Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Sahil Tripathi, Md Tabrez Nafis, Imran Hussain, Jiechao Gao

TL;DR

This work tackles the confidence paradox in DocVQA by introducing HonestVQA, a model-agnostic framework that calibrates model confidence to reflect knowledge and ethical considerations. It combines uncertainty quantification, confidence-accuracy alignment, and contrastive ethical enforcement to reduce overconfident yet incorrect outputs while maintaining multimodal grounding. The approach gains measurable improvements in accuracy and calibration across three diverse datasets, and demonstrates robust cross-domain generalization, along with theoretical guarantees for the proposed metrics H-Score and ECI. Practically, HonestVQA offers a pathway toward more trustworthy DocVQA systems with quantifiable ethical confidence, at a modest computational overhead suitable for deployment in enterprise settings.

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.

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

TL;DR

This work tackles the confidence paradox in DocVQA by introducing HonestVQA, a model-agnostic framework that calibrates model confidence to reflect knowledge and ethical considerations. It combines uncertainty quantification, confidence-accuracy alignment, and contrastive ethical enforcement to reduce overconfident yet incorrect outputs while maintaining multimodal grounding. The approach gains measurable improvements in accuracy and calibration across three diverse datasets, and demonstrates robust cross-domain generalization, along with theoretical guarantees for the proposed metrics H-Score and ECI. Practically, HonestVQA offers a pathway toward more trustworthy DocVQA systems with quantifiable ethical confidence, at a modest computational overhead suitable for deployment in enterprise settings.

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.

Paper Structure

This paper contains 20 sections, 1 theorem, 12 equations, 1 figure, 6 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Uncertainty Quantification Module
  5. Contrastive Ethical Enforcement Module
  6. Training Module
  7. Experimental Setup
  8. Datasets
  9. Evaluation Metrics
  10. Theoretical Guarantees for Evaluation Metrics
  11. Hyperparameters
  12. Experimental Analysis
  13. Comparison with Baselines
  14. Cross-Domain Generalization Testing
  15. Ablation Study
  16. ...and 5 more sections

Key Result

Lemma 4.1

Let $C(D,Q)$ be the confidence score output by a DocVQA model for a given document-question pair $(D,Q)$, and let $A(D,Q) \in \{0,1\}$ be the corresponding accuracy indicator, where $1$ denotes a correct answer and $0$ an incorrect one. Assume that $C(D,Q)$ is bounded in $[0,1]$. Then, define the Ho where $\mathcal{D}$ is the data distribution. Whereas, $\mathrm{H}$ upper-bounds the expected calib

Figures (1)

  • Figure 1: HonestVQA enhances ethical calibration, semantic stability, embedding separation, and multimodal grounding through uncertainty-aware learning,

Theorems & Definitions (3)

  • proof
  • Lemma 4.2: Discriminative Power of Ethical Confidence Index
  • proof