Revealing Epistemic Uncertainty in MLLMs via Causal-Invariant Masking

2026-10-02

AI Summary

This study addresses hallucinations induced by spurious shallow correlations and the challenge of quantifying epistemic uncertainty in multimodal large language models. To this end, we propose a Causal Invariant Masking framework that leverages semantic divergence to measure sensitivity to non-causal correlations, enabling comprehensive decomposition and detection of uncertainty. We introduce the first causal invariant masking mechanism with rigorous convergence guarantees, alongside a novel geometric proxy metric, EED, which substantially accelerates evaluation. By integrating techniques from causal inference and geometric computation within hyperspherical embedding spaces, our method achieves state-of-the-art performance across multiple benchmarks. Notably, EED improves inference speed by nearly 50% while preserving predictive accuracy.

Abstract

Multimodal Large Language Models (MLLMs) suffer from hallucinations, creating a critical need for Uncertainty Quantification (UQ) to ensure reliable deployment. However, existing approaches struggle to detect uncertainty caused by superficial associations, especially when the query-relevant signal is weak. We mainly attribute this issue to their bias toward aleatoric uncertainty arising from data ambiguity, overlooking epistemic uncertainty stemming from model limitations. To further decompose uncertainty types for a comprehensive UQ, we propose Causal-Invariant Masking (CIM), which measures the semantic shift between the original predictions and those conditioned on a causally-focused view. Based on this framework, we introduce Semantic Divergence as our core metric for UQ and provide theoretical evidence that it converges to the variance of model's sensitivity to non-causal correlations, establishing its ability to capture MLLM's limitation. To accelerate UQ in MLLMs, we further propose Expected Embedding Drift (EED), a fast geometric proxy metric that estimates semantic shift directly within the hyperspherical embedding space. Experiments show that our method achieves state-of-the-art performance on various benchmarks, while the proposed EED accelerates by nearly 50% with comparable performance.

Problem

• Multimodal Large Language Models
• Uncertainty Quantification
• Epistemic Uncertainty
• Hallucination

Innovation

• Epistemic Uncertainty
• Causal-Invariant Masking
• Semantic Divergence
• Expected Embedding Drift
• Uncertainty Quantification