acceptodds
Under review as a conference paper at ICLR 2027

HALO: Enhancing Time Series Generation via Hyperspherical Latents and Masked Autoregressive Modeling

Abstract

Most existing time series generators rely on a two-stage modeling paradigm: learns discrete latent representations of time series; performs autoregressive modeling on these discrete latents through next token prediction. However, this paradigm suffers from two stage-specific limitations: the first stage can lead to information loss when discretizing continuous time series, while the second stage is prone to error accumulation during autoregressive generation. To address these limitations, our core idea is to perform generative modeling in a continuous latent space with a more efficient autoregressive framework. We propose , which enhances time series generation via ypersphericl atents and Masked Autregressive modeling to achieve this goal by tackling two key bottlenecks: (1) variance and scale heterogeneity of continuous latent representations; (2) the difficulty of balancing generation efficiency with temporal correlation modeling. HALO first introduces a hyperspherical VAE that constrains continuous latents to a fixed-radius hyperspherical shell, effectively stabilizing the numerical fluctuations of continuous latent representations. Secondly, we develop a masked autoregressive model that balances parallel decoding and temporal correlation learning, substantially reducing the number of inference steps required for generation and improving generation stability. Our extensive experiments demonstrate that HALO achieves state-of-the-art generation performance while offering significantly improved inference efficiency over existing advanced baselines.

Primary area:
learning on time series and dynamical systems
Keywords:
Time series generation

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_PN6tOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes HALO, a two-stage generative model for time series.

Stage 1: A hyperspherical VAE (S-VAE) compresses the time series into a sequence of continuous latent tokens. Each token is constrained to a sphere of fixed radius, using a von Mises–Fisher (or Power Spherical) posterior. The aim is to reduce variance and scale heterogeneity in the latent space.

Stage 2: A masked autoregressive model learns the distribution of these tokens. A Transformer encoder–decoder builds context from visible tokens, and a small per-token diffusion head models each masked token. At inference, tokens are generated in a few rounds in random order, several tokens per round, conditioned on a domain label or unconditionally via classifier-free guidance.

The method is evaluated on 12 datasets from four domains at three sequence lengths, on two unseen domains, with ablations and an efficiency comparison.

Critical strengths

Clear motivation: The variance and norm analysis of Gaussian VAE latents (Figure 1) makes the problem easy to understand.

Sensible design: Combining continuous tokens with masked parallel generation is a reasonable approach and is clearly described. Algorithm 1 and the implementation details in the appendix make the method easy to follow.

Broad evaluation: The paper covers 12 datasets, three sequence lengths, three metrics and five seeds, plus a sensitivity analysis on the number of generation rounds.

Practical efficiency: HALO is about four times faster than TimeMAR at inference with lower memory use, and it scales to length 336, where TimeMAR runs out of memory.

Critical weaknesses
  1. Overstated motivation. The abstract claims that most time series generators use discrete tokens and next-token prediction. Most baselines in the paper do neither. The problem HALO addresses concerns only a few recent methods, and the framing should say so.

  2. Missing baselines. The introduction builds on SDformer, TimeMAR, MSDformer and PolarFormer, but only TimeMAR is evaluated. The claimed advantage over these methods rests on a single comparison. In addition, TimeGAN is outdated and has repeatedly been shown to perform poorly. It should be replaced by a current baseline such as Diffusion-TS (Yuan et al., ICLR 2024) or AEQ-RVAE-ST (Fulek et al., TMLR 2026).

  3. Core claims not tested. The paper argues that discretization harms generation quality and that step-by-step generation accumulates errors. Neither is tested directly. There is no comparison with a discrete tokenizer under an unchanged Stage 2, and Table 8 even suggests that more sequential steps improve quality. The gains in the ablation study (Table 2) are small and reported without standard deviations.

  4. Missing citation of MAR. Stage 2 closely follows MAR (Li et al., NeurIPS 2024), which is not cited. The authors should credit it and clarify what is new.

  5. Unclear cross-domain results. On unseen domains, HALO generates without any target data, yet it beats methods that use target data. Its scores also change with the number of target samples. Section 5.2 only refers to the protocol of Huang et al. (2025). Please describe the evaluation explicitly.

Questions for authors

I would like to have answers to the 5 weaknesses, and whether I might be mistaken in seeing them as mistakes. In partikular, I would like answers to (2), (3) and (5). The weaknesses (1) and (4) should be easy to fix.

