Compile once. Run locally.

Many product teams have small, fuzzy text functions that sit awkwardly between brittle rules and an expensive model call on every input. Compile by Training turns the function description into a reusable neural artifact: teacher models synthesize examples once, a LoRA adapter specializes a compact interpreter, and the packaged program handles later inputs through a local SDK. 12

The paper reports 83.6% semantic accuracy on FuzzyBench-Hard, where the fast PAW compiler had no exact matches. The higher score took 50.9 seconds to compile, compared with 3.5 seconds for the fast compiler. The benchmark uses an LLM-judged semantic metric, and the authors warn that synthetic supervision can carry teacher errors. 2

The first product build should stay narrow: choose one repeated task, write real validation cases, compile a compact candidate, then shadow it beside the current route. Measure task accuracy, p95 latency, cost, and failure types before changing traffic. Keep deterministic checks and the existing route live for cases that need guarantees. 2