MIT proposes weight-space programming: GPT-6 Astra directly analyzes and edits Qwen3.5 weights to boost performance

An MIT research team has introduced weight-space programming, allowing the frontier model GPT-6 Astra to directly analyze and modify the weights of other neural networks without gradient-based training. In experiments, Astra identified and edited just 0.016% of the weights in Qwen3.5-0.8B's feedforward networks, resolving reasoning loop issues and boosting its MATH-500 score from 7.9% to 48.1%. In addition, Astra directly generated network weights from scratch for MinAtar benchmark games, achieving a score of 517.4—outperforming a PPO baseline trained for 500 million steps by 20 times.