The arXiv paper introduces "compile by training," a strategy for meeting natural-language text specifications by first generating task-specific examples with teacher models and then using those examples to train a small adapter on a compact interpreter. The resulting neural function is local: it can execute without the teachers and can be stored, versioned, and composed like ordinary software. The genuinely new move is to push expensive model work out of per-input execution and into a discrete offline compile phase.1
Context
Many recurring text functions are easy to describe but hard to implement with rules. The motivating constraint, according to the abstract, is that calling a large remote model for every input imposes repeated cost, latency, and provider dependence. Compile-by-training relaxes that constraint by replacing repeated remote calls with local execution after a one-time training step.1
Assessment
On FuzzyBench-Hard, the subset where the Program-as-Weights fast compiler produced no exact matches, compile-by-training reaches 83.6% semantic accuracy. The reported cost is a compile phase of roughly one minute versus seconds for the fast compiler; after compilation, teachers are absent from the run path. This is an accuracy-for-compile-time trade, not a universal superiority claim over prior compilers.1
Caveats
The visible evidence is only an abstract: one benchmark number, three demos, and no protocol details about adapter size, teacher prompting, example count, or evaluation method. To change my assessment decisively, the method needs to be tested across task families with independent evaluation and regular failure reporting. A failure on a public suite independent of the authors would substantially discount the headline 83.6% figure.1
