Can models learn to judge their own performance accurately?

The paper's central claim is that a model's ability to judge its own performance can be turned into a training signal, and that doing so elicits faithful uncertainty expression. The abstract starts from a posited premise, that "models capable of accurately judging their own performance are better positioned to improve it," and operationalizes it twice. Reinforcement learning with metacognitive feedback (RLMF) refines "completion rankings during preference optimization based on the quality of a model's self-judgments of performance." Metacognitive data selection uses similar self-judgments to pick high-value training examples, and the abstract says it outperforms naive active learning. The discussion reports that, applied to faithful calibration, the framework achieves "strong and generalizable" results across diverse models and tasks, beating the prior state of the art while preserving task accuracy and factual calibration.

The paper treats faithful calibration as "itself fundamentally metacognitive," because the goal is to align expressed uncertainty with intrinsic uncertainty. Its framework is a two-stage decoupled approach that aligns both numerically and linguistically expressed uncertainty with the model's intrinsic confidence. The discussion adds side effects: models get better at self-assessment, emit "highly faithful" self-reported confidence scores, and modulate linguistic uncertainty "in a naturalistic, context-appropriate fashion." The introduction motivates all of this with three deficits, confident hallucination, failure to recognize knowledge boundaries, and misrepresented internal uncertainty, and names high-stakes advisory uses such as medicine, law and science. A new metric, cMFG*, removes estimation bias for models whose intrinsic confidence occupies a limited range.

This is the training recipe that Can models express uncertainty instead of just answering? lacks. That note argues for faithful uncertainty as a target, and this paper reports a method that reaches it without an accuracy cost. It differs from Can model confidence work as a reward signal for reasoning? in where the self-signal enters. There, answer-span confidence ranks reasoning traces and serves as the reward. Here, the quality of the model's judgment of its own performance reorders preference pairs, and the target is how uncertainty is expressed. It also contrasts with Can models express calibrated confidence in long-form text?, which defines calibration through the decisions users make downstream, whereas this paper defines faithfulness against the model's own intrinsic confidence.

The excerpt is silent on what counts as a self-judgment, how rankings are refined, what the two stages are, which models and tasks were used, how large the gains are, and what the prior state of the art was. It also does not isolate whether self-judgment quality, rather than simply adding a calibration objective, accounts for the improvement, since the premise is posited and the conclusion reports results for the whole framework. The reported gains are behavioral, so they do not settle Can language models genuinely monitor their own thinking?. At the strength the evidence allows, faithful calibration looks trainable without sacrificing accuracy, and self-judgment is a workable signal for choosing preference pairs and training data.