acceptodds
Under review as a conference paper at ICLR 2027

Smarter Than Random, or Simply Busier? Where Diffusion-LLM Scheduling Earns

Abstract

Masked diffusion language models unmask several tokens per denoising step, and inference-time schedulers decide which positions to unmask. Every published scheduler is validated against a baseline that takes no action, so its gain cannot be told apart from the gain of simply acting more. We introduce the matched-action control, a random scheduler that takes exactly the same actions at randomly chosen positions, and run three rules against it across models, tasks and step counts. Under this control the gains separate into a substrate, the confidence ordering every scheduler shares, and an overlay, what each rule adds on top of it. The substrate is where the value lies, worth on open-ended text and to on conditional tasks on two models, while the overlay is exactly zero on open-ended text at every step count and appears on conditional tasks only at the most aggressive parallelism, on tasks that differ between models. The reason is a bound. Confidence ordering already maximizes the expected number of correct tokens under the model's own distributions, so a rule can add value only through dependence among the positions unmasked together, and that value cannot exceed the number of them that would change if they were unmasked one at a time. Measured directly, the bound is real and grows faster than linearly with the number of positions unmasked per step, yet the rules realize only 4% to 13% of it, because most of the changes it counts leave correctness untouched and the rest harm as often as they help. What choice is worth therefore lies in when rather than where. Unmasking few positions early and more later, with no rule and no extra cost, improves accuracy by up to on the same tasks, and the gain of the published method closest to our own turns out to be its schedule rather than its spacing.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
dllm

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_dA1ZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper’s main point is that existing evaluations of approaches that propose strategies for decoding masked diffusion language models conflate two sources of performance gains. The first is how many tokens to unmask at a given step; the second is how to decide which tokens to unmask at a given step. The main contribution is an evaluation framework to separate the measurement of performance gains. The “substrate” defines the performance gains stemming from confidence-based selection over random token selection when the number of tokens decoded at a given step is kept fixed. The “overlay” is defined as the additional gain stemming from a given method over confidence-based selection. A divergence upper bound is also included to measure the difference between the likelihood of the final tokens being changed compared to what would be selected based on confidence-ranking. Based on the results, the claim is that confidence-based selection is the primary source of performance gains and that benefits from more sophisticated selection criteria are not statistically significant. This leads to the conclusion that deciding the number of tokens to decode at a given step may matter more than the particular strategy for selecting tokens. While the paper’s evaluation framework is useful, I find the empirical validation and scope insufficiently deep. Evaluating a wider range of tasks, datasets, and models is needed to generalise the findings. Especially given that the primary contribution of this work is on the evaluation side. I do not believe the paper has sufficient empirical validation for acceptance.

Critical strengths
  1. The method is addressing an important problem in the language diffusion area regarding performance evaluation of decoding strategies.
  2. The paper proposes simple and straightforward metrics to disentangle the performance gains.
Critical weaknesses
  1. The scope of the paper is too limited. The paper's conclusions are based on tasks where the order of token unmasking has a limited effect. Sec. 7 itself notes that strongly structured tasks such as Sudoku may behave differently, since the order of unmasking can be important. Hence, the main issue is about whether the findings generalise beyond the evaluated settings. A broader evaluation across diverse models, datasets, and tasks is needed to establish when and to what extent sophisticated selection criteria provide benefits. Without this, the framework has limited practical value.

    1. There is insufficient background information on masked diffusion models and how the existing decoding strategies work. This is essential for readers who are from outside this domain because the manuscript must be self-contained.

    2. The math notation is overloaded. The symbol t is used for both the diffusion step and the tokens, e.g. t_{<j} in Proposition 2.

    3. The code is not available for review, which is necessary for review.

Questions for authors

See weaknesses W1-W4

Reviewer_X2uTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates whether gains from masked diffusion language-model decoding schedulers arise from selecting better token positions or from changing how many tokens are decoded at each step. It introduces matched-action randomized controls and distinguishes the contribution of confidence ordering, termed the substrate, from the additional contribution of a scheduling criterion, termed the overlay. It also distinguishes step-matched comparisons from comparisons at equal total forward-pass budgets.

