ICML 2026 论文列表

Diffusion-based Q-learning has emerged as a powerful paradigm for offline reinforcement learning, but its reliance on multi-step denoising makes both training and inference computationally expensive and brittle. Recent efforts to accelerate diffusion Q-learning toward single-step action generation typically introduce auxiliary networks, policy distillation, or multi-phase training, which frequently compromise simplicity, stability, or performance. To address these limitations, we introduce Bootstrapped Flow Q-Learning (BFQ), a novel framework that enables accurate single-step action generation during both training and inference—without auxiliary networks or distillation procedures. BFQ adopts a divide-and-conquer view of the displacement vector along the flow path: it begins by learning short-range displacements that can be accurately estimated from the Flow Matching marginal velocity, and bootstraps these components to directly learn a noise-to-action mapping in a single step. This formulation eliminates multi-step denoising, resulting in a learning procedure that is substantially faster, simpler, and more robust. Extensive D4RL evaluations show that BFQ improves performance while significantly reducing computational cost compared to multi-step diffusion baselines, demonstrating that single-step action generation suffices for high-performance offline Reinforcement Learning.

Agentic multimodal models have garnered significant attention for their ability to leverage external tools to tackle complex tasks. However, it is observed that such agents often meet premature interaction collapse, caused by two primary reasons: 1) the terminal reward often appending on the last token prevents the advantage from distinguishing trajectories with exploratory behavior; 2) excessively redundant context hinders the agent from absorbing useful feedback. To address these issues, we propose the Deepening Reasoning MMSearchAgent, the framework leverages the structural proximity to derive advantage signals from the whole rollout trajectories in an entire batch, such that trajectories of different lengths are further encouraged to be generated, even when containing the same correct answer. Additionally, differentiated gaussian rewards are employed to dynamically calibrate interaction tolerance, thereby ensuring information reliability and reducing redundancy. To support multi-turn interaction training, we have constructed a multi-step deep-reasoning dataset including 3602 high-quality QA pairs with at least 3 reasoning steps. Extensive experiments demonstrate that our method achieves state-of-the-art performance, outperforming the MMSearch-R1 by 8.4\% on FVQA-test.

State-of-the-art reasoning models utilize long chain-of-thought (CoT) to solve increasingly complex problems using more test-time computation. In this work, we explore a long CoT setting where the model makes up to K successive attempts at solving a problem, in which each attempt is allowed to build on earlier ones after the model receives a hard verifier feedback. This motivates RL methods that can harness per-attempt rewards by carefully weighting individual attempts. We study optimizing the Verification@K reward (the model succeeds by the K-th attempt) and show that naively weighing the attempts by their pass/fail results in biased gradients. We introduce Calibrated Attempt-Level (CAL) GRPO by devising a weighing strategy to obtain unbiased gradients while maintaining small variance. Our theory reveals how incorporating per-attempt rewards influences the training and the eventual Verification@K performance. Experiments, baselines, and ablations on synthetic and real data corroborate our theory and the benefits of CAL-GRPO over vanilla GRPO as well as naive weighting.

Optimizing pretraining data composition is pivotal for LLM generalization. While dynamic mixing outperforms static strategies by capturing evolving training dynamics, current methods fail to reconcile computational efficiency with sample efficiency and structural flexibility for diverse pipelines.We introduce \textbf{Actor--Critic Online Data Mixing (AC-ODM)}, which approaches data mixing from a reinforcement learning perspective with a parameterized policy that we theoretically prove to act as a dynamic linear surrogate maximizing the constructive interference of gradients. To enhance practical flexibility, AC-ODM supports two operational modes: (i) a \textbf{proxy mode} for fixed, pre-prepared corpora, where a policy learned on a small model is transferred to a larger target; and (ii) a \textbf{non-proxy mode} for direct end-to-end training from scratch without priors. Empirically, AC-ODM significantly outperforms prior methods in convergence speed and downstream accuracy across various architectures. On Pythia-1B, it reaches optimal validation perplexity using up to 66\% fewer training steps than competitive baselines, delivering a 27.5\% relative improvement in MMLU accuracy and a 2.23$\times$ higher pass@1 on HumanEval, all while incurring a virtually negligible ($~$0.4\%) per-step wall-clock increase and only 2\% additional memory overhead. Code is available at \url{https://github.com/DANG-ai/AC-ODM}.

