Health-SCORE: Towards Scalable Rubrics for Improving Health-LLMs

TL;DR

Health-SCORE tackles scalable rubric-based evaluation for healthcare LLMs by clustering expert rubrics into a generalizable set and applying an adaptive selection mechanism per prompt. It demonstrates two core uses: a structured reward signal for reinforcement learning with safety-aware supervision and prompt-based in-context guidance to improve generation quality, supported by an adaptive criterion-insertion approach. Across HealthBench and CSEDB, Health-SCORE achieves alignment with physician-authored instance-level rubrics comparable to fully bespoke criteria while substantially reducing rubric-development effort and maintaining robustness under distribution shift. The approach also yields faster, more stable training and effective test-time guidance, offering a practical pathway to scalable, trustworthy health AI. The sequence-level reward in Health-SCORE is computed from selected rubric judgments, e.g., $+1$, $-1$, or $0$ per criterion, and optimized via a GRPO-based objective with adaptive KL control, enabling scalable supervision without sacrificing fidelity to expert judgment.

Abstract

Rubrics are essential for evaluating open-ended LLM responses, especially in safety-critical domains such as healthcare. However, creating high-quality and domain-specific rubrics typically requires significant human expertise time and development cost, making rubric-based evaluation and training difficult to scale. In this work, we introduce Health-SCORE, a generalizable and scalable rubric-based training and evaluation framework that substantially reduces rubric development costs without sacrificing performance. We show that Health-SCORE provides two practical benefits beyond standalone evaluation: it can be used as a structured reward signal to guide reinforcement learning with safety-aware supervision, and it can be incorporated directly into prompts to improve response quality through in-context learning. Across open-ended healthcare tasks, Health-SCORE achieves evaluation quality comparable to human-created rubrics while significantly lowering development effort, making rubric-based evaluation and training more scalable.

Figures (12)

• Figure 1: Specificity vs. cost tradeoff. Both Instance-level and Health-SCORE rubrics exhibit high specificity, but Health-SCORE is far less costly to develop.
• Figure 2: Health-SCORE rubric creation process: First, original rubrics are clustered using high-dimensional embeddings. Then the clusters are refined and a Health-SCORE rubric is proposed for each cluster, reducing redundancy while preserving core evaluative dimensions.
• Figure 3: Health-SCORE rubric Selection Process: Given a user query, an LLM-based selector scores each Health-SCORE rubric for contextual relevance. Rubrics whose scores exceed a threshold are selected. Then the adaptive Health-SCORE rubrics will be added to the system prompt and used as training rewards during RL post-training.
• Figure 4: Training using Health-SCORE as the learning objective and evaluation using human-authored rubrics as the gold standard.
• Figure 5: In-Domain and out-of-distribution (OOD) evaluation when models are trained with different rubric types. Dotted/dashed lines correspond to lower/upper bounds.