Compile by Training: Turning Natural-Language Specifications into Local Neural Functions
A research paper introduces 'compile by training', which converts natural-language specifications into reusable neural functions by using teacher models to generate examples and training a small adapter for a compact interpreter. On FuzzyBench-Hard, it achieves 83.6% semantic accuracy, with compile time of roughly a minute. The compiler is deployed in a public interactive service and demonstrated in a multi-site website helper, a language-controlled 3D avatar, and a bidirectional English-Claudish translator.
The paper presents a method to compile natural-language specifications into local neural functions, eliminating the need for repeated calls to large remote models. Teacher models generate task-specific examples at compile time, which 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, where the Program-as-Weights fast compiler produced no exact matches, compile by training reaches 83.6% semantic accuracy, but with higher compile-time cost (about a minute vs seconds). The compiler is publicly available as an interactive service, with demonstrations including a multi-site website helper, a language-controlled 3D avatar, and a bidirectional English-Claudish translator.
The approach shifts from rule-based or direct weight programming to training a small adapter on teacher-generated examples, enabling higher semantic accuracy on hard fuzzy tasks at the cost of longer compile time. The use of a compact interpreter suggests a modular architecture where the adapter modifies interpreter behavior, allowing composition and versioning of neural functions.
This method could reduce dependency on large remote models for recurring text functions, lowering per-call cost and latency. The public interactive service and diverse demos indicate early productization interest, potentially targeting developers who need customizable local AI functions.
By enabling local, reusable neural functions from natural-language specs, the technology offers cost savings on API calls, reduced latency, and greater control over AI behavior, which could appeal to businesses needing custom text processing without vendor lock-in.
Observable next signals include adoption metrics for the public service, expansion of FuzzyBench-Hard results, and further demos in consumer or enterprise contexts. Improvements in compile-time efficiency or adapter size could broaden applicability.