Program-as-Weights: Compiling Natural-Language Specifications into Compact, Locally-Executable Neural Programs
This paper introduces "fuzzy-function programming," a paradigm for compiling natural-language specifications into compact, locally-executable neural artifacts. The authors present Program-as-Weights (PAW), which uses a 4B compiler trained on a 10M-example dataset (FuzzyBench) to emit parameter-efficient adapters for a frozen, lightweight interpreter. A 0.6B Qwen3 interpreter running PAW programs matches the performance of direct prompting of Qwen3-32B while using roughly 1/50th of the inference memory and running at 30 tokens/s on a MacBook M3. The approach reframes foundation models from per-input problem solvers into tool builders — invoked once per function definition to produce a small reusable artifact for cheap, offline subsequent calls.
Key quotes
We propose fuzzy-function programming: compiling such a function from a natural-language specification into a compact, locally-executable neural artifact.
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/s on a MacBook M3.
PAW reframes the foundation model from a per-input problem solver into a tool builder: invoked once per function definition, it produces a small reusable artifact whose subsequent calls per function application are cheap and offline.
From the article
You might also wanna read
Separating Problem Solving from Code Generation: Evaluating LLMs on Competitive Programming Through Natural-Language Editorials
This paper argues that evaluations of Large Language Models (LLMs) on competitive programming tasks conflate two distinct abilities: algorit
arxiv.org·2d ago
Building Faster Parsers Through Data-Oriented Design: Flat Arrays Over Pointer Trees
This article presents a data-oriented design approach to building high-performance parsers, using the author's experience creating Yuku — a
CoT-PoT Ensembling: Efficient LLM Reasoning with Self-Consistency from Just Two Samples
This paper introduces a hybrid ensembling approach called CoT-PoT that combines Chain-of-Thought (CoT) and Program-of-Thought (PoT) reasonin
arxiv.org·1mo ago
Modular: Modular 25.7: Faster Inference, Safer GPU Programming, and a More Unified Developer Experience
Tensor Logic: A Proposed Unified Language for Neural-Symbolic AI
This paper proposes "tensor logic" as a new programming language for AI that unifies neural and symbolic AI at a fundamental level. It argue
arxiv.org·17h ago
Comments
Sign in to join the conversation.
No comments yet. Be the first.