Recent diffusion large language models (dLLMs) have demonstrated both effectiveness and efficiency in reasoning via a block-based semi-autoregressive generation paradigm. Despite their progress, the fixed-size block generations remain a critical bottleneck for effective and coherent reasoning. (I) From a global perspective, different reasoning tasks would correspond to different optimal decoding block sizes, which makes a "one-size-fits-all" assumption ineffective. (II) Even within a single reasoning task, the rigid block partitioning would break the logical flow and reduce reasoning coherence. Through empirical observations, we reveal that, for block-wise entropy, incorrect reasoning exhibits a fluctuating and unsteady trend between blocks, while the correctly generated tasks follow a consistent descending paradigm. Therefore, this paper proposes b1, a novel post-training framework that learns dynamic-size reasoning blocks via a Monotonic Entropy Descent objective with reinforcement learning to enhance reasoning coherence. b1 integrates seamlessly as a plug-and-play module with existing dLLM's post-training algorithms. Extensive experiments across various reasoning benchmarks showcase b1's consistent improvement over fixed-size block baselines. Our code has been released at https://github.com/YanJiangJerry/Block-R1.

Identifying optimal intervention sequences from offline data to guide temporal systems toward target outcomes is a critical challenge with profound implications for fields like personalized medicine. While existing methods are mostly evaluated in offline settings, practical applications demand online, adaptive strategies that can respond in real-time. To address this, we propose \textbf{G}oal-conditioned \textbf{I}ntervention via \textbf{F}actual-\textbf{T}argeted Training (\textbf{GIFT}), a novel framework for learning sequential intervention policies from observational data. GIFT learns a goal-conditioned policy by rescaling rewards with clipped importance weights, stabilizing learning and steering toward the target. Under standard assumptions, the induced operator has a unique fixed point and our procedure converges to it. We also bound the bias from clipping and approximation via the gap to the policy’s true value. Experiments show GIFT significantly outperforms existing methods in creating goal-conditioned policies for online deployment.

Model-based reinforcement learning aims to construct world models for efficient sampling. Current mainstream algorithms can be broadly categorized into two paradigms: maximum likelihood and value-aware world models. The former employs structured Recurrent/Transformer State-Space Models to capture environmental dynamics but overlooks task-relevant features. The latter prioritizes decision-critical states but suffers from sub-optimal performance. While recent efforts have sought to integrate these approaches, they typically rely on auxiliary modules or heavy external priors that significantly increase computational complexity. In this work, we propose a Value-Aligned World Model, a minimalist framework designed to synergize these two paradigms with negligible overhead. Specifically, We introduce an intrinsic latent-space value-alignment regularization that compels the world model to prioritize task-relevant features while maintaining the structural integrity of stochastic dynamics. To ensure stable optimization, we develop an adaptive weighting mechanism that acts as a self-regulating curriculum, balancing reconstruction fidelity with decision-making utility. Extensive experiments on Atari 100k and DeepMind Control benchmarks demonstrate that our algorithm consistently boosts existing methods with minimal added code and computational overhead. Code is available at supplementary material.

Parametric Computer-Aided-Design (CAD) modeling from human intent remains challenging, particularly during the conceptual design stage, where design goals are expressed through incomplete and unstructured modalities (e.g., hand-drawn sketches and textual descriptions). In this work, we rethink the human intent-to-CAD pipeline and propose a unified method that directly maps multi-level human intents to executable codes, without assuming the prior existence of target CAD models. To support our study, we construct HiCAD, the first large-scale dataset aligning hand-drawn sketches, textual descriptions, and parametric CAD codes. Based on this, we introduce HiCAD, a two-stage framework comprising Cooperative Multi-Task Alignment to bridge the representational gap between heterogeneous inputs, and Spatial-Aware Reinforcement Learning to enforce geometric and topological consistency. Extensive experiments demonstrate that our method significantly outperforms existing baselines across multiple tasks, validating its effectiveness and robustness in transforming heterogeneous human intents into high-fidelity parametric CAD models. Our project page: https://zqwlearning.github.io/HiCAD.

