Knowing When You Don’t Know: Metacognitive Uncertainty Calibration in Vision--Language Models

Mahule Roy, Subhas Roy

Track: Track 2: Papers without Workshop Proceedings

Keywords: Vision–Language Models, Uncertainty Calibration, Self-Reflective Prompting

TL;DR: We introduce a self-reflective prompting framework that improves VLM confidence calibration and reduces overconfident hallucinations by eliciting metacognitive behavior through introspection and counterfactual consistency checks.

Abstract: Vision–Language Models have achieved impressive performance across visual question answering, image captioning, and multimodal reasoning tasks. However, these models often exhibit overconfident failures, producing fluent yet incorrect responses without signaling uncertainty. In contrast, human cognition relies heavily on metacognition—the ability to monitor confidence, detect ambiguity, and recognize potential errors. In this work, we investigate whether metacognitive behaviors can be elicited in pretrained VLMs through cognitively inspired, inference-time mechanisms. We introduce a self-reflective uncertainty calibration framework that prompts VLMs to assess evidence sufficiency, ambiguity, and confidence after generating an initial response. Through experiments on 1000 challenging vision–language queries across three open-source VLMs of varying architectures, we show that structured introspective prompting significantly improves confidence–accuracy alignment and reduces overconfident hallucinations. However, substantial miscalibration remains, with models still expressing high confidence in approximately half of their errors. Our findings suggest metacognition as a promising direction for improving VLM trustworthiness, while highlighting significant limitations that require further research.

Submission Number: 2