← Daily Brief for October 20, 2026
GRAFT routes bad behavior into removable parameters
Focus: Technical AI Engineering
Date: October 19, 2026
Topics: Reward hacking, Gradient routing, Training safeguards
Evidence: Publisher Authored
Availability: Available
Summary: Anthropic-affiliated researchers propose GRAFT, which routes gradients from monitor-flagged behavior into a FORGET adapter removed after training instead of penalizing detected hacks. In a Qwen3-8B arithmetic environment, ablation nearly eliminated measured hacking while preserving gains. The study remains preliminary: monitors were simple, tasks small, and suitable anchor environments may be difficult to construct.
Why it matters: Replicate GRAFT only in a contained training study. Predefine held-out hack tests, compare penalties and prompting baselines, measure task loss after ablation, and document whether the monitor, anchor environment and model scale resemble the intended deployment.
Source: Reducing Reward Hacking by Localizing Bad Behavior Instead of Penalizing It