In-context adaptation

Adapting to a new subject or session at inference time without gradient updates, by supplying a short slice of that session's data as a context — analogous to in-context learning in large language models — or by identifying only the new session's unit embeddings from a brief calibration block. The goal is plug-and-play onboarding measured in minutes rather than a full retraining, which is exactly what a clinical or consumer device needs.

This is one of the least mature promises of neural foundation models. Some spiking and EEG models show few-shot or prompted adaptation, and freezing a backbone while fitting a tiny per-session embedding is a practical, reliable form of it. Genuinely gradient-free in-context transfer — where the frozen model reads a new subject's context and decodes correctly with no fitting at all — is an active research target rather than a solved capability, and claims of it should specify exactly what, if anything, was updated.