LLM-Guided Evolution for Medical Decision Pipelines

2026-06-05 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Emerging Technologies & Innovation · Depth: Advanced, quick

Summary

A study introduces LLM-guided MAP-Elites evolution as an inference-time method for discovering medical decision strategies, offering an alternative to costly fine-tuning or manual prompt engineering. This approach formulates urgency triage, interactive consultation, and medical image classification as evolutionary searches over executable artifacts, optimized by task-specific fitness functions. Across these three medical settings, the evolutionary method consistently improved over manually designed baselines under practical constraints. For urgency triage, evolved programs increased Semigran accuracy from 77.3% to 87.1% and emergency recall from 0.60 to 0.97, enhancing safety-weighted held-out MIMIC-ESI performance. In interactive consultation, evolved policies improved the accuracy-cost frontier across Llama-3, Qwen-3.5, and Gemma-4 models, demonstrating transferability to held-out iCRAFTMD. Furthermore, prompt-only evolution on PneumoniaMNIST improved frozen MedGemma VLMs while preserving strict JSON outputs. The improvements stem from interpretable mechanisms like calibrated triage boundaries and targeted evidence acquisition.

Key takeaway

For AI Scientists and Machine Learning Engineers optimizing LLM performance in clinical workflows, consider LLM-guided MAP-Elites evolution as an inference-time alternative to costly fine-tuning. This method can significantly improve accuracy and recall in tasks like urgency triage and interactive consultation, even transferring across models. You should explore formulating your medical decision tasks as evolutionary searches over executable artifacts, leveraging task-specific fitness functions to discover robust and interpretable strategies.

Key insights

LLM-guided evolution offers an inference-time alternative to fine-tuning for optimizing medical decision pipelines.

Principles

• Evolutionary search can optimize LLM-based medical pipelines.
• Interpretable program-level mechanisms drive performance gains.
• Task-specific fitness functions guide strategy discovery.

Method

LLM-guided MAP-Elites evolution searches over executable artifacts, optimizing them with task-specific fitness functions for medical decision tasks like triage, consultation, and image classification.

In practice

• Apply evolution to calibrate triage boundaries.
• Use targeted evidence acquisition in consultations.
• Develop finding-oriented visual decision rules.

Topics

• LLM-Guided Evolution
• Medical AI
• Clinical Decision Support
• Urgency Triage
• Interactive Consultation
• MAP-Elites

Code references

• univanxx/llm_guided_evo_medical

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• LLM-Guided Evolution for Medical Decision Pipelines — LLM-guided evolution optimizes medical decision pipelines at inference-time, outperforming manual baselines through interpretable program-level changes.
• Clinically Interpretable Sepsis Early Warning via LLM-Guided Simulation of Temporal Physiological Dynamics — LLM-guided temporal simulation improves sepsis early warning by providing interpretable physiological trajectory predictions.
• Active Evidence-Seeking and Diagnostic Reasoning in Large Language Models for Clinical Decision Support — LLMs exhibit a substantial performance gap in active, multi-turn clinical diagnosis compared to static, full-context evaluations.
• LLM-Evolved Domain-Independent Heuristics for Symbolic AI Planning — LLMs can evolve domain-independent heuristics for symbolic AI planning that outperform human-designed ones.
• Evolution Is Back: A New Way to Fine‑Tune LLMs — Evolution Strategies offer a black-box, gradient-free alternative for LLM fine-tuning, particularly effective with sparse rewards.
• Reinforcement Learning Improves LLM Accuracy and Reasoning in Disease Classification from Radiology Reports — Reinforcement learning with GRPO enhances LLM accuracy and reasoning in medical text classification without explicit reasoning supervision.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.