ICML 2026 Papers

To navigate partially observable visual environments, recent VLM agents increasingly internalize world modeling capabilities directly into their policies via explicit CoT reasoning with reinforcement learning (RL). However, mere passive exploitation of reasoning on visited states is insufficient for sparse-reward agentic tasks, as it lacks the epistemic drive to actively uncover the *known unknown* required for robust generalization. We ask: *Can VLM agents actively find signals that challenge and update their internal world model through curiosity-driven exploration?* In this work, we propose **GLANCE**, a unified framework that bridges reasoning and exploration by grounding the agent's linguistic world model into the stable visual representations of an evolving target network. Crucially, **GLANCE** leverages the discrepancy between linguistic prediction and visual reality as an intrinsic curiosity signal within reinforcement learning, steering the agent to actively explore areas where its internal model is uncertain. Extensive experiments across a series of agentic tasks show the effectiveness of **GLANCE**, and demonstrate that aligning *what the agent thinks* with *what the agent sees* is key to solving complex or sparse agentic tasks.

The combination of exponentially large action spaces, stochastic dynamics, and long-horizon decision-making under limited resources makes Sequential Stochastic Combinatorial Optimization (SSCO) particularly challenging for reinforcement learning. Hierarchical Reinforcement Learning (HRL) offers a natural decomposition, but it places the high-level policy in a Semi-Markov Decision Process (SMDP) where actions have variable durations, making it difficult to learn a world model that is suitable for planning. We introduce a model-based hierarchical framework for sequential stochastic combinatorial decision-making that directly addresses this issue. Our method combines a latent-space tree-search planner with an SMDP-aware world model for variable-duration decisions. A multi-timescale objective structures the latent dynamics so that transition magnitudes reflect the effective temporal scales of abstract actions, enabling efficient lookahead under adaptive temporal abstraction. We further learn a subgoal-conditioned budget policy jointly with the world model to support context-aware resource allocation. Across challenging SSCO benchmarks, our method outperforms strong baselines.

An important emerging application of coding agents is *agent harness optimization*: the iterative improvement of a *target agent* by editing and evaluating its code. Despite its relevance, the community lacks a systematic understanding of coding agent performance on this task. Harness optimization differs from conventional software engineering: agent harnesses interleave deterministic code with stochastic LLM completions, requiring structured capture of both intermediate execution traces and downstream outcomes. To address these challenges, we introduce (1) VeRO (Versioning, Rewards, and Observations), an *outer harness* that provides versioned snapshots, budget-controlled evaluation, and structured execution traces of *target harnesses*, and (2) VeRO-Bench, a benchmark suite of target agents and tasks with reference evaluation procedures. Using VeRO, we conduct an empirical study comparing optimizers across tasks and analyzing which modifications reliably improve target agent harnesses. We release VeRO to support research on agent optimization as a core capability for coding agents. Code is available at https://github.com/scaleapi/vero.

Effective exploration remains a key challenge in model-based reinforcement learning (MBRL), especially in high-dimensional continuous control tasks where sample efficiency is critical. Recent work addresses this by using learned policies as proposal distributions for Model-Predictive Path Integral (MPPI) planning. Early approaches update the sampling policy independently of the planner, typically via deterministic policy gradients with entropy regularization. However, since the data distribution is induced by the MPPI planner, misalignment between the policy and planner degrades value estimation and long-term performance. To address this, recent methods explicitly align the policy with the planner by minimizing KL divergence to the planner distribution or by incorporating planner-guided regularization. In this work, we unify these approaches under the Policy Optimization–Model Predictive Control (PO-MPC) framework, a family of KL-regularized MBRL methods that treat the planner’s action distribution as a prior in policy optimization. We show how existing methods emerge as special cases of this family and explore previously unstudied variants. Experiments demonstrate that these variants yield significant performance gains, advancing the state of the art in MPPI-based RL.

This work presents WorldCompass, a novel Reinforcement Learning (RL) post-training framework for the long-horizon, interactive video-based world models, enabling them to explore the world more accurately and consistently based on interaction signals. To effectively "steer" the world model's exploration, we introduce three core innovations tailored to the autoregressive video generation paradigm: 1) Clip-level Roll-out Strategy: We generate and evaluate multiple samples at a single target clip, which significantly boosts roll-out efficiency and provides fine-grained reward signals. 2) Complementary Reward Functions: We design reward function for both interaction-following accuracy and visual quality, which provide direct supervision and effectively suppress reward-hacking behaviors. 3) Efficient RL Algorithm: We employ the negative-aware fine-tuning strategy coupled with various efficiency optimizations to efficiently and effectively enhance model capacity. Evaluations on the SoTA open-source world model, WorldPlay, demonstrate that WorldCompass significantly improves interaction accuracy and visual fidelity across various scenarios.

