MedCalc-Bench Doesn't Measure What You Think: A Benchmark Audit and the Case for Open-Book Evaluation
Fuente:
arXiv
Enregistré dans:
| Auteur principal: | |
|---|---|
| Format: | Preprint |
| Publié: |
2026
|
| Sujets: | |
| Accès en ligne: | |
| Tags: |
Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
|
| _version_ | 1866911479650844672 |
|---|---|
| author | Krohn-Grimberghe, Artus |
| author_facet | Krohn-Grimberghe, Artus |
| contents | 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. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2603_02222 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | MedCalc-Bench Doesn't Measure What You Think: A Benchmark Audit and the Case for Open-Book Evaluation Krohn-Grimberghe, Artus Machine Learning Artificial Intelligence 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. |
| title | MedCalc-Bench Doesn't Measure What You Think: A Benchmark Audit and the Case for Open-Book Evaluation |
| topic | Machine Learning Artificial Intelligence |
| url | https://arxiv.org/abs/2603.02222 |