What it is
A hierarchical diffusion LM that alternates continuous latent denoising with token readout and re-noising, tying parallel decoding to valid token scaffolds. Unlike discrete diffusion that samples parallel tokens independently and continuous diffusion that only denoises latents, HC-DLM makes the latent the sole persistent state and reads tokens out at every step to scaffold the next latent update.
Key results
- Countdown CD5: 37.52% vs CCDD 25.35% at 6M params
- Countdown CD4: 84.41% vs CCDD 81.18% at 6M params
- Sudoku Hard: 72.41% vs CCDD 70.73% at 6M params
Shows that feeding token scaffolds back into a shared continuous denoising state better preserves cross-token dependencies than separate discrete or continuous chains.