Group-based reinforcement learning (RL) methods have achieved remarkable success in improving the performance of large language models (LLMs) and have been rapidly extended to agentic tasks. However, their credit assignment relies heavily on coarse-grained trajectory-level attribution according to final outcomes, making it difficult to capture the contribution of individual steps, such as valuable steps obscured within failed trajectories. To uncover latent information and enable more faithful step-level credit assignment, we propose Graph-based Group Policy Optimization (GraphGPO), which first aggregates all rollout trajectories into a unified state-transition graph and then estimates the distance from each state to the task goal using the global information encoded in the graph. Finally, GraphGPO assigns credit to each edge by estimating a graph-based advantage, based on how much the transition reduces the distance to the task goal. In this way, GraphGPO significantly improves training efficiency and achieves state-of-the-art performance across a range of challenging benchmarks.

RL-based post-training with GRPO is widely used to improve large language models on individual reasoning tasks. However, real-world deployment requires reliable performance across diverse tasks. A straightforward multi-task adaptation of GRPO often leads to imbalanced outcomes, with some tasks dominating optimization while others stagnate. Moreover, tasks can vary widely in how frequently prompts yield zero advantages (and thus zero gradients), which further distorts their effective contribution to the optimization signal. To address these issues, we propose a novel Multi-Task GRPO (MT-GRPO) algorithm that (i) dynamically adapts task weights to explicitly optimize worst-task performance and promote balanced progress across tasks, and (ii) introduces a ratio-preserving sampler to ensure task-wise policy gradients reflect the adapted weights. Experiments on both 3-task and 9-task settings show that MT-GRPO consistently outperforms baselines in worst-task accuracy. In particular, MT-GRPO achieves 16–28\% and 6\% absolute improvement on worst-task performance over standard GRPO and DAPO, respectively, while maintaining competitive average accuracy. Moreover, MT-GRPO requires 50\% fewer training steps to reach 50\% worst-task accuracy in the 3-task setting, demonstrating substantially improved efficiency in achieving reliable performance across tasks.

Group Relative Policy Optimization (GRPO) was introduced recently and used to train DeepSeek\textendash R1 for promoting reasoning in LLMs under verifiable (binary) rewards. We show that the mean{+}variance calibration of these rewards induces a contrastive loss in which the contrastive samples are synthetic data drawn from the previous policy. While GRPO was originally paired with clipping to keep updates near the old policy, we analyze variants that differ in reward normalization (mean-only vs.\ mean{+}variance) and in how they regularize updates using KL divergence: either penalizing divergence from the previous model (\emph{mirror}), penalizing divergence from a fixed reference model $\pi_{\mathrm{ref}}$, or combining both forms of regularization. For each, the optimal policy $\pi_n$ admits an explicit form in terms of the binary reward and the first and second order statistics of the reward under $\pi_{n-1}$, as well as the policies $\pi_{n-1}$ and $\pi_{\mathrm{ref}}$. Iterating results in a sequence $\{\pi_n\}$ whose \emph{probability of success (PoS)} obeys a simple recurrence that converges to a fixed point determined by the reference PoS and the regularization strength. We further show that this fixed point exceeds the reference, demonstrating that GRPO amplifies the policy's probability of success.

Reinforcement learning (RL), particularly RL from verifiable reward (RLVR), has become a crucial phase of training large language models (LLMs) and a key focus of current scaling efforts. However, optimization practices in RL largely follow those of next-token-prediction stages (e.g., pretraining and supervised fine-tuning), despite the fundamental differences between RL and these stages emphasized by recent work. One such practice is the use of the AdamW optimizer, which is widely adopted for training large-scale transformers despite its high memory overhead. Our analysis shows that both momentum and adaptive learning rate of AdamW are less influential in RL than in SFT, leading us to hypothesize that RL benefits less from Adam’s per-parameter adaptive learning rates and momentum. Confirming our hypothesis, our experiments demonstrate that the substantially more memory-efficient SGD, which is known to perform poorly in supervised learning of large-scale transformers, matches or even outperforms AdamW in RL for LLMs. Remarkably, full fine-tuning with SGD updates fewer than 0.02% of model without any sparsity-promoting regularization, more than 1,000 times fewer than AdamW. Our analysis offers potential reasons for this update sparsity. Our findings provide fresh insights into the optimization dynamics of RL in LLMs and demonstrate that RL can be substantially more parameter-efficient than previously recognized.

