Researchers have proposed a framework for multimodal large language models to make decisions that take into account the uncertainty of their output. This approach is necessary because MLLMs often fail due to issues such as poor input quality, perceptual errors, or conflicting modalities. The authors review various methods for measuring and controlling uncertainty in MLLMs, including token and logit uncertainty, semantic disagreement, and perturbation instability. They argue that evaluating uncertainty solely through confidence numbers is insufficient and propose a more comprehensive approach that considers the impact of uncertainty on system behavior. --- Why it matters: This research matters to engineers and researchers in AI because it addresses a critical issue in multimodal language models: their tendency to produce confident but incorrect answers due to uncertainty. By developing methods for evaluating and controlling uncertainty, this work has the potential to improve the reliability and trustworthiness of these systems. Source: https://arxiv.org/abs/2608.17084