BACKGROUND AND PURPOSE:

• It is challenging to measure large language model (LLM) performance on tasks requiring reasoning
• Script Concordance Testing (SCT) is a way to measure how well decisions can be adjusted when new information becomes available
• McCoy et al. (NEJM AI, 2025) created a SCT benchmark for LLM assessment to assess reasoning capabilities of these algorithms vs medical professionals  

METHODS:

• Public benchmark of 750 SCT questions
  • Drawn from 10 international, diverse datasets (9 newly releases) across multiple specialties
• SCT benchmark
  • Each item presents a clinical vignette
  • Data is added and the model is asked how the added data changed the likelihood of a diagnosis or management option
  • Model performance is scored against expert-panel responses
• Assessed
  • 10 state-of-the-art LLMs
  • 1070 medical students
  • 193 residents
  • 300 attending physicians

RESULTS:

• Overall, LLMs showed markedly lower performance on the SCT than they usually show on medical multiple-choice benchmarks
• Across prompting conditions, OpenAI’s o3 achieved the highest performance
  • OpenAI’s o3: 67.8% (SD, 1.2)
  • GPT-4o: 63.9% (SD, 1.3)
  • OpenAI’s o1-preview: 58.2% (SD, 1.3)
  • DeepSeek R1: 55.5% (SD, 1.4)
  • Google’s Gemini 2.5 Pro Preview: 52.1% (SD, 1.4)
• Reasoning-optimized models matched or exceeded student performance on multiple examinations but did not reach the level of senior residents or attending physicians
• Response-pattern analysis showed systematic overconfidence
  • Reasoning-tuned models overused strong, definitive answers
  • This suggests that training AI to explain its thinking step-by-step might actually make it less flexible when dealing with uncertain medical situations

CONCLUSION:

• Even among LLMs explicitly tuned for reasoning, SCT found that the clinical reasoning these models could display was limited
• The authors state

Our analysis of model response patterns reveals that even state-of-the-art LLMs exhibit striking overconfidence, disproportionately favoring extreme belief shifts and failing to recognize when new information should not alter clinical hypotheses

Learn More – Primary Sources:

Assessment of Large Language Models in Clinical Reasoning: A Novel Benchmarking Study