AutoRL: A Multi-Agent System towards Autonomous Reinforcement Learning
Summary
- Researchers introduce AutoRL, a multi-agent system designed to automate reinforcement learning training and support recursive self-improvement.
- AutoRL uses Perceptors to track training dynamics, Proposers to diagnose and refine interventions, and an Orchestrator to apply approved updates.
- Experiments on Qwen3 and LongCat-Flash-Lite show AutoRL outperforms static RL recipes and general agents in unattended RL supervision.
Why it matters
Automating reinforcement learning through multi-agent systems addresses unstable training dynamics and reduces the need for constant human supervision, serving as a cornerstone for recursive self-improvement in AI.