A new preprint introduces LoongSage, an agent-native, fully asynchronous reinforcement learning (RL) framework designed for large language models (LLMs) arXiv.
The framework aims to address challenges in adopting, using, and debugging existing RL frameworks for agentic tasks, which involve external tool invocation, multi-turn interactions, and long-horizon decision-making. LoongSage centralizes orchestration within a single controller process and encapsulates all GPU-resident components as Ray actors. It introduces AgentFlow, a unified execution paradigm that orchestrates both white-box and black-box agents. The framework also employs sliding-window strategies to bound trajectory staleness and TransferMesh, a topology-aware synchronization mechanism that routes weights after each step over the fastest transport available to each receiver.
According to the preprint, LoongSage incorporates a memory-efficient Teacher Manager that decouples the scaling of teacher models from GPU memory constraints, natively supporting Multi-Teacher On-Policy Distillation. The authors report extensive evaluations across four complete training runs, spanning complex reinforcement learning and multi-teacher distillation. They claim that task rewards improve steadily while the train-inference gap is kept under continuous control, and the fully asynchronous pipeline removes synchronous idle bubbles, sustaining hardware utilization.
The preprint does not include independent test results or peer review. The framework's performance and efficiency claims are based on the authors' evaluations, which have not been validated by external parties. The paper was published on arXiv on October 14, 2026.