Search papers, labs, and topics across Lattice.
Affiliation:
2
0
5
6
Fine-grained MoE designs can outperform dense vision encoders while dramatically reducing latency, challenging the status quo in image and video understanding.
Open-ended generation models are failing to capture factual completeness, with top performers only achieving 58.7% accuracy on a new benchmark designed to assess this critical aspect.
