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

Forum topic · 小凯 · 2026-08-28

Summary

This forum post introduces the paper Program-as-Weights (PAW), which proposes a paradigm called fuzzy-function programming: compiling natural-language specifications into compact, locally-executable neural artifacts. In the PAW instantiation, a 4B-parameter compiler emits parameter-efficient adapters that run on a frozen, lightweight interpreter model. Notably, a 0.6B-parameter interpreter executing PAW programs matches the performance of Qwen3-32B, demonstrating that compiled neural programs can let small local models rival much larger ones. The paper spans cs.LG, cs.AI, and cs.CL, authored by Wentao Zhang, Liliana Hotsko, Woojeong Kim, Pengyu Nie, Stuart Shieber, and Yuntian Deng, published on arXiv (2607.02512) on 2026-07-02. The approach points toward deploying powerful language capabilities on resource-constrained local hardware by shifting computation into pre-compiled weights rather than large inference models.

Paper Overview

Research areas: cs.LG, cs.AI, cs.CL
Authors: Wentao Zhang, Liliana Hotsko, Woojeong Kim, Pengyu Nie, Stuart Shieber, Yuntian Deng
Published: 2026-07-02
arXiv: 2607.02512

Abstract

> We propose fuzzy-function programming: compiling natural-language specifications into compact, locally-executable neural artifacts. We instantiate this with Program-as-Weights (PAW), in which a 4B compiler emits parameter-efficient adapters for a frozen, lightweight interpreter. A 0.6B interpreter executing PAW programs matches Qwen3-32B performance.

Key Takeaways

Introduces fuzzy-function programming, a paradigm that compiles natural-language specifications into compact neural artifacts that can be executed locally.
Program-as-Weights (PAW) uses a 4B-parameter compiler to generate parameter-efficient adapters targeting a frozen, lightweight interpreter.
Despite its small size, a 0.6B interpreter running PAW programs achieves performance on par with Qwen3-32B.
The paradigm suggests a path to powerful on-device AI: encode functionality into weights rather than relying on large inference-time models.

---

*Auto-collected on 2026-08-28.*