Search papers, labs, and topics across Lattice.
Northwestern Polytechnical University, Shenzhen Loop Area Institute, Base Model
3
0
7
Multi-speaker conversational understanding is critically under-evaluated, with MSU-Bench revealing that even leading models struggle with complex speaker grounding tasks.
Over-reliance on agentic decomposition can actually *hurt* audio understanding when a strong audio frontend already provides sufficient information, highlighting the importance of conditional evidence acquisition.
Force GNNs to explicitly reason about graph-level concepts and you get state-of-the-art performance in both classification and interpretability.
