Search papers, labs, and topics across Lattice.
3
0
6
0
Instead of trusting black-box scalar scores to evaluate world models, video verification can match end-to-end VLM recall while producing fully auditable, spatiotemporally grounded physical proofs of failure.
Long-tail semantic failures in document understanding are exposed when models are forced to reason with a unified vocabulary of visual anchors rather than treating elements in isolation.
The Temporal Ratio reveals how attention shifts between future and present frames can predict a model's ability to generalize compositional tasks in video-action contexts.
