Glossary · Term

token-wise potential

← all terms

Definition

Plain language

At any point in an AI's reasoning, the fraction of ways it could finish from there that reach the right answer.

As stated in the literature

A per-token measure of solution health estimated by forking many continuations from a position and counting correct outcomes; measured at every token to localize cliff tokens.

Also called: potential

Why it matters: Measuring this at each word reveals exactly where a model's chance of success collapses, helping locate the moment reasoning goes wrong.

For example, partway through a model's solution, if 90 out of 100 ways of finishing reach the right answer, the token-wise potential there is high.

Heard on the show

“They call it the hallucination contraction, and the setup is a checkpoint that every potential mistake has to pass before it becomes permanent.”
Episode 182 — How a Tiny Model Too Weak to Plan Cuts a Bigger Agent's Hallucinations by 80%

Mentioned in 9 episodes

  1. 182
    How a Tiny Model Too Weak to Plan Cuts a Bigger Agent's Hallucinations by 80%
  2. 172
    One Bad Token Can Sink a Model's Math, And You Can Delete It
  3. 160
    Training an AI to Take Its Own Notes, So Its Future Self Works Better
  4. 100
    How a Prompt Wrapper Lets a Frontier Model Play Poker Like an Expert
  5. 073
    When Three LLMs Talk to Each Other, Their Ideas Quietly Stop Moving
  6. 068
    The OS Trick That Makes Tree Search Practical for Coding Agents
  7. 054
    When Models Learn the Monitor Exists, the Reasoning Trace Stops Being a Window
  8. 017
    When the Agent Grades Its Own Homework: A Brutal New Benchmark for AI Workers
  9. 008
    Why Long-Horizon AI Agents Get Stuck, and a Milestone-Based Fix That Helps

Related terms