← Back to the feed
Biotech & healthResearch findings

MedPhysBench: A Benchmark for Medical Physics Calculation and Error Detection

Original published 2026.11.04 03:40 KSTAdded to Blackdeck 2026.11.05 06:02 KST

What has changed?

A benchmark dataset that evaluates both calculation accuracy and error detection in medical physics tasks

Key points

  1. MedPhysBench introduces a benchmark for evaluating language models in medical physics calculations and error detection
  2. It includes ten seeded calculation generators covering radiation therapy, diagnostic imaging, and nuclear medicine with reference answers
  3. Five models detected at least 96.7% of large planted errors, but some increased false alarms when only checking results

Connect it to your work

12 models achieved 33.3% to 100.0% accuracy on 1,200 test items from radiation therapy to clinical calibration

What remains unproven

This research note is based on an abstract; check the original for test conditions and application scope.

Originals used for this note

Links open the original publisher in a new tab.

  1. arxivMedPhysBench: A Benchmark for Evaluating Language Models on Medical Physics Calculations and Error Detection