University of Waterloo Releases PAW, Converting English Descriptions into Locally Executable Neural Programs

Researchers at the University of Waterloo have released ProgramAsWeights (PAW), an open-source project that compiles functions described in English text into Neural Programs executable locally (including on CPU).

Separation of Compilation and Inference

The core of PAW is an architecture that separates Compilation and Inference. For fixed tasks, repeated execution is handled by a small model, while a larger model generates task-specific weights. The generated functions can be stored, deployed, and combined with general code.

• Standard Compiler: A finetuned Qwen3-4B generates a LoRA adapter for a frozen Qwen3-0.6B (interpreter).
• Neural Program Composition: Consists of a task-specific LoRA adapter and a cleaned-up task description along with input/output examples (Pseudo-program) generated during compilation.
• Mechanism: Similar to the text-to-LoRA approach, it is trained on (task description, input, output) triples. Gradients flow through the frozen interpreter to the compiler and adapter generation layers.

Performance and Compile by Training

In FuzzyBench evaluations, PAW (0.6B interpreter) recorded 73.4% Exact-match accuracy, surpassing the Baseline (Qwen3-32B direct prompting) at 68.7%.

• Compile by Training: In high-precision mode, 100 steps of finetuning are performed using examples synthesized by a Teacher model, using the generated LoRA adapter as the initial value. This takes approximately 1 minute in the deployment environment.
• FuzzyBench-Hard: Achieved 83.6% Semantic accuracy on a subset where the original evaluation had no exact matches.
• Subsequent inference runs locally without calling the teacher model.

Outlook and Usage

The research team presented a long-term vision where large models act as tool builders, returning small neural programs upon input of functional descriptions for integration into software. Example code can use a hosted compiler, or if you have a GPU, you can host your own compiler and run it locally using the released model weights.