Forget retraining - AI can now tweak another AI's weights

Recent research from MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) has demonstrated that frontier AI models like OpenAI's GPT-6 Astra can directly modify the weights of other models, rather than relying on traditional training methods. Lead researcher Rikiya Takehi explained that these models can "read, write, and reason about neural network weights," treating them as editable code.

To showcase this capability, Takehi and his team tasked GPT-6 Astra with enhancing Qwen3.5-0.8B's mathematical abilities. By adjusting only 0.016 percent of the model's weights, Astra improved its math performance by up to sixfold on the MATH-500 benchmark and threefold on the IFEval benchmark for instruction following. This process, dubbed "Weight-Space Programming," does not involve traditional gradient descent training.

Instead, Astra analyzes the functions of each neural network component and adjusts the weights accordingly. The researchers believe this technique could lead to improved model performance, reduced catastrophic forgetting, and enhanced interpretability and alignment. The next step is to establish a benchmark and harness for Weight-Space Programming on LLMs to standardize and accelerate progress in this emerging field.