MetaRubric: Learning to Reward for Rubric-Based Reinforcement Learning

Reinforcement learning with verifiable rewards has made post-training highly effective when correctness can be checked automatically. However, many important model behaviors require satisfying several qualitative criteria at once. Rubric-based rewards address this setting by grading prompt-specific criteria and aggregating them into a scalar reward. Yet standard static aggregations conflate a criterion's human-assigned importance with its current usefulness as an optimization signal. We show that this assumption breaks down in rubric RL: many important criteria are already saturated or currently unreachable, while criteria that distinguish rollouts are not necessarily those with the largest human weights. We introduce POW3R, a policy-aware rubric reward framework that preserves human weights and category balance as the rubric objective while adapting criterion-level reward weights during training. POW3R uses rollout-level contrast to emphasize criteria that currently separate the policy's outputs, making the GRPO reward more informative without changing the underlying evaluation target. Across three base policies on two datasets spanning multimodal and text-only settings, POW3R wins $24$ of $30$ base-policy/metric comparisons, improving both mean rubric reward and strict completion (the fraction of prompts whose response satisfies every required rubric criterion) over vanilla GRPO with rubric rewards, and reaches the same plateau in $2.5$--$4\times$ fewer training steps. Rubric rewards should therefore distinguish what should matter in the final answer from what can teach the current policy.

Rubric-based reinforcement learning, in which an LLM judge scores responses against prompt-specific criteria, has become a standard way to post-train language models on tasks without verifiable answers. The rubric, however, is only a proxy for response quality, and a policy trained against it long enough learns to exploit the difference. We train models with GRPO on medical and science rubrics and grade their responses on out-of-distribution (OOD) benchmarks throughout training with two judges: the training judge and a stronger OOD judge. The two scores first rise together and then diverge. The policy is learning to satisfy the training judge without writing better responses, which is a form of reward hacking. To mitigate it, we propose Rubric Dropout, which applies dropout to rubric criteria: at each training step, a random subset of each rubric's criteria is removed from the reward, so the rewarded subset of criteria is resampled at each step. All responses sampled for the same prompt are scored on the same subset, so their rewards remain comparable, and evaluation always uses the full rubric. We also introduce Block Dropout, which drops semantically similar criteria together so that a near duplicate of a dropped criterion cannot keep rewarding the same behavior. Across both benchmarks and models from 1.7B to 8B, dropout improves the gold judge's score on OOD benchmarks by up to 3 percent on HealthBench-Hard and 7 percent on ResearchQA, and it reduces both the gap between the two judges and over-crediting. It keeps in-domain reward close to the baseline, and dropout rates of 30-50 percent work best. Block Dropout improves the gold judge's OOD score further at the same dropout rate. Changing which criteria are rewarded at each step is a simple, inexpensive way to make rubric rewards harder to exploit.

Many language tasks have no single answer that can be checked automatically. Rubrics provide criteria for judging responses to these tasks. For reinforcement learning, the resulting verdicts must be combined into a scalar reward. A common approach sums the points assigned to satisfied criteria. Distinct verdict patterns can thus receive the same reward, and the fixed points encode how much each criterion should count, not how strongly its verdict distinguishes the current rollouts. Beyond this aggregation problem, judging the full rubric needs more judge requests as the criterion count grows. To address these limitations, Rubric Response Theory (RRT) measures quality and selects criteria when rubric criteria are monotone indicators of a shared target. Rather than adding assigned points, RRT uses a two parameter item response model that treats the verdict pattern as evidence about scalar quality specific to the rubric. Under this model, its likelihood score maximizes the local signal-to-noise ratio for quality. Its Response Parameter Network (RPN) reads the prompt and criterion text to predict criterion difficulty and discrimination. As the policy distribution changes during training, RRT uses online expectation maximization to update the RPN from current rollout verdicts. With Qwen3.5-4B as the policy, RRT's macro criterion score across Medical, Science, Rubrics as Rewards Science, and RubricBench is 1.7 points above that of group relative policy optimization (GRPO). On hard and very hard criteria in Medical and Science, RRT gains 2.8 to 5.6 points over GRPO. At half the criterion budget, adaptive Fisher selection with a frozen RPN keeps the macro criterion score across four datasets within 0.1 points of GRPO with full judging. These results show RRT can reduce judge requests while remaining competitive with GRPO.