Researchers say GPT-6 Astra edited a small Qwen model's weights to improve its math, without training

A post on X from the account rikiyatakehi says its authors asked GPT-6 Astra to edit the weights of Qwen3.5-0.8B to improve its math. They say the model directly changed only 0.016% of the weights, fixing the small model's reasoning loops, improving math up to 6x and instruction following 3x, with no gradients or training. They call the approach Weight-Space Programming.

The authors, from MIT CSAIL, describe it in a blog post published November 2. They report MATH-500 rising from 7.9% to 48.1% and IFEval tripling. In smaller experiments they say Astra named the game a small network was trained on from its raw weights, fixed a failing car-control policy by editing 3.5% of its weights, and wrote small game-playing networks from scratch that beat PPO networks trained for 100 million steps.

The results are the authors' own claim and have not been independently checked. The post had about 9,600 likes and 419 replies on X on November 2.