Compile English function descriptions into 22 MB neural programs that run locally [P]

About this article

We built a system, ProgramAsWeights (PAW), where a neural compiler takes a plain-English function description and produces a "neural program" (a combination of a continuous LoRA adapter and a discrete pseudo-program). At inference time, these adapt a fixed interpreter to perform the specified task. The target use case is fuzzy functions: tasks that are easy to describe in language but painful to implement with rigid rules. A concrete fuzzy function: word-guessing Consider the function "given ...