Turning GLM-5.3-Flash into a Jev-like decision model

The authors demonstrate a method that forces GLM-5.3-Flash to output a decision token on the first step by crafting the prompt, enabling a single forward pass. Benchmarks show the approach matches Jev’s accuracy and speed while outperforming Laya, though Jev remains cheaper per decision. The technique also supports vision inputs.