Generative causal testing (GCT), developed by researchers at Microsoft Research, the University of California, Berkeley, the University of California, San Francisco, and Columbia University, provides a workflow that converts high-performing but opaque language-model-based brain predictors into short, testable verbal explanations, and then experimentally verifies them.
The explainability problem
Over recent years, large language models (LLMs) have become the most accurate tools for predicting how the human brain responds to language: if a person hears a story in an fMRI scanner and the same story is fed to an LLM, the model’s internal representations can often predict activity in individual patches of cortex with high fidelity. However, these models consist of millions of unreadable parameters, so a model that predicts brain activity does not itself explain what content a region is actually sensitive to — whether it is food, places, numbers, or something else.
Two steps: explanation and verification
GCT proceeds in two main steps.
-
Explanation: starting from a predictive model for a voxel or region, the method identifies short phrases that most strongly drive the model’s predicted response. An LLM then summarizes those driving words into a concise verbal explanation, often a single phrase such as “food preparation” or “location names.”
-
Verification: to test the explanation causally, the LLM writes new synthetic stories in which each paragraph is designed to activate the region according to the proposed explanation. Three subjects returned to the scanner to hear these stories. If a region’s response to its “driving” paragraphs is significantly greater than to baseline text, the explanation passes a genuine causal test rather than a merely correlational check.
Experimental findings
Across all three subjects, the synthetic GCT stories reliably drove their target regions above baseline, indicating that the short explanations capture content the cortex genuinely responds to. Explanations were most verifiable where the underlying brain-prediction models were strongest: more stable models yielded explanations that could be confirmed more reliably in the scanner.
GCT reproduced established findings — for example, the “Locations” explanation produced strong responses in the place-processing areas: retrosplenial cortex (RSC), parahippocampal place area (PPA), and occipital place area (OPA). Crucially, GCT also resolved longstanding ambiguities among these neighboring place-related regions. Initially, stories designed for one of the three regions often activated the others as well, but by generating differential stimuli (stories crafted to activate one region while leaving neighbors quiet) GCT distinguished them. For instance, RSC showed stronger responses to proper-name location mentions (like “Tokyo” or “Connecticut”) than to general location language.
Beyond known areas, the authors discovered new prefrontal “micro-regions.” Scanning a grid of candidate sites and focusing on the most stable locations, GCT revealed tiny regions that are tuned to highly specific concepts: one selective for dialogue markers (words such as “said” or “told”), one responsive to clock-time mentions (e.g., “one o’clock”), and one selective for numeric measurements (e.g., “50 feet”). These distinctions were not previously targeted by conventional searches; they emerged because the method can propose a hypothesis and immediately test it.
Broader implications
GCT’s significance extends beyond language neuroscience. It demonstrates that an accurate, data-driven model need not be the endpoint of inquiry: such a model can be distilled into readable, experimentally falsifiable theories, and those theories can be checked by generating new experiments on demand. For neuroscience, GCT points toward a faster, hypothesis-rich strategy for cortical mapping in which an AI proposes what a region might encode and a closed-loop experiment confirms or rejects that proposal within a single study. The same generate-and-verify philosophy could apply to other domains where powerful predictive models currently outpace our ability to understand them.
Acknowledgements
The work was a collaboration across Microsoft Research, UC Berkeley (Alex Huth, Bin Yu, Sihang Guo, and Aliyah Hsu), Columbia University (RJ Antonello, co-lead), and UCSF (Shailee Jain). The authors thank the study participants and the broader language-neuroscience community for tools and datasets that made the research possible. The paper, “Generative causal testing to bridge data-driven models and scientific theories in language neuroscience,” has been accepted in Nature Neuroscience, and the related code is available on Github.
This report summarizes the method and experimental findings showing how LLM-guided stimulus generation can convert opaque predictive models into testable neuroscientific explanations.
