Qwen3.5-0.8B

MIT researchers show an AI model can directly edit the weights of another model

MIT CSAIL researchers report that GPT-6 Astra◆ rewrote the weights of Qwen3.5-0.8B◆ without any gradient-descent training, changing only 0.016 percent of them. The researchers say this raised MATH-500 performance by up to sixfold and IFEval instruction-following roughly threefold, and they call the approach Weight-Space Programming. They also say the model wrote a Breakout-playing policy network from scratch.