Weight-space programming: can AI repair a model by editing its weights?
What if an AI could fix a neural network by editing the numbers inside it?
Weights are the numerical settings a model learns during training. Rikiya Takehi and colleagues at MIT CSAIL asked GPT-6 Astra to inspect or change them directly.
One vivid experiment: a simulated car kept getting stuck on a ramp. Given its controller’s weights and one recorded failure, Astra changed 3.5% of the weights in one attempt. The authors report 1,024 successful jumps in 1,024 tests.
The larger test: with Python and 50 batches of practice evaluations, Astra edited 0.016% of Qwen3.5-0.8B’s weights without gradient-based training. MATH-500 accuracy rose from 7.9% to 48.1% under the study’s sampling settings.
The crucial limit: much of that improvement came from stopping repetitive reasoning loops—a known issue with this Qwen model. A follow-up on Gemma, which rarely looped, found no established improvement. This is evidence for targeted repair; adding new knowledge remains unproven.
The include prompts, edits and evaluations. Promising early research, with independent reproduction still needed.