Multimodal reasoning for ultra-high-resolution (UHR) remote sensing (RS) is usually bottlenecked by visual evidence acquisition: the model necessities localizing tiny task-relevant regions in massive pixel spaces. While Agentic Reinforcement Learning with Verifiable Rewards (RLVR) using zoom-in tools offers a path forward, we find that standard reinforcement learning struggles to navigate these vast visual spaces without structured domain priors. In this paper, we investigate the interplay between post-training paradigms: comparing Cold-start Supervised Fine-Tuning (SFT), RLVR, and Agentic RLVR on the UHR RS benchmark. Our controlled studies yield a counter-intuitive finding: high-quality Earth-science text-only QA is a primary driver of UHR visual reasoning gains. Despite lacking images, domain-specific text injects the concepts, mechanistic explanations, and decision rules necessary to guide visual evidence retrieval. Based on this, we propose a staged knowledge injection recipe: (1) cold-starting with scalable, knowledge-graph-verified Earth-science text QA to instill reasoning structures; and (2) "pre-warming'' on the same hard UHR image–text examples during SFT to stabilize and amplify subsequent tool-based RL. This approach achieves a 60.04\% Pass@1 on XLRS-Bench, significantly outperforming larger general-purpose models (e.g., GPT-5.2, Gemini 3.0 Pro, Intern-S1) and establishing a new state-of-the-art. We provide both the automated data pipeline and the rigorous ablation studies that validate this "Text-Before-Vision'' paradigm. Datasets and code will be released.

Current reinforcement learning (RL) methods for code generation are predominantly optimized on Python, showing weak generalization to other programming languages (PLs). Although leveraging multilingual solutions offers richer semantics and a wider search landscape, naive independent training across languages suffers from optimization imbalance and fails to effectively transfer knowledge from high-resource languages. We propose Group Cross-lingual Relative Policy Optimization (GXPO), which forms training groups by generating solutions for the same problem in multiple PLs and jointly optimizes language-specific and cross-language signals, enabling more balanced optimization and improved transfer to low-resource PLs. We additionally introduce Multilingual LiveCodeBench (ML-LCB), extending LiveCodeBench to a unified multilingual evaluation setting. On ML-LCB across 8 PLs, GXPO consistently improves performance, with pronounced gains on low-resource PLs, demonstrating scalable multilingual RL for language-consistent code generation.

We introduce Debate2Create (D2C), a multi-agent LLM framework that formulates robot co-design as structured, iterative debate grounded in physics-based evaluation. A design agent and control agent engage in a thesis-antithesis-synthesis loop, while criterion-specific LLM judges provide multi-objective feedback to steer exploration. Across five MuJoCo locomotion benchmarks, D2C achieves the highest default-normalized score among the evaluated LLM-based and black-box baselines, with gains up to 3.2x on Ant and nearly 9x on Swimmer. Iterative debate yields 18-35% gains over compute-matched zero-shot generation, and D2C-generated rewards transfer to default morphologies in 4/5 tasks. These results suggest that structured, simulator-grounded multi-agent interaction is a useful mechanism for joint morphology-reward optimization under a fixed-topology, per-candidate-RL protocol. Project page: debate2create.github.io.

Active vision, also known as active perception, refers to actively selecting where and how to look in order to gather task-relevant information. It is a critical component of efficient perception and decision-making in humans and advanced embodied agents. With the rise of Multimodal Large Language Models (MLLMs) as central planners in robotic systems, the lack of methods for equipping MLLMs with active perception has become a key gap. We first provide a systematic definition of MLLM-based active perception tasks and show that GPT-o3's zoom-in strategy can be viewed as a special case, though it suffers from low efficiency and inaccurate region selection. To address these issues, we propose Active-o3, a reinforcement learning framework built on GRPO that equips MLLMs with active perception capabilities. Leveraging a modular sensing-action design and a dual-form reward, Active-o3 autonomously learns efficient and stable region selection strategies without explicit supervision. We further establish a comprehensive benchmark covering both open-world tasks (small/dense-object grounding) and domain-specific scenarios (remote sensing, autonomous driving, interactive segmentation). Experimental results demonstrate that Active-o3 significantly enhances active perception capabilities compared to Qwen2.5-VL-CoT. Moreover, we show that our RL framework not only preserves the model’s general understanding ability but can also serve as a proxy task for leveraging perception data, further improving performance on benchmarks such as RealWorldQA. We hope that our work can provide a simple codebase and unified evaluation protocol to facilitate future research on active perception with MLLMs.

