acceptodds
Under review as a conference paper at ICLR 2027

Random Remasking Scales Masked Diffusion Inference

Abstract

Masked diffusion models (MDMs) generate sequences by progressively unmasking tokens, but errors introduced during decoding are typically irreversible. Recent work has shown that iterative correction can improve sample quality, yet existing theory fails to explain this given a learned, imperfect denoiser model. We re-formulate MDM inference as a Markov chain on the clean sequence space, defined by a remask-denoise kernel that randomly re-introduces masks and then denoises. This helps us analyze the actual denoising trajectory, and we examine conditions under which each iteration step contracts the KL divergence. The analysis motivates Random Remask-Denoise (RD) inference, a purely test-time, finetuning-free, reward-free procedure. To further improve performance, we propose Trajectory Correction Training (TCT), a denoiser-training procedure that exposes the model to its own errors along the generation trajectory. On logical tasks, our approach exhibits a striking near-log-linear test-time scaling, achieving over 97% accuracy on hard Sudoku puzzles and over 99% accuracy on 3-SAT. Our approach also applies to pretrained models on actual language tasks, beating iterative correction baselines and the official Q-mode accuracy of LLaDA2.1-mini on HumanEval+.

Primary area:
foundation or frontier models, including LLMs
Keywords:
masked diffusion model, inference-time scaling

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_biNiOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies remasking in discrete diffusion models. The authors propose random remasking at inference time and trajectory correction training, which they claim and show it improves performance.

Critical strengths

The main strength of the paper is that it studies a simple idea that has been shown (in works like ReMDM) to be useful for improving discrete diffusion models with additional compute.

Critical weaknesses

My main concern is Novelty.

The paper makes minor modifications to ideas already explored in the literature, such as ReMDM. The work feels more like an ablation study, but the ablations are not extensive enough to explain why these changes work better. Instead, the paper places considerable emphasis on theory, whose connection to the practical results is not sufficiently evaluated.

Other concerns:

  • Given the limited novelty, the evaluation on a small set of tasks and benchmarks makes it difficult to assess the benefits over existing remasking methods, including random remasking, which is already widely used.
  • Large models already work well with random remasking or confidence-based remasking. Although the paper includes ablations, it is still unclear how much of the improvement comes from each proposed change.
  • The paper brings ReMDM into its own evaluation setting/setup rather than also evaluating on benchmarks used in the ReMDM paper. The model setups differ as well. This leaves it unclear whether the improvements extend to the settings where ReMDM originally showed benefits.
  • Regarding TCT, several existing works, such as ProSeCo, PRISM, and RemeDi, already modify training to improve self-correction. It is unclear what TCT adds beyond these methods if evaluated under same setup. I am also concerned about the practical cost of repeated correction, given that computational efficiency is one motivation for discrete diffusion.
Questions for authors

What do the authors consider the main new contribution beyond the relatively minor changes to existing RemeDi, ReMDM and other methods?

Reviewer_RtWVOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper approaches the topic of remasking and correction in masked diffusion models. It proposes a new denoising algorithm for masked diffusion models, which repeatedly remasks a varying proportion of tokens over the decoding process, until convergence. To further strengthen the performance under this denoising process, the authors additionally propose a variant of trajectory-aware training. The paper empirically analyses their method using two models trained from scratch on small structural tasks (Sudoku and 3-SAT), as well as on one pretrained model and one benchmark.

Critical strengths
  • The question that the paper attempts to address, of when and under which conditions remasking helps to refine the quality of the generation from the MDMs, is interesting and timely.
  • The experiments show consistent performance improvement from R^2D on the datasets and models considered, although at a very high inference cost. Compared to the baselines, R^2D allows for more effective test-time scaling.
  • I also appreciate the ablations that the authors have conducted, allowing to see which of the design choices contribute meaningfully to the observed performance.
Critical weaknesses

In my view, the main weakness of the paper is that the presented theory (Section 4) does not produce significant insights into how and when remasking works and benefits the denoising process, despite some of the claims that the authors make in the abstract and the introduction of their work. As I explain below, many of the observed behaviours remain unexplained, and the method seems only vaguely inspired by the theoretical results. Beyond the theory, I do not find the presented empirical results sufficiently informative to warrant a standalone contribution.