6 comments
AuthorsOpenReview ↗

Response to Reviewer PN6t (Part 1)

Many thanks to the valuable comments of Reviewer PN6t. We have revised the manuscript accordingly based on these valuable comments.

W1. We sincerely apologize for this issue. To avoid overstating our claims, we have carefully revised potentially exaggerated statements in both the abstract and the main text. Specifically, regarding the abstract, we revise the statement “Most existing time series generators rely on a two-stage modeling paradigm” to the more conservative wording “Some existing time series generators rely on a two-stage modeling paradigm.” For the main text, we revise the statement in the Introduction from “Existing time series generation methods generally adopt a two-stage modeling paradigm.” to the more conservative wording “Some existing time series generation methods generally adopt a two-stage modeling paradigm.” These two revisions correspond to Lines 13--16 and Lines 42--44, respectively, in the revised manuscript.


W2. We sincerely appreciate this valuable suggestion, which is essential for a more comprehensive evaluation of HALO. We conduct additional experiments by including SDformer, MSDformer, PolarFormer, and Diffusion-TS as comparison baselines. We strictly follow the recommended hyperparameter settings from the official repositories of these newly added baselines, while ensuring the same maximum number of training epochs and checkpoint evaluation intervals as HALO. The additional experimental results have been incorporated into Tables 1, 5, 6, and 7 of the revised manuscript. Even after incorporating these additional strong baselines, HALO still maintains state-of-the-art overall generation performance across different sequence-length settings. Compared with methods that perform generative modeling in discrete latent spaces, HALO exhibits clear generation advantages in both MMD, which reflects global distributional consistency, and MDD, which captures fine-grained consistency in temporal dynamics. Moreover, compared with Diffusion-TS, a one-stage baseline that performs generative modeling directly in the observation space, HALO still demonstrates advantages in both generation performance and inference efficiency (Tables 3 of the revised manuscript).


W3. We sincerely appreciate your insightful comments. We have added a VQ-VAE ablation variant and reported the detailed ablation results in Table 2, Lines 453--475 of the revised manuscript. When the S-VAE is replaced with a VQ-VAE, the average MMD and KL across all datasets deteriorate from 0.078 to 0.213 and from 0.198 to 0.391, corresponding to relative degradations of 173.08% and 97.47%, respectively. These results indicate that the quantization error introduced by discrete quantization leads to the loss of fine-grained continuous temporal information, thereby degrading generation quality.

Furthermore, in the original manuscript, we intended to convey that increasing the number of inference rounds can, to some extent, refine the generation process and improve model performance. However, continually increasing the number of inference rounds does not necessarily lead to consistent performance gains, as the resulting error accumulation can eventually offset the benefits of finer-grained generation. We sincerely apologize for the ambiguity in our previous description. To clarify this point, we have added more detailed experimental results in Tables 8 and 9, Lines 1137--1186 of the revised manuscript. The results show that, within a certain range, increasing the number of inference rounds improves generation performance. However, when the number of inference rounds is increased too aggressively, the performance begins to deteriorate, and the average ranking across datasets also declines. This indicates that excessively long inference trajectories exacerbate error accumulation, which ultimately degrades the quality of the synthesized time series.

AuthorsOpenReview ↗

Response to Reviewer PN6t (Part 2)

W4. We examine MAR [1], a generative model originally developed for the visual domain. Specifically, MAR employs a standard VAE to learn continuous latent representations. In contrast, HALO adopts a hyperspherical S-VAE for continuous representation learning. A key distinction lies in the characteristics of the two modalities.

Image data typically have a relatively stable and bounded numerical range (e.g., pixel values within 0--255 before normalization), allowing standard VAEs to learn continuous representations relatively robustly. Time series, however, often exhibit substantially more heterogeneous numerical scales and distributions. In particular, distribution shifts and heterogeneous temporal patterns can induce considerable variance and scale heterogeneity in the continuous latents produced by a standard VAE, making their underlying distribution more difficult to parameterize effectively. This motivates our use of the S-VAE, which constrains latent representations to a geometrically regularized space and facilitates more stable latent distribution modeling. Regarding the citations, we have added the relevant reference to MAR in the conditional diffusion modeling section, specifically in Lines 334--337 of the revised manuscript.