The rapid development of large language models (LLMs) has raised concerns regarding the inclusion of private or inappropriate data during training, which has led to growing interest in LLM unlearning. Many existing LLM unlearning approaches rely on prediction loss-based optimizations, such as maximizing the loss on the forget set. However, these methods often face issues such as over-forgetting and poor model utility. In this work, we address these issues by introducing a novel perspective that shifts the unlearning optimization target to reducing data attribution instead. We propose the first LLM unlearning framework based on data attribution rewards called DareU that employs reinforcement learning to update the LLM and reduce the attribution score of generated responses (i.e., de-attribute) to the forget data owners. Experimental results using an LLM classifier as an efficient approximation of attribution demonstrate that DareU outperforms existing baseline approaches, achieving effective unlearning while balancing forget quality and model utility.

Reinforcement learning (RL) has advanced LLM agents on verifiable tasks but remains challenging for open-ended tasks with vast solution spaces (e.g., complex travel planning). Lacking objective ground truth, current RL algorithms rely on reward models assigning scalar scores to individual responses. We contend such pointwise scoring induces discrimination collapse: reward model fails to distinguish subtle advantages among trajectories, compressing intra-group rewards into a narrow range. This drowns effective reward signals in reward model noise, causing optimization stagnation. To address this, we propose ArenaRL, a reinforcement learning paradigm shifting from pointwise scalar scoring to intra-group relative ranking. ArenaRL introduces a process-aware pairwise evaluation with multi-level rubrics for fine-grained relative scoring. Meanwhile, we construct an intra-group adversarial arena and devise a tournament-based ranking scheme to obtain stable advantage signals. ArenaRL achieves high-precision advantage estimation with only $O(N)$ computational complexity, striking a favourable balance between efficiency and accuracy. Furthermore, to address the lack of full-cycle benchmarks for open-ended agents, we introduce two high-quality benchmarks: Open-Travel and Open-DeepResearch, encompassing full training and multi-dimensional evaluation pipelines. Extensive experiments across three open-ended tasks validate the effectiveness of ArenaRL.

Adversarial imitation learning (AIL) achieves high-quality imitation by mitigating compounding errors inherent to behavioral cloning (BC), yet its adversarial optimization frequently leads to training instability. A class of non-adversarial Q-based imitation learning (IL) methods, exemplified by IQ-Learn, has emerged to address this instability and is widely believed to outperform BC by leveraging online environment interactions. In this paper, we revisit IQ-Learn and prove that it in fact reduces to BC: it admits an imitation gap lower bound with quadratic dependence on the horizon and therefore remains susceptible to compounding errors. Our theoretical analysis reveals why online interactions fail to help: IQ-Learn uniformly suppresses Q-values for all actions at states not covered by demonstrations, preventing generalization beyond demonstrations. To address this fundamental limitation, we introduce Dual Q-DM, a new Q-based IL method built on Bellman constraints. Crucially, Bellman constraints drive value flow: Q-values propagate from demonstrated to unvisited states through environment dynamics, enabling generalization beyond demonstrations. We prove that Dual Q-DM is equivalent to AIL and can recover expert actions at unvisited states, thereby mitigating compounding errors. To the best of our knowledge, Dual Q-DM is the first non-adversarial IL method that is theoretically guaranteed to eliminate compounding errors. Experimental results further corroborate our theoretical findings.

