Toward a Reasoning Curriculum for Brain-Trained Foundation Models

Maël Donoso

Keywords: human cognition, reasoning, foundation models, neuroimaging, curriculum learning

TL;DR: Neural signals associated with human cognition could be directly leveraged to improve AI reasoning.

Abstract: The processes of human reasoning can be predicted by computational models of human cognition, and have been associated with the neural signals recorded from several brain regions, particularly in the prefrontal cortex. Theoretically, these neural signals could be directly leveraged for training the reasoning capabilities of foundation models, in line with the emerging approach of using neuroimaging data to improve the performance of artificial intelligence systems. Practically, following this new strategy would require the development of a reasoning curriculum, a progression from simple and narrow experimental settings to more complex and open-ended reasoning experiments. This reasoning curriculum could provide valuable insights for brain-trained foundation models, and could eventually be integrated into human-written supervised fine-tuning demonstrations, which are already part of the post-training strategies of foundation models.

Paper Type: Blue Sky Paper

Submission Number: 55