NeuroEvolve: Learning What EEG to Generate for Foundation Model Adaptation

EEG foundation models (FMs) have shown strong transferability across neural decoding tasks, yet their downstream adaptation remains constrained by limited task-specific data. Synthetic EEG generation offers a promising solution, but existing methods primarily optimize for fidelity to real EEG or generic data augmentation, overlooking a critical discrepancy: high-fidelity EEG is not necessarily model-beneficial EEG. This raises a fundamental question: what EEG should be generated for effective foundation model adaptation? We propose NeuroEvolve, a feedback-driven EEG generation paradigm that learns what EEG to generate based on model needs. NeuroEvolve combines Closed-loop Strategy Discovery, where planner and evaluator agents collaboratively refine generation strategies through foundation model feedback, with Successful Experience Internalization, which distills successful interactions into reusable knowledge for unseen tasks. Extensive experiments across diverse EEG tasks, datasets, and foundation models demonstrate consistent adaptation gains and validate the effectiveness of both mechanisms. NeuroEvolve shifts EEG generation from reproducing what real EEG looks like toward learning what EEG FMs actually need.