اختبارات السبب الإيجابي يغير نموذج الدماغ البلاستيكي إلى نظرية قابلة للتحقق

LLM-based models predict fMRI brain responses to language with stunning accuracy, but those predictions are locked inside millions of inscrutable parameters - no one can read them. A collaboration between Microsoft Research, UC Berkeley, UCSF, and Columbia University just published a method that breaks that lock: generative causal testing (GCT), accepted in Nature Neuroscience.

Distilling a Neural Net Into "Food Preparation"

GCT works in two steps. First, it takes a predictive model for a single voxel or region and finds the short phrases that most strongly drive its predicted response. An LLM summarizes those words into a concise verbal explanation - often a single phrase like "food preparation" or "location names." Second, an LLM writes new stories engineered to activate that specific brain area. Three subjects heard those synthetic stories in an fMRI scanner. If the targeted region lit up more than baseline text, the explanation passed a causal test, not just a correlational one.

Across all three subjects, the method held: synthetic stories drove their target regions above baseline. The more stable the underlying brain-prediction model, the more reliably its explanation could be confirmed. That closed-loop verification is the key difference from prior correlation-only approaches.

Teasing Apart Place Regions and Finding Micro-Regions

GCT proved sharp enough to settle long-standing ambiguities. Three neighboring place-processing regions - retrosplenial cortex (RSC), parahippocampal place area (PPA), and occipital place area (OPA) - have often been treated as functionally interchangeable. By generating differential stimuli (stories designed to switch one region on while keeping its neighbors quiet), GCT teased them apart. RSC, for example, responds more strongly to proper noun location names like "Tokyo" or "Connecticut" rather than general location. That nuance is invisible in a raw predictive model.

Beyond known regions, the team discovered new prefrontal micro-regions. By scanning a grid of candidate locations and keeping only the most stable ones, GCT surfaced areas tuned to remarkably specific concepts: one selective for dialogue between people (words like "said" or "told"), one for clock times ("one o'clock"), and one for numeric measurements ("50 feet"). No one had gone looking for these distinctions; GCT proposed a hypothesis and immediately tested it.

Why This Matters Beyond Neuroscience

The same dilemma appears everywhere: a model that predicts beautifully but explains nothing. GCT shows that a data-driven model need not be the end of inquiry - it can be distilled into a readable, experimentally testable theory, and that theory can be checked against reality by generating new experiments on demand. The generate-and-verify philosophy could extend to any domain where predictive models have outrun our ability to understand them.

For neuroscience, GCT points toward a faster, more hypothesis-rich way of mapping cortex - one where an AI system proposes what a region encodes and a closed-loop experiment confirms or rejects it within a single study. The paper and code are available now.

Source: Understanding the brain with AI-driven explanations and experiments
Domain: microsoft.com