acceptodds
Under review as a conference paper at ICLR 2027

Measuring and Reducing the Train–Inference Mismatch in Discrete Diffusion Models

Abstract

In discrete diffusion, we train a denoiser model on states from a noising process (i.e., on corrupted data). At inference time, the model has to reverse this process, iteratively denoising data, and is thus exposed to states created by its own samples. This creates a train–inference mismatch: the discrepancy between the distribution over states presented to the model during training and inference. We formalize this mismatch as the Mirror Gap. We measure this gap using three measures of discrepancy: the discriminability of training and inference states, the distance between their representation centroids, and the divergence between their quantized representation distributions. Together, these measurements give us a window into how the mismatch behaves at different resolutions and across the diffusion process. Our analyses reveal the train–inference mismatch has a low-dimensional component in representation space; they also show correlations between discriminability scores early in sampling and final sample quality. These results motivate two strategies to reduce this gap: steering generations in representation space, and resampling states based on measured gap. Experimentally, we show these strategies improve the generation compute–quality curve: at the same quality (measured by MAUVE or FID), our methods show sampling-efficiency gains exceeding 8× over strong baselines while requiring no denoiser retraining, fine-tuning, or distillation; this is true in both text and image generation, and across noising processes and samplers.

Primary area:
generative models
Keywords:
discrete diffusion models, train-inference mismatch, generative modeling, sampling, exposure bias, inference-time scaling, representation steering

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_4RQaOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies the train–inference mismatch in discrete diffusion: the denoiser is trained on forward-corrupted states but at inference sees states produced by its own sampler. The mismatch is defined as a divergence between the forward and sampler marginals at each time (the "Mirror Gap") and measured in the denoiser's hidden space with three proxies: a forward-vs-reverse probe, a standardized centroid shift and a quantized-cell JS divergence. Two training-free interventions follow: Recenter, which subtracts a calibrated rank-one hidden-state shift at one block, and MirrorGate, which rejects transitions flagged by per-layer classifiers. Experiments on OpenWebText (masked and uniform-state models, several samplers), CIFAR-10 and a 3B model report large MAUVE and FID gains at small step budgets.

Critical strengths

Recenter is simple, free at inference and consistently effective: one vector addition per step improves MAUVE several-fold across two noise families, four samplers, two modalities and three model scales (Tables 1 and 7, Fig. 1), with reasonable controls (random and opposite directions, temperature, an LLM judge). The effect is large and replicated widely enough that I expect it to survive better-tuned baselines, which is why I lean toward acceptance despite the concerns below. The probe staying saturated even at T = L (Fig. 2) is a useful indication that parallel token draws are not the whole source of the mismatch.

Critical weaknesses

My main concern is that the methodological contribution is considerably simpler than the presentation suggests. The paper acknowledges that the Mirror Gap is the mismatch studied by Ning et al. and Park et al. and that Recenter is activation steering, yet the framing ("formalise", "fundamental yet overlooked"), the pushforward notation, the three lenses and Theorem 1 suggest more. No quantity measured in Sections 4–5 is a KL, so Section 3 and Theorem 1 are decoupled from what is measured and intervened on. The transfer results also undercut the measurement framework: the MDLM correction transplanted to ReMDM matches destination-calibrated Recenter (Table 3), and the frozen T = 32 correction applied at T = 64 and 128 beats it (0.119/0.222 vs 0.078/0.182, Table 4 vs Table 1). A single generic direction seems to do the work, so per-configuration calibration and the lens apparatus are not what produces the gain. I would prefer Recenter presented plainly as inference-time steering toward the forward distribution, with the analysis as motivation.

