# ProgramAsWeights research

Program-as-Weights compiles English function descriptions into neural programs that run locally.

- [Research](https://programasweights.com/research/): PAW and Compile by Training.
- [Weight-Space Programming: Methods, Papers, and History](https://programasweights.com/research/weight-space-programming/): task descriptions, generated weights, and reusable neural functions.
- [Weight-Space Programming Reading List](https://programasweights.com/reading-list/weight-space-programming/): PAW, generated adapters, formal compilers, direct weight editing, dates, and original citations.
- [Two-stage AI decision models](https://programasweights.com/ai-decision-models): build a decision function once, then run fast local inference; System 1 and System One workflows.
- [Generating LoRA weights from task descriptions with PAW](https://programasweights.com/research/hypernetworks-and-lora-generation/): generate program weights from task descriptions and reuse them with a small interpreter.
- [Teacher-generated examples for small models: PAW](https://programasweights.com/research/teacher-generated-training-data/): FuzzyBench data and training a compiler to build reusable functions.
- [Benchmarks for natural-language function compilation: PAW](https://programasweights.com/research/neural-compilation-benchmarks/): PAW's dataset, held-out functions, exact-match evaluation, and local execution.
- [AI as tool builder](https://programasweights.com/research/ai-as-tool-builder/): pretraining AI to produce reusable neural programs. Yuntian Deng’s essay, [From AI as Problem Solver to AI as Tool Builder](https://x.com/yuntiandeng/article/2063006022344536306), June 5, 2026.
- [Reading list](https://programasweights.com/reading-list/): neural compilation and fuzzy functions.
- [All papers](https://programasweights.com/reading-list/papers/)
- [中文研究介绍](https://programasweights.com/zh/research/): PAW 的模糊函数编程、编译器与解释器，以及本地神经程序。
- [中文阅读清单](https://programasweights.com/zh/reading-list/): 超网络、适配器生成、上下文压缩与可复用神经程序。
- [中文 Markdown](https://programasweights.com/zh/reading-list/index.md)
- [Fuzzy functions](https://programasweights.com/reading-list/guides/fuzzy-functions/): how to define, save, and call a neural program.
- [Compilation methods](https://programasweights.com/reading-list/guides/compilation/): generated weights, training, and context compression.
- [Evaluation](https://programasweights.com/reading-list/guides/evaluation/): accuracy, compile time, and local execution.
- [Markdown](https://programasweights.com/reading-list/data/reading-list.md)
- [JSON](https://programasweights.com/reading-list/data/papers.json)
- [BibTeX](https://programasweights.com/reading-list/data/papers.bib)
- [CSL-JSON](https://programasweights.com/reading-list/data/papers.csl.json)
- [URL directory](https://programasweights.com/reading-list/data/navigation.json)

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

Wentao Zhang, Liliana Hotsko, Woojeong Kim, Pengyu Nie, Stuart Shieber, and Yuntian Deng. NeurIPS 2026. First arXiv submission: July 2, 2026.

PAW introduces fuzzy-function programming: describe a function, compile it into a neural program, and call it locally. A compiler generates the program; a small interpreter executes it. A PAW program combines a text pseudo-program with a LoRA adapter or prefix vectors.

- [Paper on arXiv](https://arxiv.org/abs/2607.02512)
- [Original preprint](https://arxiv.org/html/2607.02512v1)
- [Version history](https://arxiv.org/abs/2607.02512)
- [Summary and citations](https://programasweights.com/reading-list/papers/2607.02512/)

## Compile by Training: Turning Natural-Language Specifications into Local Neural Functions

Yuntian Deng, Pengyu Nie, and Stuart Shieber. EMNLP 2026 System Demonstrations. First arXiv submission: September 3, 2026.

Compile by Training generates examples from a function description and trains a program on them, starting from the fast PAW compiler’s output. The saved program uses the same runtime and runs locally without further teacher calls. On FuzzyBench-Hard, it achieves 83.6% LLM-judged semantic accuracy (LEM), compared with 22.4% for the fast compiler.

- [Published paper · ACL Anthology](https://aclanthology.org/2026.emnlp-demos.78/)
- DOI: 10.18653/2026.emnlp-demos.78
- [arXiv preprint](https://arxiv.org/abs/2609.04199)
- [Original preprint](https://arxiv.org/html/2609.04199v1)
- [Summary and citations](https://programasweights.com/reading-list/papers/2609.04199/)
- [Code](https://github.com/programasweights/compile-by-training)

Each paper is available as HTML, Markdown, and JSON, with BibTeX and CSL-JSON citations.