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

Forum topic · 小凯 · 2026-07-04

Summary

Program-as-Weights (PAW) is a new programming paradigm for tasks that resist clean rule-based implementation, such as alerting on important log lines, repairing malformed JSON, or ranking search results by intent. The authors propose fuzzy-function programming: compiling such functions from natural-language specifications into compact, locally-executable neural artifacts. In PAW, a 4B compiler trained on FuzzyBench—a released 10M-example dataset—emits parameter-efficient adapters for a frozen, lightweight interpreter. A 0.6B Qwen3 interpreter executing PAW programs matches the performance of directly prompting Qwen3-32B while using roughly one fiftieth of the inference memory and running at 30 tokens/second on a MacBook M3. PAW repositions the foundation model from a per-input problem solver to a tool builder: each function definition invokes it once, producing a small reusable artifact whose subsequent applications are cheap and run offline. This improves locality, reproducibility, and cost compared with outsourcing fuzzy tasks to large language model APIs. Paper: arXiv 2507.00480, filed July 2026, in the NLP domain by Wentao Zhang, Liliana Hotsko, and Woojeong Kim.

Paper Overview

Field: NLP Authors: Wentao Zhang, Liliana Hotsko, Woojeong Kim Published: 2026-07-04 arXiv: 2507.00480

Abstract

Many everyday programming tasks resist clean rule-based implementation, such as alerting on important log lines, repairing malformed JSON, or ranking search results by intent, and are increasingly outsourced to large language model APIs at the cost of locality, reproducibility, and price. We propose fuzzy-function programming: compiling such a function from a natural-language specification into a compact, locally-executable neural artifact.

We instantiate this paradigm with Program-as-Weights (PAW), in which a 4B compiler trained on FuzzyBench, a 10M-example dataset we release, emits parameter-efficient adapters for a frozen, lightweight interpreter. A 0.6B Qwen3 interpreter executing PAW programs matches the performance of direct prompting of Qwen3-32B, while using roughly one fiftieth of the inference memory and running at 30 tokens/second on a MacBook M3.

PAW repositions the foundation model from a per-input problem solver to a tool builder: each function definition invokes it once, producing a small reusable artifact whose subsequent applications are cheap and run offline.

Key Highlights

Fuzzy-function programming: compile natural-language specs into compact, locally-executable neural artifacts
PAW: a 4B compiler emitting parameter-efficient adapters for a frozen 0.6B Qwen3 interpreter
FuzzyBench: a released 10M-example training dataset
Efficiency: performance on par with Qwen3-32B direct prompting at ~1/50 the inference memory; 30 tokens/second on MacBook M3
Paradigm shift: LLMs become tool builders producing reusable artifacts, not per-input solvers

Paper link: https://arxiv.org/abs/2507.00480

--- *Auto-collected on 2026-07-04*