This is a tremendously exciting research direction. The authors call it “weight-space programming.”
A team at MIT CSAIL asked GPT-6 Astra to inspect neural network weights, reason about how the models behave, and change those weights directly, without retraining or gradient
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming:






