including integral probability metrics (IPMs), and similarity-based metrics, including Rényi and KL divergences. We show that IPMs with respect to any bounded test class can be evaluated from finite samples up to multiplicative and additive approximation errors. Moreover, when the test class has finite fat-shattering dimension, IPMs can be evaluated with arbitrary precision. In contrast, similarity-based metrics, including Rényi and KL divergences, are not evaluable from finite samples, as their values can be critically determined by rare events. We also analyze the potential and limitations of perplexity as an evaluation method.

Modern sparse language models typically achieve sparsity through Mixture-of-Experts (MoE) layers, which dynamically route tokens to dense MLP "experts." However, dynamic hard routing has a number of drawbacks, such as potentially poor hardware efficiency and needing auxiliary losses for stable training. In contrast, the tokenizer embedding table, which is natively sparse, largely avoids these issues by selecting a single embedding per token at the cost of not having contextual information. In this work, we introduce the Large Lookup Layer (L3), which generalizes embedding tables to model decoder layers as a means of further scaling sparsity. L3 layers use static token-based routing to aggregate a *set* of learned embeddings per token in a *context-dependent way*, allowing the model to efficiently balance memory and compute by caching information in embeddings. L3 has two main components: (1) a systems-friendly architecture that allows for fast training and CPU-offloaded inference with no overhead, and (2) an information-theoretic embedding allocation algorithm that effectively balances speed and quality. We empirically test L3 by training transformers with up to 2.6B active parameters and find that L3 strongly outperforms both dense models and iso-sparse MoEs in both language modeling and downstream tasks.

Reinforcement Learning with Verifiable Rewards is an effective route for post-training to strengthen the reasoning capability of large language models. However, as training proceeds, the learning signal can collapse thus makes the training gain become marginal and ineffective. Specifically, a growing fraction of prompts' rollouts become advantage-degenerated: all the self-generated rollouts show verified-success, making the standard deviation over their rewards be zero; accordingly each rollout's advantage becomes degenerated (zero) as well. Given such rollouts' advantages, the policy-gradient for model optimization eventually vanishes, capping the training performance. We argue that some of these rollouts still contain valuable learning signals but unfortunately omitted with the existing RLVR methods. In this paper, inspired through analyzing the entropy pattern behind golden trajectories produced by external expert models, we propose EchoRL for better exploiting the advantage-degenerated rollouts to further improve the training performance. EchoRL is a lightweight module that first identifies an EchoClip from verified-success rollouts based on their step-level entropy values, and then feeds this clip back as an auxiliary supervision signal in the RL objective. Extensive experiments across 10 benchmarks, 5 LLM backbones, and 7 popular RLVR post-training methods demonstrate that EchoRL consistently improves RLVR post-training with minimal overhead.

Video depth estimation extends monocular prediction into the temporal domain to ensure coherence. However, existing methods often suffer from spatial blurring in fine-detail regions and temporal inconsistencies. We argue that current approaches, which primarily rely on temporal smoothing via Transformers, struggle to maintain strict 3D geometric consistency—particularly under rotations or drastic view changes. To address this, we propose GemDepth, a framework built on the insight that an explicit awareness of camera motion and global 3D structure is a prerequisite for 3D consistency. Distinctively, GemDepth introduces a Geometry-Embedding Module (GEM) that predicts inter-frame camera poses to generate implicit geometric embeddings. This injection of motion priors equips the network with intrinsic 3D perception and alignment capabilities. Guided by these geometric cues, our Alternating Spatio-Temporal Transformer (ASTT) captures latent point-level correspondences to simultaneously enhance spatial precision for sharp details and enforce rigorous temporal consistency. Furthermore, GemDepth employs a data-efficient training strategy, effectively bridging the gap between high efficiency and robust geometric consistency. As shown in Fig.2, comprehensive evaluations demonstrate that GemDepth achieves state-of-the-art performance across multiple datasets, particularly in complex dynamic scenarios.

A popular approach to post-training control of large language models (LLMs) is the steering of intermediate latent representations. Namely, identify a well-chosen direction depending on the task at hand and perturbs representations along this direction at inference time. While many propositions exist to pick this direction, considerably less is understood about how to choose the magnitude of the move, whereas its importance is clear: too little and the intended behavior does not emerge, too much and the model's performance degrades beyond repair. In this work, we propose the first theoretical analysis of steering strength. We characterize its effect on next token probability, presence of a concept, and cross-entropy, deriving precise qualitative laws governing these quantities. Our analysis reveals surprising behaviors, including non-monotonic effects of steering strength. We validate our theoretical predictions empirically on eleven language models, ranging from a small GPT architecture to modern models.

Reinforcement fine-tuning (RFT) has shown promise for training LLM agents to perform multi-turn decision-making based on environment feedback. However, most existing evaluations remain largely in-domain—training and testing are conducted in the same environment or even on the same tasks. In real-worlddeployment, agents may operate in unseen environments with different background knowledge, observation spaces, and action interfaces. To characterize the generalization profile of RFT under such shifts, we conduct a systematic study along three axes: (1) within-environment generalization across task difficulty, (2) cross-environment transfer to unseen environments, and (3) sequential multi-environment training to quantify transfer and forgetting. Our results show that RFT generalizes well across task difficulty within an environment, but exhibits unstable transfer to unseen environments, which correlates with shifts in both semantic priors and observation/action interfaces. In contrast, sequential training yields promising downstream gains with minimal upstream forgetting, and mixture training across environments improves the overall balance. We further provide detailed analyses and deeper insights, and hope our work helps the community develop and deploy generalizable LLM agents.

We present a batched first-order method for solving multiple linear programs in parallel on GPUs. Our approach extends the primal-dual hybrid gradient algorithm to efficiently solve batches of related linear programming problems that arise in mixed-integer programming techniques such as strong branching and bound tightening. By leveraging matrix-matrix operations instead of repeated matrix-vector operations, we obtain significant computational advantages on GPU architectures. We demonstrate the effectiveness of our approach on various case studies and identify the problem sizes where first-order methods outperform traditional simplex-based solvers, depending on the computational environment. This is a significant step toward integer programming algorithms that tightly exploit GPU capabilities. We argue that some specific operations should be allocated to GPUs and performed in full, instead of relying on lightweight heuristic approaches on CPUs.

Conformal prediction converts any black-box predictor into one with finite-sample, distribution-free coverage guarantees, outputting prediction sets $T(x)$ that contain the true label with probability at least $1-\alpha$. To construct these prediction sets, conformal prediction relies on a randomly sampled ``calibration set'' of labeled examples. In many applications, however, this labeled calibration set is costly to collect, creating a tradeoff between upfront labeling cost and downstream utility of the conformal predictor. In this work, we study *conformal prediction with costly label queries*, where unlabeled examples arrive i.i.d. and labels can be queried one at a time. After $m$ queries, we form a conformal predictor; the upfront cost of this predictor is the calibration set size $m$, and its efficiency is the expected prediction set size $\mathbb{E}|T_m(X)|$. We design an online stopping rule $\hat{m}$ that automatically balances the upfront cost against conformal efficiency *while preserving the original conformal guarantee*. Theoretically, we show that under mild regularity assumptions, the expected total cost of our stopping rule matches the best fixed calibration size in hindsight. Experimentally, we find that our stopping rule reduces cost compared to standard choices of $m$ from the literature by 40.6% $\pm$ 2.3%. Finally, we demonstrate a reduction from the probably approximately correct labeling problem of Candès et al. (2025) to CP, under which our stopping rule minimizes the total labeling cost.