When Should Diffusion Language Models Reconsider Their Commitments?
Organizations: Purdue University · UCLA · Google
Abstract
Masked Diffusion Models (MDMs) are well known for fast text generation via parallel decoding. However, independent parallel sampling can commit to inconsistent predictions, and less is known about how to selectively revise these early token commitments. In this work, we introduce lanned nmasking-and-evision Joint earning (PURL) for MDMs, a training-based framework that seamlessly unifies parallel generation with selective token revision. We derive a normalized joint transition whose likelihood factorizes into edit selection, unmask selection, and token prediction. These components are trained jointly on partially masked sequences constructed according to the model's own unmasking policy. During inference, all three decisions are supported by a single shared forward pass, enabling the model to decode masked tokens and revise noisy tokens. Across logic puzzles, mathematical reasoning, and code generation, PURL improves accuracy at matched unmask budgets, with relative gains up to 47.4% accuracy improvement and up to an reduction in scheduled decoding rounds. Beyond these gains, our analysis further reveals general principles of revisable decoding: token edits become more effective as context accumulates; early edits can steer unfinished trajectories, whereas later edits tend to remain localized; and effective edits therefore depend on both calibrating when to intervene and learning where and how to revise. Together, these suggest that effective parallel generation depends not on revising more tokens, but on making the right commitments revisable at the right time.
Figures & tables
| Train | Inference | ||||||
|---|---|---|---|---|---|---|---|
| MDLM | Random | ||||||
| TopK | |||||||
| TopK-M | |||||||
| PUMA | Random | ||||||
| TopK | |||||||
| TopK-M |
| Training configuration | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Acc | # Edits | Acc | # Edits | Acc | # Edits | Acc | # Edits | Acc | # Edits | Acc | # Edits | |
| (a) Cumulative additions | ||||||||||||
| Masked-only | 33.12 | 0.00 | 31.54 | 0.00 | 30.17 | 0.00 | 27.45 | 0.00 | 20.85 | 0.00 | 8.57 | 0.00 |
| policy-aligned pool | 40.33 | 0.00 | 42.00 | 0.00 | 39.27 | 0.00 | 32.90 | 0.00 | 23.58 | 0.00 | 9.17 | 0.00 |
| correction supervision | 41.55 | 0.00 | 41.55 | 0.00 | 38.74 | 0.00 | 33.21 | 0.00 | 21.91 | 0.00 | 9.10 | 0.00 |
| learned edit action | 42.17 | 0.41 | 41.64 | 0.61 | 41.77 | 0.89 | 32.75 | 2.30 | 26.54 | 3.91 | 15.39 | 6.49 |
Appendix figures & tables16 assets
Supplementary material from the paper’s appendix.
Appendix
| Symbol | Meaning |
|---|---|
| , | Non-mask vocabulary and mask symbol. |
| , | Sequence length and editable positions, . |
| , | Reference and current partially masked sequences. |
| , | Transient correction and successor sequences. |
| Masked positions. | |
| Unmasked positions; distinct from vocabulary . |
| Model | Inference | ||||||
|---|---|---|---|---|---|---|---|
| MDLM | Random | ||||||
| TopK | |||||||
| TopK-M | |||||||
| PUMA | Random | ||||||
| TopK | |||||||
| TopK-M |
Diffusion Language Model Decoding
Diffusion language model decoding generates text through iterative parallel refinement, balancing fewer denoising steps with accuracy, coherence, and control.
See all 267 papers