Speculative decoding can significantly accelerate LLM serving, yet most deployments today disentangle speculator training from serving, treating speculator training as a standalone offline modeling problem. We show that this decoupled formulation introduces substantial deployment and adaptation lag: (1) high time-to-serve, since a speculator must be trained offline for a considerable period before deployment; (2) delayed utility feedback, since the true end-to-end decoding speedup is only known after training and cannot be inferred reliably from acceptance rate alone due to model-architecture, diverse prompt engineering, and system-level overheads; and (3) domain-drift degradation, as the target model is repurposed to new domains and the speculator becomes stale and less effective. To address these issues, we present Aurora, a unified training–serving system that closes the loop by continuously learning a speculator directly from live inference traces. Aurora reframes online speculator learning as an asynchronous reinforcement-learning problem: accepted tokens provide positive feedback, while rejected speculator proposals provide implicit negative feedback that inherits the online traffic failure signals. Our design integrates an SGLang-based inference server with an asynchronous training server, enabling hot-swapped speculator updates without service interruption. Crucially, Aurora supports day-0 deployment: a speculator can be served immediately and rapidly adapted to live traffic, improving system performance while providing immediate utility feedback. Across experiments, Aurora achieves a 1.5$\times$ day-0 speedup on recently released frontier models (e.g., MiniMax M2.1 229B and Qwen3-Coder-Next 80B). Aurora also adapts effectively to distribution shifts in user traffic, delivering an additional 1.25$\times$ speedup over a well-trained but static speculator on widely used models (e.g., Qwen3 and Llama3).

Existing backdoor attacks on Reinforcement Learning (RL) typically rely on unrealistic white-box access to victim parameters, rewards, or observations. Inspired by real world behaviors, we introduce the Supply-Chain Backdoor (SCAB) attack to demonstrate that such assumptions are unnecessary. SCAB targets the common practice of training with third-party policies, poisoning the dataset solely through a black-box of legitimate agent-environment interactions. With only 3% data corruption, SCAB demonstrates a peak attack success rate exceeding 90% and reduces victim returns by 80%. These findings expose a critical vulnerability in the modern RL supply chain, highlighting that reliance on untrusted external agents constitutes a severe and practical security risk.

Efficient LLM inference research has largely focused on reducing the cost of each decoding step (e.g., using quantization, pruning, or sparse attention), typically applying a uniform computation budget to every generated token. In practice, token difficulty varies widely, so static compression can over-compute on easy steps and under-compute on hard ones. We study \emph{dynamic budget allocation} for autoregressive decoding: learning how much computation to spend \emph{per token} from within a single model. Self-Optimizing Language Models (SOL) pair a frozen LLM with a lightweight policy network that reads the LLM hidden state and selects a discrete \emph{efficiency action} at each decode step. Actions can jointly control (i) token-level attention sparsity, (ii) structured activation pruning in the MLP, and (iii) activation quantization bit-width, while leaving the base model weights unchanged. We train the policy with group-relative policy optimization on teacher-forced episodes: the token sequence is fixed, while we sample multiple compute schedules (i.e., “counterfactual” schedules that vary only the efficiency actions for the same token path) and compare their likelihoods under the same supervision. Our reward trades off language‑model quality against soft penalties that encourage episode‑average budget usage to match a requested target. Across model variants and compute regimes, SOL improves quality at matched budget over static allocation and strong random schedule search, offering a complementary axis for inference‑efficiency optimization. SOL discovers a better quality-efficiency pareto-front across all our experiments and improves MMLU accuracy by up to 7.3\% over uniform budget allocation strategies. \href{https://github.com/akhauriyash/SOL}

Value factorization eases non-stationarity in MARL, but its static coordination assumptions hinder generalization on long-horizon tasks with shifting dependencies. Prior VQ-VAE methods abstract trajectories yet miss time-varying inter-agent dependencies. We present TACTIC, a CTDE framework with three components: (i) VQ-VAE-based trajectory abstraction that learns discrete task-semantic classes; (ii) semantic-conditioned sparse coordination graphs that adapt dependencies by pruning edges according to variance-based pairwise payoff sensitivity; and (iii) a pretrained, frozen trajectory-class predictor that conditions local policies while decoupling task recognition from control. On SMAC and SUMO, TACTIC shows strong overall competitiveness and adaptive coordination under sparse rewards and dynamic task structures.