World action models generate actions together with visual predictions of their consequences. Sampling several pairs creates a choice. We study when these predictions provide useful evidence for making that choice.
LikePhys evaluates whether video diffusion models distinguish physically plausible events from impossible ones, without additional training. It compares model likelihoods for paired videos across twelve scenarios spanning rigid bodies, fluids, deformable materials, and optics, providing a measure that aligns with human judgements of physical plausibility.
ComBo combines features from different layers of multiple frozen vision foundation models using a lightweight, trainable adapter. It brings complementary representations together for downstream tasks and identifies which models contribute most, improving on individual backbones without fine-tuning the foundation models themselves.
Visual DNA represents images and datasets through distributions of neuron activations in a frozen neural network. These compact representations allow comparisons that focus on selected visual properties, supporting dataset analysis, evaluation of synthetic imagery, and the selection of data for transfer learning.