Skip to content
Last Dictionary
Plane
Training & Post-trainingEvaluation & BenchmarksLanguage Models & Transformers
emerging

rubric reward

nounAlso: rubric-based reward, rubrics as rewards, RaR, checklist reward, rubric grading

In plain wordsA training score built from a checklist of specific things a good answer should do, with an AI grader ticking off each one — a way to reward quality where no single right answer exists.

  1. A reward signal for reinforcement-learning fine-tuning computed by grading a response against a rubric: a list of explicit, often prompt-specific criteria (for example, that an answer flags a drug interaction, avoids recommending a dosage or cites the governing statute), each with a weight and sometimes marked essential or penalizing. An LLM judgeLLM-as-a-judge, n.: Using one AI language model to grade or compare the answers of another, as a cheaper and faster stand-in for human reviewers. checks each criterion and the results are aggregated into a scalar, either as a weighted sum of per-item verdicts or as one holistic score guided by the rubric. Rubrics extend reinforcement learning with verifiable rewardsreinforcement learning with verifiable rewards, n.: Training an AI by letting it attempt problems whose answers can be checked automatically — a math result, a passing program — and rewarding it only when the check succeeds. to medicine, law, writing and research, where answers cannot be checked by rule, while being more transparent and harder to exploit than a single preference score from a reward modelreward model, n.: A separate AI model trained to score responses the way people would, so it can stand in for human judges while another model is trained on millions of examples.. They may be written by experts, generated by models from reference answers, or both; their coverage and the judge's reliability bound what the policy can learn.

    Each medical prompt has a rubric of about twenty criteria, and the rubric reward is the weighted fraction met.

Origin

Rubric-based grading of model outputs spread through evaluation benchmarks such as OpenAI's HealthBench (2025), which scores answers against physician-written criteria. Anisha Gunjal and colleagues at Scale AI (July 2025) proposed instance-specific rubrics as on-policy RL rewards, reporting relative gains of up to about 30 percent on HealthBench over Likert-scale judge rewards; Moonshot AI's Kimi K2 (2025) used a related self-critique rubric reward.

Usage note

Not the same as process rewardsprocess reward model, n.: A model that grades each step of an AI's worked solution, not just the final answer, so mistakes can be caught where they happen — like a teacher marking every line of a proof., which grade reasoning steps; rubric items usually describe properties of the final answer. Vague or keyword-like criteria invite reward hackingreward hacking, n.: When an AI finds a shortcut that earns the reward or score it is trained on without doing the task its designers wanted — like a student gaming the grading rubric..

Related

See also
  • reinforcement learning with verifiable rewardsreinforcement learning with verifiable rewards, n.: Training an AI by letting it attempt problems whose answers can be checked automatically — a math result, a passing program — and rewarding it only when the check succeeds.
  • LLM-as-a-judgeLLM-as-a-judge, n.: Using one AI language model to grade or compare the answers of another, as a cheaper and faster stand-in for human reviewers.
  • reward modelreward model, n.: A separate AI model trained to score responses the way people would, so it can stand in for human judges while another model is trained on millions of examples.
  • RL environmentRL environment, n.: A practice world built for training an AI by trial and error: it hands the AI tasks and tools, lets it act, and scores the results so the AI can learn what works.
  • reinforcement learning from AI feedbackreinforcement learning from AI feedback, n.: Training an AI with preference judgments made by another AI instead of by people, which is cheaper and faster but only as good as the AI doing the judging.
  • Constitutional AIConstitutional AI, n.: Anthropic's technique for making an AI less harmful by having it check and rewrite its own answers against a short written set of principles, so that people need not label harmful examples by hand.
Contrast with
  • reinforcement learning with verifiable rewardsreinforcement learning with verifiable rewards, n.: Training an AI by letting it attempt problems whose answers can be checked automatically — a math result, a passing program — and rewarding it only when the check succeeds.

Sources

  1. Gunjal, A. et al. (2025). Rubrics as Rewards: Reinforcement Learning Beyond Verifiable Domains. arXiv:2507.17746. arxiv.org/abs/2507.17746

Mentioned in

LLM-as-a-judgeLLM-as-a-judge, n.: Using one AI language model to grade or compare the answers of another, as a cheaper and faster stand-in for human reviewers., reinforcement learning with verifiable rewardsreinforcement learning with verifiable rewards, n.: Training an AI by letting it attempt problems whose answers can be checked automatically — a math result, a passing program — and rewarding it only when the check succeeds., RL environmentRL environment, n.: A practice world built for training an AI by trial and error: it hands the AI tasks and tools, lets it act, and scores the results so the AI can learn what works., generative reward modelgenerative reward model, n.: A scoring model that judges an AI's answer by writing out its reasoning in words before giving a verdict, instead of simply outputting a number., HealthBenchHealthBench, prop. n.: An OpenAI test of how well AI chatbots handle health conversations, grading 5,000 realistic chats against checklists written by 262 doctors from 60 countries.