Second, the headline comparisons are not matched. γ is selected from 61 values on 5k validation generations per configuration, with budget-specific overrides, while the baselines keep their released η and temperature at every T. The one matched control, temperature (Fig. 15), reproduces the GenPPL gain but at lower MAUVE and entropy, which supports the paper but is not a tuned baseline. Table 1 reports no uncertainty for 5k-sample MAUVE, and its T = 128 MDLM and ReMDM rows coincide with Wang et al. (2025) while the Recentered rows are new runs (Table 4's own native MDLM at T = 128 is 0.017, not 0.015), which should be stated. The text 8× and image 8×/16× efficiency factors compare against numbers quoted from other papers at 1024 and 4096 NFE rather than baselines run in this pipeline, and "zero extra NFE" omits the offline calibration and γ-sweep cost per configuration. Minor: the "11× MAUVE improvement" (p. 8) matches no summary of Table 1, where ratios range from 2× to 15× over baselines near zero, and MirrorGate's gains are an order of magnitude smaller than Recenter's.

Questions for authors
  1. Does a single global forward-vs-reverse direction with constant amplitude and no time schedule match the gains in Table 1? If so, what do the three lenses and per-(sampler, T) calibration contribute?
  2. Were the ReMDM and Ψ-sampler η and the temperature tuned on validation MAUVE at each T, as γ was? What is the margin over baselines tuned this way?
  3. Please run native ReMDM and Ψ-sampler to 1024 and 4096 NFE in your pipeline, mark which Table 1 cells are quoted, and add generation-seed uncertainty.
  4. Does the denoiser's per-token NLL on reverse states actually improve under Recenter? This is the mechanism the motivation asserts, and the AR result in Fig. 30 suggests a generic quality-steering effect instead.
4 comments

We thank the reviewer for their constructive feedback and we are glad they appreciated the fact that our Recenter intervention is simple, free at inference yet consistently very effective. We are also glad that the reviewer values the empirical breadth of our study and highlights that the effects we find are large, with several-fold improvements and are widely replicated.

I would prefer Recenter presented plainly as inference-time steering toward the forward distribution, with the analysis as motivation.

We agree that the methodological contribution of our Recenter intervention is simple. We hope that its simplicity and effectiveness across settings will motivate others to adopt our method, similarly to simple inference-time methods like top-p nucleus sampling. Beyond that, we do think the lenses and Theorem 1 provide an interesting point of view and new tools to observe the mismatch that we haven't seen explored before. Notably, the centroid lens is what enabled us to discover our effective intervention. It revealed the low-rank structure of the shift between forward and reverse mean activations, without which a steering direction would have been hard to find. We value communicating to our reader the process through which we uncovered these interventions, in the hope that they may build upon it and improve it further.

No quantity measured in Sections 4–5 is a KL, so Section 3 and Theorem 1 are decoupled from what is measured and intervened on.

We thank the reviewer for pointing out the gap between the KL analysis in Section 3 and the empirical measurements in Sections 4–5, which use JS divergence. We have added corresponding figures that measure the gap using KL on the Voronoi cells, making the connection to Section 3 more explicit.

W2. We address this in Q1,2,3,4 below.

Does a single global forward-vs-reverse direction with constant amplitude and no time schedule match the gains in Table 1? If so, what do the three lenses and per-(sampler, T) calibration contribute?

Q1. The correction indeed transfers very well within noises for different samplers. Thus, a direction does work well in practice within a noise family across budgets and samplers, which we see as a strength. We additionally provide the analysis in the appendix to show its effectiveness across a wide range of configurations, as well as to detail when it works. We encourage the reader to apply the simplest setting, which we detail in the main text.

Were the ReMDM and Ψ-sampler η and the temperature tuned on validation MAUVE at each T, as γ was? What is the margin over baselines tuned this way?

Q2. We use the sampler settings studied in [2, 3] which have been extensively swept and carefully tuned. For ReMDM, [2, Appendix E.2.4, Figs. 6–8] sweeps 14 settings at each of three fast-sampling budgets (T=128,256,512T=128,256,512), for 42 setting–budget combinations. At T=128T=128, even taking the best MAUVE across all 14 settings (approximately 0.065), Recenter reaches 0.405, an absolute gain of 0.340. For Ψ, [3, Appendix D.1.2, Table 13] evaluates 21 configurations per budget across eight budgets, or 168 setting–budget combinations for Duo alone (336 including MDLM). We use the same sampling and evaluation pipeline as [3] and adopt the strongest Ψ setting the authors recommend based on the sweep.

