MIT NLP proposes Latent Lattice LLMs: recurrent latents broaden the residual stream

viksit · x · 2026-11-21

MIT NLP introduces Latent Lattice Language Models (LLLMs): instead of computing each token through a single vertical residual stream, LLLMs add a horizontal 'Lattice of Latents' axis using recurrent latents, scaling up per-token computation — blending RNN ideas with transformers.

Original post →

More from Research

Research channel →