Search papers, labs, and topics across Lattice.
4
0
6
Ego-motion ambiguity in VLMs can lead to severe spatial reasoning failures, but a new benchmark and framework show how to ground visual representations in 3D space effectively.
AdaThinkV achieves 40.79% accuracy in video reasoning while using 22.7% fewer tokens than its strongest adaptive baseline, showcasing a breakthrough in token-efficient reasoning.
ZeroSplat achieves dynamic segmentation of multiple targets in 3D without the computational burden of traditional methods, setting a new standard for efficiency and effectiveness in scene understanding.
LLMs can substantially improve their ability to follow complex instructions and constraints by explicitly auditing their own context adherence before answering.