Q1–Q2 Additional Experiments. To test the reviewer’s suggestion directly, we use a single direction with constant amplitude on a new larger checkpoint, Dream‑7B‑Base [4]. We estimate the direction at block 25 from just 512 trajectories of the original sampler and apply the same vector at every step. We set the gain to 1 throughout, with no per-(sampler,T) gain tuning and reuse it across all budgets and both the original sampler and ReMDM (remasking), with top‑p = 0.9, temperature = 1, and ReMDM η = 0.04. The table reports MAUVE (higher is better) with Dolma [5] as the reference corpus (used to pretrain Dream‑7B [4]) on 5,000 samples of length 256 per point. Parentheses give the absolute gain (margin) over the corresponding sampler.

SamplerSettingT = 32T = 64T = 128
Dream7BBase0.0400.0560.073
Recenter0.253 (+0.213)0.316 (+0.260)0.351 (+0.278)
Dream7B + ReMDMBase0.2050.2730.305
Recenter0.396 (+0.191)0.538 (+0.266)0.525 (+0.220)

A single direction therefore strongly improves MAUVE in all comparisons, with no per(sampler,T) tuning.

[continued in next comment]

Q3. We appreciate and share the reviewer’s concern for statistical rigor and reproducibility. We built our methods directly on the codebases/pipelines released by the authors of [2] and [3], we used these pipelines to rerun the native baselines underlying our sampling-efficiency comparisons, so our comparisons are indeed matched in implementation and have all been reproduced across seeds. With respect to the transfer ablation in Table 4 highlighted by the reviewer: at T=128T=128, we obtain 0.0174 MAUVE versus 0.015 in [2]. Across our 10 generation seeds, the Recenter gain was +0.2045 ± 0.0186 MAUVE (mean ± SD), so the 0.0024 discrepancy is only 1.2% of the mean gain and is consistent with generation-seed variability of MAUVE at values close to 0.

Q4. Our motivation is that the measured discrepancy between forward and reverse states can guide the choice of a steering direction that improves final sample quality. An unconditional reverse trajectory has no ground-truth target for its missing tokens.

One fixed direction and unit gain substantially improved MAUVE across both Dream‑7B samplers at every budget tested, without gain, temperature, direction tuning. We hope these gains and the paired-seed results address the reviewer's main concerns and give them enough reason to consider raising their recommendation. If there are any further questions held by the reviewer, we remain eager to engage further.

[1] Fred Zhangzhi Peng, Zachary Bezemek, Jarrid Rector-Brooks, Shuibai Zhang, Anru R. Zhang, Michael Bronstein, Alexander Tong, and Avishek Joey Bose. Planner aware path learning in diffusion language models training. In International Conference on Learning Representations, 2026. URL https://arxiv.org/abs/2509.23405. Oral.

[2] Guanghan Wang, Yair Schiff, Subham Sekhar Sahoo, and Volodymyr Kuleshov. Remasking discrete diffusion models with inference-time scaling. In Advances in Neural Information Processing Systems, 2025. URL https://arxiv.org/abs/2503.00307.

[3] Justin Deschenaux, Caglar Gulcehre, and Subham Sekhar Sahoo. The diffusion duality, chapter II: Ψ-samplers and efficient curriculum. In International Conference on Learning Representations, 2026. URL https://arxiv.org/abs/2602.21185.

[4] Jiacheng Ye, Zhihui Xie, Lin Zheng, Jiahui Gao, Zirui Wu, Xin Jiang, Zhenguo Li, and Lingpeng Kong. Dream 7B: Diffusion large language models. arXiv preprint arXiv:2508.15487, 2025. URL https://arxiv.org/abs/2508.15487.

[5] Luca Soldaini et al. Dolma: an Open Corpus of Three Trillion Tokens for Language Model Pretraining Research. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2024. URL https://aclanthology.org/2024.acl-long.840/.

Thanks, the Dream-7B result is a nice addition. The comparison with properly tuned baselines is still missing, so I keep my score.

