Wenjie Shi

Tianhao Hu, Xiangcheng Liu, Youshao Xiao et al.

Reinforcement learning (RL) has become a critical paradigm for LLM post-training, yet the rollout phase -- accounting for 50--80% of total step time -- is bottlenecked by skewed generation: long-tailed trajectories indispensable for model performance block the entire training pipeline. Asynchronous training offers a natural remedy by overlapping generation with training, but introduces a fundamental tension between efficiency and algorithmic correctness. We identify three constraints in asynchronous training to preserve convergence: intra-trajectory policy consistency, data integrity, and bounded staleness. Existing approaches fail to intrinsically address the long-tailed trajectory problem, which is further exacerbated by the imbalance characteristic of Mix-of-Experts models, or deviate from the standard RL training formulation, thereby hindering model convergence. Therefore, we propose DORA (Dynamic ORchestration for Asynchronous Rollout), which addresses this challenge through algorithm-system co-design. DORA introduces multi-version streaming rollout, a novel asynchronous paradigm that maintains multiple policy versions concurrently -- simultaneously achieving full bubble elimination without compromising algorithmic constraints. Experimental results demonstrate that our DORA system achieves substantial improvements in throughput -- up to 2--3 times higher than state-of-the-art systems on open-source benchmarks -- without compromising convergence. Furthermore, in large-scale industrial applications with tens of thousands of accelerators, DORA accelerates RL training by 2--4 times compared to synchronous training across various scenarios. The resultant open-source models, LongCat-Flash-Thinking, exhibit competitive performance on complex reasoning benchmarks, matching the capability of most advanced LLMs.

Jianing Wang, Linsen Guo, Zhengyu Chen et al.

Recent advances in agentic harness with orchestration frameworks that coordinate multiple agents with memory, skills, and tool use have achieved remarkable success in complex reasoning tasks. However, the underlying mechanism that truly drives performance remains obscured behind intricate system designs. In this paper, we propose HeavySkill, a perspective that views heavy thinking not only as a minimal execution unit in orchestration harness but also as an inner skill internalized within the model's parameters that drives the orchestrator to solve complex tasks. We identify this skill as a two-stage pipeline, i.e., parallel reasoning then summarization, which can operate beneath any agentic harness. We present a systematic empirical study of HeavySkill across diverse domains. Our results show that this inner skill consistently outperforms traditional Best-of-N (BoN) strategies; notably, stronger LLMs can even approach Pass@N performance. Crucially, we demonstrate that the depth and width of heavy thinking, as a learnable skill, can be further scaled via reinforcement learning, offering a promising path toward self-evolving LLMs that internalize complex reasoning without relying on brittle orchestration layers.

Meituan LongCat Team, Anchun Gui, Bei Li et al.

We present LongCat-Flash-Thinking, an efficient 560-billion-parameter open-source Mixture-of-Experts (MoE) reasoning model. Its advanced capabilities are cultivated through a meticulously crafted training process, beginning with long Chain-of-Thought (CoT) data cold-start and culminating in large-scale Reinforcement Learning (RL). We first employ a well-designed cold-start training strategy, which significantly enhances the reasoning potential and equips the model with specialized skills in both formal and agentic reasoning. Then, a core innovation is our domain-parallel training scheme, which decouples optimization across distinct domains (e.g., STEM, Code, Agentic) and subsequently fuses the resulting expert models into a single, nearly Pareto-optimal model. This entire process is powered by our Dynamic ORchestration for Asynchronous rollout (DORA) system, a large-scale RL framework that delivers a greater than threefold training speedup over synchronous methods on tens of thousands of accelerators. As a result, LongCat-Flash-Thinking achieves state-of-the-art performance among open-source models on a suite of complex reasoning tasks. The model exhibits exceptional efficiency in agentic reasoning, reducing average token consumption by 64.5% (from 19, 653 to 6, 965) on AIME-25, without degrading task accuracy. We release LongCat-Flash-Thinking to promote further advances in reasoning systems and agentic AI research.

Qisen Yang, Huanqian Wang, Mukun Tong et al.

