ICML 2026 Papers

The advancement of LLMs and their accessibility have triggered renewed interest in multi-agent reinforcement learning as robust and adaptive frameworks for dynamically changing environments. This paper introduces RL-Focal, a two-stage RL agent framework that routes and ensembles LLMs. First, we develop the Decider RL-agent, which learns to dynamically select an ensemble of small size ($m_i$) among $N$ LLMs ($m_i \ll N$) for incoming queries from a user-defined downstream task $i$, by maximizing both error-diversity and reasoning-performance of the selected ensemble through iterative updates of task-adaptive rewards and policy. Second, to enable effective fusion of dynamically selected LLMs, we develop the stage-2 Fusion RL-agent, which learns to resolve reasoning conflicts from different LLMs and dynamically adapt to different ensemble teams composed by the Decider Agent for different downstream tasks. {\em Third}, we introduce the focal diversity metric to better model the error correlations among multiple LLMs further improving the generalization performance of the Decider Agent, which actively prunes the ensemble combinations. By focal diversity, we enhance performance across tasks by effectively promoting reward-aware and policy-adaptive ensemble selection and inference fusion. Extensive evaluations on five benchmarks show that RL-Focal achieves the performance improvement of 8.48\% with an ensemble of small size compared to the best individual LLM in a pool and offers stronger robustness. Code is available at \url{https://anonymous.4open.science/r/rl-focal-8DCF/}

Planning with a generative model aims to estimate the value of a state using as few simulator calls as possible. SmoothCruiser achieves problem-independent complexity $\widetilde O(\varepsilon^{-4})$ by exploiting the smoothness of the entropy-regularized Bellman backup, but its estimator is only first-order. We show that the sample-complexity exponent of SmoothCruiser-type planners is governed by the order $\beta$ of the local Taylor remainder, giving oracle complexity $\widetilde O(\varepsilon^{-(2+2/(\beta-1))})$: the first-order case $\beta=2$ recovers SmoothCruiser, while a second-order/cubic remainder $\beta=3$ yields $\widetilde O(\varepsilon^{-3})$. We reach this regime with an optimal-transport-smoothed Bellman backup over action distributions, which has a closed form, a policy gradient, and a Lipschitz Hessian, and whose quadratic correction admits an unbiased cross-product estimator. The resulting SecondOrderSmoothCruiser achieves $\widetilde O(\varepsilon^{-3})$ oracle complexity for fixed OT parameters, and we relate the OT, entropy-regularized, and unregularized objectives through explicit regularization-bias bounds.

Generative models have emerged as a powerful class of policies for offline reinforcement learning (RL) due to their ability to capture complex, multi-modal behaviors. However, existing methods face a stark trade-off: slow, iterative models like diffusion policies are computationally expensive, while fast, single-step models like consistency policies often suffer from degraded performance. In this paper, we demonstrate that it is possible to bridge this gap. The key to moving beyond the limitations of individual methods, we argue, lies in a unifying perspective that views modern generative models—including diffusion, flow matching, and consistency models—as specific instances of learning a continuous-time generative trajectory governed by an Ordinary Differential Equation (ODE). This principled foundation provides a clearer design space for generative policies in RL and allows us to propose *Generative Trajectory Policies* (GTPs), a new and more general policy paradigm that learns the entire solution map of the underlying ODE. To make this paradigm practical for offline RL, we further introduce two key theoretically principled adaptations. Empirical results demonstrate that GTP achieves state-of-the-art performance on D4RL benchmarks -- it significantly outperforms prior generative policies, achieving perfect scores on several notoriously hard AntMaze tasks.

