LLM-Guided Evolution for Medical Decision Pipelines: MAP-Elites Experiments in Healthcare AI
Medical AI faces an awkward reality: LLMs making medical decisions often underperform carefully hand-crafted workflows. Fine-tuning is expensive, prompt engineering depends heavily on experience, and every new scenario means starting over.
A research team 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 simply lets long-necked individuals survive and reproduce more. Over enough generations, optimal solutions emerge.
MAP-Elites is a quality-diversity algorithm from evolutionary computation. Unlike traditional optimization that seeks only a single "optimal solution," MAP-Elites maintains an archive storing high-performing candidates across diverse behavioral characteristics. Just as natural selection preserves multiple survival strategies—some species run fast, some camouflage, some tolerate poison—MAP-Elites keeps a diverse portfolio of solutions.
The researchers combined LLMs with MAP-Elites:
- The LLM acts as a "mutation operator" (like genetic mutation), rewriting and optimizing candidate programs
- The task evaluator acts as "selection pressure" (like environmental filtering), scoring and eliminating candidates
Three Medical Scenarios
The framework was tested on three distinct medical decision scenarios:
1. Emergency Triage: Deciding whether a patient should go to the ER, a regular clinic, or self-care. The key metric is emergency recall—missing one emergency patient can be fatal.
2. Interactive Consultation: An AI doctor gathers information through step-by-step questions and delivers a diagnosis. The challenge is achieving accurate judgments with the fewest questions.
3. Medical Image Classification: Diagnosing X-rays and other medical images. Here, what evolves is the visual prompt rather than text.
The Power of Evolution
The results are impressive.
On the triage task, the evolved best program SG-c1189 achieved 87.1% accuracy, 9.8 percentage points above the hand-crafted baseline. More critically, emergency recall surged from 0.60 to 0.97—meaning the baseline missed 40% of emergency patients, while the evolved program missed only 3%.
This 87.1% result even exceeded GPT-4o (69.3%), o1 (73.3%), o3 (75.6%), and o4-mini (80.4%), trailing only human physicians (91.0%).
On the interactive consultation task, the evolved strategies exhibited advanced behaviors such as "selective self-consistency" and "evidence-balanced hypothesis probing"—not human-designed, but "discovered" by evolution itself.
On medical image classification, evolved visual prompts significantly improved MedGemma-27B's accuracy on PneumoniaMNIST.
What Evolution Discovered
The most interesting part is analyzing what strategies evolution "found."
In triage, the evolved program learned "safety-biased voting"—when information is uncertain, it tends to assign patients to more urgent categories. This rule wasn't explicitly programmed by humans; it emerged naturally under evolutionary pressure.
In consultation, the evolved program learned "utility-oriented question selection"—asking not random questions, but those that maximally reduce uncertainty. This strikingly resembles human physicians' interviewing strategies.
In image classification, evolved visual prompts contained "structured output constraints"—forcing the model to output in specific formats to reduce hallucinations.
Cross-Model Transfer
An unexpected finding: programs evolved on gpt-oss-120b transfer to other LLMs without any modification, with almost no performance loss. This suggests evolution discovered not a model-specific "shortcut," but some form of general decision logic.
Why This Matters
The paper's core insight: rather than having humans manually 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 in a new scenario. The MAP-Elites + LLM approach is more like industrialized production: set the objective function, let evolution search automatically, and obtain a diverse collection of solutions.
More importantly, evolved strategies are interpretable—they are concrete program code that you can read, audit, and verify. This 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 isn't "trained"—it's "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)