- 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
-
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.
-
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 In addition, TimeGAN is outdated and has repeatedly been shown to perform poorly.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).
-
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.
-
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.
-
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
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.
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.