GPT-6 Astra enables model reprogramming by directly reading and writing neural network weights
GPT-6 Astra sees the matrix and reprograms other AI models directly through their weights
Researchers have proposed a weight-space programming method that leverages the Astra language model to directly analyze and modify the raw numerical weights of other neural networks, bypassing traditional iterative optimization training. In experiments on Qwen3.5-0.8B, Astra mitigated reasoning loops by altering just 0.016% of the weights, boosting the model's accuracy on the MATH-500 benchmark from 7.9% to 48.1%. Experimental logs have been published on GitHub, and the modified model has been uploaded to Hugging Face.
This research demonstrates the feasibility of directly reading and modifying neural network weights, offering a new approach for model diagnostics and targeted capability repair without retraining.
Source:The Decoder · the-decoder.com