MIT lets GPT-6 Astra directly modify Qwen weights

- MIT CSAIL researchers have shown that GPT-6 Astra can directly read and edit the weights of a small language model, improving its math performance without training or gradient descent.
- Modifying just 0.016% of Qwen3.5-0.8B's weights boosted its MATH-500 score from 7.9% to 48.1%, a roughly 6x improvement, and instruction-following ability also improved 3x.
- In a separate test, Astra wrote weights from scratch for a small game network in a single response, achieving a score 20x higher than a PPO baseline trained for 500 million steps.
Source details
- First seen here
- 2 Nov 2026 21:03 UTC
- Outlets
- 2
