Large Language Model-Guided NAS
NAS: Neural Architecture Search
Overview
This field studies how language models can design neural network architectures for a task, rather than relying only on manually designed models or fixed search spaces. Common approaches have LLMs generate or edit model code, then use execution feedback, evolutionary search, memory of prior trials, or conventional NAS to select and refine candidates. Systems are being tested across vision, time-series forecasting, reinforcement learning, scientific modeling, and specialized settings such as clinical prediction, often using proxy training or weight sharing to reduce evaluation costs. A central challenge is making generated designs reliably executable and ensuring that proxy scores and benchmark gains translate to strong results under full training and real deployment constraints such as latency and energy use.
Latest papers
24 papers3 papers in the last four weeks, against 2 the four weeks before. 0.0% of all new papers.