A new training framework lets medical AI models revise the rules that guide their own reasoning as they learn.
Researchers introduced Evo-PI, a system that treats reasoning principles as living, language-based instructions rather than fixed reward signals. Standard training locks in a set of rules from the start; Evo-PI runs a feedback loop in which the model's behavior reshapes those rules, and the revised rules then sharpen the model's next round of reasoning. The team tested the approach on medical visual question answering — tasks that require a model to interpret images alongside text — across eight benchmarks and several model architectures. Accuracy gains reached as high as 24.6%.
Most alignment research focuses on what a model outputs, not how it reasons to get there. Evo-PI targets the reasoning process directly, which matters in high-stakes domains like medicine where a plausible-sounding wrong answer can cause real harm. If the approach transfers beyond medical imaging, it could offer a general alternative to the static reward models that currently bottleneck complex AI reasoning.
The code is public, so the claims are testable — though benchmark gains on curated datasets have a habit of shrinking when models meet messier real-world data.