Search papers, labs, and topics across Lattice.
2
0
5
Standard hybrid architectures order their layers backwards for multilingual data: placing a full-attention layer first triggers immediate cross-lingual representation alignment and accelerates distillation by up to 2.5X.
Injecting structured visual perturbations into multimodal RL significantly stabilizes training and boosts video reasoning performance, even when answers are superficially similar.
