acceptodds
Under review as a conference paper at ICLR 2027

When Does State-Dependent Denoising Preserve Its Target?

Abstract

Diffusion language models adaptively choose which masked tokens to reveal, revise, or remask from the current model state. Even when every local update uses a valid conditional distribution, their state-dependent composition can change the distribution of completed sequences. We formalize this gap as a target-transport problem and identify two distinct boundaries. Specifically, for monotone generation, any predictable sequence of exact singleton or joint-block conditionals preserves the target, whereas independent parallel block updates are exact only when the selected conditional factorizes. For revisable generation, a state-dependent mixture of target-reversible kernels preserves the target exactly when its selector commutator vanishes. An action-ratio acceptance rule restores invariance for arbitrary positive selectors under reversible component kernels, while leaving within-block factorization error untouched. Empirically, exact finite systems verify these identities and exhibit total-variation defects up to 0.79 outside the target-preserving regimes. Normalized neural language targets reproduce the same separation: experiments with frozen LLaDA and Dream models further show endpoint divergence that grows with revision horizon, selector concentration, and remasking width, while state-independent controls remain matched. Together, these results distinguish routing error from block approximation and provide explicit criteria for target-preserving adaptive denoising.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Diffusion Language Models, Target Preservation, State-Dependent Denoising

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ejSWOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies whether adaptive decoding in dLLMs preserves the intended distribution of completed sequences. They show that joint-block reveal can preserve the target distribution while independently sampling tokens within a block may introduce error. So the authors apply a acceptance rule based on the ratio of action-selection probabilities before and after a remask in dLLMs. The experiments mainly measure target-distribution differences in the exact settings and endpoint differences between naive and corrected decoding in dLLMs. I highlighted the key sentence below.

Critical strengths
  1. I think the paper raises an interesting question about adaptive decoding in dLLMs. The distinction between selecting which tokens to update and predicting their new values is useful.
  2. The theoretical analysis is reasonably clear. The authors distinguish between monotone reveal and revision and provide explicit conditions for target preservation. I also like the separation between routing error and block factorization error, as these are two different issues that cannot necessarily be addressed by the same correction.
Critical weaknesses
  1. Unclear motivation for preserving the target distribution. My main concern is regarding the motivation part. I understand that adaptive revision can change the target distribution, but I am not convinced that why preserving this distribution is necessarily the right objective for dLLM refinement.
  • If revision is intended to correct incorrect tokens and improve reasoning, isn’t changing the original generation distribution may be actually desirable? For example, a revision may successfully replace an incorrect token with a correct one, making that position less likely to be revised again. However, your proposed MH correction could reject such a beneficial revision.

  • Thus, I think the paper needs to explain the logic for why preserving the target distribution is a must for practical dLLM decoding, rather than just assuming that distribution changes caused by adaptive revision are necessarily undesirable in the introduction, which is circular to some extent.

  1. The evidence that the proposed correction is actually beneficial is somewhat limited. Although the paper reports endpoint differences between naive and corrected revision, I do not think a larger Hamming distance directly demonstrates better target preservation or improved generation quality.
  • Specifically, Appendix Table 7 shows that surprise-based correction slightly reduces GSM8K accuracy compared with naive revision across all three checkpoints in the main setting: from 18.4 to 17.4 for LLaDA Instruct, 2.3 to 2.1 for LLaDA Base, and 3.7 to 3.5 for Dream. The 64-token LLaDA-Instruct result also decreases from 30.5 to 28.1. These differences are really small and not convincing enough to me and they do not demonstrate a practical benefit from the proposed correction. I wonder why the task-level results are not discussed more directly in the main text.
  1. The connection between token-level and block-level revision is not sufficiently demonstrated. The paper discusses token and block updates under a unified framework in block diffusion models yet the main revision experiments appear to update one token at a time, while block remasking is evaluated separately.
  • In my understanding, revising the tokens in one block can change the conditional distributions of other blocks and previously revised blocks may need to be revisited. It is unclear how these cross-block interactions affect the proposed correction, especially when blocks overlap or are reconstructed in parallel.

  • I can see appendix I.10 discusses composite block actions, which is good, but I would like to see a clearer connection between this formulation and the actual block remasking implementation. For example, when selecting 2 or 4 tokens from a candidate pool, how is the joint selection probability of the entire block computed? Is it calculated from the actual block selection process, or simply approximated by multiplying individual token selection probabilities?