Experiments use LLaDA-8B and Dream-7B, including instruction-tuned variants, across arithmetic, translation, extraction, code, and prose reconstruction settings. The paper reports substantial benefits from confidence ordering in several settings, limited additional gains from its delay heuristic, and stronger gains from adaptive decoding widths or combined slow-start and spatial-spacing schedules.

Critical strengths
  1. Separating the effect of which positions are selected from the effect of how many positions are decoded at each step is a useful methodological contribution. Recording and replaying per-input action profiles is more informative for this purpose than merely comparing methods at similar aggregate compute.

  2. The paper provides a useful empirical analysis of why additional token-selection heuristics often fail to improve decoding accuracy. Rather than only reporting negative results, the authors examine how often deferred tokens would actually change under sequential decoding and whether these changes improve correctness. The finding that most prediction changes do not translate into accuracy gains provides valuable insight into the limitations of dependency-based decoding heuristics.

Critical weaknesses

The paper’s own results also show that position choice can matter substantially at a fixed schedule. In Table 7, dilated positions improve GSM8K accuracy over confidence ordering by 0.301 and 0.442 for the two instruction-tuned models decoded in blocks; random ordering also outperforms confidence ordering in those settings. The authors discuss this regime dependence, but it is difficult to reconcile with the stronger wording in Section 5 that confidence ordering supplies the entire benefit of choosing positions. Spatial separation is itself a position-selection decision. The evidence therefore supports a more conditional conclusion about the interaction between schedule, confidence ordering, and spatial selection—not a general conclusion that additional position choice has negligible value.

Questions for authors
  1. Please provide pseudocode defining ktk_t, the gate, the random candidate pool, and the treatment of deferred positions in later and final steps. For the delay experiment, does the control randomize positions to be deferred or positions to be unmasked? How is the confidence-ordering component preserved while matching the actual unmasking counts?

  2. How should the substantial fixed-schedule position-selection gains in Table 7 be reconciled with Section 5’s characterization of confidence ordering as providing the entire benefit of position choice? Are spatially constrained policies outside the intended class of “criteria,” and, if so, what mathematical or operational distinction excludes them?

Reviewer_ukveOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates the value of position selection and unmasking schedules in diffusion LLM decoding. It introduces matched-action controls and a substrate/overlay decomposition, evaluates threshold and delay rules on LLaDA and Dream, and uses a parallel-versus-sequential divergence bound to explain their limited gains. The authors conclude that per-step token counts and spatial spacing matter more than additional ordering criteria, and propose combining slow-start scheduling with spaced confidence decoding.

Critical strengths

The paper makes an effort to control computational budgets and assess statistical reliability, using matched per-step unmasking counts, paired bootstrap confidence intervals, and checks for copying artifacts.

Critical weaknesses

1. The central distinction between "which tokens" and "how many" is not well justified.

Position selection and per-step unmasking counts are inherently coupled, as both determine subsequent contexts, confidence rankings, and decoding decisions. A policy may succeed precisely because it selects particular tokens at particular stages. The matched-action control replays the counts from one policy's trajectory while replacing its selections with random ones, providing a conditional comparison rather than an independent attribution of performance gains. The paper builds its central argument around this decomposition without sufficiently justifying its significance.

2. Confidence-greedy ordering is not necessarily optimal, even when decoding one token at a time.

The paper claims that "no criterion can improve on confidence ordering" when decoding one token per step. However, a zero divergence bound only means that there is no parallel-versus-sequential discrepancy within a single-token step. It does not mean that confidence is the best position to select first. Even when decoding one token at a time, different ordering policies can lead to different final outputs, and a less confident first choice can produce a better result.

Proposition 1 only addresses immediate expected token correctness, not the final decoding outcome. Appendix A.9 also acknowledges that the accumulated bound does not apply to the full decoding process, since alternative policies change subsequent states. The theoretical analysis therefore does not support the paper's broader conclusions about the limited value of alternative ordering policies.

3. The evaluated heuristics are too limited to support general conclusions about token ordering.

Section 5 mainly evaluates confidence thresholding and an adjacency-based delay rule. Thresholding primarily changes decoding counts, while the delay rule is a simple local modification of confidence ordering. Their limited gains say little about other decoding policies. For example, KLASS [1] uses prediction stability, DAPD [2] explicitly considers token dependencies, and learned ordering methods [3, 4] optimize position selection. These approaches are not represented by the confidence-based overlays evaluated in this paper, yet the authors draw much broader conclusions about token ordering.