We thank the reviewer for their follow-up. We have now completed the requested comparison against extensively tuned baselines:

We swept η over the dense grid {0, 0.0025, …, 0.15}, comprising 61 configurations per sampler and per budget, for both ReMDM and Duo++ with the Ψ-sampler at T ∈ {32, 64, 128}: 366 sampler–budget–η combinations in total. As for γ, we selected η separately at each budget using validation MAUVE on 5,000 generations. (Throughout the main text, temperature is fixed at 1 for all models and samplers and is not swept in any comparison.)

For ReMDM, none of the swept configurations exceeded the MAUVE of the η = 0.04 setting selected through Wang et al.’s [2] own sweep and used in Table 1. We therefore retain this setting. For the Ψ-sampler, the selected η values were 0.0975, 0.125, and 0.1025 at T = 32, 64, and 128, respectively.

We evaluated each setting using 10 fresh generation seeds, each producing 5,000 samples. The table below reports mean MAUVE (averaged over 20 reference-sampling seeds and 10 clustering seeds per generation replicate).

SamplerTηBaseline MAUVERecenter MAUVEAbsolute margin
ReMDM320.04000.0070.055+0.048
ReMDM640.04000.0160.211+0.195
ReMDM1280.04000.0570.405+0.348
Duo++ + Ψ320.09750.0320.368+0.336
Duo++ + Ψ640.12500.1050.448+0.343
Duo++ + Ψ1280.10250.2210.489+0.268

Across all baseline settings, the generation-seed standard deviation was at most 7.7% of the corresponding Recenter margin shown above, indicating that baseline generation-seed variability is small relative to the observed improvements.

This extensive baseline-tuning sweep confirms the large gains from Recenter: for instance, 13.2× the MAUVE of ReMDM at T = 64 and 11.5× for the Ψ-sampler at T = 32.

We thank the reviewer for their patience, as this was a long experiment to run. We hope this now addresses the reviewer’s last reservation and gives them enough reason to consider raising their score.

Reviewer_UFotOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies the train–inference mismatch in discrete diffusion models: the discrepancy between the forward (noised-data) states seen during training and the reverse states produced by the model's own sampling. It formalizes this as the Mirror Gap and measures it with three representation-based discrepancies (discriminability, centroid distance, quantized distribution divergence). It finds that the mismatch has a low-dimensional component in representation space. Based on this, it proposes two training-free interventions: Recenter, which steers hidden representations toward the forward centroid along the leading drift direction, and MirrorGate, which resamples states rejected by per-layer discriminators. Experiments on masked and uniform diffusion, multiple samplers, text and images report improved compute–quality trade-offs.

Critical strengths
  • The train–inference mismatch, i.e., exposure bias, is an important problem for diffusion models. Extending its study to discrete diffusion is meaningful to the community.
  • The paper is well written, and the formulation of the mismatch at the level of state marginals is clean.
  • Recenter is simple and adds almost no inference cost, yet it is effective across noise types, samplers, modalities and model scales.
  • The experiments are comprehensive, with sufficient ablations (direction, rank, layer, schedule, transfer, temperature control).
Critical weaknesses
  • Limited novelty. The mismatch studied here is the well-known exposure bias, previously studied in autoregressive and continuous diffusion models, and also in discrete diffusion with training-side remedies. Recenter is essentially contrastive activation steering with a time-dependent magnitude. The main new elements are the marginal-level measurement and its application to discrete diffusion.
  • Relatively weak evaluation. Most results are in a low-quality regime where baseline MAUVE is close to zero (e.g., 0.006–0.015 for MDLM at T=32–128), so the headline efficiency multipliers are inflated. In higher-quality settings the gains are clearly smaller (e.g., 0.599→0.695 on the 3B masked model at T=64). More evidence that the method substantially improves stronger models or higher-quality regimes would make the paper much stronger.
  • Link between mismatch and prediction error is missing. The paper shows that reverse states are distinguishable from forward states. However, it does not show how this mismatch affects the denoiser's prediction error, or how such errors propagate during sampling. I would encourage the authors to provide more discussion along this line.
Questions for authors

Please address weaknesses 2 and 3 above.

