Search papers, labs, and topics across Lattice.
The Hong Kong University of Science and Technology Hong Kong SAR
4
0
5
MAFIA reveals that memory-augmented LLMs can be compromised with a staggering 90.7% success rate, even under rigorous auditing conditions.
Achieving 100x parameter efficiency, LoopWM redefines how we approach long-horizon world modeling by introducing iterative latent depth as a new scaling axis.
Shifting from answer-centric to evidence-centric reasoning, CoVER-7B outperforms even leading closed-source models in long-video understanding tasks.
MLLMs can gain surprisingly strong 3D spatial reasoning abilities simply by tapping into the latent knowledge already present in video generation models.
