The paper proposes fuzzy-function programming, where natural-language specifications are compiled into compact neural artifacts that run locally. Its Program-as-Weights (PAW) system uses a 4B compiler trained on a released 10M-example FuzzyBench dataset to emit parameter-efficient adapters for a frozen lightweight interpreter. The authors report that a 0.6B Qwen3 interpreter executing PAW programs matches direct prompting of Qwen3-32B, while using about one-fiftieth of the inference memory and reaching 30 tokens/s on a MacBook M3. The artifact is generated once per function definition and reused offline.
No heat snapshots are available in the last 24 hours.