Gibbs Sampling Method to Reconstruct Multimodal LLM Perceptual Priors

This method uses Gibbs sampling to reconstruct the implicit prior distributions of Multimodal Large Language Models. It moves beyond fixed-input probing to sample from the high-dimensional input spaces that influence model behavior and expectations.