Hongyu Zang

Riashat Islam, Hongyu Zang, Manan Tomar et al. · mila

Several self-supervised representation learning methods have been proposed for reinforcement learning (RL) with rich observations. For real-world applications of RL, recovering underlying latent states is crucial, particularly when sensory inputs contain irrelevant and exogenous information. In this work, we study how information bottlenecks can be used to construct latent states efficiently in the presence of task-irrelevant information. We propose architectures that utilize variational and discrete information bottlenecks, coined as RepDIB, to learn structured factorized representations. Exploiting the expressiveness bought by factorized representations, we introduce a simple, yet effective, bottleneck that can be integrated with any existing self-supervised objective for RL. We demonstrate this across several online and offline RL benchmarks, along with a real robot arm task, where we find that compressed representations with RepDIB can lead to strong performance improvements, as the learned bottlenecks help predict only the relevant state while ignoring irrelevant information.

Riashat Islam, Manan Tomar, Alex Lamb et al. · mila

Learning to control an agent from data collected offline in a rich pixel-based visual observation space is vital for real-world applications of reinforcement learning (RL). A major challenge in this setting is the presence of input information that is hard to model and irrelevant to controlling the agent. This problem has been approached by the theoretical RL community through the lens of exogenous information, i.e, any control-irrelevant information contained in observations. For example, a robot navigating in busy streets needs to ignore irrelevant information, such as other people walking in the background, textures of objects, or birds in the sky. In this paper, we focus on the setting with visually detailed exogenous information, and introduce new offline RL benchmarks offering the ability to study this problem. We find that contemporary representation learning techniques can fail on datasets where the noise is a complex and time dependent process, which is prevalent in practical applications. To address these, we propose to use multi-step inverse models, which have seen a great deal of interest in the RL theory community, to learn Agent-Controller Representations for Offline-RL (ACRO). Despite being simple and requiring no reward, we show theoretically and empirically that the representation created by this objective greatly outperforms baselines.

Riashat Islam, Hongyu Zang, Anirudh Goyal et al. · mila

Goal-conditioned reinforcement learning (RL) is a promising direction for training agents that are capable of solving multiple tasks and reach a diverse set of objectives. How to \textit{specify} and \textit{ground} these goals in such a way that we can both reliably reach goals during training as well as generalize to new goals during evaluation remains an open area of research. Defining goals in the space of noisy and high-dimensional sensory inputs poses a challenge for training goal-conditioned agents, or even for generalization to novel goals. We propose to address this by learning factorial representations of goals and processing the resulting representation via a discretization bottleneck, for coarser goal specification, through an approach we call DGRL. We show that applying a discretizing bottleneck can improve performance in goal-conditioned RL setups, by experimentally evaluating this method on tasks ranging from maze environments to complex robotic navigation and manipulation. Additionally, we prove a theorem lower-bounding the expected return on out-of-distribution goals, while still allowing for specifying goals with expressive combinatorial structure.

Hongyu Zang, Xin Li, Jie Yu et al.

Offline reinforcement learning (RL) struggles in environments with rich and noisy inputs, where the agent only has access to a fixed dataset without environment interactions. Past works have proposed common workarounds based on the pre-training of state representations, followed by policy training. In this work, we introduce a simple, yet effective approach for learning state representations. Our method, Behavior Prior Representation (BPR), learns state representations with an easy-to-integrate objective based on behavior cloning of the dataset: we first learn a state representation by mimicking actions from the dataset, and then train a policy on top of the fixed representation, using any off-the-shelf Offline RL algorithm. Theoretically, we prove that BPR carries out performance guarantees when integrated into algorithms that have either policy improvement guarantees (conservative algorithms) or produce lower bounds of the policy values (pessimistic algorithms). Empirically, we show that BPR combined with existing state-of-the-art Offline RL algorithms leads to significant improvements across several offline control benchmarks. The code is available at \url{https://github.com/bit1029public/offline_bpr}.

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

We introduce LongCat-Flash-Thinking-2601, a 560-billion-parameter open-source Mixture-of-Experts (MoE) reasoning model with superior agentic reasoning capability. LongCat-Flash-Thinking-2601 achieves state-of-the-art performance among open-source models on a wide range of agentic benchmarks, including agentic search, agentic tool use, and tool-integrated reasoning. Beyond benchmark performance, the model demonstrates strong generalization to complex tool interactions and robust behavior under noisy real-world environments. Its advanced capability stems from a unified training framework that combines domain-parallel expert training with subsequent fusion, together with an end-to-end co-design of data construction, environments, algorithms, and infrastructure spanning from pre-training to post-training. In particular, the model's strong generalization capability in complex tool-use are driven by our in-depth exploration of environment scaling and principled task construction. To optimize long-tailed, skewed generation and multi-turn agentic interactions, and to enable stable training across over 10,000 environments spanning more than 20 domains, we systematically extend our asynchronous reinforcement learning framework, DORA, for stable and efficient large-scale multi-environment training. Furthermore, recognizing that real-world tasks are inherently noisy, we conduct a systematic analysis and decomposition of real-world noise patterns, and design targeted training procedures to explicitly incorporate such imperfections into the training process, resulting in improved robustness for real-world applications. To further enhance performance on complex reasoning tasks, we introduce a Heavy Thinking mode that enables effective test-time scaling by jointly expanding reasoning depth and width through intensive parallel thinking.

