Researchers have developed Generative Causal Testing (GCT), a framework that translates uninterpretable LLM-based brain-prediction models into concise, testable verbal hypotheses about cortical function. This method distills model parameters into short phrases describing what specific brain regions respond to, such as "food preparation," and then verifies these explanations through targeted fMRI experiments.
- GCT identifies the phrases driving a predictive model for a brain region and summarizes them into a concise explanation using an LLM.
- An LLM generates synthetic stories engineered to activate that specific region based on the generated explanation.
- Subjects listen to these stories in an fMRI scanner, allowing researchers to confirm if the targeted area responds significantly above baseline.
- The approach confirmed known selectivity, distinguished neighboring place-processing regions previously thought interchangeable, and identified prefrontal micro-regions tuned to concepts like dialogue and clock times.
This method bridges the gap between high-accuracy black-box predictions and scientific understanding by providing hypotheses that can be directly confirmed or refuted in follow-up experiments.