NOVA: A Verification-Aware Agent Harness That Evolves Industrial Recommender Architectures

arXiv 2606.27243·medium signal

Industrial ad-recommender upgrades like RankMixer, TokenMixer-Large, and MixFormer deliver real gains but are expert-intensive and hard to scale; AutoML only tunes hyperparameters while real wins need cross-module changes under strict constraints. NOVA is a verification-aware agent harness for automating architecture evolution in production recommender systems. It's a window into how large platforms are starting to let agents propose and verify model-architecture changes rather than just configs.