VideoLVR: Latent Visual Reasoning over Grounded Video Evidence
Keeps video evidence reachable during reasoning, so multimodal models stop forgetting what they saw.
Explainer · on the wire
The explainer for this story is still on the wire.
The claim, the mechanism, and the figures land here once generated.
Source · arXiv