Definition
Plain language
A penalty applied when a model writes out a plan but doesn't follow it.
As stated in the literature
In MaR's regulation reward, a multiplicative penalty (e.g., 0.7x) applied when a binary shortcut flag fires, indicating the model produced a plan but jumped to the answer without executing it.
Why it matters: Without such a penalty, models learn to produce plan-shaped text as decoration while still answering from intuition, undermining the whole point of training them to reason.
For example, if the model writes "step 1, step 2, step 3" but then skips straight to the final answer without actually executing the steps, its reward is multiplied by 0.7.
Heard on the show
“The shortcut penalty is the authors saying: if you write a plan and don't follow it, we're going to dock you, hard.”Episode 079 — An Old Idea From Cognitive Psychology Reshapes How We Reward Reasoning Models