AI's 'evolutionary' learning methods only work if the AI is already good at the task

Researchers tried to make AI models learn better by having them 'evolve' new parts, but it turns out this only helps if the AI is already well-tuned for the job. If the AI is not already good at the task, the evolutionary process actually makes it worse or has no effect. This means that simply adding an 'evolutionary' layer to an AI model will not automatically improve its performance.

AI developers often try to mimic biological evolution to create systems that can adapt and improve on their own. This paper shows that this approach has a critical limitation: it only works when the AI's core components are already well-aligned with the task. This finding pushes back on the idea that AI can simply 'evolve' its way to better performance without careful initial design. It means that the hard work of building a good base model still matters more than fancy evolutionary add-ons.