Probing Multimodal LLM Perceptual Priors with Gibbs Sampling

Summary

Researchers propose a method to directly sample the perceptual priors of multimodal large language models. By steering a generative model along interpretable axes and using the target model as a judge in Gibbs sampling, the approach explores visual expectations that fixed-input tests and direct prompting can miss. Experiments on facial trustworthiness and perceived cheapness in art images recover established biases and reveal surprising new priors, highlighting possible effects on real-world model behavior.