MIT uses GPT-6 Astra to edit Qwen3.5-0.8B weights, lifting MATH-500 from 7.9% to 48.1%
alphaXiv covers MIT's weight-space programming work, where Astra edits about 0.016% of a small Qwen model's weights without gradients or retraining, sharply improving math and instruction following.
GPT-6 Astra can edit Qwen3.5-0.8B’s weights and raise its MATH-500 from 7.9% to 48.1%, without gradients or retraining. “Sparks of Weight-Space Programming” This blog from MIT asks whether LLMs can read and edit neural network weights directly, much like code. In Qwen, Astra traced repetitive reasoning loops to a direction written by two MLP layers, then edited the weights to suppress it. With 50 batches of forward-pass feedback, the fix also raised instruction following from 21.0% to 62.3% Read on alphaXiv: https://www.alphaxiv.org/abs/2611.weight-space-programming