acceptodds
Under review as a conference paper at ICLR 2027

DIAL: One Diffusion Language Model, Multiple Inference-Time Compute Budgets

Abstract

Block diffusion language models trade generation quality for inference cost by varying the number of tokens committed per denoising step, yet this block size is typically fixed during training, requiring separate specialized models for different deployment settings. We present DIAL, a training framework for block-size-elastic diffusion language models in which the inference block size is a runtime choice shared by a single parameter set. The core mechanism, mixed block-size training, samples block sizes throughout training, exposing one network to a range of denoising granularities from autoregressive-like factorization to highly parallel generation. To preserve performance across this range, DIAL combines per-block noise conditioning with a zero-initialized recurrent draft-refinement loop, a two-stream cached parameterization, and argmax-agreement early stopping. Experiments show that mixed block-size training substantially reduces the degradation that otherwise occurs when a model is evaluated away from its training block size, while a single DIAL model achieves competitive perplexity across all tested inference granularities. By treating block size as an inference-time control, dial exposes multiple quality–compute operating points from one trained model, allowing deployment to adapt generation cost to the available compute budget without retraining or maintaining separate block-size specialists.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Diffusion Language Models, Block Diffusion, Inference-Time Compute, Dynamic Decoding.

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

This paper proposes DIAL, a block-size-elastic diffusion language model that supports different block sizes at inference time using one set of parameters. The main idea is mixed block-size training (MBT), where the training block size is sampled from {1, 2, 4, 8, 16}. The method further introduces per-block noise conditioning, recurrent draft refinement, a two-stream cached architecture, and argmax-based early stopping. Experiments show that one DIAL checkpoint maintains relatively stable perplexity across block sizes, while fixed-block BD3-LM models degrade when evaluated away from their training block size.

Critical strengths
  • The paper studies a useful deployment problem: supporting multiple inference block sizes with a single checkpoint. The full train/evaluation grid gives clear evidence that fixed-block models degrade off-design, while DIAL remains substantially more stable.

  • The empirical analysis is careful. The ablations help separate the role of MBT from the additional architectural components, and the paper is unusually explicit about limitations and negative results. In particular, the benefit of recurrent draft refinement grows with block size in the expected direction.

Critical weaknesses
  • The novelty relative to concurrent Adaptive Block Diffusion (ABD; Jain, 2026) is not sufficiently clear to me. Both methods randomize block configurations during training and aim to make one model work across decoding scales. The manuscript states that ABD was evaluated at a single operating point, but the public ABD manuscript appears to make a broader across-scale claim. The extra DIAL components and more extensive evaluation may still constitute a useful contribution, but the current novelty distinction seems overstated.

  • The main quality-compute trade-off is not fully validated for the actual M4-enabled sampler. Early stopping changes the generative distribution because remaining masked positions are committed using the final argmax, while the reported perplexity is sampling-free and therefore does not measure this change. In addition, most specialists in the main comparison do not use M1-M3, so the main grid is not fully architecture-matched.

Questions for authors
  1. Could the authors clarify the distinction from ABD, especially regarding the claim that ABD only evaluates a single operating point? More broadly, the related-work discussion on adaptive inference in diffusion LMs could include recent methods operating on complementary axes, such as DAEDAL (arXiv:2508.00819), Prophet (arXiv:2508.19982), and Sequential Diffusion Language Models (arXiv:2509.24007). A clearer positioning relative to this line of work would materially affect my novelty assessment.

  2. Can the authors evaluate the quality of the actual M4-enabled generation process against M4-disabled decoding, preferably at matched compute or latency? Since M4 changes the sampling distribution, a generation-based comparison would materially affect my confidence in the claimed quality-compute frontier. Architecture-matched fixed-block controls at additional block sizes would also help separate elasticity from the gains of the other modules.

1 comment
AuthorsOpenReview ↗

Response to Reviewer EWUU

We thank Reviewer EWUU for the positive assessment of the problem formulation and experimental analysis. We especially appreciate the comments on the novelty distinction from ABD and the need to validate the actual generation behavior of M4.

