Definition
Plain language
A training-free routing system that decides whether a model should reason out loud or just answer directly.
As stated in the literature
An entropy-dynamics routing method that probes the first 64 tokens of decoding, computes cumulative entropy, trend, and smoothness, and routes between chain-of-thought, Standard, and Direct decoding modes based on hand-tuned or learned thresholds.
Why it matters: It can cut inference cost substantially by reserving expensive reasoning for queries that actually need it, without any fine-tuning.
For example, EDRM watches the first 64 tokens of a model's response, sees high but smoothly falling entropy, and decides this query doesn't need chain-of-thought — short-circuiting to a direct answer.
Heard on the show
“The authors call the system EDRM, but for our purposes the name matters less than the shape.”Episode 077 — Reading a Model's Confidence Curve to Decide When Chain-of-Thought Is Worth It