Selective Few-Shot Improves LLM Clinical Chains-of-Thought

In a blinded comparative study published in 2026, researchers evaluated large language model–generated chains-of-thought (CoTs) across 200 standardized assisted reproductive technology (ART) cases, comparing zero-shot, random few-shot (five random examples), and selective few-shot (six curated examples) prompting strategies. Selective few-shot prompting significantly outperformed other strategies on logical clarity, use of key information, and clinical accuracy (P<.001), while a GPT-4o automated evaluator failed to detect these differences, indicating prompt design and human-in-the-loop assessment are critical for trustworthy clinical CoT generation.