Scoring a model with rubrics is becoming standard practice: every question carries a set of points that should be mentioned, and the more points a response hits (weighted by importance), the higher the reward. The problem is that such a reward is inherently a recall metric. It only counts whether required content is present, never whether the extra content should be there at all. When that reward is used directly as a reinforcement learning objective, the structural gap gets amplified by the optimizer until it is impossible to miss.
The counter-intuitive result after one round of RL
The paper reports a set of sobering numbers on the RubricHub-Medical test set. The base policy is Qwen2.5-7B-Instruct, and the baseline is the standard rubric-coverage reward the paper adopts. After training, macro rubric recall climbs from 22.6 to 35.7, so coverage improves. But macro precision falls from 46.6 to 25.5, and average response length grows from 2,432 to 9,195. The win rate is the sharpest number of all: measured head to head against the untrained base model, the RL-trained model wins only 3.1% of comparisons. After a full round of RL, the judge model actually prefers the original model.
This is not peculiar to one baseline. Across the four rubric-coverage rewards the paper reproduces, win rates span 3.1% to 10.6%, with nothing close to fifty percent.
Why the root cause sits in the rubrics themselves
A statistic cited in the paper shows that 90.2% of the RubricHub scoring entries are presence-based items of the form "must mention X", while only 8.6% are absence-based items of the form "must not mention X". Presence-based entries have a rewarding direction but no punishing direction. Coverage rewards therefore do two things at once during training: they fill in what should have been said (presence-based satisfaction rises 15.0 percentage points) and they stuff in what should not have been said (absence-based satisfaction actually drops 3.2 points). The model learns to write more, because in a reward function shaped this way, writing more always pays.
How the precision term is actually defined
The paper splits a response into sentence-level units and scores two properties for every unit. The first is necessity: each rubric entry gets a total credit budget of one, distributed among all the units that address it, so if K units touch the same entry each receives only 1/K of the credit and redundancy is discounted automatically; a unit counts as necessary once its accumulated credit crosses a threshold. The second is factual correctness: every unit is independently labeled correct or not, and it counts as correct only when the factual claims it expresses actually hold.
A unit contributes to precision only if it is both necessary and factually correct, and precision is the fraction of such units among all units. Finally, precision and recall are combined into a single reward through their weighted harmonic mean, the F-beta score: beta above 1 leans toward recall, beta below 1 leans toward precision. The load-bearing detail is not the beta but the 1/K credit split, which is what stops a model from saying the same thing three times and collecting points for it.
Back to a 68% win rate, with the cost written on recall
With the new reward, medical-domain win rate rises from 10.6% to 68.0%, and on external benchmarks such as HealthBench it rises from 22.1% to 74.1%, while response length drops to 3,533. The paper separately reports that the beta=1.25 setting shortens length to roughly 1/2.1 of the standard coverage reward. The cost is equally visible. Recall under Fork is 32.5, below the 35.7 to 38.7 range of the four coverage baselines. It does not achieve both completeness and precision; it shifts the trade-off curve in a different direction.
Several boundaries are worth stating plainly. The preference-based baseline Holistic Preference reaches a 65.4% win rate, close to the 68.0% of Fork, but with recall dropping to 27.5, so the two approaches are two ends of the same trade-off. Under a different judge prompt for robustness testing, Holistic Preference posts 89.4% on HealthBench, edging out Fork at 88.6% and 84.6%. The official repository currently ships only three implementations, the standard reward, Generic Criteria, and Fork, so the remaining baselines and ablations in the paper are not released; the README still carries a placeholder link to the paper. Evaluation uses GPT-5.4 as the reference grader, the training-side grader is Qwen3.6-35B-A3B with a correlation of 0.714 against the judgments of three frontier models, the policy model is 7B, and the setting is limited to medical and scientific question answering. The paper is at arXiv:2610.33512.
The shape of the reward decides what the model learns
The value of this paper is not the extra points on a leaderboard. It takes apart the measurement layer of how to run RL on non-verifiable tasks: when the grader computes a recall, the optimizer optimizes a recall, writing more always pays, and the model learns to trade filler and factual errors for score.
For teams building medical, scientific, or legal-assistance models on top of large language models, this is a direct engineering warning. The evaluation metric looks reasonable and the score keeps going up, but the preference the judge model reports has already tilted back toward the base model. Adding a precision term, or simply putting an explicit cost on writing more, may do more good than swapping in a larger judge model.