Search papers, labs, and topics across Lattice.
2
0
4
0
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.
Chimera achieves 7.3x compute efficiency over traditional models while enabling zero-shot extrapolation from short video clips to significantly longer sequences.
