Xujie Shen

Wei Fu, Jiaxuan Gao, Xujie Shen et al. · tsinghua

Reinforcement learning (RL) has become a dominant paradigm for training large language models (LLMs), particularly for reasoning tasks. Effective RL for LLMs requires massive parallelization and poses an urgent need for efficient training systems. Most existing large-scale RL systems for LLMs are synchronous, alternating generation and training in a batch setting where rollouts in each training batch are generated by the same model. This approach stabilizes RL training but suffers from severe system-level inefficiency: generation must wait until the longest output in the batch is completed before model updates, resulting in GPU underutilization. We present AReaL, a fully asynchronous RL system that completely decouples generation from training. Rollout workers in AReaL continuously generate new outputs without waiting, while training workers update the model whenever a batch of data is collected. AReaL also incorporates a collection of system-level optimizations, leading to substantially higher GPU utilization. To stabilize RL training, AReaL balances the workload of rollout and training workers to control data staleness, and adopts a staleness-enhanced PPO variant to better handle outdated training samples. Extensive experiments on math and code reasoning benchmarks show that AReaL achieves up to 2.77$\times$ training speedup compared to synchronous systems with the same number of GPUs and matched or improved final performance. The code of AReaL is available at https://github.com/inclusionAI/AReaL/.

Ziqiang Dang, Tianxing Fan, Boming Zhao et al.

Incorporating temporal information effectively is important for accurate 3D human motion estimation and generation which have wide applications from human-computer interaction to AR/VR. In this paper, we present MoManifold, a novel human motion prior, which models plausible human motion in continuous high-dimensional motion space. Different from existing mathematical or VAE-based methods, our representation is designed based on the neural distance field, which makes human dynamics explicitly quantified to a score and thus can measure human motion plausibility. Specifically, we propose novel decoupled joint acceleration manifolds to model human dynamics from existing limited motion data. Moreover, we introduce a novel optimization method using the manifold distance as guidance, which facilitates a variety of motion-related tasks. Extensive experiments demonstrate that MoManifold outperforms existing SOTAs as a prior in several downstream tasks such as denoising real-world human mocap data, recovering human motion from partial 3D observations, mitigating jitters for SMPL-based pose estimators, and refining the results of motion in-betweening.

Lin Zeng, Boming Zhao, Jiarui Hu et al.

Novel view synthesis with neural models has advanced rapidly in recent years, yet adapting these models to scene changes remains an open problem. Existing methods are either labor-intensive, requiring extensive model retraining, or fail to capture detailed types of changes over time. In this paper, we present GaussianUpdate, a novel approach that combines 3D Gaussian representation with continual learning to address these challenges. Our method effectively updates the Gaussian radiance fields with current data while preserving information from past scenes. Unlike existing methods, GaussianUpdate explicitly models different types of changes through a novel multi-stage update strategy. Additionally, we introduce a visibility-aware continual learning approach with generative replay, enabling self-aware updating without the need to store images. The experiments on the benchmark dataset demonstrate our method achieves superior and real-time rendering with the capability of visualizing changes over different times