Neural architecture search has long been constrained by hand-designed search spaces. A revised paper on arXiv argues that this limitation holds NAS back, and notes that recent LLM-assisted methods can open up the search space. The abstract says those methods "often suffer" from some issue, though the provided excerpt cuts off before specifying it.

The paper, titled "Structuring Open-Ended NAS," appears to address that gap by using LLMs to semi-automatically structure design knowledge. The framing suggests a middle path: keeping the flexibility of open-ended search while making the search itself more efficient through structured design knowledge. The provided abstract does not include experimental results or further implementation details.