Agentic Neural Architecture Search
06:00 · July 11, 2026 · arXiv cs.AI RSS
Neural architecture search (NAS) methods have grown increasingly efficient, yet they remain bounded by manually engineered search spaces that require substantial domain expertise and must be rebuilt for every new task. Large language models (LLMs) can generate architectures in an open-ended space, but how to optimally divide the labor between LLM-driven design and NAS-driven search remains unexplored. We propose a mechanism that bridges these two paradigms: an LLM produces a high-quality seed architecture, then decomposes it into a "slotted architecture", a scaffold with named, interchangeable module slots that automatically defines a bounded, task-specific search space for conventional NAS to explore, without manual engineering. We instantiate this mechanism in AgentNAS, a modular three-phase pipeline in which each component's contribution can be measured independently. On 17 tasks spanning classification, dense regression, segmentation, and multi-label tagging across diverse modalities (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. Ablation studies show that the two search mechanisms are broadly complementary: the LLM-generated seed already surpasses published baselines on the majority of tasks, and NAS delivers additional gains in most cases through combinatorial recombination across slots, a mode of search that independent LLM samples cannot replicate. These patterns hold across three LLMs of different capability levels, confirming that the division of labor is robust. Our code is available at https://github.com/alroimfebruary/AgentNAS.
Summary
Neural architecture search has long been constrained by the need for manually designed search spaces that embed domain expertise and must be rebuilt for each new task. AgentNAS addresses this limitation through a hybrid pipeline that lets a large language model first propose a seed architecture and then decompose it into a slotted scaffold. The scaffold consists of named, interchangeable module slots whose alternatives are automatically enumerated, thereby creating a bounded yet task-specific search space that conventional NAS algorithms can explore without further human engineering.
The resulting three-phase AgentNAS system separates the contributions of the LLM and the NAS stage so each can be measured independently. In the first phase the LLM iteratively proposes, implements, and evaluates candidate networks until performance saturates. The second phase converts the best seed into the slotted form, exposing combinatorial degrees of freedom at the module level while preserving the LLM’s macro-level decisions on depth, width progression, and backbone type. A standard NAS procedure then recombines the slot alternatives in the third phase.
Evaluated across the 17 tasks of NAS-Bench-360 and Unseen NAS—covering image classification, dense regression, segmentation, and multi-label tagging in multiple modalities—AgentNAS reaches state-of-the-art accuracy on 11 tasks and surpasses both published baselines and task-specific expert designs. Ablation experiments conducted with three LLMs of varying capability show that the LLM-generated seed already exceeds most baselines on the majority of tasks, while the subsequent NAS recombination step supplies additional gains that independent LLM sampling at matched compute budgets cannot replicate. The observed complementarity between the two mechanisms remains consistent across model scales.
Why it matters
High technical depth, novelty in bridging open-ended LLM generation with combinatorial NAS, full reproducibility via public code, and direct applicability for Dutch researchers advancing AutoML and agentic systems.