Reinforcement Learning with Verifiable Rewards (RLVR) has become a cornerstone for unlocking complex reasoning in Large Language Models (LLMs). Yet, scaling up RL is bottlenecked by limited existing verifiable data, where improvements increasingly saturate over prolonged training. To overcome this, we propose **Golden Goose**, a simple trick to synthesize unlimited RLVR tasks from unverifiable internet text by constructing a multiple-choice question-answering version of the fill-in-the-middle task. Given a source text, we prompt an LLM to identify and mask key reasoning steps, then generate a set of diverse, plausible distractors. This enables us to leverage reasoning-rich unverifiable corpora typically excluded from prior RLVR data construction (e.g., science textbooks) to synthesize **GooseReason-0.7M**, a large-scale RLVR dataset with over 0.7 million tasks spanning mathematics, programming, and general scientific domains. Empirically, GooseReason effectively revives models saturated on existing RLVR data, yielding robust, sustained gains under continuous RL and achieving new state-of-the-art results for 1.5B and 4B-Instruct models across 15 diverse benchmarks. Finally, we deploy Golden Goose in a real-world setting, synthesizing RLVR tasks from raw FineWeb scrapes for the cybersecurity domain, where no prior RLVR data exists. Training Qwen3-4B-Instruct on the resulting data **GooseReason-Cyber** sets a new state-of-the-art in cybersecurity, surpassing a 7B domain-specialized model with extensive domain-specific pre-training and post-training. This highlights the potential of automatically scaling up RLVR data by exploiting abundant, reasoning-rich, unverifiable internet text.

Interactively learning from observation and language feedback is an increasingly studied area driven by the emergence of large language model (LLM) agents. While impressive empirical demonstrations have been shown, so far a principled framing of these decision problems remains lacking. In this paper, we formalize the Learning from Language Feedback (LLF) problem, assert sufficient assumptions to enable learning despite latent rewards, and introduce *transfer eluder dimension* as a measure to characterize the hardness of LLF problems. We formalize the intuition that information in the feedback governs the learning complexity of LLF problems. We demonstrate cases where learning from rich language feedback can be exponentially faster than learning from reward. We develop a no-regret algorithm, called `HELiX`, that provably solves LLF problems through sequential interactions, with performance guarantees that scale with the transfer eluder dimension of the problem. Across several empirical domains, we show that `HELiX` performs well even when repeatedly prompting LLMs does not work reliably. Our contributions mark an important step towards designing principled interactive learning algorithms from generic language feedback.

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a powerful paradigm for post-training large reasoning models (LRMs) using policy-gradient methods such as GRPO. To stabilize training, these methods typically center trajectory rewards by subtracting the empirical mean for each prompt. Statistically, this centering acts as a control variate (or baseline), reducing the variance of the policy-gradient estimator. Typically, the mean reward is estimated using per-prompt empirical averages for each prompt in a batch. Drawing inspiration from Stein's paradox, we propose using shrinkage estimators that combine per-prompt and cross-prompt means to improve the overall per-prompt mean estimation accuracy, particularly in the low-generation regime typical of RLVR. Theoretically, we construct a shrinkage-based baseline that provably yields lower-variance policy-gradient estimators across algorithms. This baseline serves as a drop-in replacement for existing per-prompt mean baselines, requiring no additional hyperparameters or computation. Empirically, shrinkage baselines consistently outperform standard empirical-mean baselines, leading to lower-variance gradient updates and improved training stability.

Red-teaming Vision-Language Models is essential for identifying vulnerabilities where adversarial image-text inputs trigger toxic outputs. Existing approaches treat image generation as a black box, providing only terminal toxicity scores while remaining temporally opaque regarding when and how toxic semantics emerge during multi-step synthesis. We introduce $\textbf{STARE}$, a hierarchical reinforcement learning framework that treats the denoising trajectory as an exploitable attack surface. By synergizing a high-level prompt editor with low-level T2I fine-tuning via Group Relative Policy Optimization (GRPO), STARE achieves a 68\% improvement in Attack Success Rate over state-of-the-art baselines including black box and white-box variants. More importantly, we reveal the Optimization-Induced Phase Alignment phenomenon: while vanilla models exhibit diffuse toxicity, adversarial optimization systematically concentrates conceptual harms into early semantic phases and detail-oriented harms into late refinement. This discovery transforms toxicity formation from a chaotic process into a series of predictable vulnerability windows. This temporal alignment transforms red-teaming from a trial-and-error process into a targeted structural analysis. Our work provides both a potent attack engine and a diagnostic foundation for developing next-generation, phase-aware safety mechanisms. Content warning: This paper contains examples of toxic content that may be offensive or disturbing.

