Definition
Plain language
Deciding per-query whether to use chain-of-thought reasoning or a direct answer.
As stated in the literature
In EDRM, the per-query decision pattern using a 64-token entropy probe to select between Direct, Standard, and CoT decoding modes for each individual input.
Why it matters: It saves compute on easy queries while preserving reasoning quality on hard ones, instead of forcing every query through the same expensive path.
For example, a quick entropy probe might decide that 'what is 2+2' goes straight to a direct answer while a multi-step word problem gets routed to a full chain-of-thought.