Comment 1: Novelty relative to ABD and positioning among adaptive inference methods

We agree that our original distinction from ABD was not sufficiently precise. Both ABD and DIAL train a shared model across multiple block sizes, and ABD does report results across multiple block sizes. We have corrected the related discussion and no longer characterize ABD as evaluating only a single block-size operating point.

Our distinction is instead in the training and architectural approach. ABD uses a two-stage procedure consisting of MDLM pretraining followed by adaptive block-size fine-tuning, whereas DIAL trains from scratch using uniform mixed block-size sampling. DIAL introduces per-block noise conditioning and recurrent draft refinement to support this training regime, together with a two-stream attention implementation for efficient context reuse.

Table 4 reports the available benchmark comparison. We emphasize that these are reported results rather than a fully controlled reproduction. Table 5 further isolates the contribution of MBT: compared with a fixed-(L'=4) DIAL specialist retaining M1–M3, full DIAL performs better on LAMBADA and PTB, although the specialist has a small WikiText-103 advantage at its training operating point.

We have also expanded the related-work discussion to position DIAL alongside the methods suggested by the reviewer:

  • DAEDAL: adapts total output length.

  • Prophet: adapts the number of inner denoising/refinement steps.

  • Sequential Diffusion Language Models (SDLM): adapts prefix acceptance for block-diffusion-style generation from pretrained autoregressive models.

ABD and DIAL: address adaptation across block granularity.

This distinction makes clear that DIAL addresses block-size elasticity rather than claiming to introduce adaptive inference in diffusion language models generally.

Comment 2: Does M4 preserve actual generation quality, and are the controls architecture-matched?

We agree that this question cannot be answered using sampling-free likelihood metrics alone. M4 changes the generation process, so its effect must be evaluated on actual generated samples.

We therefore conducted the direct M4-on/M4-off generation comparison in Table 2. At (L'=8), argmax commitment reduces decoding time by approximately 18.4%, while Distinct-4 increases from 0.884 to 0.953 and Gen-PPL changes from 32.72 to 34.36. At (L'=4), the results show a more substantial trade-off, and at (L'=16), M4 rarely activates.

We also added the architecture-matched fixed-(L'=4) control in Table 5. This comparison retains M1–M3 in both models and changes whether mixed block-size training is used. It therefore helps separate the effect of MBT from the additional architectural components.

We have revised the manuscript to state the scope of these results accurately: M4 can reduce computation with a configuration-dependent quality trade-off, while MBT provides benefits beyond those of a fixed-block model using the same supporting architecture.

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

This paper observes that block dLLMs fix the block size during training and degrade severely when decoded at any other block size. To handle this problem, this paper proposes DIAL, a training framework that enables dLLMs to adapt to different block sizes during decoding. DIAL uses MBT as its core mechanism and is complemented with four modules, which are (M1) per-block noise conditioning, (M2) Recurrent Draft Refinement, (M3) a two-stream parameterization, and and (M4) argmax-agreement early stopping. Experiments show that a single DIAL model degrades gracefully on block sizes unseen while training.

Critical strengths
  1. Simple core mechanism. The core mechanism MBT only changes which attention mask is used at each training step. The supporting modules are also lightweight.
  2. Comprehensive evaluation. The experiments include a three seed replication, a model-size ladder from 15M to 170M parameters, and per-request latency and memory measurements.
Critical weaknesses
  1. Questionable generalization. All experiments use a single model and configuration: a 170M-parameter DiT-style block diffusion model.
  2. No evaluation of generation quality. All quality metrics are likelihood-based. The paper reports no generative perplexity, diversity metrics, human evaluation or samples.
  3. Limited practicality. The most widely used dLLMs, LLaDA and Dream, are trained as full-sequence masked diffusion models without any block structure. For those models, the same weights can be decoded at different block sizes without retraining. And they are commonly used this way without catastrophic degradation.
Questions for authors
  1. The paper contains several passages referring to earlier versions of the paper and to claims that have been withdrawn. Could the authors clarify the reason for these passages?
  2. Does elasticity remain nearly free at larger scales, or when fine-tuning an existing pretrained dLLMs?
  3. Could you compare against the concurrent block-randomization approach under matched model size, data and evaluation protocol?
1 comment
AuthorsOpenReview ↗

Response to Reviewer ZXTd

We thank Reviewer ZXTd for recognizing the simplicity of MBT and the breadth of the original evaluation. We address the concerns about practicality, generalization, and comparison with concurrent work below.

