GPT-6 Astra “Fixed” Another Neural Network Without Retraining: MIT Demonstrates Weight-Space Programming
MIT researchers have shown that GPT-6 Astra can find and fix errors in other neural networks by directly modifying their weights without retraining. In an experiment with Qwen3.5-0.8B, accuracy on MATH-500 rose from 7.9% to 48.1%.
Read more →