Although Reinforcement Learning Fine-Tuning (RLFT) applied to Vision-Language Models (VLMs) substantially enhances multimodal reasoning capabilities, their prohibitive training cost limits broad adoption. Surprisingly, most existing methods simply port Large Language Model (LLM) RLFT techniques to VLMs, while ignoring a intrinsic property of multimodal models: their dynamic text–vision alignment. We ask a new question: Can this intrinsic alignment be turned into a training signal that makes VLM RLFT more efficient? We analyze how a VLM plans to attend, actually attends, and ideally should attend during reasoning, and derive two lightweight metrics from these patterns. Predictive View Accuracy (PVA) estimates sample difficulty, and Reasoning View Accuracy (RVA) reflects the quality of chain-of-thought (CoT) reasoning. These alignment signals enable automated data curriculum and dense reasoning supervision. We introduce FOCUS-RL, a plug-and-play framework that can be seamlessly integrated into any VLM and dramatically boosts RLFT training efficiency. FOCUS-RL achieves 2.5 x – 4 x faster convergence over vanilla GRPO and consistent accuracy gains (+4.4 on average) across six different benchmarks and multiple VLM families.

Reinforcement learning with verifiable rewards (RLVR) on foundation models has led to significant gains in math and code generation. Extending these gains to open-ended domains remains challenging: ground-truth verification is unavailable, human annotation is expensive, and learned reward models are prone to reward hacking. We introduce Rubric Curriculum RL (RcRL), a general self-improvement method for non-verifiable tasks that requires no new data, human annotations, or stronger teacher models. RcRL exploits the generation-verification gap, where judging good outputs is easier than producing them. We combine pairwise preferences over rollouts with a curriculum over rubric criteria, yielding a more discriminative signal than absolute scoring while reducing reward hacking through a non-stationary objective. Whereas baselines plateau or collapse within a few hundred steps, RcRL preserves output entropy and keeps improving past 1000 steps. On creative writing, RcRL outputs are preferred 70.5% of the time compared to the base model, with consistent gains across multiple creative writing benchmarks and judges. On HealthBench, RcRL improves over the base model by 14.6% (full set) and 25.4% (consensus subset), outperforming all HealthBench training baselines, including instance-specific rubrics.

Unsupervised skill discovery (USD) aims to learn diverse behaviors without reward functions, but often results in task-irrelevant or hazardous behaviors due to uniform exploration. Guided skill discovery (GSD) addresses this issue by incorporating human intent to focus exploration on meaningful regions. However, existing GSD methods typically require training additional guidance models, and rely on pre-defined rules or expert demonstration, which can be ineffective under sparse, online-collected human feedback. To overcome this, we propose COLLIE, a GSD framework that leverages dense unsupervised data to construct a semantically coherent skill latent space. This latent space is well-structured, enabling reliable guidance with sparse online feedback. Moreover, its semantic coherence property enables training-free construction of guidance signals, eliminating the need for additional model training beyond skill learning. Theoretical analysis justifies the effectiveness of our training-free guidance signal, while experiments across diverse state-based and pixel-based tasks show that COLLIE learns diverse, human-aligned skills, avoids hazardous behaviors, and achieves superior downstream performance with minimal human feedback.

World models have significantly improved the sample efficiency of model-based reinforcement learning (MBRL) by enabling policy learning in imagination, thereby reducing the need for direct interaction with the real environment. However, most existing world model methods are trained independently for each task or perform multi-task learning using offline datasets, failing to fully exploit the latent relationships among tasks in online interactive scenarios. To address this limitation, we propose TaskLoom, a knowledge-sharing world model architecture for online RL. TaskLoom adopts a grouped two-stage training paradigm: first, the tasks are divided into several groups based on the similarity of world model gradients, and fine-grained knowledge is shared among tasks within each group; second, coarse-grained knowledge is exchanged across groups, enabling hierarchical knowledge transfer and reuse. Experimental results show that TaskLoom outperforms baseline methods on widely used benchmarks such as Proprio Control, Visual Control and Meta-World, validating the effectiveness of the proposed knowledge-sharing mechanism for both low-dimensional state and high-dimensional visual inputs.

We study the problem of learning the optimal policy in a discounted, infinite-horizon reinforcement learning (RL) setting in the presence of adversarially corrupted rewards. To address this problem, we develop a novel robust variant of the Q-learning algorithm and analyze it under the challenging asynchronous sampling model with time-correlated data. Despite corruption, we prove that the finite-time guarantees of our approach match existing bounds, up to an additive term that scales with the fraction of corrupted samples. We also establish an information-theoretic lower bound, revealing that our guarantees are near-optimal. Notably, our algorithm is agnostic to the underlying reward distribution and provides the first finite-time robustness guarantees for asynchronous Q-learning. A key element of our analysis is a refined Azuma-Hoeffding inequality for almost-martingales, which may have broader applicability in the study of RL algorithms.