The scheduling analysis also overlooks UGC [5], which provides a theoretical perspective on data-dependent unmasking schedules. The paper does not adequately establish how its conclusions extend beyond the specific heuristics considered.

[1] S. H. Kim et al., KLASS: KL-Guided Fast Inference in Masked Diffusion Models, in NeurIPS 2025.

[2] B. Kim et al., DAPD: Dependency-Aware Parallel Decoding via Attention for Diffusion LLMs, in ICML 2026.

[3] M. Jazbec et al., Learning Unmasking Policies for Diffusion Language Models. in ICML 2026.

[4] J. Xu et al., Scheduling Thoughts: Learning the Order of Thought in Diffusion Language Models. in ICML 2026.

[5] M. J. Wainwright. The Data Geometry of Masking Diffusion: Certified-Optimal Schedules via Unmasking Growth Complexity. 2026.

4. The proposed solution makes the main conclusion less convincing.

The paper argues that additional ordering criteria offer limited benefits beyond confidence, yet Section 6.4 reports substantial gains from spatial spacing, which is itself a position-selection rule. Under the same dilated schedule, spacing improves instruction-tuned GSM8K accuracy over confidence ordering by 30.1 and 44.2 percentage points.

This raises a question about the distinction between ordering criteria and spatial spacing, since both determine which positions are decoded together. Moreover, the strong interaction between slow start and spacing reported in Section 6.4 shows that the value of position selection depends on the schedule, directly contradicting Section 5's claim that confidence alone captures the benefit of token selection. Meanwhile, the proposed solution largely builds on ideas already present in dilated scheduling.

5. The treatment of EOS failures raises concerns about the experimental comparison.

The paper uses whole-block decoding for base models but 32-token semi-autoregressive decoding for instruction-tuned models, motivated by premature EOS and padding predictions under whole-block confidence decoding. Rainbow Padding [6] has already analyzed this failure, yet is not cited despite its direct relevance to this experimental choice.

More importantly, blockwise decoding imposes a left-to-right ordering constraint that prevents some of these failures. Confidence is therefore evaluated within an already restricted set of positions rather than across the full output. Since this block restriction masks the global failure mode of unrestricted confidence ordering, evaluating within blocks—even against block-restricted random baselines—does not support general claims about confidence-based selection.

[6] B. Kim et al. Rainbow Padding: Mitigating Early Termination in Instruction-Tuned Diffusion LLMs. in ICLR 2026.

6. The paper is difficult to follow.

The introduction presents detailed numerical results and exceptions before clearly motivating the research question. The decoding rules, controls, and comparisons are explained across multiple sections, with frequent cross-references that interrupt the logical flow. In particular, understanding what each reported gain measures often requires looking elsewhere for the corresponding baseline and experimental setting. This fragmented presentation makes the paper's central argument unnecessarily difficult to follow.

Questions for authors
  1. What justifies treating position selection and per-step unmasking counts as separable sources of improvement when each step's selection alters subsequent contexts and step counts—and why should comparisons conditioned on a specific policy's count trajectory generalize?

  2. At one token per step, a zero divergence bound only means there is no parallel-versus-sequential discrepancy within a single step. How does a within-step bound of zero rule out alternative ordering policies achieving better final accuracy across the full decoding trajectory?

  3. On what grounds do the authors distinguish spatial spacing from a position-selection criterion, given that both decide which tokens are decoded together? How do the authors reconcile the large gains from spacing under fixed per-step counts with the claim that confidence ordering already captures the value of token selection?

  4. Conditioning can shift confidence rankings and change which positions are selected in subsequent steps, even when current argmax predictions remain unchanged. Since the divergence bound only counts within-step argmax flips, what justifies interpreting its recovery rate as the remaining potential for improving token ordering?

  5. Whole-block confidence ordering fails on instruction-tuned models due to premature EOS/padding predictions, forcing the authors to rely on a 32-token left-to-right block constraint. When confidence ordering requires an imposed semi-autoregressive structure to prevent global failures, how can block-restricted evaluations support general conclusions about its sufficiency?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.