One model can now predict how any experiment will behave — without retraining for each dataset

Researchers built a single AI model that learns causal relationships from data instead of requiring a separate model for each new dataset. This means you can feed it observational data from a novel experiment and it predicts what will happen if you intervene, without any retraining.

Causal inference—figuring out what causes what from messy real-world data—has always required building a custom model for each specific problem. This is slow, expensive, and often fails when data is scarce. If this approach works at scale, it collapses the cost and time of causal discovery for scientific research. The real test is whether laboratories and companies actually adopt it instead of their existing workflows, and whether the amortized model stays accurate when it encounters data fundamentally different from its training distribution.