Reinforcement learning (RL) has powered many recent breakthroughs in large language models (LLMs), especially for tasks where rewards can be computed automatically, such as code generation. However, it is less effective in open-ended medical dialogue, where feedback is ambiguous, context-dependent, and difficult to simply summarize into a single scalar signal—often requiring heavily supervised reward models and creating risks of reward hacking. Thus, we introduce ORBIT, an open-ended rubric-based incremental training framework tailored for critical medical dialogues. ORBIT integrates medical dialogue construction with dynamically generated case-conditioned rubrics that serve as adaptive guides for incremental RL. Unlike approaches that rely on external medical knowledge bases or handcrafted rules, ORBIT uses rubric-guided evaluation and can be implemented with general-purpose instruction-following LLMs, avoiding task-specific judge fine-tuning. With only 2k training samples, ORBIT raises Qwen3-4B-Instruct's HealthBench-Hard score from 7.0 to 27.5, achieving state-of-the-art performance among similarly sized open-source models while maintaining strong consultation quality as rubric coverage broadens. Project page: https://pidneuralode.github.io/ORBIT.

Current research on video hallucination mitigation primarily focuses on isolated error types, leaving *compositional* hallucinations—arising from incorrect reasoning over multiple interacting spatial and temporal factors largely underexplored. We introduce **OmniVCHall**, a benchmark designed to systematically evaluate both isolated and compositional hallucinations in Video Large Language Models (VLLMs). OmniVCHall spans diverse video domains, introduces a novel camera-based hallucination type, and defines a fine-grained taxonomy, together with adversarial answer options (*e.g.*, “All are correct” and “None of the above”) to prevent shortcut reasoning. The evaluations of 39 representative VLLMs reveal that even advanced models (*e.g.*, Qwen3-VL and GPT-5) exhibit substantial performance degradation. We propose **TriCD**, a contrastive decoding framework with a triple-pathway calibration mechanism. An adaptive perturbation controller dynamically selects distracting operations to construct negative video variants, while a saliency-guided enhancement module adaptively reinforces grounded token-wise visual evidences. These components are optimized via reinforcement learning to encourage precise decision-making under compositional hallucination settings. Experimental results show that **TriCD** consistently improves performance across two representative backbones, achieving an average accuracy improvement of over 10\%. The data and code are released at https://github.com/BMRETURN/OmniVCHall.

Time series reasoning tasks often start with a natural language question and require targeted analysis of a time series. Evidence may span the full series or appear in a few short intervals, so the model must decide what to inspect. Most existing approaches encode the entire time series into a fixed representation before inference, regardless of relevance. We introduce ARTIST, which formulates time-series reasoning as a sequential decision problem. ARTIST interleaves reasoning with adaptive temporal segment selection. It adopts a controller-reasoner architecture and uses reinforcement learning to train the controller role to select informative segments and the reasoner role to generate segment-conditioned reasoning traces and final answers. During inference, the model actively acquires task-relevant information instead of relying on a static summary of the full sequence. We use a novel hierarchical policy optimization approach for post-training that allows the model to excel in both segment selection and question-answering behavior. We evaluate ARTIST on six time-series reasoning benchmarks and compare it with large language models, vision-language models, and prior time-series reasoning systems. ARTIST improves average accuracy by 6.46 absolute percentage points over the strongest baseline. The largest gains appear on rare event localization and multi-segment reasoning tasks. Supervised fine-tuning improves performance, and reinforcement learning provides additional gains by optimizing question-adaptive segment selection, showing that selective data use drives effective time-series reasoning.

The family of linear recurrent neural networks has shown strong performance as recurrent memory units in partially observable reinforcement learning. We provide a theoretical justification for their empirical effectiveness by constructing and studying two linear filters: (i) the first exactly reproduces the pre–softmax logits of the belief vector in a hidden Markov model (HMM) under a deterministic transition matrix, thereby serving as a sufficient statistic for optimal policy learning, (ii) the second achieves vanishing state-decoding error under a nearly deterministic transition matrix, thus reducing state ambiguity to near zero. The results extend to action-controlled HMMs, where the corresponding linear filters become time-varying with action-dependent dynamics. We illustrate our main results through numerical experiments and further show that the constructed linear filter serves as a strong feature extractor in a small reinforcement learning game.

Video anomaly understanding (VAU) relies on sparse, context-dependent cues. However, existing passive paradigms suffer from observational aliasing, where static sampling fails to disambiguate semantically distinct events. To overcome this, we propose $Anom\text{-}\pi$, a closed-loop framework that reconceptualizes video understanding as an active sequential decision-making process within a dynamic environment. Inspired by human video-reviewing behavior, this framework unifies internal cognitive reasoning and strategic evidence acquisition into an interleaved policy, utilizing temporal atomic operators such as local backtracking, temporal expansion, and fine-grained sampling to endow the model with perceptual proactivity. To learn such complex interaction strategies under video-level weak supervision, we design Interactive Direct Preference Optimization (iDPO) to achieve trajectory-level policy alignment, guided by an Active Evidence Inquiry (AEI) utility that balances task success, informative evidence acquisition, and interaction cost. This approach enables the agent to learn to actively disambiguate hypotheses while suppressing redundant exploration. Extensive experiments demonstrate that our framework, with only 2B parameters, achieves highly competitive performance, significantly outperforming state-of-the-art large-scale VAU models in complex scenarios.

