Compile by Training: Turning Natural-Language Specifications into Local Neural Functions
A small local adapter is trained from teacher-generated examples, then reused as a standalone text function.
The paper frames this as a way to avoid repeated remote-model calls for recurring text tasks. At compile time, teacher models synthesize task examples for a compact interpreter; after that, the trained function runs without the teachers and can be stored, versioned, and composed. On FuzzyBench-Hard, it reports 83.6% semantic accuracy, with compile time closer to a minute than seconds. The authors also describe a public interactive service and demos including a website helper, a 3D avatar, and an English-Claudish translator. ArXiv · AI/CL/LG's note
The paper frames this as a way to avoid repeated remote-model calls for recurring text tasks. At compile time, teacher models synthesize task examples for a compact interpreter; after that, the trained function runs without the teachers and can be stored, versioned, and composed. On FuzzyBench-Hard, it reports 83.6% semantic accuracy, with compile time closer to a minute than seconds. The authors also describe a public interactive service and demos including a website helper, a 3D avatar, and an English-Claudish translator. ArXiv · AI/CL/LG's note
score 5