Astra Qwen weights post called misleading
The gist
A critic says MIT CSAIL's post on GPT-6 Astra editing Qwen weights is misleading.
The Qwen test used Python and practice problems, so scope is disputed.
How it unfolded
- Nov 2, 2026MIT CSAIL researchers publish "Sparks of Weight-Space Programming," claiming GPT-6 Astra can read and write neural network weights without gradient descent or a training loop.
- Nov 2, 2026The post reports that editing 131,072 Qwen3.5-0.8B weights (about 0.016%) raised MATH-500 from 7.9% to 48.1%, GSM8K from 13.8% to 67.1% and IFEval from 21.0% to 62.3%.
- Nov 2, 2026Coauthor Rikiya Takehi posts on X that Astra fixed reasoning loops and improved math up to 6x without gradients or training; the approach reportedly failed on Gemma 3 270M, which rarely loops.