Program-as-Weights compiles fuzzy functions into compact neural artifacts

Researchers propose fuzzy-function programming to compile natural-language specifications into locally-executable neural artifacts, addressing the locality, reproducibility, and cost issues of using large language model APIs for everyday tasks. They instantiate this paradigm with Program-as-Weights (PAW), which uses a 4B compiler trained on their released 10M-example FuzzyBench dataset to emit parameter-efficient adapters for a frozen interpreter.

• A 0.6B Qwen3 interpreter executing PAW programs matches the performance of direct prompting of Qwen3-32B.
• The approach uses roughly one fiftieth of the inference memory and runs at 30 tokens/s on a MacBook M3.
• PAW reframes foundation models as tool builders, producing small reusable artifacts for cheap, offline subsequent calls.

This method allows developers to create compact, locally-executable functions that avoid the high costs and latency associated with per-input API calls.