Yuheng Zhao

Han Fang, Yize Huang, Yuheng Zhao et al.

Robot manipulation has increasingly adopted data-driven generative policy frameworks, yet the field faces a persistent trade-off: diffusion models suffer from high inference latency, while flow-based methods often require complex architectural constraints. Although in image generation domain, the MeanFlow paradigm offers a path to single-step inference, its direct application to robotics is impeded by critical theoretical pathologies, specifically spectral bias and gradient starvation in low-velocity regimes. To overcome these limitations, we propose the One-step MeanFlow Policy (OMP), a novel framework designed for high-fidelity, real-time manipulation. We introduce a lightweight directional alignment mechanism to explicitly synchronize predicted velocities with true mean velocities. Furthermore, we implement a Differential Derivation Equation (DDE) to approximate the Jacobian-Vector Product (JVP) operator, which decouples forward and backward passes to significantly reduce memory complexity. Extensive experiments on the Adroit and Meta-World benchmarks demonstrate that OMP outperforms state-of-the-art methods in success rate and trajectory accuracy, particularly in high-precision tasks, while retaining the efficiency of single-step generation.

Wei Gao, Yuheng Zhao, Tianyuan Wu et al.

Agentic Reinforcement Learning (RL) enables Large Language Models (LLMs) to perform autonomous decision-making and long-term planning. Unlike standard LLM post-training, agentic RL workloads are highly heterogeneous, combining compute-intensive prefill phases, bandwidth-bound decoding, and stateful, CPU-heavy environment simulations. We argue that efficient agentic RL training requires disaggregated infrastructure to leverage specialized, best-fit hardware. However, naive disaggregation introduces substantial synchronization overhead and resource underutilization due to the complex dependencies between stages. We present RollArc, a distributed system designed to maximize throughput for multi-task agentic RL on disaggregated infrastructure. RollArc is built on three core principles: (1) hardware-affinity workload mapping, which routes compute-bound and bandwidth-bound tasks to bestfit GPU devices, (2) fine-grained asynchrony, which manages execution at the trajectory level to mitigate resource bubbles, and (3) statefulness-aware computation, which offloads stateless components (e.g., reward models) to serverless infrastructure for elastic scaling. Our results demonstrate that RollArc effectively improves training throughput and achieves 1.35-2.05\(\times\) end-to-end training time reduction compared to monolithic and synchronous baselines. We also evaluate RollArc by training a hundreds-of-billions-parameter MoE model for Qoder product on an Alibaba cluster with more than 3,000 GPUs, further demonstrating RollArc scalability and robustness. The code is available at https://github.com/alibaba/ROLL.

Yuheng Zhao, Yu-Hu Yan, Amit Attia et al.

Parameter-free stochastic optimization aims to design algorithms that are agnostic to the underlying problem parameters while still achieving convergence rates competitive with optimally tuned methods. While some parameter-free methods do not require the specific values of the problem parameters, they still rely on prior knowledge, such as the lower or upper bounds of them. We refer to such methods as ``partially parameter-free''. In this work, we target achieving ``fully parameter-free'' methods, i.e., the algorithmic inputs do not need to satisfy any unverifiable condition related to the true problem parameters. We propose a powerful and general grid search framework, named \textsc{Grasp}, with a novel self-bounding analysis technique that effectively determines the search ranges of parameters, in contrast to previous work. Our method demonstrates generality in: (i) the non-convex case, where we propose a fully parameter-free method that achieves near-optimal convergence rate, up to logarithmic factors; (ii) the convex case, where our parameter-free methods are competitive with strong performance in terms of acceleration and universality. Finally, we contribute a sharper guarantee for the model ensemble, a final step of the grid search framework, under interpolated variance characterization.

Yuheng Zhao, Yu-Hu Yan, Kfir Yehuda Levy et al.

Smoothness is known to be crucial for acceleration in offline optimization, and for gradient-variation regret minimization in online learning. Interestingly, these two problems are actually closely connected -- accelerated optimization can be understood through the lens of gradient-variation online learning. In this paper, we investigate online learning with Hölder smooth functions, a general class encompassing both smooth and non-smooth (Lipschitz) functions, and explore its implications for offline optimization. For (strongly) convex online functions, we design the corresponding gradient-variation online learning algorithm whose regret smoothly interpolates between the optimal guarantees in smooth and non-smooth regimes. Notably, our algorithms do not require prior knowledge of the Hölder smoothness parameter, exhibiting strong adaptivity over existing methods. Through online-to-batch conversion, this gradient-variation online adaptivity yields an optimal universal method for stochastic convex optimization under Hölder smoothness. However, achieving universality in offline strongly convex optimization is more challenging. We address this by integrating online adaptivity with a detection-based guess-and-check procedure, which, for the first time, yields a universal offline method that achieves accelerated convergence in the smooth regime while maintaining near-optimal convergence in the non-smooth one.

Zhijun Zheng, Tian Qiu, Yuheng Zhao et al.

In visual analytics, applying filters to drill-down and extract higher-value insights is a common and important data analysis method. When the drill-down space becomes excessively large, analysts may lose orientation, leading to decreased efficiency in the drill-down process. To tackle these challenges, we propose the Intelligent Drill-Down Framework, in which a large language model (LLM) facilitates the generation of visual insights, leverages user interaction data to interpret user intent, and generates appropriate drill-down paths. Our method is designed to assist users in identifying valuable drill-down paths when exploring multidimensional data, thereby reducing the cognitive burden of data interpretation and facilitating the generation of insights. Specifically, we propose a drill-down path recommendation method, in which the LLM is trained to approximate a validated greedy algorithm. Secondly, we analyze the user's intent to construct a drill-down chart. Finally, we design a branch management method. Building upon this framework, we designed a system that includes a hybrid interface providing hierarchical navigation to monitor users and manage parallel branches, a visualization panel for interactive data exploration, and an insight panel to present analytical findings and generate drill-down recommendations. We evaluated the effectiveness of our method through a demonstrative use case and a user study.

