The term foundation model is useful when it points to a concrete capability: a pretrained representation that can be reused across tasks and settings. In neuroscience, however, the label should not substitute for evidence.
Three levels of transfer
1. Task transfer
A shared backbone should support forecasting, behavior decoding, and representation analysis without requiring a separate model family for each objective.
2. Dataset transfer
Adaptation to an independently collected dataset should reuse most of the pretrained model. Frozen-backbone evaluation is particularly informative because it prevents target-only retraining from being mistaken for transfer.
3. Representation transfer
The strongest claim is not merely that predictions improve, but that independently collected neural data can be embedded into a shared space that supports biological comparison.
Scale is not the only criterion
Scaling behavior is important, but early neural foundation-style models often investigate only a subset of the full agenda: cross-session generalization, multi-task reuse, transfer efficiency, or shared representations. A model does not need to demonstrate every possible foundation-model property in one paper; it does need to state clearly which properties are supported.
For neural data, the key question is not simply “How large is the model?” but “What scientific structure survives transfer?”
This note is part of an evolving research notebook. Publication claims should be cited from the linked papers rather than from this informal summary.