2 comments

We thank the reviewer for their comprehensive feedback. We are glad they see our work as a valuable contribution to the community.

W1. Recenter is essentially contrastive activation steering with a time-dependent magnitude. The main new elements are the marginal-level measurement and its application to discrete diffusion.

We agree that Recenter builds on activation steering. Our contribution is how we derive the steering direction. Conventional contrastive steering uses examples of desired and undesired behaviours, while Recenter uses the model’s own train–inference mismatch. Theorem 1 motivates the marginal-level measurement the reviewer highlights by showing that identical expected stepwise KL errors can lead to different discrepancies between forward and reverse state distributions. We find that the shift between their mean activations across diffusion times is low-rank, and Recenter uses its leading direction to steer toward the forward mean. We also find that higher mismatch scores early in sampling predict lower final MAUVE across trajectory groups (Fig. 3), motivating MirrorGate’s use of these scores to decide which transitions to resample.

More evidence that the method substantially improves stronger models or higher-quality regimes would make the paper much stronger.

We follow the reviewer's advice by providing further experiments on Dream‑7B‑Base [4], a widely used 7B-parameter masked diffusion language model. We are happy to share that our results persist at that larger scale as detailed below. We thank the reviewer for this suggestion and believe it materially strengthens the generality of our findings.

We report MAUVE (higher is better) with Dolma [5] as the reference corpus (used to pretrain Dream‑7B [4]) on 5,000 samples of length 256 per point. We use block 25 as our intervention site. Gain 0 is the baseline for each sampler. Bold indicates the highest MAUVE among the gains shown.

SamplerRecenter gainT = 32T = 64T = 128T = 256 (= L)
Dream7B0 (base)0.0400.0560.0730.086
10.2390.3030.3550.335
20.3680.3760.3440.358
Dream7B + ReMDM0 (base)0.2050.2730.3050.298
10.4580.5500.6030.630
20.6420.6510.7040.595

We use top‑p = 0.9, temperature = 1, and ReMDM η = 0.04. The improvement persists at T = L = 256 where the numbers of steps matches the sequence length, a high quality regime, and also on ReMDM (remasking).

On a related note:

W2. “Relatively weak evaluation. Most results are in a low-quality regime where baseline MAUVE is close to zero (e.g., 0.006–0.015 for MDLM at T=32–128), so the headline efficiency multipliers are inflated.”

The evaluations provided in our paper are, to our knowledge, the standard regime of comparison in the literature and current state-of-the-art in DDMs. We have chosen them for comparability to contemporary works ([1], [2], [3]). One of the main appeals of DDMs is to be able to generate in fewer steps than AR which is restricted to one token per step. E.g., in Table 1, the regime of T32 - 128 out of a 1024 token sequence corresponds to a fast sampling regime, which aims to generate several tokens per step in parallel [1]. We think this regime matters because control over inference-time compute is one of the main motivations behind DDMs, and on Duo++ with the Ψ-sampler, Recenter achieves in 32 steps a MAUVE score that the baseline does not reach with 512 steps. The gains also persist at higher budgets. For instance at 512 steps (T=L/2), Recenter raises MAUVE from 0.350 to 0.651 on ReMDM and from 0.340 to 0.606 on Duo++ with the Ψ-sampler (Fig. 1, right).

The point raised by the reviewer about comparing raw values of MAUVE is also important. While such multipliers have been often reported as headline metrics ([1], [2]), it is worth clarifying that since MAUVE scores are strictly between 0 and 1, such multipliers can indeed feel arbitrary. Our gains reported on the 3B models (for instance 0.599->0.695 at T64 on the masked model) are thus still meaningful since they bridge about 24% of the remaining gap to the metric's saturation ceiling of 1.

[continued in next comment]

W3. The paper shows that reverse states are distinguishable from forward states. However, it does not show how this mismatch affects the denoiser's prediction error, or how such errors propagate during sampling. I would encourage the authors to provide more discussion along this line.

