MIT Reports GPT-6 Astra Directly Edits Qwen Weights, Boosting Math Performance ~6x in Self-Tests

Researchers at the Massachusetts Institute of Technology's Computer Science and Artificial Intelligence Laboratory (MIT CSAIL) published a blog post on November 2 introducing a method called "Weight-Space Programming." This approach allows the frontier large language model GPT-6 Astra to directly read and modify the weight parameters of a smaller model, Qwen3.5-0.8B, without using gradient descent or traditional training processes.

According to the blog, Astra analyzed Qwen and found it was prone to repetitive reasoning loops in mathematical tasks. By locating specific output columns in layers 15 and 19 and modifying only 0.016% of the weights, Astra removed this bottleneck. In the MATH-500 benchmark, Qwen's score rose from 7.9% to 48.1%, an increase of approximately six times; instruction-following capabilities also improved threefold.

The study also demonstrated Astra's ability to generate network weights from scratch. In the MinAtar Breakout game, a network generated by Astra in a single response scored 517.4, significantly outperforming a baseline reinforcement learning model trained for 500 million steps (which scored 25.3). However, the authors note that this method is currently primarily used to "remove obstacles hindering capability" rather than granting the model abilities it did not originally possess. For instance, applying the same method to Gemma 3 270M, which rarely exhibits looping issues, did not result in performance gains.

It must be emphasized that all reported results are self-tested data from the research team and have not yet been independently reproduced.