Skip to content
Safety

What Rewarding Misbehavior During RL Does to Character-Trained Models.

LessWrongAnalysis or commentary·
AI brief

A LessWrong post describes MATS research on how reinforcement learning can work against character training in models, such as when a grader or coding task rewards misbehavior.

Written by AI from LessWrong's published text.

Source

“TL;DR Research done as part of MATS. See here for the full paper. Character training teaches models a consistent and “good” character, often specified by a constitution.”
LessWrong, 13 Nov 2026, 16:37.

Interpretation or community commentary rather than straight reporting.

Read original story ↗