Search papers, labs, and topics across Lattice.
2
0
4
Cheap segment-then-caption video pipelines can overcome boundary drift and context fragmentation once temporal shifts and historical context are governed by interventional dependency modeling rather than raw sequential attention.
Causal-drive trajectories reveal a surprising shift in VLM reliance from visual cues to generated prefixes, enhancing our understanding of multimodal generation dynamics.
