# Scalable Stewardship of an LLM-Assisted Clinical Benchmark with Physician Oversight

arXiv: https://arxiv.org/abs/2512.19691 · HTML: https://mlsanity.com/paper/2512.19691/

## Key facts

- Cited by 2 papers in the map (data through 14 Oct 2026).
- Builds on 20 of its 57 references that are in the map.
- First posted on arXiv 22 Dec 2025.

## Builds on (most cited first, 20 of 20)

- [Training language models to follow instructions with human feedback](https://mlsanity.com/paper/2203.02155/) (2022)
- [DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models](https://mlsanity.com/paper/2402.03300/) (2024)
- [Qwen3 Technical Report](https://mlsanity.com/paper/2505.09388/) (2025)
- [HybridFlow: A Flexible and Efficient RLHF Framework](https://mlsanity.com/paper/2409.19256/) (2024)
- [High-Dimensional Continuous Control Using Generalized Advantage Estimation](https://mlsanity.com/paper/1506.02438/) (2015)
- [What Disease does this Patient Have? A Large-scale Open Domain Question Answering Dataset from Medical Exams](https://mlsanity.com/paper/2009.13081/) (2020)
- [Understanding R1-Zero-Like Training: A Critical Perspective](https://mlsanity.com/paper/2503.20783/) (2025)
- [Training Language Models to Self-Correct via Reinforcement Learning](https://mlsanity.com/paper/2409.12917/) (2024)
- [Pervasive Label Errors in Test Sets Destabilize Machine Learning Benchmarks](https://mlsanity.com/paper/2103.14749/) (2021)
- [Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data](https://mlsanity.com/paper/2404.01413/) (2024)
- [MedCalc-Bench: Evaluating Large Language Models for Medical Calculations](https://mlsanity.com/paper/2406.12036/) (2024)
- [MedHELM: Holistic Evaluation of Large Language Models for Medical Tasks](https://mlsanity.com/paper/2505.23802/) (2025)
- [A Preliminary Study of o1 in Medicine: Are We Closer to an AI Doctor?](https://mlsanity.com/paper/2409.15277/) (2024)
- [General Scales Unlock AI Evaluation with Explanatory and Predictive Power](https://mlsanity.com/paper/2503.06378/) (2025)
- [Are LLMs Better than Reported? Detecting Label Errors and Mitigating Their Effect on Model Performance](https://mlsanity.com/paper/2410.18889/) (2024)
- [Training LLMs for EHR-Based Reasoning Tasks via Reinforcement Learning](https://mlsanity.com/paper/2505.24105/) (2025)
- [AI-based Clinical Decision Support for Primary Care: A Real-World Study](https://mlsanity.com/paper/2507.16947/) (2025)
- [From Scores to Steps: Diagnosing and Improving LLM Performance in Evidence-Based Medical Calculations](https://mlsanity.com/paper/2509.16584/) (2025)
- [Med-U1: Incentivizing Unified Medical Reasoning in LLMs via Large-scale Reinforcement Learning](https://mlsanity.com/paper/2506.12307/) (2025)
- [Rethinking Test-Time Scaling for Medical AI: Model and Task-Aware Strategies for LLMs and VLMs](https://mlsanity.com/paper/2506.13102/) (2025)

## Built on by (most cited first, 2 of 2)

- [MedCalc-Bench Doesn't Measure What You Think: A Benchmark Audit and the Case for Open-Book Evaluation](https://mlsanity.com/paper/2603.02222/) (2026)
- [Unknown is not normal: separating language-model extraction from rule-based decision logic for clinical risk scores](https://mlsanity.com/paper/2609.34112/) (2026)

## Abstract (arXiv)

Reference labels for machine-learning benchmarks are increasingly synthesized with LLM assistance, but their reliability remains underexamined. We audit MedCalc-Bench, a clinical benchmark for medical score computation whose labels were partly derived with LLM assistance, and develop a scalable physician-in-the-loop stewardship pipeline to reassess them. At least 27% of test labels are likely erroneous or incomputable. On a 50-instance subset validated by physicians, our recomputed labels agree with physician ground truth 74% of the time (95% CI, 60-84%) versus 20% for the originals (95% CI, 11-33%). Using original labels to evaluate frontier LLMs underestimates accuracy by 16-23 percentage points. In a controlled reinforcement-learning experiment, a model trained on recomputed labels outperforms one trained on originals by 13.5 percentage points (95% CI, 10.6-16.6%) on physician-labeled instances, and this advantage extends to related medical tasks. LLM-assisted benchmarks can propagate systematic errors into both evaluation and post-training unless actively stewarded.