Program-as-Weights: A Programming Paradigm for Fuzzy Functions

Program-as-Weights (PAW), from researchers at the University of Waterloo, Cornell University, and Harvard University, reframes fuzzy software tasks as compile-once, run-locally neural programs. A natural-language specification is first converted into a discrete pseudo-program by an off-the-shelf Qwen3-4B-Instruct-2507 compiler, then into a per-function PEFT artifact by a trained 4B Qwen3 LoRA compiler; the resulting program is executed by a frozen Qwen3 0.6B interpreter. Trained on FuzzyBench, a 10M-example dataset spanning 29 thematic versions and 800+ task categories, PAW’s default LoRA instantiation reaches 65.7% exact match with rank 64, versus 50.4% for prefix-tuning and 9.8% for direct prompting. On the verified FuzzyBench test set, the Qwen3 0.6B PAW system achieves 73.78% exact match, surpassing direct prompting of Qwen3-32B at 68.70% while using about 50× less inference memory, and it runs at 30 tokens/s on a MacBook M3 with a ∼430 MB shared base plus a 23 MB per-program adapter. The paper also shows a multimodal extension using Qwen3-VL-4B as compiler, where PAW LoRA reaches 0.274, 0.414, and 0.552 on Circuit, Chemical, and Music diagram tasks, respectively, and reports only minor degradation under heavy noisy-specification perturbations because the pseudo-program denoises the input before execution.