Search papers, labs, and topics across Lattice.
1
0
3
MLLMs struggle with multi-hop reasoning in document-style images, achieving only 62.83% accuracy compared to over 90% for human annotators.