Reported evidence
Compile by training → enables → Reusable neural functions
Compile by training uses teacher-generated task examples to train small adapters for a compact interpreter, producing neural functions that can run without the teachers and can be stored, versioned and composed. This mechanism does not specify the final deployment location.
Read the evidence
Compile by training provides an explicit mechanism: teacher models first generate task examples from natural-language specifications, then train small adapters for use by a compact interpreter; once completed, the resulting neural functions can run without teacher models [S1]. The products of this method can be stored, versioned and composed [S1].Read the full research