Researchers at MIT CSAIL have demonstrated a method for a frontier AI model to diagnose and repair the underlying logic of another neural network by reading and writing its weights directly. Using the technique they call Weight-Space Programming, the team used GPT-6 Astra to identify and remove reasoning loops in the Qwen3.5-0.8B model. Astra analyzed the weights to find the signal causing the repetition and edited 131,000 weights in two MLP layers, raising the target model’s MATH-500 score from 7.9% to 48.1% without using traditional gradients or retraining.
The experiments were conducted with a budget of 50 forward passes on practice problems and cost approximately $1,000. Beyond fixing errors, the researchers found Astra could distinguish between networks trained on different tasks, such as Pacman versus Pong, simply by examining the weight matrices. While some industry observers describe the results as early “signs of life” rather than a dramatic shift, the capability suggests a native path for AI to recursively self-improve without the high compute costs of launching new training jobs.
