Beyond Correctness: Validity-Oriented Evaluation of Biomedical LLM Judges
Artificial Intelligence
Summary
The authors created a new way to test how well biomedical language models judge medical tasks, especially when there aren't many expert human ratings available. They improved existing tests by adding clear, checkable changes to data that help compare model preferences. Their method checks if models are accurate, consistent, and follow instructions. Using this, they tested different training methods on a model called Llama-3.1-8B-Instruct and found that combining supervised fine-tuning followed by reinforcement learning worked best, especially on complex medical tasks like extracting specific information and calculations.
biomedical benchmarkslanguage modelssupervised fine-tuningreinforcement learningPICO extractionclinical calculationsmodel evaluationpreference pairsrobustnesscompliance
Authors
Rodrigo de Oliveira, Federico Pittino, James Gwinnutt, Jay Nanavati
Abstract
We propose a scalable, validity-oriented pipeline for evaluating biomedical LLM judges when high-quality human judgments are scarce. First, we augment existing human-labelled biomedical benchmarks with deterministic, metric-grounded mutations that produce auditable preference pairs. Second, we evaluate judges beyond aggregate correctness using three deployment-relevant dimensions: correctness against metric-derived gold labels, robustness under repeated stochastic sampling, and compliance with the requested output format. We use this pipeline to assess Llama-3.1-8B-Instruct under four regimes: (1) base, using the instruct model as is; (2) SFT, distillation-based supervised fine-tuning only; (3) RL, GRPO-based reinforcement learning only; and (4) SFT$\rightarrow$RL, SFT followed by RL. The base and single-stage regimes struggle on structured medical discrimination such as PICO extraction and clinical calculations, whereas SFT$\rightarrow$RL performs best across correctness, compliance, and robustness; gains concentrate on decomposable tasks (PICO, MedCalc), at times matching or outperforming frontier models.