Theory

  • The contraction ratio depends primarily on the distribution pp which does not depend explicitly on the model parameters or the model error, but rather seems to be an arbitrary choice, not really connected to the denoising process. This is further exacerbated by Proposition 3, where the KL divergence between pp and pdatap_{data} in the numerator does not seem particularly meaningful (given that pp is some arbitrary distribution). As such, the KL contraction bound offers limited insights into how the remasking process can be optimised.
  • Consequently, the proposed denoising algorithm is only very loosely related to the theory, using qualitative and imprecise justifications such as ”ϵt\epsilon_t typically increases with t” (l.233), or “early small-tt steps are wasted while ρt(p)KL(pdata∣∣p)\rho_t(p)KL(p_{data}||p) still dominates” (l.268). It is unclear whether R^2D is in any way optimal, or why it outperforms the variant of the ReMDM algorithm the authors consider.
  • Further, the proposed theory does not allow to answer important and informative questions such as: based on the properties of the decoding model fθf_\theta, what would be the optimal remasking schedule? why does the confidence-based remasking remove the scaling effect (l. 458)? given an imprecise model and the denoising trajectory, when can we expect the unmasked sequence x^0\hat{x}_0 to converge?
  • Regarding the comparison to ReMDM: in ReMDM, the expected number of masks at time t is L(1−αt)L(1 - \alpha_t), where αt\alpha_t is the noise schedule of the diffusion process. Then, σt\sigma_t (which is kept constant) further dictates how many of these masks correspond to remasked tokens. In contrast, in R^2D, the number of masks at time tt is LtLt, and this value depends on tt but does not depend on αt\alpha_t. In my view, the existing theory does not manage to explain why lack of dependence of R^2D on αt\alpha_t leads to worse results.

Experiments

  • Given that R^2D is an inference-only method which does not require additional training, running additional experiments validating its effectiveness should not be compute-heavy. I think the results could be strengthened by extending the evaluation to other pre-trained models and tasks.
  • Existing MDLM decoding methods often rely on block-wise decoding. Does the proposed method also offer significant performance improvements when using block-wise decoding?
Questions for authors
  • In section 5.5, the authors show that stochastic sampling loses 10.4 points on accuracy compared to argmax sampling. The authors claim that this is because of the distribution shift, as TCT sampler is trained along the argmax trajectory. However, the authors then also train a variant of the TCT algorithm with stochastic sampling. For this checkpoint, what would be the performance when using stochastic sampling at inference time? Does the distribution shift indeed explain the performance gap from stochastic sampling?
  • What are the exact decoding settings used by ReMDM and by the Standard baseline (e.g. block size, number of decoding steps, unmasking rule (random or confidence based))?
Reviewer_9pnROpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies inference-time scaling for masked diffusion models through repeated random remasking and denoising. The method is motivated by a Markov chain analysis with an imperfect denoiser.

Critical strengths

The paper writing sounds plausible, and the theoretical/empirical scopes are also well - justified.

Critical weaknesses

My main concerns are the scope of the empirical evidence, the computational cost, and the distinction from prior work.

  1. The strongest results require substantial inference-time computation. Although R²D uses fewer model evaluations than the evaluated ReMDM configurations, its cost remains much higher than standard decoding.

  2. A positive aspect is the sustained improvement as the correction budget increases. The experiments demonstrate a useful way to trade additional inference computation for accuracy. However, remasking-based inference-time scaling already exists in ReMDM, so the contribution should be framed around the improved scaling behavior and adaptive allocation of computation rather than the existence of this scaling axis itself.

  3. The applicability beyond structured reasoning tasks remains unclear. Most evidence comes from Sudoku and 3-SAT, supplemented by one coding benchmark. In open-ended generation, repeated identical predictions do not necessarily indicate higher quality, and multiple distinct outputs may be equally valid.

  4. The novelty of TCT relative to prior self-correction training methods, particularly PRISM, needs clearer discussion. Training on model-generated tokens to support self-correction is already established. The paper should isolate what these differences contribute through controlled comparisons.

Overall, I would need to see stronger evidence that the gains extend beyond specialized tasks and remain compelling under practical computational constraints.

Questions for authors

See the weakness section.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.