Can LLMs redesign themselves as they train?
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
An article on arXiv-style research describes letting coding agents redesign a neural network's architecture during training, by inserting code-written layers and using a "phantom gradient" to find where the model needs help.