Claude read the weights of the small model that pulls 11 fields out of invoices and purchase orders, found what got in its way, and edited the numbers directly. It used no optimiser and no labelled training. Every edit is shown here: what it changed in the architecture and which fields it fixed or broke. The test documents use layouts Claude never saw.
A trained neural network is a program written in numbers. Training writes that program indirectly: show it millions of examples and let gradient descent nudge every number a little. Weight-space programming treats the numbers as code someone can read and edit. An LLM (here Claude) looks at the weights and at how the network behaves, works out what a few specific numbers are doing wrong, and changes those numbers directly. It uses no gradients and no new data, and it gets only a small budget of test runs to check its work.
| Training | Weight-space programming | |
|---|---|---|
| Who changes the numbers | an optimiser, following gradients | an LLM, deciding on purpose |
| How many numbers change | all of them, a little | a few, by a lot |
| Why each change was made | no stated reason | written down, one hypothesis per edit |
| What it can do | teach new abilities | remove obstacles; it cannot add an ability the network lacks |
This page has two demos: Qwen3-VL-2B reading invoices (2.13 billion numbers; the first four tabs) and small networks playing a game (3 to 401 numbers; on its own page ↗), where every edit can be seen working.
Each row is one field, out of the number of documents. Hover a bar for exact counts. The 4B and 8B are the same model family, run in the SOTA IDP with the same prompt.
Method from Takehi, Teoh, Zhang and Khattab, “Weight-Space Programming” (2026). Documents, prompt and scoring from the SOTA IDP / Legacy IDP stops of this series. Wingbeat is an original game made for this page; the idea of small networks learning a one-button flying game comes from the open-source neuroevolution demos.