OmniMouse Brain Model Achieves State-of-the-Art Performance with 150 Billion Neural Tokens

OmniMouse, a brain model capable of multi-modal and multi-task analysis, has set a new benchmark by examining a dataset comprising 3.1 million neurons sourced from the visual cortex of 73 mice. The findings, presented in arXiv preprint 2604.18827, involved over 150 billion neural tokens collected during 323 sessions in which the mice viewed natural movies, images, parametric stimuli, and engaged in various behaviors. Contrary to the typical AI trend where larger models yield better results, this research revealed that increasing model parameters reached a performance plateau. Instead, improvements were linked to the volume of data. The models demonstrated three adaptable testing modes: neural prediction, behavioral decoding, and neural forecasting, either separately or in combination, surpassing specialized baselines in nearly all assessment categories. This study investigates whether the scaling principles that revolutionized AI in language and vision can be applied to brain activity modeling, utilizing extensive neural recordings to explore this idea.