What it is
Compiles natural-language fuzzy function specs into small neural adapters for a frozen local model, turning repeated LLM calls into reusable offline programs. Instead of prompting a large model on every input, compile a spec once into a pseudo-program plus LoRA adapter for a frozen 0.6B interpreter; this makes fuzzy functions reusable, local, and cheap.
Key results
- 73.78% EM on FuzzyBench vs 68.70% for prompted Qwen3-32B
- ~50x lower inference memory: ~1.2 GB bf16 vs ~60 GB for Qwen3-32B
- Runs at 30 tok/s on a MacBook M3 with a 23 MB per-program adapter
Shows many fuzzy application functions can be compiled once into small adapters, letting tiny local models match or beat much larger prompted models.