Hongyu Zang, Xin Li, Leiji Zhang et al.

While bisimulation-based approaches hold promise for learning robust state representations for Reinforcement Learning (RL) tasks, their efficacy in offline RL tasks has not been up to par. In some instances, their performance has even significantly underperformed alternative methods. We aim to understand why bisimulation methods succeed in online settings, but falter in offline tasks. Our analysis reveals that missing transitions in the dataset are particularly harmful to the bisimulation principle, leading to ineffective estimation. We also shed light on the critical role of reward scaling in bounding the scale of bisimulation measurements and of the value error they induce. Based on these findings, we propose to apply the expectile operator for representation learning to our offline RL setting, which helps to prevent overfitting to incomplete data. Meanwhile, by introducing an appropriate reward scaling strategy, we avoid the risk of feature collapse in representation space. We implement these recommendations on two state-of-the-art bisimulation-based algorithms, MICo and SimSR, and demonstrate performance gains on two benchmark suites: D4RL and Visual D4RL. Codes are provided at \url{https://github.com/zanghyu/Offline_Bisimulation}.

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.

Ruixiang Sun, Hongyu Zang, Xin Li et al.

Visual Model-Based Reinforcement Learning (MBRL) promises to encapsulate agent's knowledge about the underlying dynamics of the environment, enabling learning a world model as a useful planner. However, top MBRL agents such as Dreamer often struggle with visual pixel-based inputs in the presence of exogenous or irrelevant noise in the observation space, due to failure to capture task-specific features while filtering out irrelevant spatio-temporal details. To tackle this problem, we apply a spatio-temporal masking strategy, a bisimulation principle, combined with latent reconstruction, to capture endogenous task-specific aspects of the environment for world models, effectively eliminating non-essential information. Joint training of representations, dynamics, and policy often leads to instabilities. To further address this issue, we develop a Hybrid Recurrent State-Space Model (HRSSM) structure, enhancing state representation robustness for effective policy learning. Our empirical evaluation demonstrates significant performance improvements over existing methods in a range of visually complex control tasks such as Maniskill \cite{gu2023maniskill2} with exogenous distractors from the Matterport environment. Our code is avaliable at https://github.com/bit1029public/HRSSM.

Chen Liu, Hongyu Zang, Xin Li et al.

Image-based Reinforcement Learning is a practical yet challenging task. A major hurdle lies in extracting control-centric representations while disregarding irrelevant information. While approaches that follow the bisimulation principle exhibit the potential in learning state representations to address this issue, they still grapple with the limited expressive capacity of latent dynamics and the inadaptability to sparse reward environments. To address these limitations, we introduce ReBis, which aims to capture control-centric information by integrating reward-free control information alongside reward-specific knowledge. ReBis utilizes a transformer architecture to implicitly model the dynamics and incorporates block-wise masking to eliminate spatiotemporal redundancy. Moreover, ReBis combines bisimulation-based loss with asymmetric reconstruction loss to prevent feature collapse in environments with sparse rewards. Empirical studies on two large benchmarks, including Atari games and DeepMind Control Suit, demonstrate that ReBis has superior performance compared to existing methods, proving its effectiveness.

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.

Wanli Yang, Hongyu Zang, Junwei Zhang et al.

Reinforcement learning (RL) has achieved remarkable success in LLM reasoning, but whether it can also improve direct recall of parametric knowledge remains an open question. We study this question in a controlled zero-shot, one-hop, closed-book QA setting with no chain-of-thought, training only on binary correctness rewards and applying fact-level train-test deduplication to ensure gains reflect improved recall rather than reasoning or memorization. Across three model families and multiple factual QA benchmarks, RL yields ~27% average relative gains, surpassing both training- and inference-time baselines alike. Mechanistically, RL primarily redistributes probability mass over existing knowledge rather than acquiring new facts, moving correct answers from the low-probability tail into reliable greedy generations. Our data-attribution study reveals that the hardest examples are the most informative: those whose answers never appear in 128 pre-RL samples (only ~18% of training data) drive ~83% of the gain, since rare correct rollouts still emerge during training and get reinforced. Together, these findings broaden the role of RL beyond reasoning, repositioning it as a tool for unlocking rather than acquiring latent parametric knowledge.

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.

Zeyu Wang, Yao-Hui Li, Xin Li et al.

