MIT: GPT-6 Astra can directly modify Qwen weights, math capability surges 6x

What happened: MIT researchers reported that GPT-6 Astra edited the weights of a 0.8B Qwen3.5 model directly, without training or gradient descent, under a method they call Weight-Space Programming. Key numbers: 0.016% of weights changed; MATH-500 rose from 7.9% to 48.1%; instruction following improved 3x; length-limit hits fell from 93.0% to 33.1%; a from-scratch Why it matters: The work suggests a general frontier model can diagnose and rewrite small model weights from natural language goals, which the researchers frame as an alternative to retraining when a small model underperforms.

MIT researchers reported that GPT-6 Astra edited the weights of a 0.8B Qwen3.5 model directly, without training or gradient descent, under a method they call Weight-Space Programming.

The work suggests a general frontier model can diagnose and rewrite small model weights from natural language goals, which the researchers frame as an alternative to retraining when a small model underperforms.