Ray Pan

Zengyi Qin, Jinyuan Chen, Yunze Man et al.

Training computer use agents requires full-featured OS sandboxes with GUI environments, which consume substantial hardware resources as the number of sandboxes scales. Stochastic errors arising from diverse software execution within these sandboxes further demand robust infrastructure design and reliable error recovery. We present OSGym, a scalable OS environment infrastructure for computer use agents, built around these key optimization strategies: (1) Decentralized OS state management, which isolates failures to individual replicas and significantly enhances overall system reliability; (2) Hardware-aware OS replica orchestration, which addresses CPU-bounded scaling bottlenecks and substantially reduces compute overhead; (3) KVM virtualization with copy-on-write disk management, which shares a common bootable disk across VM instances and provisions only instance-specific modifications, reducing physical disk consumption by 88% and increasing disk provisioning speed by 37 times; and (4) Robust container pool with multi-layer fault recovery. Together, these optimizations yield strong scalability and resource efficiency: OSGym manages over a thousand OS replicas under constrained resources, supports parallel trajectory generation at 1420 multi-turn trajectories per minute, and reduces per-replica cost to 0.2-0.3 USD per day, a 90% reduction over standard deployment. Our experiments validate OSGym across end-to-end pipelines for data collection and training for computer use agents. We believe OSGym establishes a new foundation for scalable, general-purpose computer use agent research.

Hongye Zheng, Yichen Wang, Ray Pan et al.

This paper presents a gradient-informed fine-tuning method for large language models under few-shot conditions. The goal is to enhance task adaptability and training stability when data is limited. The method builds on a base loss function and introduces two gradient-related regularization terms. The first enforces gradient direction consistency to guide parameter updates along task-relevant directions and prevent drift. The second controls gradient magnitude to avoid abnormal updates. Together, these components support a more efficient and stable optimization path. To further improve cross-task generalization, the method incorporates a gradient alignment mechanism. This mechanism measures the consistency between optimization directions of the source and target tasks. It enhances fine-tuning performance in multi-task and cross-domain scenarios. Across various natural language understanding tasks, the method outperforms existing fine-tuning strategies in average accuracy, gradient stability, and directional alignment. Empirical evaluations under different sample sizes and domain-specific tasks confirm the method's robustness and broad applicability in low-resource environments. In particular, the method shows clear advantages in controlling parameter update paths. The results demonstrate that a gradient-based fine-tuning framework can effectively leverage the representational power of large language models. It ensures training stability while reducing dependence on large volumes of labeled data.