LoongSage paper describes asynchronous agentic RL framework for LLM training
The arXiv abstract reports a roughly 14K-line agent-native reinforcement learning framework built on Megatron and SGLang, with validated training recipes.
In plain words
Training large language models with reinforcement learning, where models learn by acting over many steps, is getting harder as models take on agentic tasks like calling tools and making long multi-turn decisions. LoongSage is a proposed framework meant to make that training simpler to run and debug. Its authors say it runs asynchronously, meaning parts work in parallel without waiting for each other, and they report steady reward gains in their test runs.
What was announced
The authors, writing in an arXiv abstract, argue that as LLM training shifts from static reasoning toward agentic work with tool use, multi-turn interaction and long-horizon decisions, existing RL frameworks have become unwieldy, raising adoption and debugging overhead. LoongSage is presented as an agent-native, fully asynchronous framework with a lightweight core of roughly 14,000 lines of code, built on Megatron for distributed training and SGLang for high-throughput inference.
Technically, the abstract describes a single controller process for orchestration, GPU components packaged as Ray actors, and an AgentFlow paradigm covering both white-box and black-box agents. Asynchronous execution uses sliding-window strategies to bound trajectory staleness, while TransferMesh routes updated weights after each step over the fastest transport available to each receiver. A memory-efficient Teacher Manager decouples teacher model scaling from GPU memory, supporting multi-teacher on-policy distillation. Across four complete training runs, the authors report steadily improving task rewards, a controlled train-inference gap, and hardware utilization sustained by removing synchronous idle bubbles.
What it does not show
- Evidence is limited to the paper's abstract; no independent evaluation, benchmark detail, or availability information is provided.
For security and IT teams
- Scope
- The claims come from a single arXiv abstract covering four self-reported training runs; no external validation or release details appear in the supplied evidence.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.