AutoRL: Multi-Agent System Automates RL Training, Lifts Qwen3 AIME from 46.7% to 66.7%

omarsar0 · x · 2026-11-17

A paper from Meituan, CUHK, ShanghaiTech and Fudan introduces AutoRL, a multi-agent system that supervises live RL runs like an on-call RL researcher: Perceptor agents watch training curves and evals, three Proposers diagnose issues, and an Orchestrator applies hot updates or checkpoint restarts. Qwen3 improves on AIME 2024 from 46.67% to 66.67% (+11.67 on AIME/HMMT 2025); LongCat-Flash-Lite agentic coding pass@1 rises from 52.29% to 65.13%, sustaining 14.7 days of unattended training with 11 runtime failure fixes.

Original post →

More from coding & agent

coding & agent channel →