XNews

AutoRL Multi-Agent System Automates Reinforcement Learning Training

AutoRL Multi-Agent System Automates Reinforcement Learning Training

Researchers from Meituan, CUHK, ShanghaiTech and Fudan introduced AutoRL, a multi-agent system that supervises live reinforcement learning training through Perceptor, Proposer and Orchestrator roles. On Qwen3, AutoRL raised AIME 2024 from 46.67% to 66.67%, improved AIME 2025 and HMMT 2025 by 11.67 points each, and lifted LongCat-Flash-Lite agentic coding pass@1 from 52.29% to 65.13%.

Event version 1.

Key insights

Three agentic roles supervise training

Perceptors track training dynamics, rollouts, evaluation and efficiency, triggering deliberation on anomalies, degradation or stagnation. Three Proposers diagnose problems and refine interventions through shared evidence, while an Orchestrator reviews recommendations and applies approved changes via online hot updates or checkpoint-based resumption.

Math benchmark gains on Qwen3

Against a static GRPO recipe on Qwen3, AIME 2024 rises from 46.67% to 66.67%, AIME 2025 and HMMT 2025 each improve by 11.67 points, and IMO-AnswerBench improves by 9.75 points. Qwen3-1.7B improves AIME 2025 from 18.33% to 30.00%.

Agentic coding improvement with LongCat-Flash-Lite

On SWE-related tasks with LongCat-Flash-Lite, pass@1 rises from 52.29% to 65.13% across an unattended long RL run with interruptions.

Flawed recipe repair

Given a deliberately flawed recipe with a weak grader, no KL and 131K-token responses, AutoRL detects and fixes the problems and beats a general-purpose agent on most tasks.

Validation remains compute-intensive

The research records show existing agents can handle unattended RL supervision and dynamic adjustment. The difficulty for future recursive self-improvement lies in developing fundamental methods valuable for resource-consuming validation.

Source posts