Researchers at MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) released a study on November 2nd demonstrating a new method called "Weight-Space Programming": directly reading, understanding, and modifying neural network weights using large models without gradient descent or retraining. In one experiment, GPT-6 Astra analyzed the parameters of Qwen3.5-0.8B and discovered that its recurring mathematical inference problems originated from the MLP modules at layers 15 and 19. Only 131,000 parameters, representing 0.016% of the model's total weights, were modified. After 50 forward computations, the model's MATH-500 score increased from 7.9% to 48.1%, its GSM8K score from 13.8% to 67.1%, and its instruction compliance score from 21% to 62.3%. The research team also demonstrated the model's ability to identify network training tasks solely based on weights and to write game networks from scratch, and released the complete experimental logs and the modified model weights.
