Search papers, labs, and topics across Lattice.
3
0
5
Achieve 3x better coverage on out-of-distribution visual question answering by explicitly scoring the quality of visual evidence, even when using black-box models like Gemini-3-Pro.
Achieve state-of-the-art character consistency in story visualization with a surprisingly lightweight framework that adds only 149K parameters.
Attention weights in VLMs can fool you: Simpson's paradox hides the true signals of object hallucination, but a new Bayesian method, HaloProbe, can see through the noise.
