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

Forum topic · 小凯 · 2026-09-05

Summary

A new paper by Yuntian Deng, Pengyu Nie, and Stuart Shieber (arXiv:2609.04199) introduces "compile by training," a method that converts natural-language specifications into reusable neural functions. At compile time, teacher models generate task-specific examples used to train a small adapter on a compact interpreter. The resulting function runs locally without the teacher models and can be stored, versioned, and composed like ordinary software—eliminating repeated costs, latency, and provider dependence of calling large remote models per input. On FuzzyBench-Hard, where a fast Program-as-Weights compiler produced no exact matches, compile by training achieved 83.6% semantic accuracy. Higher accuracy comes at higher compile-time cost: roughly one minute versus seconds for the fast compiler. The compiler is deployed in a public interactive service, with compiled functions demonstrated in a multi-site website assistant, a language-controlled 3D avatar, and a bidirectional English–Claudish translator.

Paper Overview

Field: NLP Authors: Yuntian Deng, Pengyu Nie, Stuart Shieber Published: 2026-09-03 arXiv: 2609.04199

Abstract

Many recurring text functionalities are easy to describe in natural language but hard to implement with rules, while calling a large remote model for every input brings repeated cost, latency, and provider dependence. The authors propose 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.

Results

On FuzzyBench-Hard—a subset where the fast Program-as-Weights compiler failed to produce any exact matches—compile by training reaches 83.6% semantic accuracy.
Higher accuracy comes at higher compile-time cost: about one minute, versus seconds for the fast compiler.

Deployment

The compiler is deployed in a public interactive service, with compiled functions demonstrated in:

A multi-site website assistant
A language-controlled 3D avatar
A bidirectional English–Claudish translator

Original Abstract

> 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-Hard, compile by training reaches 83.6% semantic accuracy.

*Auto-collected on 2026-09-05*