LLM-guided evolution boosts medical triage accuracy to 87.1%

A team of seven researchers (Sviridov et al.) from institutions including Skoltech and AI labs introduced LLM-guided MAP-Elites evolution as an inference-time method for optimizing medical decision pipelines without costly fine-tuning or manual prompt engineering. They formulated three clinical tasks—urgency triage, interactive consultation, and medical image classification—as evolutionary searches over executable artifacts optimized by task-specific fitness functions. In triage, the evolved programs improved Semigran accuracy from 77.3% to 87.1% and emergency recall from 0.60 to 0.97, while also boosting safety-weighted performance on the held-out MIMIC-ESI dataset. For interactive consultation, the method improved the accuracy-cost frontier across three open-source LLMs (Llama-3, Qwen-3.5, Gemma-4) and transferred to the held-out iCRAFTMD dataset. In PneumoniaMNIST classification, prompt-only evolution improved frozen MedGemma VLMs while preserving strict JSON output formatting.

Qualitative analysis revealed the gains came from interpretable program-level mechanisms—calibrated triage boundaries, targeted evidence acquisition, selective commitment to reliable LLM outputs, and finding-oriented visual decision rules—rather than superficial prompt rewording. The authors provide an open-source implementation repository. This approach offers a practical, cost-effective alternative to fine-tuning for deploying LLMs in clinical workflows, enabling safer and more accurate decision-making without retraining models. It demonstrates that evolutionary algorithms can systematically discover better strategies for LLM-based medical pipelines, potentially accelerating adoption in resource-constrained healthcare settings.