Search papers, labs, and topics across Lattice.
1
0
3
5
Zero-shot VLMs falter at geometric reasoning, with only one out of five models surpassing random performance on basic jigsaw puzzles, revealing a scaling cliff in their capabilities.
