MedCalc-Bench Doesn't Measure What You Think: A Benchmark Audit and the Case for Open-Book Evaluation
About
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.
Key insights
What problem does this audit address?
MedCalc-Bench evaluates LLMs on clinical calculator tasks but conflates formula memorization with clinical reasoning. The paper audits benchmark implementations and tests open-book prompting, where models receive calculator specifications at inference time.
How does open-book prompting compare to prior work?
Unlike RL or agentic approaches, simply appending calculator specifications raises GLM-4.6V from 51.9% to 81.5% on 1,100 Verified rows, surpassing MedCalc-R1's 74% without fine-tuning. The audit also found over 20 implementation errors across 55 calculators.
What are the practical implications?
The findings suggest MedCalc-Bench should be reframed as a tool-use benchmark evaluating parameter extraction rather than formula recall. No production deployment is documented; results are benchmark prototypes on the Verified test split.
What limitations do the authors acknowledge?
Primary evaluation uses only mid-range GLM models; frontier models were not run due to cost. The 275-row GLM-4.7 subsample limits precision. The composite upper bound assumes GPT-5.2-Thinking outperforms GLM-4.6V on all rows, which may not always hold.