When Should Diffusion Language Models Reconsider Their Commitments?

Haoyang Zheng, Yidong Ouyang, Blake Jianhang Chen, Tian Tan, Ziqian Lin, Liqian Peng, Alec Go, Guang Lin

Parallel decoding makes Masked Diffusion Models (MDMs) incredibly fast, but it comes with a catch: because the model predicts multiple tokens simultaneously and independently, it often makes contradictory "commitments" that it can't take back.

A new paper from researchers at Purdue, UCLA, and Google introduces PURL (Planned Unmasking-and-Revision Joint Learning), a framework that gives these models a "delete" key. By learning to selectively revise its own predictions during the generation process, PURL significantly boosts accuracy and efficiency in complex reasoning tasks.

TL;DR

PURL enables Masked Diffusion Models to reconsider and fix early token commitments during parallel decoding. By jointly training the model to unmask new tokens and revise noisy ones in a single forward pass, PURL achieves up to a 47.4% accuracy boost and an 8× reduction in decoding rounds.

What Problem Does This Paper Address?

Standard MDMs generate text by starting with a sequence of [MASK] tokens and progressively filling them in. The strength of this approach is parallelism—you can fill in many tokens at once. The weakness is irreversible commitment.

Because tokens in a single step are sampled independently, the model might choose a variable name in one part of a code block that contradicts a logic choice made simultaneously in another part. In traditional "absorbing" diffusion, once a token is unmasked, it is frozen. These early errors propagate, often leading to "collisions" where the final output is nonsensical or logically inconsistent.

The challenge is: how can a model learn to identify its own mistakes and fix them without needing a slow, separate "verifier" model or an expensive autoregressive backup?

Methods

The core of PURL is a training-based framework that unifies generation and revision into a single joint transition.

  1. Joint Likelihood Factorization: The authors derive a mathematical framework where the model's objective is split into three parts:
  2. Edit Selection: Deciding where to revise existing tokens.
  3. Unmask Selection: Deciding where to reveal new tokens.
  4. Token Prediction: Deciding what tokens to actually write.
  5. Single Shared Forward Pass: Unlike "draft-and-verify" methods that use two different models, PURL uses one backbone. During inference, a single forward pass generates scores for two "planners" (Edit and Unmask) and the standard token logits. This is illustrated in the inference pipeline:
PURL Inference Pipeline
  1. Transient Correction Sequences: To train the model to fix errors, the researchers use a clever trick: they take a "clean" partially generated sequence, intentionally inject "noisy" (incorrect) tokens, and then task the model with identifying and correcting those specific errors while retaining the correct tokens.

Datasets

The researchers evaluated PURL across several challenging domains:

  • Logic Puzzles: Sudoku (10,000 test puzzles).
  • Mathematical Reasoning: GSM8K and MATH500.
  • Code Generation: HumanEval+ and MBPP+.
  • General Knowledge: MMLU (to ensure the model maintains general capabilities after specialized post-training).
  • TinyGSM: A large-scale dataset of 11.8 million Python math programs used for training the smaller 125M parameter models.

Key Insights and Ideas

  • Revision is a Timing Problem: One of the most important findings is that "more revision" isn't always better. Revision is most effective when delayed. Early in the generation, the model lacks the context to know if a token is wrong. As context accumulates, edits become more surgical and effective.
  • Learning to Retain is as Important as Learning to Edit: A major risk of revision is "undoing" correct work. PURL includes a specific "retention" loss that trains the model to keep valid predictions stable even when it's looking for errors elsewhere.
  • Steering vs. Localizing: The authors found that early edits tend to "steer" the entire trajectory of the generation, while later edits are more localized, fixing specific syntax or arithmetic errors. This is visible in the "Action Density" heatmaps:
Action Density Heatmaps

Key Results

PURL consistently outperforms non-revisable baselines across the board, especially when the "budget" (the number of tokens unmasked per step) is high.

  • Sudoku: At a high parallelism budget (K=32), PURL maintains 71.60% accuracy, while the best baseline (PUMA) collapses to 57.35%.
  • Mathematical Reasoning: On GSM8K, PURL (using an 8B LLaDA backbone) achieved 81.71% accuracy, a significant jump over the 76.83% baseline.
  • Efficiency: PURL can match the accuracy of standard models while using significantly fewer decoding rounds. For example, moving from K=1 to K=8 provides an 8× reduction in rounds without sacrificing quality.
  • Qualitative Success: Figure 1 shows a clear example where a standard model gets stuck on an incorrect variable assignment (_price), whereas PURL's "Selective Revision" identifies the error and swaps it for the correct house_cost.
PURL vs No Token Edit

Limitations and Future Work

  • Heuristic Gating: The current "gate" that decides when to start revising is a simple threshold based on how much of the sequence is unmasked. The authors suggest that future work could use the actual structure of the task (e.g., the syntax tree of a code block) to better time these interventions.
  • Structured vs. Open-Ended Text: The paper focuses heavily on structured reasoning (math, code, logic). While these are the areas where MDMs struggle most, it remains to be seen how PURL performs on more creative or open-ended prose where "correctness" is subjective.
  • Computational Overhead: While PURL uses a single forward pass, the planners do add a small number of parameters (e.g., 0.5M to 1M), though this is negligible compared to the billions of parameters in the backbone.

For a deep dive into the math behind the joint transition and the full ablation studies, you can find the paper on arXiv at 2611.11551v1.

Summary updated

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 $\underline{\mathrm{P}}$lanned $\underline{\mathrm{U}}$nmasking-and-$\underline{\mathrm{R}}$evision Joint $\underline{\mathrm{L}}$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 $8\times$ 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.
BibTeX
@misc{zheng2026when,
  author = {Haoyang Zheng and Yidong Ouyang and Blake Jianhang Chen and Tian Tan and Ziqian Lin and Liqian Peng and Alec Go and Guang Lin},
  title = {{When Should Diffusion Language Models Reconsider Their Commitments?}},
  year = {2026},
  eprint = {2611.11551},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.11551},
}
Figure