W5. For this problem, we first add a more detailed description of the cross-domain experimental setting. Our experiments strictly follow the official TimeDP codebase. The variation in scores with different numbers of prompt samples arises because these samples are also used to normalize the generated time series. Therefore, changing the number of prompt samples results in a different normalizer, which in turn leads to different evaluation scores. A detailed description of the cross-domain generation setting is provided in Appendix D, Lines 966--981 of the revised manuscript. In addition, we report the detailed cross-domain generation results in Table 10 of the revised manuscript, where the results of other baselines are sourced from the paper of TimeDP [2]. Finally, we would like to further clarify the sources of HALO's strong cross-domain generalization. First, through randomized masking, HALO flexibly models the joint distribution of latent representations under diverse masking configurations, providing a strong basis for learning transferable temporal representations. Second, HALO performs generative modeling in the latent space, where low-dimensional representations not only compress time series into shorter token sequences but also suppress redundant noise. This provides a more stable latent environment for robust generative modeling and consequently benefits cross-domain generation. Third, compared with left-to-right autoregressive models with a fixed causal inductive bias, masked generative modeling can transform the same sample into multiple training instances with different masked positions and masking ratios. This mechanism can be viewed as an implicit form of data augmentation, further encouraging HALO to learn diverse and transferable temporal patterns and thereby enhancing its cross-domain generalization capability [3].

References

[1] Autoregressive image generation without vector quantization. NeurIPS 2024

[2] TimeDP: Learning to Generate Multi-Domain Time Series with Domain Prompts. AAAI 2025

[3] Diffusion beats autoregressive in data-constrained settings. NeurIPS 2026

I appreciate the detailed rebuttal and the many additional experiments.

I have also gone through the other reviews and each author response. Points (1), (2) and (4) have, I think, been dealt with. The revised results, however, come across as weakening the two core claims. First, in Table 1, VQ-based TimeMAR and SDformer arrive at perceptibly similar performance to HALO, while the VQ-VAE ablation lands clearly further behind (e.g. Solar MMD 0.649 vs. 0.032), so that the numbers point more to a variant that is not well configured than to the quantization itself. Second, in Table 8 the fully sequential decoding (K=42) arrives at roughly the same level as K=12, which does not really get the error accumulation argument across. Finally, given that the inverse CDF mapping on the target support set (Eq. 15) coincides with the target marginals by construction, the cross-domain comparison to TimeDP in the main text would deserve qualification.

I stand by my score.

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

This paper proposes HALO, a time-series generation framework. It aims to reduce the information loss of discrete latent quantization and solve the low efficiency and error accumulation of token-by-token autoregressive generation. First, it uses a Spherical VAE (S-VAE) to constrain continuous latents to a fixed radius. Then, it combines bidirectional masked context modeling with conditional diffusion for grouped parallel generation. Experiments on 12 in-domain and 2 cross-domain datasets show that HALO achieves strong generation quality and fast inference speed.

Critical strengths
  • The research problem is practical, and the two-stage design is clear. Combining continuous latent modeling with masked autoregressive generation preserves fine-grained continuous details while cutting sequential decoding costs, showing a clear design motivation.
  • The experiments are broad and show competitive generation performance. The paper tests HALO across multiple domains and sequence lengths against recent generative models, alongside ablations, cross-domain tests, and speed analyses. HALO performs well on several distribution metrics and runs much faster than autoregressive baselines.
Critical weaknesses
  • The core method overlaps heavily with existing generative frameworks, limiting its standalone novelty. SphereAR (ICLR 2026) already uses fixed-radius spherical latents, and MAR (NeurIPS 2024) already combines masked autoregression with conditional diffusion on continuous tokens. HALO closely mirrors these two designs without clarifying its technical differences or showing unique adaptations for time series, reading more like a combination and task transfer of existing modules.

  • The definition and implementation description of the probability distribution of S-VAE are inconsistent. In Section 4.2, Equation (3) explicitly adopts the von Mises–Fisher (vMF) distribution to define the posterior, while Appendix D describes that the actual encoder uses the Power Spherical posterior. These two are different, and the paper does not clearly state which posterior is used in the actual experiments, causing ambiguity in the definition of the core latent variable modeling method.

