Glossary Entry

Weight-Space Programming

Using a general-purpose language model to read, edit, or write the weights of another neural network directly from a plain-language goal, without gradient descent or a training loop.

Interpretability LLMs

Also called: weight space programming

Seed source: Sparks of Weight-Space Programming (Takehi et al., 2026)

Weight-space programming treats a network’s weights as something a language model can inspect and rewrite by reasoning, in place of an optimiser. The term was introduced in November 2026 by researchers at MIT CSAIL, who used a frontier model to identify what small networks had been trained to do, to write small game-playing networks from zeros, and to remove a looping fault from a 0.8-billion-weight language model by changing 0.016% of its weights.

Early evidence suggests it works as automated interpretability surgery: it can remove a specific fault that hides an ability a model already has, but it has not been shown to add abilities a model lacks, and results on larger models are small and vary between runs.