Search papers, labs, and topics across Lattice.
2
0
6
KDFlow unlocks 6x faster LLM knowledge distillation by decoupling student training and teacher inference, using hidden state transfer to minimize communication overhead.
Latent visual reasoning in multimodal LLMs is largely ineffective, as the "imagination" happening in latent space doesn't actually attend to the input or influence the output, making explicit text-based imagination a surprisingly better alternative.