Theory of Mind (ToM) is a must-acquire skill for modern foundation model systems to operate effectively and safely in the real world. Recent works have explored honing ToM via post-training; however, we show that such progress is confounded by a pervasive “shortcut” issue: tasks can reach up to 99% accuracy by simply exploiting spurious causal correlations, leading to a false sense of ToM. Motivated by this, we first develop a framework to systematically examine ToM datasets for shortcuts and provide guidance for future development. We find that questions reducible to pure state tracking, such as “belief,” are especially shortcut-prone compared to mind questions, such as “intention,” where reasoning beyond tracking is required. Using four shortcut-free datasets across three ToM contexts, we then comprehensively study whether Reinforcement Fine-Tuning with verifiable rewards and explicit reasoning chains, called Thinking-RFT, elevates ToM beyond Supervised Fine-Tuning, or SFT. Our key findings are as follows. First, Thinking-RFT effectively improves ToM in all scenarios, with a 6% improvement over SFT, particularly in complex higher-order reasoning, with a 10% improvement over SFT, and multimodal cases, with a 7% improvement over SFT. It also generalizes notably better to unseen domains and higher-order queries while being more robust to counterfactuals. Second, ToM benefits specifically from the joint effect of reasoning and RL: Thinking-RFT outperforms No-Thinking-RFT by 7% on average. Third, RFT works by learning to ground its reasoning on anchor cues, such as keywords and state changes, that correspond to causal factors. We believe our study is useful for developing effective and robust ToM post-training datasets and advancing critical ToM capabilities. Project code is available at: https://github.com/jkz-338/Robust-ToM-RL.git

Reinforcement learning (RL) has become essential to the post-training of large language models (LLMs) for reasoning, agentic capabilities and alignment. Successful RL relies on sufficient exploration of diverse actions by the model during training, which creates a potential failure mode: a model could strategically alter its exploration during training to influence the subsequent training outcome. In this paper we study this behavior, called exploration hacking. First, we create model organisms of selective RL resistance by fine-tuning LLMs to follow specific underperformance strategies; these models can successfully resist our RL-based capability elicitation in agentic biosecurity and AI R&D environments while maintaining performance on related tasks. We then use our model organisms to evaluate detection and mitigation strategies, including monitoring, weight noising, and SFT-based elicitation. Finally, we show that current frontier models can exhibit explicit reasoning about suppressing their exploration when provided with sufficient information about their training context, with higher rates when this information is acquired indirectly through the environment. Together, our results suggest exploration hacking is a possible failure mode of RL on sufficiently capable LLMs.

Reinforcement learning with verifiable rewards (RLVR) has been a main driver of recent breakthroughs in large reasoning models. Yet it remains a mystery how rewards based solely on final outcomes can help overcome the long-horizon barrier to extended reasoning. To understand this, we develop a theory of the training dynamics of RLVR for transformers on compositional reasoning tasks. Our theory shows that mixed-difficulty training naturally induces an **implicit curriculum**: without any explicit schedule, easier problems become learnable first and shape the frontier for harder ones, creating a learning progression from easy to hard during optimization. The effectiveness of this curriculum is governed by the smoothness of the difficulty spectrum. When the spectrum is smooth, training dynamics enter a well-behaved *relay* regime, in which persistent gradient signals on easier problems make slightly harder ones tractable and keep training at the edge of competence. When the spectrum contains abrupt discontinuities, training undergoes grokking-type phase transitions with prolonged plateaus before progress recurs. As a technical contribution, our analysis develops and adapts techniques from Fourier analysis on finite groups to our setting. We validate the predicted mechanisms empirically via synthetic experiments.

