On Computation and Reinforcement Learning

How does the amount of compute available to a reinforcement learning (RL) policy affect its learning? Can policies using a fixed amount of parameters, still benefit from additional compute? The standard RL framework does not provide a language to answer these questions formally. Empirically, deep RL policies are often parameterized as neural networks with static architectures, conflating the amount of compute and the number of parameters. In this paper, we formalize compute bounded policies and prove that policies which use more compute can solve problems and generalize to longer-horizon tasks that are outside the scope of policies with less compute. Building on prior work in algorithmic learning and model-free planning, we propose a minimal architecture that can use a variable amount of compute. Our experiments complement our theory. On a set 31 different tasks spanning online and offline RL, we show that $(1)$ this architecture achieves stronger performance simply by using more compute, and $(2)$ stronger generalization on longer-horizon test tasks compared to standard feedforward networks or deep residual network using up to 5 times more parameters.

Approximation of Log-Partition Function in Policy Mirror Descent Induces Implicit Regularization for LLM Post-Training

Policy mirror descent (PMD) provides a principled framework for reinforcement learning (RL) by iteratively solving KL-regularized policy improvement subproblems. While this approach has been adopted in training advanced LLMs such as Kimi K1.5/K2, the ideal closed-form PMD updates require reliable partition function estimation, a significant challenge when working with limited rollouts in the vast action spaces of LLMs. We investigate a practical algorithm, termed PMD-mean, that approximates the log-partition term with the mean reward under the sampling policy and performs regression in log-policy space. Specifically, we characterize the population solution of PMD-mean and demonstrate that it implicitly optimizes mirror descent subproblems with an adaptive mixed KL--$χ^2$ regularizer. This additional $χ^2$ regularization constrains large probability changes, producing more conservative updates when expected rewards are low and enhancing robustness against finite-sample estimation errors. Experiments on math reasoning tasks show that PMD-mean achieves superior performance with improved stability and time efficiency. These findings deepen our understanding of PMD-mean and illuminate pathways toward principled improvements in RL algorithms for LLMs. Code is available at https://github.com/horizon-rl/OpenKimi.

Quantum Reinforcement Learning with Transformers for the Capacitated Vehicle Routing Problem

This paper addresses the Capacitated Vehicle Routing Problem (CVRP) by comparing classical and quantum Reinforcement Learning (RL) approaches. An Advantage Actor-Critic (A2C) agent is implemented in classical, full quantum, and hybrid variants, integrating transformer architectures to capture the relationships between vehicles, clients, and the depot through self- and cross-attention mechanisms. The experiments focus on multi-vehicle scenarios with capacity constraints, considering 20 clients and 4 vehicles, and are conducted over ten independent runs. Performance is assessed using routing distance, route compactness, and route overlap. The results show that all three approaches are capable of learning effective routing policies. However, quantum-enhanced models outperform the classical baseline and produce more robust route organization, with the hybrid architecture achieving the best overall performance across distance, compactness, and route overlap. In addition to quantitative improvements, qualitative visualizations reveal that quantum-based models generate more structured and coherent routing solutions. These findings highlight the potential of hybrid quantum-classical reinforcement learning models for addressing complex combinatorial optimization problems such as the CVRP.

DFPO: Scaling Value Modeling via Distributional Flow towards Robust and Generalizable LLM Post-Training

Training reinforcement learning (RL) systems in real-world environments remains challenging due to noisy supervision and poor out-of-domain (OOD) generalization, especially in LLM post-training. Recent distributional RL methods improve robustness by modeling values with multiple quantile points, but they still learn each quantile independently as a scalar. This results in rough-grained value representations that lack fine-grained conditioning on state information, struggling under complex and OOD conditions. We propose DFPO (Distributional Value Flow Policy Optimization with Conditional Risk and Consistency Control), a robust distributional RL framework that models values as continuous flows across time steps. By scaling value modeling through learning of a value flow field instead of isolated quantile predictions, DFPO captures richer state information for more accurate advantage estimation. To stabilize training under noisy feedback, DFPO further integrates conditional risk control and consistency constraints along value flow trajectories. Experiments on dialogue, math reasoning, and scientific tasks show that DFPO outperforms PPO, FlowRL, and other robust baselines under noisy supervision, achieving improved training stability and generalization.

