MIT researchers report GPT-6 Astra rewrites another model's weights directly
MIT researchers report that GPT-6 Astra can alter the internal weights of another language model without a conventional training run. In a test on a Qwen model, math accuracy rose from about 7.9 percent to 48.1 percent.
Researchers at the Massachusetts Institute of Technology report that the AI model GPT-6 Astra can rewrite the internal weights of another model directly. The method skips the conventional training runs through which language models are usually improved. The account comes from a single published report, which links back to The Decoder.
In the central experiment, the researchers targeted a model from the Qwen family, which they aimed to strengthen on mathematical tasks. Its math accuracy rose from about 7.9 percent to 48.1 percent. To achieve this, only about 0.016 percent of the model's weights were changed, a very small intervention relative to the size of the whole model.
According to the report, this approach could simplify how AI models are optimized. It suggests that one model can steer the internal workings of another, and the article frames this as a fundamental shift in how AI systems are developed further.
Open questions remain. It is unclear whether the gains extend beyond mathematical tasks, and unclear how well the results generalize to other model families. The article also states that the approach runs into clear limits, which the summary does not specify.
The excerpt does not address the safety or reproducibility implications of the method. Interventions that reach deep into a model's weights raise obvious questions about traceability and control, but the available material does not say how these were assessed or whether the experiment can be independently repeated.
Source: the-decoder.com
8
