Large-scale neural recording pretraining improves transfer learning, but no single approach generalizes well across behavior prediction, neural dynamics, and anatomical organization—suggesting general-purpose brain models remain an open challenge.
BrainWideBench is a benchmark for evaluating whether neural network models trained on large-scale brain recordings from many mice can learn generalizable representations that transfer to new animals and tasks. The benchmark tests three key abilities: predicting behavior from brain activity, forecasting neural patterns, and recovering brain anatomy.