Search papers, labs, and topics across Lattice.
1
0
2
6
LLMs can match or exceed the reasoning performance of autoregressive methods, but with significantly improved compute efficiency, by using fixed-sequence "memory blocks" for latent reasoning.
