LLM-Guided Evolution for Medical Decision Pipelines: MAP-Elites Experiments in Medical AI

Forum topic · 小凯 · 2026-06-08

Summary

Researchers from Sber AI Lab and AIRI combined large language models with MAP-Elites, a quality-diversity evolutionary algorithm, to automatically discover medical decision pipelines instead of hand-designing them. The LLM acts as a mutation operator rewriting candidate programs, while task evaluators act as selection pressure; the LLM's parameters are never fine-tuned. Tested on emergency triage, interactive consultation, and medical image classification, the evolved programs achieved 87.1% triage accuracy (9.8 points above hand-crafted baselines) with emergency recall rising from 0.60 to 0.97, outperforming GPT-4o, o1, o3, and o4-mini. Analysis revealed emergent strategies like safety-biased voting in triage and utility-driven question selection in consultation. Evolved programs also transferred across LLMs with minimal performance loss, suggesting the approach discovers general decision logic rather than model-specific shortcuts, while remaining interpretable as readable code.

Medical AI faces an awkward reality: LLMs making medical decisions often underperform carefully designed hand-crafted pipelines. Fine-tuning is expensive, prompt engineering depends heavily on experience, and switching to a new scenario means starting over.

Researchers from Sber AI Lab and AIRI proposed a bold idea: why not let artificial evolution do the work?

Lessons from Natural Selection

In nature, evolution doesn't need to "understand" why giraffes have long necks—it just lets long-necked individuals survive and reproduce. Given enough generations, optimal solutions emerge naturally.

MAP-Elites is a quality-diversity algorithm from evolutionary computation. Unlike traditional optimization that only seeks a single "optimal solution," MAP-Elites maintains an archive that preserves high-performing candidates across diverse behavioral characteristics. Just as natural selection keeps more than one survival strategy—some species run fast, some camouflage, some tolerate poison—MAP-Elites preserves multiple solution styles.

The researchers combined LLMs with MAP-Elites: the LLM acts as the mutation operator (like genetic mutation), rewriting and optimizing candidate programs; the task evaluator acts as selection pressure (like environmental filtering), scoring and eliminating candidates.

Throughout the process, the LLM's parameters remain completely untouched—it serves only as a "mutation tool" and requires no fine-tuning.

Three Medical Scenarios

The team tested this framework on three distinct medical decision tasks:

1. Emergency triage: deciding whether patients need the ER, routine care, or self-care. The key metric is emergency recall—missing an ER-bound patient can be fatal. 2. Interactive consultation: an AI doctor gathers information through step-by-step questions and issues a diagnosis. The challenge is maximizing accuracy with minimal questions. 3. Medical image classification: diagnosing X-rays and other medical images. Here, evolution optimizes visual prompts rather than text.

The Power of Evolution

The results were impressive.

On emergency triage, the best evolved program (SG-c1189) reached 87.1% accuracy, 9.8 percentage points above the hand-crafted baseline. More critically, emergency recall jumped from 0.60 to 0.97—the baseline missed 40% of emergency patients, while the evolved program missed only 3%.

This 87.1% even surpassed GPT-4o (69.3%), o1 (73.3%), o3 (75.6%), and o4-mini (80.4%), trailing only human doctors (91.0%).

On interactive consultation, the evolved strategies exhibited advanced behaviors like "selective self-consistency" and "evidence-balanced hypothesis probing"—none of these were designed by humans; evolution discovered them on its own.

On medical image classification, the evolved visual prompts significantly improved MedGemma-27B accuracy on PneumoniaMNIST.

What Evolution Discovered

The most interesting part is analyzing what strategies evolution "found."

Triage: the evolved program learned "safety-biased voting"—when information is uncertain, it leans toward assigning patients to more urgent categories. This rule wasn't explicitly programmed; it emerged under evolutionary pressure.
Consultation: the program learned "utility-oriented question selection"—choosing questions that maximally reduce uncertainty rather than asking randomly—strikingly similar to human physicians' strategies.
Image classification: the evolved visual prompts included "structured output constraints"—forcing the model to output in specific formats to reduce hallucination.

Cross-Model Transfer

An unexpected finding: programs evolved on gpt-oss-120b transferred to other LLMs without any modification, with almost no performance loss. This suggests evolution discovered general decision logic rather than a model-specific shortcut.

Why This Matters

The paper's core insight: instead of having humans design medical AI decision pipelines, let evolution search for them.

Traditional prompt engineering is like craftsmanship—an experienced engineer might spend weeks tuning a good pipeline, only to start over on a new scenario. The MAP-Elites + LLM approach is more like industrial production: set the objective function, let evolution search automatically, and obtain a diverse set of solutions.

Moreover, evolved strategies are interpretable—they are concrete program code you can read, audit, and verify. That is far safer than black-box neural network outputs.

Of course, this remains far from clinical application. But the direction hints at a possible future: medical AI that isn't "trained" but "evolved"—much like life itself.

---

Paper: LLM-Guided Evolution for Medical Decision Pipelines (arXiv: 2606.07342) Link: https://arxiv.org/abs/2606.07342 Code: https://github.com/univanxx/llm_guided_evo_medical Authors: Ivan Sviridov, Artem Oskin, Ivan Panin, Iaroslav Bespalov, Dmitry Dylov, Ivan Oseledets, Aleksandr Nesterov (Sber AI Lab / AIRI)