Understanding the brain with AI-driven explanations and experiments
| Source: Microsoft Research Blog
Tags: Microsoft Research, neuroscience, LLMs, fMRI, interpretability, brain modeling, Nature Neuroscience
Microsoft Research and UC Berkeley introduce Generative Causal Testing (GCT), a framework that converts opaque LLM-based brain models into short verbal hypotheses about what each cortical region responds to — then verifies them in live fMRI scans, uncovering micro-regions tuned to dialogue, clock times, and measurements.
Details
LLMs predict how the human brain responds to language with remarkable accuracy, but their internal representations are uninterpretable — they tell us that a brain region responds to language, not what it is responding to. Microsoft Research, UC Berkeley, UCSF, and Columbia University introduce Generative Causal Testing (GCT) to close that gap. Accepted in Nature Neuroscience. GCT works in two stages. First, it distills brain-prediction models into short readable descriptions of what each cortical patch responds to — phrases like "food preparation" or "location names." Second, an LLM generates new stories specifically engineered to activate the targeted region. If the explanation is correct, the region lights up in a follow-up fMRI scan, providing experimental confirmation. In practice, GCT confirmed known selectivity patterns and — more interestingly — distinguished between neighboring place-processing regions previously considered interchangeable. It also revealed tiny prefrontal micro-regions tuned to specific concepts: dialogue, clock times, and measurement units. The framework represents a shift in how LLMs serve science: not just as prediction engines but as hypothesis generators whose outputs can be tested in physical experiments. The loop from model to explanation to experimental verification had not previously been closed at this granularity.