Program-as-Weights compiles natural-language specs into compact neural artifacts

Researchers propose fuzzy-function programming to replace large language model APIs for tasks like log alerting and JSON repair, aiming to improve locality, reproducibility, and cost. They introduce Program-as-Weights (PAW), which uses a 4B compiler trained on the released FuzzyBench dataset to generate parameter-efficient adapters for a frozen interpreter.

• The system compiles natural-language specifications into compact, locally-executable neural artifacts.
• A 0.6B Qwen3 interpreter executing PAW programs matches the performance of direct prompting with Qwen3-32B.
• This 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, allowing users to invoke the compiler once per function definition to produce small, reusable artifacts for cheap, offline subsequent calls.