We agree that more discussion of the relationship between mismatch and error is a valuable addition, and we thank the reviewer for this suggestion. During training, we corrupt a known clean sequence, so we can evaluate a prediction against the original token. Since inference states have no associated ground-truth clean sequence, directly measuring denoising prediction error is not well-defined. Therefore, since the objective of generative modeling is to reproduce the target distribution, we use MAUVE to assess how closely the generated distribution matches the reference. Figure 3 shows that greater mismatch early in sampling is associated with lower final MAUVE. We designed our interventions specifically to reduce the mismatch, applying them reduces the measured gap and improves generation quality, providing evidence that targeting the mismatch improves the samples the model generates. If the reviewer has a different notion of prediction error in mind, we would be happy to discuss what ground truth would be appropriate here. Figure 13 shows how this mismatch develops over successive sampling steps. Even when sampling starts from forward-corrupted data, representations drift away from the corresponding forward reference at every noise level, including when the sampler reveals only one token at a time.

Additionally, we provide a new experiment on how mismatch relates to errors in fluency. We sample 5,000 native MDLM continuations from clean OpenWebText passages forward-noised to t=0.75, with the number of remaining steps sampled uniformly from {1,...,128}. We group trajectories into deciles by their current mismatch classifier score. We then measure external GenPPL on the tokens generated at positions that were still masked when we measured mismatch. As the table below shows, trajectories with higher mismatch generally have higher downstream GenPPL.

Mismatch decileGenPPL ↓
1 (lowest)32.90
246.29
348.74
449.71
552.17
654.43
755.26
857.85
960.89
10 (highest)67.72

We observe a clear ordering: downstream GenPPL increases across all ten deciles. The mismatch measured by the classifier thus predicts the fluency of tokens the model has yet to generate.

We hope the Dream‑7B gains at higher baseline MAUVE and our clarification of the link between mismatch and prediction error address the two points the reviewer identified as material to their assessment and give them enough reason to consider raising their recommendation. We would be happy to discuss any further questions the reviewer may have.

[1] Fred Zhangzhi Peng, Zachary Bezemek, Jarrid Rector-Brooks, Shuibai Zhang, Anru R. Zhang, Michael Bronstein, Alexander Tong, and Avishek Joey Bose. Planner aware path learning in diffusion language models training. In International Conference on Learning Representations, 2026. URL https://arxiv.org/abs/2509.23405. Oral.

[2] Guanghan Wang, Yair Schiff, Subham Sekhar Sahoo, and Volodymyr Kuleshov. Remasking discrete diffusion models with inference-time scaling. In Advances in Neural Information Processing Systems, 2025. URL https://arxiv.org/abs/2503.00307.

[3] Justin Deschenaux, Caglar Gulcehre, and Subham Sekhar Sahoo. The diffusion duality, chapter II: Ψ-samplers and efficient curriculum. In International Conference on Learning Representations, 2026. URL https://arxiv.org/abs/2602.21185.

[4] Jiacheng Ye, Zhihui Xie, Lin Zheng, Jiahui Gao, Zirui Wu, Xin Jiang, Zhenguo Li, and Lingpeng Kong. Dream 7B: Diffusion large language models. arXiv preprint arXiv:2508.15487, 2025. URL https://arxiv.org/abs/2508.15487. [5] Luca Soldaini et al. Dolma: an Open Corpus of Three Trillion Tokens for Language Model Pretraining Research. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2024. URL https://aclanthology.org/2024.acl-long.840/.

Reviewer_Nzx8OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The authors introduce the concept of a "Mirror Gap" which they observe from the train-inference mismatch in discrete diffusion models. They extensively validate their claim and observation through mathematical and empirical analysis. Based on this insight they notice a low rank similarity which they use to address the problem through steering and re-sampling. One approach is pure centering and another requires a classifier for re-sampling.