Multi-View Reinforcement Learning (MVRL) seeks to provide agents with multi-view observations, enabling them to perceive environment with greater effectiveness and precision. Recent advancements in MVRL focus on extracting latent representations from multiview observations and leveraging them in control tasks. However, it is not straightforward to learn compact and task-relevant representations, particularly in the presence of redundancy, distracting information, or missing views. In this paper, we propose Multi-view Fusion State for Control (MFSC), firstly incorporating bisimulation metric learning into MVRL to learn task-relevant representations. Furthermore, we propose a multiview-based mask and latent reconstruction auxiliary task that exploits shared information across views and improves MFSC's robustness in missing views by introducing a mask token. Extensive experimental results demonstrate that our method outperforms existing approaches in MVRL tasks. Even in more realistic scenarios with interference or missing views, MFSC consistently maintains high performance.

Wanli Yang, Fei Sun, Rui Tang et al.

Fine-tuning, a foundational method for adapting large language models, has long been considered ineffective for model editing. Here, we challenge this belief, arguing that the reported failure arises not from the inherent limitation of fine-tuning itself, but from adapting it to the sequential nature of the editing task, a single-pass depth-first pipeline that optimizes each sample to convergence before moving on. While intuitive, this depth-first pipeline coupled with sample-wise updating over-optimizes each edit and induces interference across edits. Our controlled experiments reveal that simply restoring fine-tuning to the standard breadth-first (i.e., epoch-based) pipeline with mini-batch optimization substantially improves its effectiveness for model editing. Moreover, fine-tuning in editing also suffers from suboptimal tuning parameter locations inherited from prior methods. Through systematic analysis of tuning locations, we derive LocFT-BF, a simple and effective localized editing method built on the restored fine-tuning framework. Extensive experiments across diverse LLMs and datasets demonstrate that LocFT-BF outperforms state-of-the-art methods by large margins. Notably, to our knowledge, it is the first to sustain 100K edits and 72B-parameter models,10 x beyond prior practice, without sacrificing general capabilities. By clarifying a long-standing misconception and introducing a principled localized tuning strategy, we advance fine-tuning from an underestimated baseline to a leading method for model editing, establishing a solid foundation for future research.

Hongyu Zang, Xin Li, Mingzhong Wang

This work explores how to learn robust and generalizable state representation from image-based observations with deep reinforcement learning methods. Addressing the computational complexity, stringent assumptions and representation collapse challenges in existing work of bisimulation metric, we devise Simple State Representation (SimSR) operator. SimSR enables us to design a stochastic approximation method that can practically learn the mapping functions (encoders) from observations to latent representation space. In addition to the theoretical analysis and comparison with the existing work, we experimented and compared our work with recent state-of-the-art solutions in visual MuJoCo tasks. The results shows that our model generally achieves better performance and has better robustness and good generalization.

Hongyu Zang, Zhiwei Yu, Xiaojun Wan

In this paper, we explore a new approach for automated chess commentary generation, which aims to generate chess commentary texts in different categories (e.g., description, comparison, planning, etc.). We introduce a neural chess engine into text generation models to help with encoding boards, predicting moves, and analyzing situations. By jointly training the neural chess engine and the generation models for different categories, the models become more effective. We conduct experiments on 5 categories in a benchmark Chess Commentary dataset and achieve inspiring results in both automatic and human evaluations.

Hongyu Zang, Xiaojun Wan

Recently, encoder-decoder neural models have achieved great success on text generation tasks. However, one problem of this kind of models is that their performances are usually limited by the scale of well-labeled data, which are very expensive to get. The low-resource (of labeled data) problem is quite common in different task generation tasks, but unlabeled data are usually abundant. In this paper, we propose a method to make use of the unlabeled data to improve the performance of such models in the low-resourced circumstances. We use denoising auto-encoder (DAE) and language model (LM) based reinforcement learning (RL) to enhance the training of encoder and decoder with unlabeled data. Our method shows adaptability for different text generation tasks, and makes significant improvements over basic text generation models.

Hongyu Zang, Xiaojun Wan

Language style transfer has attracted more and more attention in the past few years. Recent researches focus on improving neural models targeting at transferring from one style to the other with labeled data. However, transferring across multiple styles is often very useful in real-life applications. Previous researches of language style transfer have two main deficiencies: dependency on massive labeled data and neglect of mutual influence among different style transfer tasks. In this paper, we propose a multi-agent style transfer system (MAST) for addressing multiple style transfer tasks with limited labeled data, by leveraging abundant unlabeled data and the mutual benefit among the multiple styles. A style transfer agent in our system not only learns from unlabeled data by using techniques like denoising auto-encoder and back-translation, but also learns to cooperate with other style transfer agents in a self-organization manner. We conduct our experiments by simulating a set of real-world style transfer tasks with multiple versions of the Bible. Our model significantly outperforms the other competitive methods. Extensive results and analysis further verify the efficacy of our proposed system.