Comment 1: Can full-sequence models such as LLaDA and Dream simply be decoded at different block sizes?

We agree that full-sequence masked diffusion models can be adapted to use different inference block sizes without retraining. However, this does not establish that such decoding is equivalent to training a model across block sizes.

We identify two important differences:

Generation quality: Table 1 shows that a full-sequence MDLM's WikiText-103 perplexity increases from 30.79 at its native block size to 74.04 at (L'=4) and 143.36 at (L'=1). Its Distinct-4 also drops from 0.978 to 0.013 at (L'=4). In contrast, DIAL maintains substantially more stable performance across its supported block sizes.

Caching efficiency: Full bidirectional attention allows prefix representations to depend on future masked positions, which can change during denoising. This prevents the same exact reuse of committed prefix representations available under block-causal attention. Table 6 and our caching measurements show the practical benefit of the latter structure.

We have clarified that DIAL addresses a different problem from merely allowing a full-sequence model to be decoded in blocks: it trains one model to operate across block granularities while supporting efficient reuse of committed context.

Comment 2: Passages referring to earlier versions and withdrawn claims

We agree that these passages were distracting and could create uncertainty about the final claims. They originated as notes documenting hypotheses that were revised during manuscript development. We have removed or rewritten them so that the paper presents only the final method, evidence, and conclusions.

Comment 3: Does elasticity remain effective at larger scales and when fine-tuning pretrained models?

We extended the model-size evaluation from 15M to 170M parameters. As shown in Table 3, DIAL consistently achieves lower WikiText-103 perplexity than the corresponding fixed-block BD3-LM baseline across all five tested scales.

Regarding fine-tuning existing pretrained models, we agree that this is an important practical direction. Our current experiments focus on training DIAL from scratch. We therefore do not claim that the same results have been established for arbitrary pretrained full-sequence checkpoints. We have clarified this scope and discuss adaptation of existing pretrained models as a separate direction.

Comment 4: Comparison with concurrent ABD under matched settings

We added Table 4 to compare DIAL against the publicly reported ABD results on shared benchmarks. The results favor DIAL on the listed comparisons. However, ABD uses a different two-stage training procedure and sampling distribution, and we were unable to complete a full controlled reproduction under identical configurations during the rebuttal period.

We have therefore avoided describing this as a strictly matched comparison and have clarified the experimental differences.

Reviewer_7touOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes DIAL, a block-size-elastic diffusion language model that trains a single model to support variable block sizes. DIAL samples a block size at each training iteration and trains the model using the corresponding block-wise denoising objective. In addition to the training algorithm, the authors introduce per-block noise conditioning, recurrent draft refinement, and a two-stream architecture. The paper also proposes an argmax-agreement early stopping mechanism, which enables dynamic halting during inference by checking whether the argmax predictions remain identical across consecutive denoising steps. Experiments demonstrate improved perplexity over fixed-block BD3-LM baselines across different block sizes.

Critical strengths
  1. The proposed early stopping mechanism reduces the number of function evaluations (NFEs) during inference.
  2. DIAL outperforms BD3-LM, which uses a fixed block size.
Critical weaknesses
  1. Lack of generation quality evaluation. The authors did not measure actual generation quality. The paper's quality evaluation is mainly based on zero-shot perplexity and likelihood-based downstream evaluations. As the reported perplexity of discrete diffusion models is based on an NELBO rather than exact likelihood, it may not accurately reflect actual generation quality. Therefore, generative perplexity (Gen-PPL), together with sample entropy or MAUVE, should also be reported to assess the proposed method. Moreover, the evaluation protocol for perplexity is not clearly specified. For instance, it is unclear how self-conditioning is handled during perplexity evaluation.

