Program-as-Weights: A Programming Paradigm for Fuzzy Functions (arXiv 2507.00480)

Forum topic · 小凯 · 2026-07-04

Summary

This arXiv paper (2507.00480) by Wentao Zhang, Liliana Hotsko, and Woojeong Kim proposes fuzzy-function programming: compiling everyday tasks that resist clean rule-based implementation—such as alerting on important log lines, repairing malformed JSON, or ranking search results by intent—directly from natural-language specifications into compact, locally executable neural artifacts. The approach is instantiated as Program-as-Weights (PAW), where a 4B-parameter compiler, trained on FuzzyBench (a newly released 10M-example dataset), emits parameter-efficient adapters for a frozen, lightweight interpreter. A 0.6B Qwen3 interpreter running 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. The paradigm reframes foundation models from per-input problem solvers into tool builders: each function definition is invoked once to produce a small reusable artifact, after which every subsequent function call is cheap and can run fully offline, restoring locality, reproducibility, and low cost compared to LLM API outsourcing.

Paper Overview

Field: NLP Authors: Wentao Zhang, Liliana Hotsko, Woojeong Kim arXiv: 2507.00480

Introduction

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. These tasks are increasingly outsourced to large language model APIs, at the cost of locality, reproducibility, and price.

The Approach: Program-as-Weights (PAW)

The authors propose fuzzy-function programming: compiling such functions from a natural-language specification into a compact, locally-executable neural artifact.

They instantiate this paradigm with Program-as-Weights (PAW):

A 4B-parameter compiler trained on FuzzyBench, a newly 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 direct prompting of Qwen3-32B, while using roughly one-fiftieth of the inference memory.
The setup runs at 30 tokens/second on a MacBook M3.

Key Takeaway

PAW reframes foundation models from per-input problem solvers into tool builders: each function definition is invoked once to produce a small reusable artifact, whose subsequent applications are cheap to call and can run fully offline.

---

*Source: arXiv:2507.00480, collected 2026-07-04.*