We show how to "compile" human-readable programs into standard decoder-only transformer models.

General decision intelligence needs to solve tasks in many distinct domains.

CaST, which consistently outperforms existing methods with good interpretability.

X trains an interpretable surrogate to mimic the behavior of a pretrained time series model.

The ability to perform causal reasoning is widely considered a core feature of intelligence.

We shed light on the current progress of GAD, setting a robust groundwork for subsequent investigations in this domain.