tmlr
A 50K-turn voice pipeline benchmark and an 85K-param meta-controller that cuts P95 latency 10.3% and energy 71% vs fixed cloud. TMLR 2026.