"Compile by Training: Turning Natural-Language Specifications into Local Neural Functions" by Yuntian Deng , Pengyu Nie , Stuart Shieber
TLDR:
The text discusses "compile by training," a method that converts natural-language specifications into reusable neural functions by distilling teacher-generated examples into small adapters for efficient deployment. This process eliminates the need for repeated calls to a large remote model, reducing costs and latency. The resulting neural function can be independently stored and used like regular software. Despite a longer compile time compared to other methods, "compile by training" achieves a high semantic accuracy of 83.6% on the FuzzyBench-Hard subset. The compiler has been successfully deployed in various applications like a multi-site website helper, a language-controlled 3D avatar, and an English-Claudish translator.
Read Paper / Blog
huggingface.co
Paper page - Compile by Training: Turning Natural-Language Specifications into Local Neural Functions
Join the discussion on this paper page
1September 5, 2026 62
