Researchers have taught GPT-6 Astra to directly edit the weights of other models

Researchers from MIT have introduced a method that allows the GPT-6 Astra model to directly modify the weights of other neural networks without the need for a traditional fine-tuning process.

During the experiment, the Astra model successfully optimized the performance of Qwen, increasing its accuracy in solving mathematical problems from 7.9% to 48.1%. To achieve this result, only 0.016% of the total weights of the target model needed to be changed. This approach allows for precise adjustments to neural network behavior, bypassing the standard computational costs of training.

Despite its high efficiency, the technology has clear limitations and is not a universal solution for all types of architectures. The method demonstrates the possibility of direct intervention in the structure of neural networks, which could simplify model optimization and configuration processes in the future.