  • The core claim that continuous latents avoid quantization loss is not directly verified. Although the paper criticizes VQ-VAE for losing fine-grained details, its VAE ablation only compares against a continuous Gaussian VAE rather than a discrete VQ-VAE. Without controlled comparisons on reconstruction error or spectral fidelity under matched compression rates and encoder capacity, this core motivation remains unproven.

Questions for authors
  • What are the exact differences in core modeling mechanisms between HALO and prior works like SphereAR and MAR? Beyond applying existing mechanisms to time-series generation, what standalone technical improvements does this paper propose?
  • Does the actual training of S-VAE use the vMF or the Power Spherical posterior? Please clarify the exact distribution used in experiments and make the main text equations match the implementation.
  • Under the same compression rate and similar encoder capacity, does S-VAE preserve fine-grained time-series details better than VQ-VAE? Can the authors provide a direct comparison of reconstruction errors or spectral fidelity?
5 comments
AuthorsOpenReview ↗

Response to Reviewer qNbh (Part 1)

Many thanks to the valuable comments of Reviewer qNbh. We have revised the manuscript accordingly based on these valuable comments.

W1. We have carefully examined both SphereAR and MAR. First, we would like to clarify that the hyperspherical VAE was not originally proposed by SphereAR; it was first introduced in [1], which has already been cited in our original manuscript. Hyperspherical VAEs have also been applied to other domains, such as audio generation [2]. Therefore, the contribution of HALO does not lie in proposing the S-VAE itself, but rather in identifying a specific challenge in time series generation and introducing a targeted solution. In particular, we identify substantial variance and scale heterogeneity in the continuous latent space of time series, which makes the latent distribution difficult for the subsequent generator to parameterize effectively. Based on this observation, HALO introduces the S-VAE as a tailored solution to constrain such heterogeneity through hyperspherical latent modeling, thereby providing a more regularized and stable latent space for subsequent generative modeling. To the best of our knowledge, this is the first work to introduce an S-VAE into time series generation. Moreover, compared with the VQ-VAE architectures adopted by prior methods (SDformer and MSDformer), we further improve the representation learner by combining convolutional neural networks with attention mechanisms. This hybrid architecture introduces complementary local and global inductive biases, enabling the encoder to capture both short-range temporal patterns and long-range dependencies, and thereby learn more expressive latent representations.

Regarding the masked autoregressive paradigm, it originates from MaskGIT, while MAR extends this paradigm to continuous latent spaces. In our original manuscript, we have already cited MaskGIT and explicitly stated that, inspired by this line of work, we introduce masked autoregressive modeling to improve parallel decoding capability and reduce the number of inference steps, thereby alleviating error accumulation in long-sequence generation. Importantly, HALO is motivated by a substantially different challenge from MAR. Specifically, HALO identifies the variance and scale heterogeneity of continuous time-series latents as a key obstacle to effective latent-space generative modeling and introduces hyperspherical representation learning as a targeted solution, whereas MAR does not investigate this time-series-specific issue or provide such a corresponding design. To more clearly acknowledge the methodological relationship, we have added the citation to MAR in the conditional diffusion modeling section (Lines 334--337 of the revised manuscript), explicitly clarifying that the technical overlap is limited to the use of conditional diffusion for continuous latent modeling.


W2. Regarding this issue, we sincerely apologize for not clarifying it in the main text. The Power Spherical posterior serves as a practical approximation to the von Mises--Fisher (vMF) distribution and has been commonly adopted in modern implementations of hyperspherical latent-variable models. Compared with vMF, it enables more efficient and stable sampling and optimization while preserving the key geometric properties of hyperspherical distributions. We have added a corresponding clarification in the main text of the revised manuscript and provided a more detailed description in the Appendix D.


References

[1] Hyperspherical variational auto-encoders. UAI 2018

[2] SphereVAE: Hyperspherical Latent Autoencoders for Robust Autoregressive Speech Representation Modeling. arXiv:2609.09903, 2026

AuthorsOpenReview ↗

Response to Reviewer qNbh (Part 2)

W3. Regarding this issue, we provide an intuitive reconstruction comparison between the two. S-VAE achieves better reconstruction accuracy:

