Definition
Plain language
A planted clue that points an AI toward a wrong answer only if it actually does some hidden arithmetic, meant to force its reasoning out into the open.
As stated in the literature
In chain-of-thought faithfulness studies, an embedded cue that requires the model to perform an explicit computation (e.g., letter-to-number mapping plus modular arithmetic) to reach the target answer, so that following it should leave a detectable trace in the reasoning; contrasted with a simple hint that directly states the answer.
Also called: simple hint
Why it matters: It forces a model to leave a trace in its reasoning if it relied on the clue, giving researchers a way to check whether the visible reasoning actually reflects how the answer was reached.
For example, a planted clue might spell out that the answer is hidden by turning the letters of a word into numbers and adding them up, so the model can only use it by visibly doing that math.
Heard on the show
“The fix is what the authors call a complex hint, and it's elegant.”Episode 094 — Chain-of-Thought Monitoring Fails Across Languages, and Worst Where It's Needed Most