Automated Architecture Synthesis Via Targeted Evolution

Liquid AI's STAR method searches for neural-network designs against chosen quality and resource requirements. It encodes a model's computational components and their connections as numerical sequences, builds and evaluates candidate networks, then selects and recombines successful designs for another round. The objectives can include model size or memory requirements, as well as measured response latency on target hardware. That makes deployment constraints part of the design search itself: a team can seek an architecture suited to its hardware instead of selecting solely for a quality score. Research account.