MedCalc-Bench Doesn't Measure What You Think: A Benchmark Audit and the Case for Open-Book Evaluation

TL;DR

This paper audits the MedCalc-Bench benchmark, uncovers errors, demonstrates that open-book prompting significantly improves model accuracy, and argues the benchmark mainly measures memorization rather than clinical reasoning.

Contribution

It systematically audits and corrects errors in MedCalc-Bench, shows that open-book prompting greatly enhances performance, and advocates for re-framing the benchmark as a tool-use evaluation.

Findings

Over 20 errors identified and fixed in the benchmark

Open-book prompting raises accuracy to 81-85%

GPT-5.2 achieves 95-97% accuracy, limited by dataset issues

Abstract

MedCalc-Bench is a widely used benchmark for evaluating LLM performance on clinical calculator tasks, with state-of-the-art direct prompting scores plateauing around 35% on the Verified split (HELM MedHELM leaderboard) and the best published approach-RL with verifiable rewards-reaching 74%. We present three contributions that challenge the benchmark's current framing. First, we conduct a systematic audit of the benchmark's calculator implementations, identifying and fixing over 20 errors ranging from critical formula inaccuracies to runtime bugs in a NeurIPS-published dataset. Second, we show that a simple intervention-providing the model with the calculator specification at inference time ("open-book" prompting)-raises accuracy from ~52% to 81-85% on GLM-4.6V and GLM-4.7, surpassing all published results including RL-trained systems, without any fine-tuning. Third, we establish an upper bound of 95-97% using GPT-5.2-Thinking, with residual errors attributable primarily to ground-truth issues and dataset ambiguities. Our findings suggest that MedCalc-Bench predominantly measures formula memorization and arithmetic precision rather than clinical reasoning, and would be better framed as a tool-use evaluation.