Critical strengths
  • The core observation is well observed and mathematically and empirically analyzed
  • The evaluation covers a set of different input modalities
  • Carefully and well executed empirical study
Critical weaknesses
  • Limited novelty: train-inference (exposure-bias) mismatch is a well-known problem [1] (maybe not in the area of discrete diffusion but definitely in the autoregressive ml area) and the strategies used to mitigate the problem are usually similar to the proposed solution. The authors cite the work but the distinction between the prior work is not quite clear for me - this is the main limiter for my rating and needs to be addressed for a higher evaluation. Seeing that the contribution seems to be a well-engineered solution it requires the previous point to be addressed for a higher rating
  • Please provide a clear distinction between dLLM, LLMs solution for the training-inference mismatch and stronger contrast between your contribution and the prior steering/re-sampling works to mitigate the issue

Minor:

  • How does this discrepancy affect low-rank representation of the models (does it affect common quantization strategies for example etc.)
  • Limited scale of models in the evaluation

Writing quality:

  • The abstract does not motivate the topic on addressing the main concerns similar to a technical report rather than an academic work.
  • In general, the paper addresses the core problem of which is being addressed in this work, but never actually motivates as to why this discrepancy is actually negative/positive or in which scenario this problem lead to large decay etc.
  • Acronyms like KL are used in the introduction, but not defined until approach section.

[1] Fred Zhangzhi Peng, Zachary Bezemek, Jarrid Rector-Brooks, Shuibai Zhang, Anru R. Zhang, Michael Bronstein, Alexander Tong, and Avishek Joey Bose. Planner aware path learning in diffusion language models training. In International Conference on Learning Representations, 2026. URL https://arxiv.org/abs/2509.23405. Oral.

I am generally open to changing my ranking if valid points are made by the authors.

Questions for authors

Have been asked under weakness section.

4 comments

We thank the reviewer for their constructive feedback, for highlighting the breadth of our experiments, and for acknowledging that our work is well executed.

train–inference mismatch is a well-known problem

Please provide a clear distinction between dLLM, LLMs solution for the training-inference mismatch and stronger contrast between your contribution and the prior steering/re-sampling works to mitigate the issue

We agree that this should be made clearer in our work. As the reviewer mentions, exposure bias is indeed a well-known problem.

Train-inference mismatch in autoregressive vs. diffusion models. Autoregressive (AR) models generate text from left to right. In discrete diffusion models (DDMs), the noising process determines which states the model sees during training at each noise level, and the sampler determines which states it visits at inference. Thus, for a denoiser, beyond simply model quality, the choice of noising process and sampler has a strong impact on the train–inference mismatch. For example, under uniform noise, the training distribution can contain plausible but incorrect tokens, which may enable self-correction at inference and have less mismatch. Standard left-to-right AR models have no analogous choice of noising process/sampler---with the mismatch deriving directly from model quality---so findings from AR do not always transfer directly to diffusion. To design a noising process or a sampler that reduces the mismatch, we first need to measure and understand how it evolves across diffusion time. Our work takes a step in that direction.

Comparison to prior work in diffusion models. With respect to the literature within discrete diffusion, we compare our work to Peng et al. [1] and Park et al. [2] in Section 3.2 and with Theorem 1. Notably, as referred to by the reviewer, [1] addresses part of this mismatch with their method, PAPL, but: (i) through a different pathwise quantity and (ii) in the mask noise setting only. Further, as shown in our Table 1, we outperform PAPL's headline metrics (0.182 MAUVE against 0.067 at T128) even though their method (unlike ours) requires retraining the checkpoint.

Mismatch over time. We also measure how this mismatch evolves over time. Prior work recognizes the train–inference mismatch in discrete diffusion but provides no way to measure it across generation/time. We introduce three such measures. That the mismatch is measurable and strong is itself nontrivial, especially under uniform noise, where one might have expected a more subtle effect. This raises the question of how we might design better noising kernels to mitigate it. Our measurements also reveal a structure in the mismatch: the shift between forward and reverse mean activations could span up to T directions over T steps, yet it concentrates in just a few. To our knowledge, this low-rank structure has not previously been reported in discrete diffusion or AR models, nor exploited for inference-time correction. Finally, while definitely related, unlike conventional activation steering [3], which typically uses contrasts between desired and undesired behaviours and specific heuristics, Recenter derives its direction from the discrepancy between the model's own training and inference distributions. The scope of our work also extends beyond language models to discrete diffusion more broadly, as shown by our experiments on images.

