Definition
Plain language
The point in an AI's reasoning where its chance of success has already hit zero and stays there.
As stated in the literature
From prior work, the token after which success probability is pinned at zero; contrasted with the earlier cliff token that triggers the collapse rather than marking its aftermath.
Also called: critical tokens
Why it matters: Distinguishing the point of no return from the earlier word that caused it helps researchers locate where a fix should actually be applied.
For example, several words after a model makes a fatal arithmetic slip, it reaches a point where no continuation can possibly reach the right answer.
Heard on the show
“There's a nice contrast here with earlier work, too — the "critical tokens" idea from Lin and colleagues.”Episode 172 — One Bad Token Can Sink a Model's Math, And You Can Delete It