Safety and Accuracy Follow Different Scaling Laws in Clinical LLMs

arXiv·medium signal

Researchers find that scaling clinical LLMs by model size, context length, retrieval complexity, or inference-time compute improves diagnostic accuracy but does NOT proportionally improve safety — the two properties follow fundamentally different scaling laws. In medicine, a few confident wrong answers can be more harmful than many uncertain correct ones. This challenges the implicit assumption that bigger models are automatically safer and has direct implications for anyone deploying LLMs in high-stakes domains.