  2. Poor clarity and presentation. The method is difficult to follow. The explanation of the algorithm relies heavily on textual descriptions with ambiguous or overly compressed technical terms. For instance, Lines 149–156 introduce "order-averaging slack of the any-order bound" and "joint commitment" without clearly explaining their relationship to the degradation caused by larger block sizes. More generally, the paper frequently mixes discussions of likelihood bounds, sampling behavior, and computational efficiency without clearly distinguishing them. The manuscript requires substantial polishing and more self-contained explanations of the proposed method.

  3. Concerns about manuscript quality control and compliance. The authors omitted the mandatory AI Use Statement required by the ICLR 2027 Author Guidelines, and the paper contains an incorrect reference to a central baseline. For instance, BD3-LM is cited as "Block discrete denoising diffusion language models" (Lines 438–440), whereas its actual title is "Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Models." These issues raise concerns about the thoroughness of the manuscript's quality control. The authors should take responsibility for the accuracy of the references and technical claims presented in the paper.

  4. Lack of generation quality evaluation at the reported throughput. While the authors report inference throughput across different block sizes, the generation quality achieved at those throughput levels is not evaluated. In particular, the proposed argmax-agreement early stopping mechanism changes the sampling distribution by committing the remaining masked tokens using argmax predictions. However, the authors do not evaluate how this affects generation quality. Therefore, the reported perplexity–throughput trade-off does not necessarily establish a meaningful generation quality–throughput trade-off.

Questions for authors
  1. Could the authors report actual generation quality using metrics such as Gen-PPL with sample entropy or MAUVE across different block sizes?
  2. How is recurrent draft refinement (self-conditioning) handled during the MC-ELBO perplexity evaluation?
  3. Could the authors clarify the explanation in Lines 149–156, particularly how "order-averaging slack" and "joint commitment" contribute to performance degradation as the block size increases?
  4. Could the authors report generation quality at the measured throughput levels and evaluate the effect of argmax-agreement early stopping on generation quality?
2 comments
AuthorsOpenReview ↗

Response to Reviewer 7tou

We thank Reviewer 7tou for the careful review and for identifying important gaps in the generation evaluation and presentation.

Comment 1: Lack of actual generation quality evaluation

We agree that likelihood-based metrics alone do not establish the quality of generated text, particularly when M4 changes the sampling procedure by committing remaining masked tokens using argmax predictions.

To address this concern, we conducted generation experiments and report Gen-PPL, sample entropy, Distinct-4, and actual decoding time in Table 2. The results show that M4's effect depends on the block size and commitment strategy. At (L'=8), argmax commitment reduces decoding time from 141.0s to 115.0s while maintaining comparable diversity (Distinct-4: 0.884 to 0.953). However, the (L'=4) results show a more substantial trade-off.

We therefore revised the manuscript to present M4 as a compute–quality trade-off rather than claiming that it universally preserves generation quality.

Comment 2: Handling of recurrent draft refinement during MC-ELBO evaluation

We have clarified the evaluation protocol.

During training, recurrent draft refinement is used with probability psc=0.5p_{\mathrm{sc}}=0.5. A detached draft pass produces the previous clean-token estimate: arg⁡max⁡fθ(xt,σ,x0,0)\arg\max f_\theta( \mathbf{x}_t,\boldsymbol{\sigma},\mathbf{x}_0,\mathbf{0} ). The estimate is incorporated into the main forward pass through Embed(xt)+WscEmbed(x^0prev)\mathrm{Embed}(\mathbf{x}_t) + W_{\mathrm{sc}} \mathrm{Embed}(\hat{\mathbf{x}}_0^{\mathrm{prev}}).

With probability (1−psc)(1-p_{\mathrm{sc}}), the previous draft is set to zero.

During MC-ELBO evaluation, we use single-pass evaluation with (x^0prev=0)(\hat{\mathbf{x}}_0^{\mathrm{prev}}=\mathbf{0}) for all 1024 Monte Carlo draws per sequence. This matches the NFE cost of the BD3-LM baseline scorer and avoids introducing an additional draft pass into the likelihood evaluation. We have explicitly added this distinction to the evaluation description.

