Language agents are AI programs that reason and call tools. They often use , which generate many words at once. But current methods waste computation by treating all words equally and generating beyond tool calls.
The authors created a metacognitive controller that monitors the agent's interaction and the model's confidence. It decides which tokens to finalize early and how many positions to compute. It uses a cascaded policy: one part finalizes predictable tokens quickly, another shrinks the active computation to stop after a tool call.
On several benchmarks, this approach reduced inference FLOPs by 20.6% versus the most efficient baseline while keeping task accuracy similar. This makes diffusion agents faster and cheaper without harming performance.