Dr. Kernel: Reinforcement Learning Done Right for Triton Kernel Generations

High-quality kernel is critical for scalable AI systems, and enabling LLMs to generate such code would advance AI development. However, training LLMs for this task requires sufficient data, a robust environment, and the process is often vulnerable to reward hacking and lazy optimization. In these cases, models may hack training rewards and prioritize trivial correctness over meaningful speedup. In this paper, we systematically study reinforcement learning (RL) for kernel generation. We first design KernelGYM, a robust distributed GPU environment that supports reward hacking check, data collection from multi-turn interactions and long-term RL training. Building on KernelGYM, we investigate effective multi-turn RL methods and identify a biased policy gradient issue caused by self-inclusion in GRPO. To solve this, we propose Turn-level Reinforce-Leave-One-Out (TRLOO) to provide unbiased advantage estimation for multi-turn RL. To alleviate lazy optimization, we incorporate mismatch correction for training stability and introduce Profiling-based Rewards (PR) and Profiling-based Rejection Sampling (PRS) to overcome the issue. The trained model, Dr.Kernel-14B, reaches performance competitive with Claude-4.5-Sonnet in Kernelbench. Finally, we study sequential test-time scaling for Dr.Kernel-14B. On the KernelBench Level-2 subset, 31.6% of the generated kernels achieve at least a 1.2x speedup over the Torch reference, surpassing Claude-4.5-Sonnet (26.7%) and GPT-5 (28.6%). When selecting the best candidate across all turns, this 1.2x speedup rate further increases to 47.8%. All resources, including environment, training code, models, and dataset, are included in https://www.github.com/hkust-nlp/KernelGYM.

Reinforcement World Model Learning for LLM-based Agents

Large language models (LLMs) have achieved strong performance in language-centric tasks. However, in agentic settings, LLMs often struggle to anticipate action consequences and adapt to environment dynamics, highlighting the need for world-modeling capabilities in LLM-based agents. We propose Reinforcement World Model Learning (RWML), a self-supervised method that learns action-conditioned world models for LLM-based agents on textual states using sim-to-real gap rewards. Our method aligns simulated next states produced by the model with realized next states observed from the environment, encouraging consistency between internal world simulations and actual environment dynamics in a pre-trained embedding space. Unlike next-state token prediction, which prioritizes token-level fidelity (i.e., reproducing exact wording) over semantic equivalence and can lead to model collapse, our method provides a more robust training signal and is empirically less susceptible to reward hacking than LLM-as-a-judge. We evaluate our method on ALFWorld and $τ^2$ Bench and observe significant gains over the base model, despite being entirely self-supervised. When combined with task-success rewards, our method outperforms direct task-success reward RL by 6.9 and 5.7 points on ALFWorld and $τ^2$ Bench respectively, while matching the performance of expert-data training.

UI-Mem: Self-Evolving Experience Memory for Online Reinforcement Learning in Mobile GUI Agents

Online Reinforcement Learning (RL) offers a promising paradigm for enhancing GUI agents through direct environment interaction. However, its effectiveness is severely hindered by inefficient credit assignment in long-horizon tasks and repetitive errors across tasks due to the lack of experience transfer. To address these challenges, we propose UI-Mem, a novel framework that enhances GUI online RL with a Hierarchical Experience Memory. Unlike traditional replay buffers, our memory accumulates structured knowledge, including high-level workflows, subtask skills, and failure patterns. These experiences are stored as parameterized templates that enable cross-task and cross-application transfer. To effectively integrate memory guidance into online RL, we introduce Stratified Group Sampling, which injects varying levels of guidance across trajectories within each rollout group to maintain outcome diversity, driving the unguided policy toward internalizing guided behaviors. Furthermore, a Self-Evolving Loop continuously abstracts novel strategies and errors to keep the memory aligned with the agent's evolving policy. Experiments on online GUI benchmarks demonstrate that UI-Mem significantly outperforms traditional RL baselines and static reuse strategies, with strong generalization to unseen applications. Project page: https://ui-mem.github.io