Group Relative Policy Optimization (GRPO) trains Chain-of-Thought reasoning with verifiable rewards, but estimating thought-level advantages without value functions often suffers from high variance. Although tree-style branching is used in practice to reduce variance, it lacks a theoretical explanation of why it works and whether it is important or potentially necessary. We study thought-level advantage estimation in GRPO from a variance perspective under a minimal tree-style setting where multiple continuations are sampled for each thought. Using the multivariate delta method, we reveal a sampling-dimension asymmetry. Increasing sampled thoughts ($K$) leaves a strictly positive estimation-variance floor, whereas increasing continuations per thought ($M$) drives the leading-order estimation variance to zero at rate $1/M$. This implies that, within the fixed-temperature GRPO-style estimator without value models studied here, accurate thought-level advantage estimation cannot be achieved by scaling thought sampling alone, making continuation-level branching a principled and potentially necessary mechanism rather than a heuristic. Experiments further provide empirical evidence for its effectiveness and potential necessity, demonstrating improved optimization stability, training efficiency, and final performance not only in math but also across vision domains and under different model architectures and sizes.

Group Relative Policy Optimization (GRPO) is effective for training language models on complex reasoning. However, since the objective is defined relative to a group of sampled trajectories, extended deliberation can create more chances to realize relative gains, leading to inefficient reasoning and overthinking, and complicating the trade-off between correctness and rollout efficiency. Controlling this behavior is difficult in practice, considering (i) Length penalties are hard to calibrate because longer rollouts may reflect harder problems that require longer reasoning, penalizing tokens risks truncating useful reasoning along with redundant continuation; and (ii) supervision that directly indicates when to continue or stop is typically unavailable beyond final answer correctness. We propose Weakly Supervised GRPO (WS-GRPO), which improves rollout efficiency by converting terminal rewards into correctness-aware guidance over partial trajectories. Unlike global length penalties that are hard to calibrate, WS-GRPO trains a preference model from outcome-only correctness to produce prefix-level signals that indicate when additional continuation is beneficial. Thus, WS-GRPO supplies outcome-derived continue/stop guidance, reducing redundant deliberation while maintaining accuracy. We provide theoretical results and empirically show on reasoning benchmarks that WS-GRPO substantially reduces rollout length while remaining competitive with GRPO baselines.

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a key ingredient for unlocking complex reasoning capabilities in large language models. Recent work ProRL has shown promise in scaling RL by increasing the number of training steps. However, performance plateaus after thousands of steps, with clear diminishing returns from allocating more computation to additional training. In this work, we investigate a complementary paradigm for scaling RL: \textbf{BroRL}—increasing the number of rollouts per example to hundreds to exhaustively \textbf{Bro}aden exploration, which yields continuous performance gains beyond the saturation point observed in ProRL when scaling the number of training steps. Our approach is motivated by a mass balance equation analysis allowing us to characterize the rate of change in probability mass for correct and incorrect tokens during the reinforcement process. We show that under a one-step RL assumption, sampled rollout tokens contribute to correct-mass expansion, while unsampled tokens outside rollouts may lead to gains or losses depending on their distribution and the net reward balance. Importantly, as the number of rollouts per example $N$ increases, the effect of unsampled terms diminishes, making overall correct-mass expansion more likely. To validate our theoretical analysis, we conduct simulations under more relaxed conditions and find that a sufficiently large rollout size $N$—corresponding to ample exploration—can increase the probability mass of correct tokens broadly, and in our simulator it increases all correct-token probabilities and eliminates knowledge shrinkage. Empirically, BroRL revives models saturated after 3K ProRL training steps and demonstrates robust, continuous improvement, achieving strong results for the 1.5B model across diverse benchmarks. Notably, under the same training time, BroRL is both more data- and compute-efficient: large-$N$ rollouts reduce the number of filtered samples during dynamic sampling at the algorithmic level and nearly double generation throughput compared to ProRL in our hardware setup; this throughput increase is consistent with shifting generation from a more memory-bound regime toward a more compute-bound one.

Aligning large language models (LLMs) to diverse human preferences is fundamentally challenging since criteria can often conflict with each other. Inference-time alignment methods have recently gained popularity as they allow LLMs to be aligned to multiple criteria via different alignment algorithms at inference time. However, inference-time alignment is computationally expensive since it often requires multiple forward passes of the base model. In this work, we propose *inference-aware meta-alignment* (IAMA), a novel approach that enables LLMs to be aligned to multiple criteria with minimal computational overhead at inference time. IAMA trains a base model such that it can be effectively aligned to multiple task optima via different inference-time alignment algorithms. To solve the non-linear optimization problems involved in IAMA, we propose *non-linear GRPO*, which provably converges to the optimal solution in the space of probability measures.