slime merges Score Centering, a first-principles fix for RL instability when train and sampling policies diverge
hsu_byron · x · 2026-09-23
The open-source RL training framework slime (THUDM, 8.5k stars) has merged Score Centering support, enabled via --use-score-centering.
- Background: RL with LLMs is very unstable when training and sampling policies differ; standard fixes like matching numerics and importance sampling only work around the problem
- The authors find the root cause from first principles and propose a way to directly cancel it; Score Centering is competitive and compatible with existing approaches while being simple to implement
- The PR is merged into main — ready to try in your RL runs today
More from coding & agent
- mcp-server-devutils: Zero-auth MCP server bundling base64, UUID, JWT decode, cron and more — modelcontextprotocol · 2026-09-23
- Kiso: an open-source agent runtime that makes crash-window tool executions explicitly uncertain — niurenwangdadan · 2026-09-23
- Open-source GUI agents top out at 8% task success on composite cross-device tasks — maier_ak · 2026-09-23
- JarvisGUI benchmark tests GUI agents across Android, Windows and Ubuntu in one workflow — maier_ak · 2026-09-23
- Cross-DURIAN: A New Benchmark for Multi-Device GUI Agents — maier_ak · 2026-09-23
- "Proactive" personal agents: reminder, proposal, or acting without asking? — sujingshen · 2026-09-23