Neural Architecture Optimization for Low-Latency Large Language Model Inference. JAAIR 2026, 5 (1).