Multimodal language models better detect uncertain answers using causally focused masking
Revealing Epistemic Uncertainty in MLLMs via Causal-Invariant Masking
Computer Vision and Pattern RecognitionArtificial Intelligence
Summary
Multimodal large language models sometimes give wrong or made-up answers because they can’t always tell when they are unsure. The authors found that previous methods mostly detect uncertainty caused by unclear data but miss uncertainty stemming from the models’ own knowledge gaps. They created a new way called Causal-Invariant Masking that compares normal predictions to those focused only on important causes, helping reveal when the model is unsure due to its own limitations. They also made a faster measurement to estimate this uncertainty, which works well on tests.
What this means in practice
- •For ai platform engineers: Improve reliability of AI systems by detecting when multimodal models are uncertain due to their own knowledge limits.
- •For autonomous vehicle developers: Use the method to better identify uncertain visual and language inputs that may lead to incorrect decisions in self-driving systems.
Authors
Haoyang Luo, Linwei Tao, Jie Gui, Xinghao Chen, Chang Xu, Jianyuan Guo, Minjing Dong
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.