Sparse-LaViDa
Summary
Sparse-LaViDa is an efficiency-oriented continuation of LaViDa-O. It truncates masked output positions that are not decoded in the current step, caches prompt and revealed tokens, and uses register tokens plus a step-causal mask to preserve useful context.
Official Artifacts
- Preprint: arXiv 2512.14008
- Official project page: Sparse-LaViDa
- Release caveat: the project page’s GitHub link is explicitly a placeholder pointing to LaViDa-O; no separate official checkpoint was verified at ingest time.
Role In The Wiki
Sparse-LaViDa belongs in the diffusion-LM serving and sparse-decoding branch. Its main lesson is conditional: removing redundant masked tokens helps long generations, while short outputs leave too little work to truncate.