Definition
Plain language
A training technique that rewards an AI for building up confidence gradually rather than locking in an answer too soon.
As stated in the literature
A confidence-trajectory-shaped reward term layered onto GRPO that penalizes front-loaded confidence (favoring monotonically increasing mid-chain commitment) to discourage premature confidence and improve faithfulness.
Why it matters: Training against front-loaded confidence pushes models toward chains of thought that genuinely contribute to the answer rather than rationalize it.
For example, the reward penalizes runs where the model is already 90 percent sure of the answer in the first sentence, and rewards ones whose confidence climbs steadily.
Heard on the show
“With progressive confidence shaping, it jumps to sixty-one percent.”Episode 081 — When Reasoning Models Decide Before They Think: Detecting and Fixing Premature Confidence