Comment 3: Clarifying order-averaging slack and joint commitment

We agree that the original explanation did not sufficiently distinguish likelihood-bound effects from generation-time sampling effects. We have rewritten this discussion around two separate phenomena:

  • Order-averaging slack in the likelihood bound. The absorbing-diffusion NELBO over a block averages over possible token-unmasking orders. At (L'=1), there is only one order. As (L') increases, a single model must account for an increasingly large set of possible orders, increasing the variational slack of the bound.

  • Parallel commitment error during generation. When multiple masked positions are committed in the same denoising step, their predictions are made without conditioning on the newly sampled values at the other positions. This creates a coordination challenge that increases with block size.

These are distinct effects: the first concerns likelihood estimation, while the second concerns the reverse sampling process. Recurrent Draft Refinement (M2) is designed to address the latter by allowing predictions within a block to benefit from an intermediate draft.

Comment 4: Generation quality at reported throughput and the effect of M4

Table 2 directly compares M4-enabled and M4-disabled generation, including actual decoding time and generation metrics. As discussed above, the results show that the trade-off depends on block size and commitment strategy. We have revised the claims to reflect this evidence and avoid generalizing the favorable (L'=8) result to every operating point.

Manuscript quality and compliance

We corrected the BD3-LM reference to its official ICLR 2025 proceedings title, Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Models, and added the required ICLR 2027 AI Use Disclosure section. We also revised the relevant method explanations and removed distracting references to earlier manuscript versions.

Thank you for the response and the additional experiments. However, my main concern about generation quality is not resolved, and the response reveals issues that weaken the paper's main contributions.

1. Generation quality is still not compared with the baselines. The new Gen-PPL and entropy results (Table 2) are limited to an ablation of M4 within DIAL, so they do not show whether DIAL's perplexity gains over BD3-LM translate into better samples. Since Gen-PPL depends on the entropy of the samples, generation quality is typically compared at matched entropy or along generative frontiers [4], whereas the settings in Table 2 differ in entropy (5.01–5.39). The results are also based on only 16 sequences, whereas prior work evaluates Gen-PPL on at least 100 samples [1, 2, 3], and no distributional metric such as MAUVE [5] is reported. Without a comparison of this kind, the paper does not establish that its likelihood gains translate into better generation.

2. The response weakens the main contributions.

  • Mismatch between perplexity evaluation and generation. The response clarifies that MC-ELBO evaluation uses a single pass with the previous draft set to zero (x^0prev=0\hat{\mathbf{x}}_0^{\text{prev}} = \mathbf{0}), whereas generation uses recurrent draft refinement. The reported perplexities therefore evaluate the model without the self-conditioning path that it relies on during generation. This widens the gap between the perplexity results and actual generation quality.
  • Early stopping (M4). M4 reduces decoding time only where it degrades quality (e.g., Gen-PPL rises from 29.37 to 52.99 at L′=4L'=4 with sampled commit), and where quality is preserved (L′=16L'=16), the time barely changes (76.7 s vs. 76.2 s).

3. The revisions cannot be verified. The current PDF still contains the incorrect BD3-LM reference, has no AI Use Statement. The claimed revisions are not reflected in the manuscript.

For these reasons, I maintain my score.

[1] Lou et al., Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution, 2024.
[2] Hayakawa et al., Distillation of Discrete Diffusion through Dimensional Correlations, 2025.
[3] Pynadath et al., CANDI: Hybrid Discrete-Continuous Diffusion Models, 2025.
[4] Pynadath et al., Generative Frontiers: Why Evaluation Matters for Diffusion Language Models, 2026.
[5] Pillutla et al., MAUVE: Measuring the Gap Between Neural Text and Human Text using Divergence Frontiers, 2021.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank the Area Chair and Reviewers 7tou, ZXTd, and EWUU for their careful reviews and constructive feedback. We appreciate the reviewers' recognition of the practical importance of block-size elasticity, the simplicity of mixed block-size training (MBT), and the potential of DIAL's early-stopping mechanism.

