Confidence From Graded Past Episodes Instead of the Current Inference Alone

arXiv 2609.17708·medium signalXBlueskyLinkedInCopy link

arXiv 2609.17708 (15 Sep 2026) argues every existing confidence estimator shares one premise, that confidence can be read off the current inference by introspection, token probabilities or resampling. XConf stores a record of the model's own graded past episodes, each holding the task, the model's reflection, its stated confidence, the outcome and a lesson written after grading; its Recall stage retrieves past episodes on similar tasks met with a similar stated confidence and reads their historical success rate, and its Reflect stage shows that record to the model to name its recurring failure mode and restate confidence. The framing matters for anyone routing work by confidence, since it decides what to ship, escalate or retry.