HC-DLM Couples Discrete Tokens with Continuous Latent Trajectories
October 2, 2026
Hierarchical Continuous Diffusion Language Models (HC-DLM) address the dependency problem in diffusion-based generation by coupling discrete token generation with a continuous latent trajectory. The training objective is derived from a variational bound on token likelihood to ensure valid token configurations during denoising.
HOW THIS AFFECTS YOU
●
researcherThis approach provides a principled way to maintain statistical dependencies in parallel-decoded diffusion language models.