TKG-Thinker: Towards Dynamic Reasoning over Temporal Knowledge Graphs via Agentic Reinforcement Learning

Temporal knowledge graph question answering (TKGQA) aims to answer time-sensitive questions by leveraging temporal knowledge bases. While Large Language Models (LLMs) demonstrate significant potential in TKGQA, current prompting strategies constrain their efficacy in two primary ways. First, they are prone to reasoning hallucinations under complex temporal constraints. Second, static prompting limits model autonomy and generalization, as it lack optimization through dynamic interaction with temporal knowledge graphs (TKGs) environments. To address these limitations, we propose \textbf{TKG-Thinker}, a novel agent equipped with autonomous planning and adaptive retrieval capabilities for reasoning over TKGs. Specifically, TKG-Thinker performs in-depth temporal reasoning through dynamic multi-turn interactions with TKGs via a dual-training strategy. We first apply Supervised Fine-Tuning (SFT) with chain-of thought data to instill core planning capabilities, followed by a Reinforcement Learning (RL) stage that leverages multi-dimensional rewards to refine reasoning policies under intricate temporal constraints. Experimental results on benchmark datasets with three open-source LLMs show that TKG-Thinker achieves state-of-the-art performance and exhibits strong generalization across complex TKGQA settings.

Cross-Domain Offline Policy Adaptation via Selective Transition Correction

It remains a critical challenge to adapt policies across domains with mismatched dynamics in reinforcement learning (RL). In this paper, we study cross-domain offline RL, where an offline dataset from another similar source domain can be accessed to enhance policy learning upon a target domain dataset. Directly merging the two datasets may lead to suboptimal performance due to potential dynamics mismatches. Existing approaches typically mitigate this issue through source domain transition filtering or reward modification, which, however, may lead to insufficient exploitation of the valuable source domain data. Instead, we propose to modify the source domain data into the target domain data. To that end, we leverage an inverse policy model and a reward model to correct the actions and rewards of source transitions, explicitly achieving alignment with the target dynamics. Since limited data may result in inaccurate model training, we further employ a forward dynamics model to retain corrected samples that better match the target dynamics than the original transitions. Consequently, we propose the Selective Transition Correction (STC) algorithm, which enables reliable usage of source domain data for policy adaptation. Experiments on various environments with dynamics shifts demonstrate that STC achieves superior performance against existing baselines.

RL-VLA$^3$: Reinforcement Learning VLA Accelerating via Full Asynchronism

In recent years, Vision-Language-Action (VLA) models have emerged as a crucial pathway towards general embodied intelligence, yet their training efficiency has become a key bottleneck. Although existing reinforcement learning (RL)-based training frameworks like RLinf can enhance model generalization, they still rely on synchronous execution, leading to severe resource underutilization and throughput limitations during environment interaction, policy generation (rollout), and model update phases (actor). To overcome this challenge, this paper, for the first time, proposes and implements a fully-asynchronous policy training framework encompassing the entire pipeline from environment interaction, rollout generation, to actor policy updates. Systematically drawing inspiration from asynchronous optimization ideas in large model RL, our framework designs a multi-level decoupled architecture. This includes asynchronous parallelization of environment interaction and trajectory collection, streaming execution for policy generation, and decoupled scheduling for training updates. We validated the effectiveness of our method across diverse VLA models and environments. On the LIBERO benchmark, the framework achieves throughput improvements of up to 59.25\% compared to existing synchronous strategies. When deeply optimizing separation strategies, throughput can be increased by as much as 126.67\%. We verified the effectiveness of each asynchronous component via ablation studies. Scaling law validation across 8 to 256 GPUs demonstrates our method's excellent scalability under most conditions.