Weight-space programming uses a language model to inspect another model and directly change the numerical weights that control its behavior, according to the article. In the reported Qwen experiment, GPT-6 Astra edited about 0.016% of the weights and raised MATH-500 accuracy from 7.9% to 48.1%. The evidence points mainly to fixing repetitive reasoning that prevented answers from finishing, while a Gemma follow-up produced much smaller gains. Fine-tuning usually learns weight updates through gradients; this procedure used a language model to design them, supported by practice data and repeated evaluations. The relationship is between an editor and a target, while teacher-student distillation trains one model to imitate another model’s behavior. Continual learning would require successive edits to add capabilities while preserving earlier ones, and recursive self-improvement would require improved models to become better at producing further improvements. The article states it would treat this as a promising tool for model repair and test its…