MIT uses GPT-6 Astra to edit Qwen weights, boosting math up to 6x
MIT researchers had GPT-6 Astra inspect and modify Qwen3.5-0.8B's weights, editing just 0.016% to improve math up to 6x and instruction following up to 3x, and also wrote a game-playing network's weights from scratch, dubbing it Weight-Space Programming.
The level of AI advancement we’ve attained this year has just been crazy. Researchers at MIT just got GPT-6 Astra to directly modify Qwen’s weights & make it up to 6x better at math. And it did this without any training. a Recap: ⟣ They gave Astra access to Qwen3.5-0.8B’s weights & asked it to improve the model’s math ability. ⟣ Astra noticed Qwen kept getting stuck in reasoning loops, then figured out which part of its weights was responsible. ⟣ It edited just 0.016% of the weights, improving math performance by up to 6x & instruction following by up to 3x. They also tested Astra on other neural networks. And in another experiment, it wrote the weights for a game-playing network from scratch, outperforming a RL baseline trained for 500 MILLION steps by 20x. It’s super crazy. The researchers are calling this Weight-Space Programming. Basically, getting an AI to inspect, understand & directly modify another neural network. Still early research, but imagine how far this could go if models become good at finding & fixing problems inside other models. We’re getting into some ridiculous territory here.
This post quotes another post: “GPT-6 Astra edits model weights directly: Weight-Space Programming”