The paper and repository publish dSpark training and evaluation methods, while checkpoints and training data remain unreleased.
dSpark learns token reveal schedules for faster diffusion language models
2026-10-27Frontier watchWATCH
dSpark distills token-unmasking schedules that reveal positions which reduce uncertainty elsewhere, rather than simply revealing the most confident token first. The authors report 10.12 tokens per forward pass on LLaDA-8B-Instruct and 5.12 on Dream-7B-Instruct while preserving benchmark accuracy. Training and evaluation code is public, but model checkpoints and training data are still promised for a later release and the repository has no declared license.
Official · PaperLearning When to Commit: Schedule Distillation for Fast Diffusion Language ModelsOfficial · CodedSpark research repository
Availability: Training/evaluation code public; checkpoints and training data promised later
What this means for you
For diffusion-language-model serving, scheduling can change wall-clock cost without changing the base architecture. Missing checkpoints, data and license metadata make the current release unsuitable for a bounded production trial.
FRONTIER WATCH
Public artifacts and validated results are different.
The recorded fact is the public work and its artifacts. Author-reported outcomes do not establish independent reproduction. Next checks include usable code, task conditions and third-party validation.
Evidence & related material
Technical details & scores
| Method | Schedule distillation with a KL anchor on failed trajectories |
|---|---|
| Reported Parallelism | 10.12 tokens/forward on LLaDA-8B-Instruct; 5.12 on Dream-7B-Instruct |
| Reported Wall Clock | 249 TPS versus 15 TPS for base LLaDA on GSM8K, A100 batch size 1 |
| Compute | Training scripts specify four A100 GPUs |
| Availability | Training/evaluation code public; checkpoints and training data promised later |
| License | No repository license declared at check time |
| Time Basis | arXiv v1 submitted 2026-10-26T17:35:58Z; repository pushed 2026-10-26T17:40:30Z |
Technical impact 4
Engineering value 4
Adoption signal 1
Maturity 2
Verification cost 4
Risk 4