NeuroGenesysWhat is a brain foundation model?
Last updated • By Will Schulz
A brain foundation model is a large neural network pretrained, often self-supervised, on many hours of neural recordings across sessions, subjects or devices. It can then be adapted with little data to new decoding tasks or new people, instead of training a bespoke decoder for every patient and session.
What are examples of brain foundation models?
POYO (Azabou et al., NeurIPS 2023) is the canonical early example: it tokenizes individual spikes, uses cross-attention with a PerceiverIO backbone to train across recordings from different animals and neurons, and transfers to new sessions. On the encoding side, Wang et al. (Nature, 2025) trained a foundation model of mouse visual cortex that generalized to new mice with minimal training.
In industry, Neuralink reported on October 1, 2026 that self-supervised encoders pretrained on more than 50,000 hours of unlabeled intracortical data kept decoders working for weeks without recalibration and set an 11.32 bits-per-second cursor record. In Neuralink's own words, multi-participant models so far "perform no better than their single-participant counterparts" online. This is a company blog post, not peer-reviewed.
Is every large neural decoder a foundation model?
No. The term is used loosely. Pretrained decoders such as POYO and Neuralink's encoders differ from pretrained encoding or digital-twin models such as Wang et al.'s. Meta's Brain2Qwerty decodes typed sentences from MEG and EEG in 35 healthy volunteers, with a 32% average character error rate for MEG and 19% for the best participants, but it is a task-specific decoder rather than a foundation model.
Why do brain foundation models matter?
They turn session-to-session drift, the classic failure of clinical BCIs, into a domain-adaptation problem. And because the corpus becomes the model, they tie decoding performance to neural-data governance, as we argue in Decoders that generalize.
Where can I read more on NeuroGenesys?
Sources
- A Unified, Scalable Framework for Neural Population Decoding — arXiv:2310.16046 (Azabou et al.), NeurIPS 2023
- Foundation model of neural activity predicts response to new stimulus types — Nature (Wang et al.), 2025
- Pretraining on 50,000 Hours of Unlabeled Brain Data — Neuralink, October 2026 (company blog)
- Brain-to-Text Decoding: A Non-invasive Approach via Typing — arXiv:2502.17480 (Meta AI), 2025
