Search papers, labs, and topics across Lattice.
2
0
5
Streaming video MLLMs perform far better when historical context is actively internalized into evolving latent tokens rather than queried as passive external visual buffers.