Reducing Reward Hacking by Localizing Bad Behavior Instead of Penalizing It

Published 24 October 2026

GRAFT confines learning updates from detected reward hacks to parameters deleted after training, helping prevent even reward hacks that monitors fail to detect.