Tag
Researchers propose Latent Thought Flows, compressing 256 text tokens into 8 continuous latents for one-step generation, achieving a better Pareto frontier than autoregressive baselines.