Skip to content
conftrace_
2026 EMNLP EMNLP 2026

Compile by Training: Turning Natural-Language Specifications into Local Neural Functions

AbstractMany recurring text functions are easy to describe but difficult to implement with rules, while calling a large remote model for every input introduces repeated cost, latency, and dependency on a provider. We present compile by training, which turns a natural-language specification into a reusable neural function. At compile time, teacher models generate task-specific examples that are used to train a small adapter for a compact interpreter. The resulting function runs without the teachers and can be stored, versioned, and composed like ordinary software. On FuzzyBench-Focus, compile by training reaches 83.6% accuracy, compared with 22.4% for the previous Program-as-Weights compiler, which predicts function weights in a single forward pass. A representative cold compile takes 50.9 seconds on a B300 GPU. We deploy the compiler in a public interactive service and demonstrate compiled functions in a multi-site website helper and a language-controlled 3D avatar.

More from EMNLP 2026

All EMNLP 2026 papers →