Compiling from Words

ProgramAsWeights(PAW) takes a plain English function description as input and compiles it into a neural program that combines a LoRA adapter + discrete pseudo-program.

The key idea is adapting a fixed interpreter at inference time. Task-specific weights are not trained directly into the interpreter; instead, the compiler exports the spec as a program in a single pass.

• interpreter: Qwen3 0.6B or GPT-2 124M
• compiler: a finetuned Qwen3-4B that takes a spec as input and outputs LoRA weights + pseudo-program
• training data: 10 million (description, input, output) triples synthesized with GPT-5.2
• compile time: the program is generated in a single forward pass without gradient descent

As a representative example, a fuzzy function that guesses a secret word from hints was implemented and released as the browser game Alien Taboo.

On the FuzzyBench benchmark, performance was as follows.

• PAW + Qwen3 0.6B interpreter: 73.4%
• Qwen3 0.6B raw prompting: 9.8%
• Qwen3 32B raw prompting: 68.7%

In other words, the 0.6B model with PAW applied achieved higher accuracy than 32B raw prompting, which is 50x larger. In addition, the GPT-2 interpreter combination is small enough to run in the browser via WebAssembly, at around 134MB base + about 5MB per program, demonstrating the potential for local execution.