Liverpool-Huawei Hybrid Decision Model Cuts JevBench Latency 72.5% by Routing Between Zero-or-T Thinking Modes

Researchers from University of Liverpool, UCL and Huawei Noah's Ark Lab propose a Hybrid Decision Model that routes between implicit one-forward-pass answers and full autoregressive reasoning via a lightweight self-routing block estimating the 'value of thinking.' On JevBench, HDM-4B scores 0.857 vs Jev's 0.851 at a 0.179s median versus 0.652s — a 72.5% latency cut. The paper introduces Calibrated Decision Policy Optimisation for episodic tasks and shows HDM-27B cuts MiniWoB token cost 80.1%.