Search papers, labs, and topics across Lattice.
1
0
3
LVLMs know more than they say: critical information for visual document understanding is often better represented in intermediate layers than in the final output, suggesting a bottleneck in the generation process.
