AI NewsMIT Researchers Show LLMs Can Read and Edit Neural Network Weights Without Retraining

MIT Researchers Show LLMs Can Read and Edit Neural Network Weights Without Retraining

12:18 PM IST · November 3, 2026

By modifying just 0.016% of the model’s weights, Astra increased its score on the MATH-500 benchmark from 7.9% to 48.1%.