AgentNAS: The LLM Draws the Blueprint, Search Does the Grind
Neural architecture search quietly died of a labor problem: someone had to hand-engineer the search space before the search could find anything, and that someone was an expensive human expert. AgentNAS (arXiv 2607.07984, July 8) redivides the labor. An LLM designs a high-quality seed architecture for the task; the system decomposes that seed into a scaffold with named, interchangeable module slots; classical NAS then grinds through the combinatorial recombinations of those slots. The model supplies the taste, the search supplies the brute force.
The results earn the attention: state of the art on 11 of 17 tasks spanning classification, segmentation, regression and tagging. Two findings stand out. The LLM-generated seed alone already beats published baselines on most tasks — before any searching happens. And the search still adds real gains on top, because recombining across slots explores a space that independently re-sampling an LLM never reaches. The result holds across three LLMs of different capability levels, so it is the recipe that works, not one lucky model. Code is on GitHub.
The bigger thread: another entry in models-design-the-system. Agents already optimize kernels, evolve training recipes, and rewrite their own harnesses; now the hand-crafted search space — the actual bottleneck that killed NAS the first time around — becomes generated output. Every field still stuck on a we-need-an-expert-to-define-the-space problem should read this one.
Paper: https://arxiv.org/abs/2607.07984
Code: https://github.com/alroimfebruary/AgentNAS
← Back to all articles
The results earn the attention: state of the art on 11 of 17 tasks spanning classification, segmentation, regression and tagging. Two findings stand out. The LLM-generated seed alone already beats published baselines on most tasks — before any searching happens. And the search still adds real gains on top, because recombining across slots explores a space that independently re-sampling an LLM never reaches. The result holds across three LLMs of different capability levels, so it is the recipe that works, not one lucky model. Code is on GitHub.
The bigger thread: another entry in models-design-the-system. Agents already optimize kernels, evolve training recipes, and rewrite their own harnesses; now the hand-crafted search space — the actual bottleneck that killed NAS the first time around — becomes generated output. Every field still stuck on a we-need-an-expert-to-define-the-space problem should read this one.
Paper: https://arxiv.org/abs/2607.07984
Code: https://github.com/alroimfebruary/AgentNAS
Comments