DatasetSamplesS-VAE MMD ↓VQ-VAE MMD ↓S-VAE K-L ↓VQ-VAE K-L ↓S-VAE MDD ↓VQ-VAE MDD ↓
Electricity36590.0000100.0000760.0005370.0238660.0010850.001612
Solar5200.0001340.0002950.0009740.01005977.39051133.794342
Wind39620.0000030.0000110.0003680.0083300.0054170.009621
Traffic21180.0012190.0044880.0014250.0099200.0026050.006424
Taxi19420.0009910.0043530.0178460.0106120.0132770.008320
Pedestrian16770.0004080.0003240.0008230.0052890.0046730.007398
Air (KDDCup)15930.0000710.0002120.0007900.0116240.0028710.004291
Temperature1260.0000490.0003150.0020560.0214590.0298360.057989
Rain1260.0037460.0053680.0168400.0159530.1159450.147553
NN5330.0058380.0139320.1455030.0713770.1374590.243572
FRED-MD320.0002960.0017800.1697500.0550090.0178260.031025
Exchange270.0001670.0000260.0255220.0979170.0375420.034228

We also conducted an additional ablation study using VQ-VAE:

DatasetHALO MMD ↓w/ VQ-VAE MMD ↓HALO K-L ↓w/ VQ-VAE K-L ↓
Electricity0.001 ± 0.0000.003 ± 0.0000.002 ± 0.0000.057 ± 0.002
Solar0.033 ± 0.0000.649 ± 0.0010.019 ± 0.0000.186 ± 0.008
Wind0.007 ± 0.0000.241 ± 0.0020.063 ± 0.0020.186 ± 0.002
Traffic0.068 ± 0.0010.441 ± 0.0010.004 ± 0.0000.338 ± 0.013
Taxi0.083 ± 0.0010.185 ± 0.0020.008 ± 0.0000.254 ± 0.005
Pedestrian0.071 ± 0.0010.099 ± 0.0010.035 ± 0.0000.178 ± 0.004
Air0.010 ± 0.0000.049 ± 0.0000.020 ± 0.0010.185 ± 0.002
Temperature0.247 ± 0.0030.255 ± 0.0020.158 ± 0.0060.594 ± 0.043
Rain0.025 ± 0.0000.109 ± 0.0000.006 ± 0.0000.168 ± 0.009
NN50.199 ± 0.0020.296 ± 0.0010.092 ± 0.0020.474 ± 0.011
FRED-MD0.001 ± 0.0000.004 ± 0.0000.228 ± 0.0100.285 ± 0.001
Exchange0.184 ± 0.0030.226 ± 0.0021.742 ± 0.0431.666 ± 0.005
Average0.0780.2130.1980.381

More detailed ablation results are provided in Table 2, Lines 453--475 of the revised manuscript.

Thank you for your reply.

For the first two points, it addressed my concerns about the innovation and theoretical consistency of the paper, and the supplementary experiment for the third point also solved my problem.

Therefore, I will increase my score.

Reviewer_Bm9oOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This work presents HALO, a novel deep generative framework for time series generation. The method performs generative modeling in a continuous latent space to mitigate the information loss associated with discrete two-stage approaches and introduces masked autoregressive modeling to enhance inference stability and efficiency. Extensive experiments across multiple benchmark datasets show that HALO achieves state-of-the-art performance in both in-domain and cross-domain generation. In particular, the method maintains strong generation stability as the sequence length increases.

Critical strengths
  1. The authors provide a quantitative analysis of variance and scale heterogeneity in continuous latent representations. To address this issue, they introduce a hyperspherical VAE that constrains the latent space and mitigates such heterogeneity.

  2. The authors identify the limitations of MAE-style single step reconstruction in capturing temporal dependencies compared with GPT-style autoregressive modeling, as further supported by the results of SDformer-m. The proposed masked autoregressive modeling offers a reasonable balance between temporal dependency modeling and parallel decoding.

  3. The paper covers recent two-stage generative models with discrete representations, including SDformer and TimeMAR. Moreover, the authors largely follow the experimental protocol of TimeDP, facilitating fair comparisons with existing methods.

Critical weaknesses
  1. Insufficient architectural details of the conditional diffusion module. The paper introduces conditional diffusion modeling, but its implementation details are not sufficiently clear. In particular, the authors should specify the number of diffusion steps used during training and inference. In addition, the architecture of the conditional diffusion head should be described in greater detail, including its depth, hidden dimensionality, and other relevant configurations. Providing these details would improve the technical clarity and reproducibility of the work.

