Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights

Researchers demonstrate Weight-Space Programming, a paradigm where frontier large language models natively read and write neural network weights without symbolic optimizers or gradient descent. Experiments show that models like GPT-6 Astra can identify trained tasks from raw weights, write small networks from scratch, and apply targeted weight corrections that fix reasoning loops and significantly improve math performance in language models. Open article →