  1. The experiments are limited. The real dLLM experiments only use three checkpoints on only three datasets WikiText, GSM8K, and MBPP. Most experiments focus on 32 token continuations, with a 64 token extension only for LLaDA.
  • I think these generation lengths are relatively limited for evaluating reasoning and code generation compared to existing dLLM studies. Although MBPP is included, the MBPP results are all N/A in Table 10 and I don’t know why there is no pass1 accuracy for MBPP. Therefore, I think the current experiments are largely limited compared to existing dLLM studies and not convincing.
  1. Limited method novelty. The proposed action ratio correction seems to follow the same MH detailed balance principle in previous work. I think the main contribution is applying these existing ideas to the context of dLLM decoding, rather than proposing a fundamentally new dLLM correction method. Given the simplicity of the correction, I would expect stronger evidence that it provides meaningful benefits in practical dLLM applications.
Questions for authors

Q1 Why is preserving the target distribution necessarily the right objective for adaptive dLLM refinement? If the goal is to correct incorrect tokens and improve reasoning quality, wouldn't changing the original generation distribution sometimes be beneficial? Could the MH correction reject useful revisions?

Q2 How do the authors explain the GSM8K results in Appendix Table 7, where surprise-based correction slightly reduces accuracy compared with naive revision across all three checkpoints?

Q3 Could the authors clarify how the proposed method handles dependencies between different blocks during iterative revision? In particular, if revising one block changes the conditional distributions of other blocks, how is this interaction handled? How is the joint selection probability of a block implemented and computed in the actual remasking experiments?

Q4 How exactly is surprise computed for an already generated token in LLaDA and Dream? Is the token masked before estimating its probability?

Q5 Could the authors provide experiments on longer reasoning and fix the NA results in coding tasks, and report the computational overhead and effective number of accepted revisions? This would help establish whether the proposed correction offers practical benefits. Also, I noticed that some experiments are 2 seed, some are 4 seeds and all results lack std. I strongly suggest using a consistent number of seeds across experiments and reporting the mean and standard deviation especially for GSM8K and MBPP to make the results more reliable and easier to compare.

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

The paper studies when adaptive decoding preserves the original target distribution. It distinguishes predictable monotone reveal using exact conditionals, state-dependent mixtures of reversible revision kernels, and independent block reconstruction.

Critical strengths
  1. The distinction between selection error and reconstruction error is useful and mathematically clear.
  2. The exact-target experiments provide a clean validation setting.
Critical weaknesses
  1. The frozen-model measurements do not identify the distributional errors claimed in several passages. Coupled endpoint Hamming measures sensitivity to a decoding intervention; it does not establish target bias or separately measure the routing and factorization terms. Although the main text acknowledges that Hamming is not TV, Appendix D.2 says the block panel measures the factorization term, and Appendix J claims to separate routing from component approximation. Crucial implementation details are also missing: candidate-pool generation, full reverse action probabilities, and the comparator used for remasking. The positive uniform-remasking Hamming values in Table 11 make the last ambiguity consequential. The promised paired bootstrap intervals are absent from the PDF.
  2. The theoretical contribution needs more positioning and applicability. The action-ratio correction is from previous cited papers. The finite-horizon bound is correct, but its strict one-step TV-contraction assumption fails for singleton updates on the stated product spaces: states differing at three positions have disjoint one-step supports. Consequently, the Hamming sweep is not a “direct finite-step test of Theorem 3.7.” These issues limit both the claimed theoretical advance and its demonstrated connection to practical decoding.
Questions for authors
  1. What are the two endpoints compared in Figure 6 and Table 11? Why does uniform remasking produce positive Hamming, whereas the matched naive-versus-corrected uniform control requires exact agreement?
  2. How are candidate pools constructed, and what complete forward and reverse action probabilities enter the acceptance ratio? If pools are state dependent or random, how are their probabilities handled? How are replacement conditionals queried?
  3. Which experimental kernels satisfy the contraction assumption in Theorem 3.7? Is the intended contracting kernel a single revision event or a multistep composition?
  4. Where are the promised paired uncertainty intervals, and what population and resampling unit do they represent? How are repeated seeds within a prompt handled?
  5. Which theoretical results are intended as new beyond the cited locally informed MCMC correction, and which frozen-model conclusions remain supported when interpreted strictly as coupled trajectory sensitivity?
Reviewer_vwkxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies when adaptive denoising preserves a declared target distribution. It distinguishes predictable monotone reveal, state-dependent revision, and independent parallel block updates. Exact singleton or joint-block reveal preserves the target under predictable selection, whereas independent block sampling requires conditional factorization to be exact. For revision with target-reversible component kernels, the authors characterize invariance through a selector-commutator criterion and apply an action-ratio acceptance rule to restore it. The paper also separates routing and within-block factorization errors and provides a finite-horizon bound under a contraction assumption. Exact finite systems and normalized language-model distributions on finite candidate lattices verify these preservation properties by complete enumeration. Experiments with frozen LLaDA and Dream models measure endpoint disagreement between uncorrected and corrected operators across selectors, revision horizons, and remasking configurations.

