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

Summary

This paper introduces fuzzy-function programming, a new paradigm that compiles natural-language specifications into compact, locally-executable neural artifacts. The authors instantiate the idea with Program-as-Weights (PAW), where a 4B-parameter compiler emits parameter-efficient adapters for a frozen, lightweight interpreter model. Notably, a 0.6B-parameter interpreter executing PAW programs matches the performance of Qwen3-32B, demonstrating that natural-language programs can be encoded directly into neural weights rather than token sequences. The work spans machine learning (cs.LG), artificial intelligence (cs.AI), and computational linguistics (cs.CL), and is authored by Wentao Zhang, Liliana Hotsko, Woojeong Kim, Pengyu Nie, Stuart Shieber, and Yuntian Deng. It was released on arXiv as 2607.02512 in July 2026. The approach points toward efficient local execution of language-defined functionality with small models.