Search papers, labs, and topics across Lattice.
3
0
8
8
Multimodal models can better understand degraded images by dropping pixel-perfect reconstruction and directly optimizing the latent space for reasoning, leading to higher perceptual quality.
Current robot manipulation benchmarks fail to capture the messy reality of real-world deployment, so this work introduces a new benchmark, ManipArena, to close the sim2real gap.
Forget scaling depth and width鈥擬OUE unlocks a new "virtual width" dimension for Mixture-of-Experts by cleverly reusing a single expert pool across layers.
