MIT CSAIL researchers on November 2 reported a Weight-Space Programming experiment in which GPT-6 Astra edited model weights directly instead of using gradient descent or fine-tuning. In the research write-up, the team says Astra changed 131,072 weights in Qwen3.5-0.8B, about 0.016% of the model, targeting MLP layers 15 and 19.
With a 50-forward-pass budget, the edited model improved from 7.9% to 48.1% on MATH-500 and from 13.8% to 67.1% on GSM8K, according to the post. The share of MATH-500 answers hitting the 32,768-token limit also fell from 93.0% to 33.1%, which the authors cite as evidence that the edit removed a looping failure mode.
The caveat is central to the claim. As the post puts it, "Fixing loops, however, is not the same as making a model better at math: it lets the model finish answers it could already work out." The authors also report a negative control on Gemma 3 270M, where Astra did not meaningfully improve results.