AgentNAS: LLM seed + NAS beats baselines on 11 of 17 tasks

AgentNAS, described in the arXiv paper "Agentic Neural Architecture Search" (submitted 8 Jul 2026, arXiv:2607.07984), pairs LLM-driven design with conventional neural architecture search. The pipeline has three modular phases: an LLM generates a high-quality seed architecture, the seed is decomposed into a "slotted architecture" scaffold, and conventional NAS explores the resulting bounded, task-specific search space.

What is AgentNAS?

AgentNAS is a modular three-phase pipeline that combines LLM design and NAS by first producing an LLM-generated seed architecture, then decomposing it into a slotted architecture that automatically defines a bounded search space for conventional NAS. The paper frames the slotted architecture as a scaffold with named, interchangeable module slots, removing the need for manually engineered search spaces for each task.

The authors emphasize that each component of the pipeline can be measured independently. They instantiate the mechanism in AgentNAS to show how responsibilities split between an LLM and a NAS procedure: the LLM provides a strong starting design and NAS performs combinatorial recombination across slots to find additional gains.

How does AgentNAS perform on benchmarks?

On 17 tasks spanning classification, dense regression, segmentation, and multi-label tagging across NAS-Bench-360 and Unseen NAS, AgentNAS establishes a new state of the art on 11 tasks, outperforming published baselines including task-specific expert designs. The paper reports that the LLM-generated seed alone already surpasses published baselines on the majority of tasks, and that NAS produces further improvements in most cases through combinatorial recombination across slots.

The authors evaluated the pipeline across three LLMs of different capability levels and observed the same broad pattern: the division of labor between LLM design and NAS was robust across models. The experiments therefore claim both that LLMs can propose strong architectures in an open-ended space and that conventional NAS still contributes uniquely by searching the bounded slot combinations that independent LLM samples cannot replicate.

Why does this approach matter?

AgentNAS addresses a persistent friction in NAS workflows: the need to handcraft a search space for each new task. By using an LLM to generate a seed and then automatically deriving a slotted, task-specific search space, the pipeline reduces reliance on domain expertise to design search spaces for every task. That matters because it can lower the engineering cost of launching NAS on diverse problems and because the paper shows measurable performance gains: 11 out of 17 tasks reached new best results.

The split also highlights a practical division of labor: LLMs contribute creative, open-ended architectural proposals while NAS delivers structured combinatorial search where it excels. The authors' ablation studies attribute distinct, complementary roles to each component and confirm the pattern across three LLMs.

What to watch

Watch for the authors' code link included in the paper at a "https URL" and for independent reproductions on NAS-Bench-360 and Unseen NAS to verify the reported 11-of-17 state-of-the-art result. Follow-up work that evaluates AgentNAS across additional benchmarks or with more LLM families will test how generally the LLM-plus-NAS division holds.

Paper and record: arXiv:2607.07984 (submitted 8 Jul 2026), DOI https://doi.org/10.48550/arXiv.2607.07984. The authors name Seokhoon Jeong, Mijung Kim, and Taehwan Kim.