The black-box nature of deep reinforcement learning (RL) hinders them from real-world applications. Therefore, interpreting and explaining RL agents have been active research topics in recent years. Existing methods for post-hoc explanations usually adopt the action matching principle to enable an easy understanding of vision-based RL agents. In this paper, it is argued that the commonly used action matching principle is more like an explanation of deep neural networks (DNNs) than the interpretation of RL agents. It may lead to irrelevant or misplaced feature attribution when different DNNs' outputs lead to the same rewards or different rewards result from the same outputs. Therefore, we propose to consider rewards, the essential objective of RL agents, as the essential objective of interpreting RL agents as well. To ensure reward consistency during interpretable feature discovery, a novel framework (RL interpreting RL, denoted as RL-in-RL) is proposed to solve the gradient disconnection from actions to rewards. We verify and evaluate our method on the Atari 2600 games as well as Duckietown, a challenging self-driving car simulator environment. The results show that our method manages to keep reward (or return) consistency and achieves high-quality feature attribution. Further, a series of analytical experiments validate our assumption of the action matching principle's limitations.

Wenjie Shi, Shiji Song, Hui Wu et al.

Model-free deep reinforcement learning (RL) algorithms have been widely used for a range of complex control tasks. However, slow convergence and sample inefficiency remain challenging problems in RL, especially when handling continuous and high-dimensional state spaces. To tackle this problem, we propose a general acceleration method for model-free, off-policy deep RL algorithms by drawing the idea underlying regularized Anderson acceleration (RAA), which is an effective approach to accelerating the solving of fixed point problems with perturbations. Specifically, we first explain how policy iteration can be applied directly with Anderson acceleration. Then we extend RAA to the case of deep RL by introducing a regularization term to control the impact of perturbation induced by function approximation errors. We further propose two strategies, i.e., progressive update and adaptive restart, to enhance the performance. The effectiveness of our method is evaluated on a variety of benchmark tasks, including Atari 2600 and MuJoCo. Experimental results show that our approach substantially improves both the learning speed and final performance of state-of-the-art deep RL algorithms.

Wenjie Shi, Shiji Song, Cheng Wu et al.

This paper investigates trajectory tracking problem for a class of underactuated autonomous underwater vehicles (AUVs) with unknown dynamics and constrained inputs. Different from existing policy gradient methods which employ single actor-critic but cannot realize satisfactory tracking control accuracy and stable learning, our proposed algorithm can achieve high-level tracking control accuracy of AUVs and stable learning by applying a hybrid actors-critics architecture, where multiple actors and critics are trained to learn a deterministic policy and action-value function, respectively. Specifically, for the critics, the expected absolute Bellman error based updating rule is used to choose the worst critic to be updated in each time step. Subsequently, to calculate the loss function with more accurate target value for the chosen critic, Pseudo Q-learning, which uses sub-greedy policy to replace the greedy policy in Q-learning, is developed for continuous action spaces, and Multi Pseudo Q-learning (MPQ) is proposed to reduce the overestimation of action-value function and to stabilize the learning. As for the actors, deterministic policy gradient is applied to update the weights, and the final learned policy is defined as the average of all actors to avoid large but bad updates. Moreover, the stability analysis of the learning is given qualitatively. The effectiveness and generality of the proposed MPQ-based Deterministic Policy Gradient (MPQ-DPG) algorithm are verified by the application on AUV with two different reference trajectories. And the results demonstrate high-level tracking control accuracy and stable learning of MPQ-DPG. Besides, the results also validate that increasing the number of the actors and critics will further improve the performance.

Wenjie Shi, Shiji Song, Cheng Wu

Maximum entropy deep reinforcement learning (RL) methods have been demonstrated on a range of challenging continuous tasks. However, existing methods either suffer from severe instability when training on large off-policy data or cannot scale to tasks with very high state and action dimensionality such as 3D humanoid locomotion. Besides, the optimality of desired Boltzmann policy set for non-optimal soft value function is not persuasive enough. In this paper, we first derive soft policy gradient based on entropy regularized expected reward objective for RL with continuous actions. Then, we present an off-policy actor-critic, model-free maximum entropy deep RL algorithm called deep soft policy gradient (DSPG) by combining soft policy gradient with soft Bellman equation. To ensure stable learning while eliminating the need of two separate critics for soft value functions, we leverage double sampling approach to making the soft Bellman equation tractable. The experimental results demonstrate that our method outperforms in performance over off-policy prior methods.