rubrics-as-rewards-paper

Rubrics as Rewards: Reinforcement Learning Beyond Verifiable Domains

Authors: Anisha Gunjal, Anthony Wang, Elaine Lau, Vaskar Nath et al.
Date: 2026-04-22

Summary

Proposes using rubrics (structured evaluation criteria) as reward signals for RL training in domains where verification is difficult. Extends RL beyond math/code into open-ended tasks.

Key Claims

• Rubrics can serve as effective reward signals for RL training
• Enables RL in domains without clear verifiers (creative writing, analysis, reasoning)
• Bridges the gap between verifiable and non-verifiable reward modeling

Concepts

• Reward Hacking — rubrics may reduce reward hacking vs simple scalar rewards
• RL Infrastructure

See Also

• tweet-rubrics-as-rewards — earlier tweet coverage
• openrubrics-reward-modeling — companion paper on synthetic rubric generation