Compile by Training turns text specs into local neural functions

A new EMNLP 2026 system-demo paper from Yuntian Deng, Pengyu Nie and Stuart Shieber proposes what they call "compile by training": you write a natural-language description of a small text function, and the system produces a compact model that runs locally, with no further calls to a large teacher. The paper on arXiv is short and unambiguous about the pipeline.

"At compile time, teacher models generate task-specific examples that are used to train a small adapter for a compact interpreter," the authors write. "The resulting function runs without the teachers and can be stored, versioned, and composed like ordinary software."

The headline number is 83.6% semantic accuracy on FuzzyBench-Hard, a subset on which the "Program-as-Weights fast compiler produced no exact matches." The gain comes with a bill. Compile time moves from seconds to roughly a minute.

Alongside the paper the authors have deployed the compiler as a public interactive service, and point to three demonstrations of compiled functions: "a multi-site website helper, a language-controlled 3D avatar, and a bidirectional English-Claudish translator." The abstract lists no per-demo accuracy figures, does not disclose the size of the compact interpreter, and does not name the teacher models used at compile time. A pair of researchers on our tracker list had already picked up the arXiv link.