AstraFlow: Dataflow-Oriented Reinforcement Learning for Agentic LLMs

Researchers introduced AstraFlow, a dataflow-based RL system for agentic LLMs that separates rollout, dataflow management, and training into independent components.

The design supports multi-policy training, elastic scaling, and heterogeneous cross-region execution without any system-level code changes.

Across math, coding, search, and AgentBench benchmarks, AstraFlow matched or outperformed existing systems and reduced training time by 2.7x.

The system aims to make large-scale RL for LLM agents faster to engineer and easier to scale on mixed compute resources.