The Role of Model Confidence on Bias Effects in Measured Uncertainties

ACL ARR 2025 February Submission4942 Authors

Abstract: With the growing adoption of Large Language Models (LLMs) for open-ended tasks, accurately measuring epistemic uncertainty, a key indicator of a model's lack of knowledge or confidence, has become crucial for ensuring reliable outcomes. However, quantifying epistemic uncertainty in such tasks remains challenging due to the presence of aleatoric uncertainty, which arises from inherent randomness among multiple valid answers. Building on previous work showing that LLMs are more likely to copy information from input when model confidence is low, we empirically analyze how text-based and image-based biases in input affect the behavior of GPT-4o and Qwen2-VL across varying confidence levels in Visual Question Answering (VQA) tasks. Our findings reveal that all considered biases induce greater changes in measured uncertainties, when model confidence after bias mitigation is lower. Moreover, lower model confidence leads to greater underestimation of epistemic uncertainty (i.e. overconfidence) due to the presence of bias, whereas it has no significant effect on the direction and smaller effect on the magnitude of aleatoric uncertainty changes. Based on these observations, we hypothesize that biases degrade the ranking performance of measured uncertainty, motivating our exploration of bias mitigation as a potential uncertainty quantification approach. This approach improves uncertainty quantification in the presence of aleatoric uncertainty with GPT-4o.

Paper Type: Long

Research Area: Interpretability and Analysis of Models for NLP

Research Area Keywords: uncertainty

Contribution Types: Model analysis & interpretability

Languages Studied: English

Submission Number: 4942