To address the three core questions raised: (1) actual generation quality beyond likelihoods, (2) block-size elasticity vs. simply altering inference block sizes in full-sequence models, and (3) differences from concurrent Adaptive Block Diffusion (ABD), we added new experiments and revised the manuscript:

1. Generation quality & M4. We evaluated actual generations using generative perplexity, sample entropy, Distinct-4, and decoding time. We also directly compared M4-enabled and M4-disabled decoding. These results show that the quality–compute trade-off depends on block size and commitment strategy; we therefore avoid claiming that M4 universally preserves generation quality.

2. Full-sequence vs. block-causal diffusion. Evaluating a full-sequence MDLM at smaller inference block sizes without retraining causes substantial degradation in perplexity and diversity due to train–inference mismatch. Two-stream implementation profiling also shows lower per-step compute and memory usage.

3. Scaling & ABD. We extended evaluation across five scales and compared against publicly reported ABD numbers. While ABD also supports multiple block sizes, DIAL's distinction lies in its specific training and architectural design rather than multi-block training in general. We revised novelty claims and clarified that the ABD comparison is not a controlled reproduction.

4. Manuscript revisions. We clarified the MC-ELBO protocol, separated likelihood-bound effects from sampling effects, corrected the BD3-LM reference, added the ICLR 2027 AI Use Disclosure, and removed references to prior manuscript versions.


Additional Experimental Evidence

Table 1. Full-sequence MDLM across inference block sizes (no retraining, L′=1024L'=1024). Train–inference mismatch degrades MDLM as block size shrinks, whereas DIAL remains stable.

Metric1024168421
MDLM WT103 PPL30.7946.7258.7974.04102.90143.36
DIAL WT103 PPL—30.6029.7028.9028.4028.00
MDLM LAMBADA PPL42.8258.5968.6485.40109.35162.69
MDLM PTB PPL78.56135.39163.12207.42256.90338.12
MDLM Distinct-40.9780.6530.2210.0130.0170.031
DIAL Distinct-4—0.9670.9530.915——

Table 2. M4 generation quality & cost (16 sequences, length 1024, T=50T=50). Benefits are configuration-dependent: at L′=8L'=8, M4 cuts decoding time by 18.4% with comparable diversity; at L′=16L'=16 impact is minimal; at L′=4L'=4 quality–diversity trade-offs are larger.

L′L'M4 settingNFETime (s)Gen-PPLEntropyDistinct-4
4Off12,714293.329.375.39.969
4Sampled commit8,571159.752.995.33.970
4Argmax commit11,025206.319.705.10.915
8Off6,391141.032.725.01.884
8Argmax, K=2K=25,576115.034.365.21.953
16Off3,20076.738.765.27.961
16Argmax, K=2K=23,19476.240.025.33.967

Table 3. Scaling on WikiText-103 (PPL). DIAL outperforms BD3-LM at every scale.

Scale15M34M62M100M170M
BD3-LM62.448.141.736.230.4
DIAL58.944.738.532.828.9
Change−5.6%−7.1%−7.7%−9.4%−4.9%

Table 4. DIAL vs. publicly reported ABD PPL (not controlled reproductions).

Benchmark (Block Size)DIALABDDiff
WT103 (L′=1L'=1)28.0029.69−1.69
WT103 (L′=4L'=4)28.9030.16−1.26
LAMBADA (L′=1L'=1)45.9051.00−5.10
PTB (L′=4L'=4)84.1094.34−10.24

Table 5. Fixed-block control (full DIAL vs. fixed-L′=4L'=4 specialist with M1–M3, no MBT). The specialist slightly edges out DIAL in-domain (WT103), while MBT improves zero-shot transfer (LAMBADA, PTB).

ModelWT103LAMBADAPTB
Fixed L′=4L'=4 (M1–M3)28.450.488.6
Full DIAL (MBT)28.946.384.1

Table 6. Two-stream per-step profiling (per-step cost, not total training time-to-convergence).

Scale15M34M62M100M170M355M
Fused (ms)52.391.5158.2246.1481.91146.2
DIAL (ms)25.946.684.0135.9245.8661.9
Speedup2.02×1.96×1.88×1.81×1.96×1.73×
VRAM saved61.7%55.7%49.8%45.0%39.1%32.4%

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.