Search papers, labs, and topics across Lattice.
Seoul National University
5
0
8
Even leading MLLMs struggle with object hallucination, revealing critical integration-stage weaknesses that standard benchmarks fail to capture.
MLLMs struggle to generalize in Video Temporal Grounding not just due to unseen concepts, but because visual domain shift breaks their ability to link temporal localization with entity attention – a problem EVIDENT solves by explicitly routing adaptation through visual entity evidence.
Natural images aren't Euclidean—modeling them on a hypersphere unlocks better generative performance.
RIS models struggle with motion-based queries, but a new data augmentation and contrastive learning approach closes the gap without sacrificing performance on appearance-based descriptions.
Ditch the noisy diffusion detour: ASBM finds a surprisingly direct route from data to noise, slashing sampling steps and boosting image fidelity.