Yuheng Zhao, Andrew Jacobsen, Nicolò Cesa-Bianchi et al.

We develop parameter-free algorithms for unconstrained online learning with regret guarantees that scale with the gradient variation $V_T(u) = \sum_{t=2}^T \|\nabla f_t(u)-\nabla f_{t-1}(u)\|^2$. For $L$-smooth convex loss, we provide fully-adaptive algorithms achieving regret of order $\widetilde{O}(\|u\|\sqrt{V_T(u)} + L\|u\|^2+G^4)$ without requiring prior knowledge of comparator norm $\|u\|$, Lipschitz constant $G$, or smoothness $L$. The update in each round can be computed efficiently via a closed-form expression. Our results extend to dynamic regret and find immediate implications to the stochastically-extended adversarial (SEA) model, which significantly improves upon the previous best-known result [Wang et al., 2025].

Wei Gao, Yuheng Zhao, Dilxat Muhtar et al.

Agentic reinforcement learning (RL) has emerged as a key driver for improving the multi-step reasoning and tool-use capabilities of LLMs. However, its efficiency is bottlenecked by long-tail rollouts with multi-turn environment interactions, making static GPU provisioning a poor fit: overprovisioning wastes GPUs on stragglers, while underprovisioning increases contention and slows training. We observe that production serving clusters routinely leave substantial GPU compute and memory headroom. Based on this observation, we argue for cooperative elasticity: opportunistically repurposing underutilized serving GPUs to execute rollouts. Realizing cooperative elasticity is non-trivial because it must preserve serving Service Level Objectives (SLOs) under bursty traffic and minimize communication overhead. To address these challenges, we present ROSE, a cooperative, resource-elastic post-training system that safely harvests idle compute and memory on serving GPUs to accelerate agentic RL rollouts. ROSE consists of three components: (1) an SLO-safe co-serving executor that improves rollout throughput while preserving serving SLOs through efficient GPU memory and compute sharing; (2) a cross-cluster weight transfer engine that leverages weight shards and sparsity for fast weight synchronization across clusters; and (3) an elastic rollout scheduler that dynamically provisions cooperative capacity and routes trajectory rollouts across dedicated rollout GPUs and opportunistic serving GPUs. Experiments across multiple model sizes and cluster scales show that ROSE improves average end-to-end throughput by 1.20-3.31 x compared with state-of-the-art resource-fixed and elastic baselines.

Wei Gao, Yuheng Zhao, Dakai An et al.

Reinforcement Learning (RL) is a pivotal post-training technique for enhancing the reasoning capabilities of Large Language Models (LLMs). However, synchronous RL post-training often suffers from significant GPU underutilization, referred to as bubbles, caused by imbalanced response lengths within rollout steps. Many RL systems attempt to alleviate this problem by relaxing synchronization, but this can compromise training accuracy. In this paper, we introduce tail batching, a novel rollout scheduling strategy for synchronous RL that systematically consolidates prompts leading to long-tail responses into a small subset of rollout steps (long rounds), while ensuring that the majority of steps (short rounds) involve only balanced, short rollouts. By excluding long responses from short rounds and rescheduling them into a few designated long rounds, tail batching effectively reduces GPU idle time during rollouts and significantly accelerates RL training without sacrificing accuracy. We present RollPacker, a system that fully harnesses the benefits of tail batching through holistic optimizations across all three RL stages: elastic parallelism adaptation for rollout, dynamic resource allocation and scheduling for reward, and stream-based training. Empirical results show that RollPacker achieves a 2.03x-2.56x end-to-end training time reduction compared to veRL and up to 2.24x speedup compared to RLHFuse for the Qwen2.5 family of LLMs on up to 128 H800 GPUs.

Chongke Bi, Xin Gao, Baofeng Fu et al.

Large-scale simulations on supercomputers have become important tools for users. However, their scalability remains a problem due to the huge communication cost among parallel processes. Most of the existing communication latency analysis methods rely on the physical link layer information, which is only available to administrators. In this paper, a framework called PCLVis is proposed to help general users analyze process communication latency (PCL) events. Instead of the physical link layer information, the PCLVis uses the MPI process communication data for the analysis. First, a spatial PCL event locating method is developed. All processes with high correlation are classified into a single cluster by constructing a process-correlation tree. Second, the propagation path of PCL events is analyzed by constructing a communication-dependency-based directed acyclic graph (DAG), which can help users interactively explore a PCL event from the temporal evolution of a located PCL event cluster. In this graph, a sliding window algorithm is designed to generate the PCL events abstraction. Meanwhile, a new glyph called the communication state glyph (CS-Glyph) is designed for each process to show its communication states, including its in/out messages and load balance. Each leaf node can be further unfolded to view additional information. Third, a PCL event attribution strategy is formulated to help users optimize their simulations. The effectiveness of the PCLVis framework is demonstrated by analyzing the PCL events of several simulations running on the TH-1A supercomputer. By using the proposed framework, users can greatly improve the efficiency of their simulations.