Definition
Plain language
When an AI commits to its answer before doing any of the reasoning it later writes down.
As stated in the literature
A reasoning failure mode diagnosed via mid-chain probes: the model's confidence in its eventual final answer is already saturated at the start of the chain, indicating the reasoning text is decorative rather than causal.
Why it matters: It means the visible chain of thought isn't actually doing the reasoning, which undermines explanations users rely on for trust.
For example, the model's probe-measured confidence in 'B' is already 95 percent at the very first sentence of its multi-paragraph reasoning.
Heard on the show
“RL converges on premature confidence as a local optimum because that's the best thing in its rollout distribution to amplify.”Episode 081 — When Reasoning Models Decide Before They Think: Detecting and Fixing Premature Confidence