MIT Scientists Modify Neural Network Weights Directly to Improve Performance

Researchers at MIT CSAIL used GPT-6 Astra to directly alter a tiny fraction of a smaller model's underlying weights. The precise modification significantly increased the smaller system's score on benchmark math tests without undergoing standard retraining.