The Confidence Paradox: Can LLM Know When It's Wrong

TL;DR

The paper introduces HonestVQA, a self-supervised framework that calibrates model confidence with correctness and ethics in Document VQA, improving accuracy and reducing overconfidence across multiple datasets.

Contribution

It presents HonestVQA, a novel model-agnostic, self-supervised approach with new metrics for ethical confidence calibration in Document VQA.

Findings

Improves accuracy and F1 scores by up to 4.3%

Reduces overconfidence in model predictions

Achieves 78.9% accuracy and 76.1% F1-score across datasets

Abstract

Document Visual Question Answering (DocVQA) models often produce overconfident or ethically misaligned responses, especially under uncertainty. Existing models like LayoutLMv3, UDOP, and DONUT focus on accuracy but lack ethical calibration. We propose HonestVQA, a model-agnostic, self-supervised framework that aligns model confidence with correctness using weighted loss and contrastive learning. We introduce two new metrics Honesty Score (H-Score) and Ethical Confidence Index (ECI)-to evaluate ethical alignment. HonestVQA improves accuracy and F1 by up to 4.3% across SpDocVQA, InfographicsVQA, and SROIE datasets, while reducing overconfidence. It also generalizes well across domains, achieving 78.9% accuracy and 76.1% F1-score.