Hugging Face fixes TRL CISPO default that shrank gradients fivefold in its reproduction

Hugging Face shipped TRL v1.18.0 on November 13th, 2026, raising the CISPO default for the 'epsilon_high' setting to 5.0 after an inherited 0.2 ceiling clipped every on-policy importance weight to 0.2 when users left the parameter unset, shrinking gradients fivefold in its reproduction. The fix corrects a reinforcement-learning loss default that had been silently capping on-policy weights near 1.

Hugging Face fixes TRL CISPO default that shrank gradients fivefold in its reproduction Hugging Face's TRL v1.18.0 sets the CISPO default to 5.0 after an inherited 0.2 ceiling clipped on-policy weights near 1. By RuntimeWire Staff https://runtimewire.com/author/runtimewire-staff · Published Hugging Face, co-founded by Clement Delangue, has corrected a default in TRL's CISPO reinforcement-learning loss that clipped every on-policy importance weight to 0.2 when users left 'epsilon high' unset. The fix shipped in TRL v1.18.0 on November 13th, 2026, and raises the CISPO default for that setting to 5.0.…