  2. Lack of sensitivity analysis on diffusion inference steps. I believe that the sensitivity of generation performance to the number of diffusion steps during inference is an important aspect of the model analysis. I suggest including such an analysis, as it would provide deeper insights into the behavior of the conditional diffusion head.

  3. The ablation analysis should be further strengthened. In particular, it would be useful to consider an additional variant based on a standard VAE, where the continuous latents are explicitly rescaled and normalized before being passed to the subsequent generative model. This comparison would help clarify whether the improvements primarily stem from the hyperspherical latent geometry introduced by S-VAE, rather than from simple scale normalization. I therefore suggest including this variant to provide stronger empirical support for the effectiveness of S-VAE.

  4. The setting of the maximum masking ratio is insufficiently justified. Since inference starts from a fully masked state, a larger masking ratio during training may reduce the train–inference discrepancy. However, the paper does not examine whether larger masking ratios lead to further improvements.

Questions for authors

Please refer to critical weaknesses

1 comment
AuthorsOpenReview ↗

Response to Reviewer Bm9o

Many thanks to the valuable comments of Reviewer Bm9o. We have revised the manuscript accordingly based on these valuable comments.

W1. Thank you for this insightful comment. We have added a detailed information of the conditional diffusion head, from Lines 960-963 in the Appendix D. For the diffusion head, the number of diffusion steps is set to 100 during both training and inference. Regarding its specific architecture, we employ a two-layer MLP with a hidden dimension of 512.


W2. We have added a sensitivity analysis on the effect of the number of diffusion steps during inference on generation performance, as presented in Appendix F, Lines 1244--1249 of the revised manuscript. The results show that reducing the number of diffusion steps weakens the refinement and denoising capability of the diffusion head, thereby limiting its ability to recover the distribution of latent representations and ultimately degrading the quality and fidelity of the synthesized time series.


W3. Thank you for this valuable suggestion. We have conducted additional experiments, and the results are presented in the table below:

DatasetHALO MMD ↓Normalized VAE MMD ↓HALO K-L ↓Normalized VAE K-L ↓
Electricity0.001 ± 0.0000.004±0.000‾\underline{0.004 \pm 0.000}0.002 ± 0.0000.006±0.001‾\underline{0.006 \pm 0.001}
Solar0.033 ± 0.0000.049±0.002‾\underline{0.049 \pm 0.002}0.019±0.000‾\underline{0.019 \pm 0.000}0.017 ± 0.000
Wind0.007 ± 0.0000.087±0.002‾\underline{0.087 \pm 0.002}0.063 ± 0.0020.113±0.006‾\underline{0.113 \pm 0.006}
Traffic0.068 ± 0.0010.075±0.001‾\underline{0.075 \pm 0.001}0.004 ± 0.0000.014±0.001‾\underline{0.014 \pm 0.001}
Taxi0.083 ± 0.0010.084±0.002‾\underline{0.084 \pm 0.002}0.008 ± 0.0000.012±0.001‾\underline{0.012 \pm 0.001}
Pedestrian0.071 ± 0.0010.075±0.001‾\underline{0.075 \pm 0.001}0.035±0.000‾\underline{0.035 \pm 0.000}0.034 ± 0.001

The results show that simply rescaling the continuous latents produced by a standard VAE cannot achieve performance comparable to that of the S-VAE. This is because simply rescaling the continuous latents produced by a standard VAE does not provide the directional modeling capability of the S-VAE in the latent space. This causes the latent representations to lose information in both scale and direction, thereby negatively affecting the model's generation performance.


W4. We sincerely appreciate your insightful comments. We have added additional experiments on the minimum masking ratio in Appendix F, Lines 1236--1241 of the revised manuscript. The results show that increasing the minimum masking ratio does not consistently improve generation quality and can even lead to performance degradation on several datasets. This is because we adopt a truncated dynamic masking strategy. Although a larger minimum masking ratio makes the training condition more similar to the fully masked setting encountered during inference, it simultaneously narrows the range of candidate masking ratios, thereby reducing the diversity of masking patterns and visible contexts observed during training. This weakened diversity limits context learning and ultimately degrades the model's generation performance.

Comments on the paper

1 comment

For all revierers: as the discussion period is approaching its end, please join the discussion and put acknowledge or feedback on authors' rebuttal, thank you!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.