MIT Team Uses GPT-6 Astra to Directly Rewrite Qwen Open Weights, Boosting Math Performance

Original sourceVentureBeat

Summary

A MIT CSAIL research team used OpenAI's GPT-6 Astra to directly modify the weights of the open-source Qwen model, raising MATH-500 scores from 7.9% to 48.1% — roughly six times — while changing only about 0.016% of weights. This method, called weight-space programming, does not rely on traditional…

Key points

Those following post-training and open-weight reuse can learn about an alternative path that does not rely on gradient training, along with its clear limitations.
If weights can be edited directly, the cost structure and methodology of model customization may branch in new directions.
Research teams and model maintainers can explore weight repairs targeting specific capabilities, but there is no commercial deployment yet and practical use still requires validation alongside traditional training methods.

模型後訓練權重編輯

Editorial note

This page is Code & Chain's editorial summary of public sources. It may be prepared with AI assistance and published through an automated workflow. Refer to the original sources; this content is not investment, legal, or tax advice.