Async RL recipes converge: score centering and masked importance sampling now default in TRL
_lewtun · x · 2026-10-27
The slime RL framework highlights that score centering + masked importance sampling are now the default in TRL's async GRPO trainer, with reportedly strong stability across various staleness and sampler quantization settings — evidence that the community is converging on a common recipe for stable asynchronous RL fine-tuning.
More from coding & agent
- Querio launches: analytics platform that puts your data stack in a git repo for AI agents — thetripathi58 · 2026-10-28
- Stop conflating "harness": labs build the pilot, companies build the mech — robleclerc · 2026-10-28
- MiRL: error-targeted RL cuts terminal agent misbehaviors 43-57% — _akhaliq · 2026-10-28
- Vjeux's AI agents hit 2.6x speedups on Open Kernel Challenge kernels, then stalled — Vjeux · 2026-10-28
- Judgment Labs opens access as a monitoring layer for long-horizon AI agents — alifcoder · 2026-10-28
- Anthropic Reveals 12 New Prompting Rules for Claude Opus 5.5 — JeremyCMorgan · 2026-10-28