Understanding the Brain Through AI-Based Explanation and Experimentation

Recently, LLMs have been predicting how the human brain responds to language with very high accuracy. However, the internal structure of models made up of millions of parameters remains a 'black box' that humans cannot understand, which has limited our ability to grasp exactly which concepts specific brain regions respond to.

To address this problem, researchers from Microsoft Research and several universities introduced the GCT (Generative Causal Testing) framework. GCT analyzes brain prediction models and summarizes, in short, readable language, which concepts a given cortical region responds to—such as 'food preparation' or 'place names.'

This framework goes beyond simple explanation to perform experimental verification as well.

• An LLM generates new stories designed to activate a specific brain region.
• Subjects listen to that story while inside a scanner.
• If the generated explanation is correct, the researchers check whether the corresponding brain region actually activates, thereby verifying the causal relationship.

As a result of the experiments, GCT was able to distinguish neighboring regions that had previously been thought to perform the same function, and achieved the feat of newly discovering micro-regions in the prefrontal cortex that respond precisely to specific concepts such as conversation, time, and measurements.