- Introduces fuzzy-function programming: compiling natural-language specifications into compact neural artifacts for tasks that resist rule-based solutions.
- Presents Program-as-Weights (PAW): a 4B compiler trained on FuzzyBench dataset emits parameter-efficient adapters for a frozen lightweight interpreter.
- Achieves performance matching Qwen3-32B with a 0.6B Qwen3 interpreter, using 1/50th inference memory and running at 30 tokens/s on MacBook M3.
- Reframes foundation models as tool builders: invoked once per function to produce reusable, cheap, offline-executable artifacts.