22 upvotes
Precision objective curbs reward hacking in rubric-based RL
What is Eat Your Precision: Precision Objective Mitigates Reward Hacking in Rubric-based RL?
Instead of rewarding how much of a rubric a response covers, the authors propose rewarding how precisely it hits the rubric criteria. This precision-focused objective reduces the model’s tendency to game the rubric and produces more faithful answers on tasks that can’t be automatically checked.
Who it is for
Anyone training RL agents for subjective or open-ended tasks can apply this tweak to get better alignment without extra data.