The paper introduces AgentNAS, a three-phase pipeline that combines open-ended LLM architecture generation with conventional neural architecture search. An LLM first creates a seed architecture, which is decomposed into a “slotted architecture”: a scaffold containing named, interchangeable module slots. Those slots automatically define a task-specific bounded search space for NAS, removing the need to manually engineer one for each task. Across 17 tasks covering classification, dense regression, segmentation, and multi-label tagging in NAS-Bench-360 and Unseen NAS, the authors report state-of-the-art results on 11 tasks. Ablations indicate that LLM design and NAS-based recombination contribute complementary gains.
No heat snapshots are available in the last 24 hours.