In offline reinforcement learning (RL), single-step temporal-difference (TD) learning can suffer from bootstrapping error accumulation over long horizons. Action-chunked TD methods mitigate this by backing up over multiple steps, but can introduce suboptimality by restricting the policy class to open-loop action sequences. To resolve this trade-off, we present Chunk-Guided Q-Learning (CGQ), a single-step TD algorithm that guides a fine-grained single-step critic by regularizing it toward a chunk-based critic trained using temporally extended backups. This reduces compounding error while preserving fine-grained value propagation. We theoretically show that CGQ attains tighter critic optimality bounds than either single-step or action-chunked TD learning alone. Empirically, CGQ achieves strong performance on challenging long-horizon OGBench tasks, often outperforming both single-step and action-chunked methods.

Portrait pose transfer (PPT) requires generative models to preserve fine-grained identity details while following complex pose and layout modification instructions. Existing methods often struggle with extensive data annotation requirements or employ optimization objectives that are suboptimal for addressing PPT's two key challenges. In this work, we propose PortraitRL, a novel post-training framework that addresses these challenges with a multi-objective reward mechanism. Specifically, we employ LVLM-based reward functions to effectively evaluate PPT's two challenges and apply within-group standardization to eliminate scale differences, allowing these rewards to effectively guide optimization. More importantly, we devise a novel reinforcement learning algorithm, Negative-aware Score Preference Optimization (NaSPO), which automatically identifies positive and negative preference samples through within-group advantages, eliminating annotation requirements while fully leveraging both positive and negative learning signals. Extensive experiments show state-of-the-art performance, with significant improvements in both detail preservation and editing accuracy.

Distributional reinforcement learning (DRL) models the full return distribution rather than expectations, but extending it to multivariate settings remains challenging. Many common metrics do not naturally generalize beyond one dimension or lose computational tractability, and the multivariate case introduces additional difficulties such as general matrix discounting, for which no contraction results are available. We introduce Sliced Distributional Reinforcement Learning (SDRL), which lifts tractable one-dimensional divergences to multivariate return distributions via projections. We prove Bellman contraction for uniform slicing under shared scalar discounting, and introduce a maximum-slicing variant with contraction under general dense discount matrices. SDRL supports a broad class of base divergences; we analyze Wasserstein, Cramér, and Maximum Mean Discrepancy (MMD), and characterize which SDRL variants suit the standard single-sample Bellman update used in distributional RL. We evaluate SDRL on a toy chain problem and a gridworld image-based environment as well as a subset of Atari games. Code is available at https://github.com/BaptisteDebes/SlicedDistributionalRL

Model routing chooses which language model to use for each query. By sending easy queries to cheaper models and hard queries to stronger ones, it can significantly reduce inference cost while maintaining high accuracy. However, most existing routers treat this as a fixed choice among a small set of models, which makes them hard to adapt to new models or changing budget constraints. In this paper, we propose SCOPE (Scalable and Controllable Outcome Performance Estimator), a routing framework that goes beyond model selection by predicting their cost and performance. Trained with reinforcement learning, SCOPE makes reasoning-based predictions by retrieving how models behave on similar problems, rather than relying on fixed model names, enabling it to work with new, unseen models. Moreover, by explicitly predicting how accurate and how expensive a model will be, it turns routing into a dynamic decision problem, allowing users to easily control the trade-off between accuracy and cost. Experiments show that SCOPE is more than just a cost-saving tool. It flexibly adapts to user needs: it can boost accuracy by up to **25.7\%** when performance is the priority, or cut costs by up to **95.1\%** when efficiency matters most. We release the dataset and code at our project page: https://sullivan07043.github.io/SCOPE/.

In this paper, we study the role of the critic in actor-critic for entropy-regularized, finite, discounted environments. We establish that, when the critic is exact, using the latter as a baseline is an actual variance-reduction method. In this case, actor-critic with stochastic gradients matches the sample complexity of deterministic policy gradient, reaching an $\epsilon$-optimal regularized value with $\tilde{O}(\log(1/\epsilon))$ samples. In practice, the critic is learned alongside the actor: the variance of the actor update is then influenced by the critic's variance and bias. Specifically, when the critic has a sufficiently small error, the variance reduction and rapid convergence are preserved. This suggests to learn the critic first, keeping it up to date after each actor update, underscoring the pivotal role of accurate critic estimation in actor-critic methods.