We will make these comparisons clearer in our paper.

Limited scale of models in the evaluation

We agree with this point, and have now evaluated Recenter on Dream‑7B‑Base [4], a widely used 7B-parameter masked diffusion language model. We are happy to share that our results transfer well, as detailed below.

We report MAUVE scores (higher is better) on Dolma [5] as the reference corpus on 5,000 samples of length 256 per point. (Dolma is a dataset used to pretrain Dream‑7B [4].) Recenter is applied at layer 25. Gain 0 is the baseline for each sampler. Bold indicates the highest MAUVE among the gains shown. We use top‑p = 0.9, temperature = 1, and ReMDM η = 0.04.

SamplerRecenter gainT = 32T = 64T = 128T = 256 (= L)
Dream7B0 (base)0.0400.0560.0730.086
10.2390.3030.3550.335
20.3680.3760.3440.358
Dream7B + ReMDM0 (base)0.2050.2730.3050.298
10.4580.5500.6030.630
20.6420.6510.7040.595

Writing quality

We will try to improve on those points for CR, and we will introduce the KL acronym.

We hope these additions and clarifications resolve the main objection to a higher score. Please let us know if any substantive concerns remain. We will address them during the discussion or camera-ready at the latest.

[continued in next comment]

[1] Fred Zhangzhi Peng, Zachary Bezemek, Jarrid Rector-Brooks, Shuibai Zhang, Anru R. Zhang, Michael Bronstein, Alexander Tong, and Avishek Joey Bose. Planner aware path learning in diffusion language models training. In International Conference on Learning Representations, 2026. URL https://arxiv.org/abs/2509.23405. Oral.

[2] Yong-Hyun Park, Chieh-Hsin Lai, Satoshi Hayakawa, Yuhta Takida, and Yuki Mitsufuji. Jump your steps: Optimizing sampling schedule of discrete diffusion models. In International Conference on Learning Representations, 2025. URL https://arxiv.org/abs/2410.07761.

[3] Jinghao Zhang, Yuting Liu, Wenjie Wang, Qiang Liu, Shu Wu, Liang Wang, and Tat-Seng Chua. Personalized text generation with contrastive activation steering. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2025. URL https://aclanthology.org/2025.acl-long.353/.

[4] Jiacheng Ye, Zhihui Xie, Lin Zheng, Jiahui Gao, Zirui Wu, Xin Jiang, Zhenguo Li, and Lingpeng Kong. Dream 7B: Diffusion large language models. arXiv preprint arXiv:2508.15487, 2025. URL https://arxiv.org/abs/2508.15487.

[5] Luca Soldaini et al. Dolma: an Open Corpus of Three Trillion Tokens for Language Model Pretraining Research. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2024. URL https://aclanthology.org/2024.acl-long.840/.

Dear Reviewer Nzx8, as the discussion period ends today (AoE), we would appreciate your thoughts on our response. We particularly want to emphasize that, although train–inference mismatch has been studied in AR models, discrete diffusion presents a substantially different setting that warrants its own study. We believe your concerns are all addressable and welcome any remaining questions; we will incorporate the necessary revisions by camera-ready at the latest. Thank you!

Thank you for the detailed reponse!

Due to the general lack in confidence level in this area I will increase my rating but ultimately leave the decision to the Area Chair to evaluate if the novelty is sufficient enough from his understanding.

From my readings of the response and from the comments of other reviewers (and the corresponding author responses) the novelty concern still remains open, the author accept that the work does relate significantly to prior work as such it seems my claim stands to a certain degree. But, as the other reviewer seem to show a higher confidence rating and overall seem to consider the work novel enough I will not stand against the overall accept tendency!

Issues that have not been addressed, but will not affect my rating since I did not respond in a timely manner for the authors to respond (at least not written in the review comments or notified changes in the main work):

  • discrepancy affect low-rank representation and on quantized models
  • The core motivation problem: "why this discrepancy is actually negative/positive or in which scenario this problem lead to large decay"
  • 7B is still a small model so scale up has not been addressed

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.