Reproduce “Unveiling Uncertainty: A Deep Dive into Calibration and Performance of Multimodal Large Language Models” on OpenResearch

About this paper

Zijun Chen, Wenbo Hu, Guande He, Zhijie Deng, Zheng Zhang, Richang Hong

A study by Chen et al. investigates the uncertainty calibration of Multimodal Large Language Models (MLLMs), revealing persistent overconfidence despite calibration remaining stable through visual fine-tuning. The research demonstrates MLLMs' ability to integrate multimodal information to reduce uncertainty and shows their self-assessment significantly improves with explicit prompting and techniques like Temperature Scaling.

Key insights

• MLLM calibration remains relatively consistent after visual fine-tuning, and multimodal training minimally impacts linguistic task calibration, yet persistent miscalibration (overconfidence) is a common issue.
• MLLMs exhibit significantly higher uncertainty when processing image-derived information compared to text-derived information, especially in pre-trained stages, but can effectively integrate both modalities to reduce overall uncertainty.
• MLLMs tend to provide answers regardless of knowledge, but their ability to self-assess uncertainty (truthfulness) significantly improves when explicitly prompted to express doubt.

Summary by alphaXiv.