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

arXiv: https://arxiv.org/abs/2506.23464 · HTML: https://mlsanity.com/paper/2506.23464/

## Key facts

- Cited by 11 papers in the map (data through 14 Oct 2026).
- Builds on 15 of its 20 references that are in the map.
- First posted on arXiv 30 Jun 2025.

## Builds on (most cited first, 15 of 15)

- [CLIP-Adapter: Better Vision-Language Models with Feature Adapters](https://mlsanity.com/paper/2110.04544/) (2021)
- [When Does Label Smoothing Help?](https://mlsanity.com/paper/1906.02629/) (2019)
- [DocLLM: A layout-aware generative language model for multimodal document understanding](https://mlsanity.com/paper/2401.00908/) (2023)
- [On the Calibration of Large Language Models and Alignment](https://mlsanity.com/paper/2311.13240/) (2023)
- [Ethical Reasoning over Moral Alignment: A Case and Framework for In-Context Ethical Policies in LLMs](https://mlsanity.com/paper/2310.07251/) (2023)
- [Understanding Softmax Confidence and Uncertainty](https://mlsanity.com/paper/2106.04972/) (2021)
- [Making Large Language Models Better Reasoners with Alignment](https://mlsanity.com/paper/2309.02144/) (2023)
- [Calibrating Language Models with Adaptive Temperature Scaling](https://mlsanity.com/paper/2409.19817/) (2024)
- [Adaptation with Self-Evaluation to Improve Selective Prediction in LLMs](https://mlsanity.com/paper/2310.11689/) (2023)
- [Enhancing Visual Document Understanding with Contrastive Learning in Large Visual-Language Models](https://mlsanity.com/paper/2402.19014/) (2024)
- [Advancing Multimodal Large Language Models in Chart Question Answering with Visualization-Referenced Instruction Tuning](https://mlsanity.com/paper/2407.20174/) (2024)
- [LayoutLLM: Large Language Model Instruction Tuning for Visually Rich Document Understanding](https://mlsanity.com/paper/2403.14252/) (2024)
- [Document Visual Question Answering Challenge 2020](https://mlsanity.com/paper/2008.08899/) (2020)
- [Large Language Models for Failure Mode Classification: An Investigation](https://mlsanity.com/paper/2309.08181/) (2023)
- [DocVideoQA: Towards Comprehensive Understanding of Document-Centric Videos through Question Answering](https://mlsanity.com/paper/2503.15887/) (2025)

## Built on by (most cited first, 11 of 11)

- [Failure Modes in LLM Systems: A System-Level Taxonomy for Reliable AI Applications](https://mlsanity.com/paper/2511.19933/) (2025)
- [Rewarding the Journey, Not Just the Destination: A Composite Path and Answer Self-Scoring Reward Mechanism for Test-Time Reinforcement Learning](https://mlsanity.com/paper/2510.17923/) (2025)
- [FalseCrashReducer: Mitigating False Positive Crashes in OSS-Fuzz-Gen Using Agentic AI](https://mlsanity.com/paper/2510.02185/) (2025)
- [Do Large Language Models Know What They Don't Know II? A Fully Behavioral, Non-Cognitive Measure of Epistemic Honesty](https://mlsanity.com/paper/2609.07879/) (2026)
- [Do LLMs Know They Are Being Tested? Evaluation Awareness and Incentive-Sensitive Failures in GPT-OSS-20B](https://mlsanity.com/paper/2510.08624/) (2025)
- [Generative AI for Managerial Decision-Making under Ambiguity and Sycophancy](https://mlsanity.com/paper/2603.03970/) (2026)
- [Confidence Calibration in Large Language Models](https://mlsanity.com/paper/2605.23909/) (2026)
- [Are Two LLMs Better Than One? A Student-Teacher Dual-Head LLMs Architecture for Pharmaceutical Content Optimization](https://mlsanity.com/paper/2602.11957/) (2026)
- [LLM-Guided Semantic Bootstrapping for Interpretable Text Classification with Tsetlin Machines](https://mlsanity.com/paper/2604.12223/) (2026)
- [When Confidence Fails: Overconfidence in LLMs under Uncertainty and Missing Clinical Information](https://mlsanity.com/paper/2608.09080/) (2026)
- [TRACE: Trustworthy Retrieval-Augmented Conversational Engine](https://mlsanity.com/paper/2608.10176/) (2026)

## Abstract (arXiv)

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.