Weight-space programming

This page is password protected.

lasal.aiOCR story · weight-space programming
No training run: an LLM edits the weights

Weight-space programming on Qwen3-VL-2B

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.

Fields right, per field
Original 2BAfter the edits4B (same family, more weights)8B

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.

Accuracy after each kept step
Dev (the agent saw these scores)Test (held out, scored afterwards)
Table view

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.