Critical strengths

Clear Distinction Between Routing and Factorization Errors: The paper clearly distinguishes predictable monotone reveal from state-dependent revision, and routing error from within-block factorization error. This distinction helps explain why adaptive selection can preserve the target during monotone reveal but require additional balance conditions during revision. It also clarifies why correcting action selection cannot recover dependencies discarded by independent parallel updates.

Controlled Evaluation with Exactly Known Targets: The finite-state experiments and normalized language-model targets allow direct evaluation of target preservation by complete enumeration. Comparing joint and independent block updates under corrected and uncorrected routing isolates the two sources of error. These experiments provide clear support for the theoretical distinctions without relying on sampled output differences as a substitute for distributional measurements.

Critical weaknesses

Limited Connection Between Frozen-Model Diagnostics and Target Preservation: The theoretical guarantees require component kernels that are reversible with respect to a common target, but this property is not established for the frozen LLaDA and Dream updates. The paper acknowledges that endpoint Hamming disagreement is not target TV. Nevertheless, the matched-tape comparisons by themselves establish that the acceptance rule changes realized outputs, rather than showing which operator is closer to an intended target distribution. Increasing disagreement with revision horizon can also arise from rejection-induced trajectory divergence under matched random tapes. The horizon sweep therefore does not distinguish accumulation of target-distribution bias from accumulation of path differences, and does not directly validate the TV bound in Eq. (12). A distributional diagnostic on a tractable restriction of the actual diffusion-model conditionals, together with a control that randomly rejects proposals at a matched acceptance rate, would help separate these explanations.

Limited Additional Theory Beyond Existing Locally Informed MCMC: The action-ratio acceptance rule in Eq. (10) and the reversibility guarantee in Theorem 3.5 coincide with Algorithm 1, Eq. (2), and Proposition 1 of Maire and Vandekerkhove (2022), which the paper appropriately cites. The concern is that the paper does not sufficiently distinguish inherited results from new contributions. The monotone-reveal result follows from the chain rule, the routing-factorization bound is a triangle inequality, and the finite-horizon bound follows a standard contraction recursion. The synthesis for adaptive denoising is useful, but the paper should explain more precisely what the selector-commutator characterization adds beyond the existing framework. Given the limited distributional validation on actual diffusion language models, the additional theoretical and empirical contribution is not yet sufficiently established.

Questions for authors

Target Preservation in Frozen Models: What common target, if any, are the frozen LLaDA and Dream revision kernels reversible with respect to? Can the authors evaluate conditional compatibility and target preservation on a tractable restricted state space using the actual diffusion-model conditionals? This would help establish whether the action-ratio correction improves target fidelity in these models.

Interpretation of the Horizon Sweep: How do the matched-tape results compare with a control that randomly rejects proposals at a matched acceptance rate? Can the authors provide a distribution-level measurement that distinguishes accumulation of target-distribution bias from rejection-induced trajectory divergence, and clarify the relationship between the Hamming horizon sweep and the TV bound in Eq. (12)?

Additional Theoretical Contribution: Could the authors explicitly distinguish results inherited from existing locally informed MCMC from new results? In particular, how do Eq. (10) and Theorem 3.5 relate to Algorithm 1 and Proposition 1 of Maire and Vandekerkhove (2022), and what additional insight or guarantee does the selector-commutator characterization provide beyond that framework?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.