acceptodds
Under review as a conference paper at ICLR 2027

FROM SCORES TO ACTIONS: CERTIFIED PROMOTION AND K > 2 CONTROL OF JOINT TIME-SERIES GENERATORS

Abstract

Joint time-series generators should create new trajectories while allowing rela- tionships to evolve and protecting what a task requires. We call this goal selective structural fidelity. It raises three distinct questions: are the generated paths plau- sible, is a requested change executed correctly, and is there evidence to adopt it? We develop a reference-aware evaluation contract that separates these questions. Our rank-adaptive K >2 path controller (KRPC) realizes matrix requests through thin products, with endpoint and complement preservation, basis covariance and explicit refusal. The evidence distinguishes execution from benefit. On a locked 2024 weather panel, event-preserving correction improves six quality measures over its calibrated reference; path conditioning also improves Energy and incre- ment Energy over a matched history-conditioned control. A distinct, locked 2025 expert policy improves Energy by 0.0600% over its protected reference, but does not lead the full roster. Rank-4 requests on K = 500 streamflow ensembles es- tablish high-dimensional execution. EEG and controlled decision studies expose quality–coverage and information limits. These are task-specific gains, not uni- versal dominance.

Primary area:
learning on time series and dynamical systems
Keywords:
multivariate time-series generation, controllable generation, structure-preserving generation, low-rank path control, probabilistic forecasting, proper scoring rules

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_WxuiOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper is concerned with synthetic time series generation and how generation can be edited post-hoc to enforce certain characteristics while changing others. It proposes 2 ideas: 1. A three question protocol that separates plausibility (How close is the generator before post-hoc editing to the true distribution?) correctness (Is the requested characteristic properly enforced while others are preserved?) and adoption (Does it actually improve the target metric) aiming to improve the protocol that is typically used to evaluate post-hoc approaches. 2. It proposes a method (KRPC) that enforces exact specific preservation with a refusal semantic. The paper then demonstrates applications of the framework in 4 studies from which 2 use KRPC.

Critical strengths

The paper focuses on improving research protocols and battling the general issue of overclaiming experimental results. In my opinion, work in this direction is quite important, especially given the current submission numbers.

The conceptual framework is intuitive and seems to tackle a real evaluation problem for post-hoc time-series generator editing.

The paper flags its own gaps very consistently which is transparent and intellectually honest.

Critical weaknesses

I have a number of concerns that I am listing below (not ordered by importance):

I believe the abstract and the introduction need some revisions as they are failing to provide a clear entry into this work. I base this judgement on the following points:

  • The abstract is oddly specific about the results of the proposed method which, at this point, are not particularly insightful.
  • Certain terms like "base law", "correct action" "justified" are comprehensible given the full manuscript but are not clearly defined in the introduction.
  • Figure 1 is referenced directly on the first page but is in itself not comprehensible without Formula 3 which is only introduced much later.
  • The literature review is extremely dense and does not really provide a good categorization of the work as it is not really clear how KRPC differs from other post-hoc methods and whether other works proposed similar protocols for post-hoc editing.

I would recommend to include some actual problematic examples of post-hoc edits and corresponding judgements here that motivate the proposed protocol.

As far as i understand KRPC is a concrete (but not a unique) executor that allows the framework to be implemented in practice. However, given the results in the appendix that show that the method is not necessarily better than other established approaches, I wonder why it takes such a prominent role in the manuscript. Making both the method and the framework equal contributions seems odd.

The conclusion seems rushed and is not particularly informative by itself.

Foundation models (e.g. Sundial as mentioned in the introduction) seem to be the strongest practical target for post-hoc methods. It seems like all experimental results are based on generators that are generally retrainable.

Questions for authors

It seems like only Question 2 (correctness) can be fully verified. How does this affect the usability of the proposed framework in practice?

How does one choose which properties of a generated time-series should be protected? I assume, in practice, it is often unclear what should actually be preserved exactly?

Given that we have a generator that we can actually retrain or change. When is the proposed framework still preferable?

Would it be possible to include an example of a post-hoc method in combination with a foundation model?

1 comment

I thank the reviewers for their effort to address my concerns. With the state of the current manuscript I chose to keep my scoring.

Reviewer_FPiQOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes a framework for evaluating and controlling generators of joint multivariate time series. A "request" names an editable relation, and a "protection" names a property that must remain fixed. The authors want three questions to be answered separately: whether the generated paths are plausible, whether the requested change was executed correctly, and whether adopting it is justified ("certified promotion", an ordered conjunction of gates). The method, KRPC, is a rank-rr linear edit of the time-centered increments in a low-dimensional channel subspace that leaves the complement and the endpoint unchanged (Theorem 2) and can return NO-OP or ABSTAIN. Theorem 1 is the standard Energy-score identity. The evidence comes from locked MERRA-2 weather panels, an EEG generator, a K=500K=500 streamflow study and a synthetic decision study.

Critical strengths
  1. Careful and honest reporting. The paper separates the latent law from the finite ensemble and reports unfavorable results, e.g., the 2023 correction improves Energy but worsens the q99 error by 16.7%, and the half-mix ablation has a lower finite-ensemble Energy than the EEG candidate. The coupling-bias identities for the ensemble Energy score are correct and practically relevant.
  2. KRPC is cheap (O(NHKr+NHr2)O(NHKr + NHr^2)), satisfies exact identities, has an explicit refusal path, and is evaluated up to K=500K=500.
Critical weaknesses
  1. The contribution is small and not shown to matter. Theorem 1 is standard (as the authors note), Theorem 2 follows from U⊤U=IU^\top U=I and PH2=PHP_H^2=P_H, and the paper itself reports that SVD projection, protected whiten–color and dense OT reproduce or tie KRPC. The examples in which the contract changes a decision concern an earlier version of the method or a synthetic study, and the headline gains are very small (0.097% and 0.0600% in Energy, ranks 8/21 and 3/22).
  2. The evidence is hard to trust and hard to check. The MERRA-2 comparisons use a different method version each year on data used before for development (in the paper's words, "not unchanged-model replications"). The EEG baselines are single-seed local configurations on a previously used cohort, and the primary metric EoffE_{\mathrm{off}} is declared only in the post-hoc Appendix F.1, whereas Fig. 2 shows VV, on which the half-mix ablation beats the candidate. The released material replays only Table 4 and the EEG aggregate statistics, so the main tables cannot be reproduced end to end.
Questions for authors
  1. What is the technical contribution beyond a rank-rr linear rescaling (Weakness 1), and what can a practitioner do with it that was not possible before? Can you give one example on a standard benchmark in which "certified promotion" overturns the decision of a conventional leaderboard?
  2. Why should Energy differences of 0.097% and 0.0600% be considered meaningful? What is the year-to-year variability, and can you show a confirmatory result for one unchanged method on data that were not used for development?
  3. Which headline numbers are confirmatory, which are post-hoc or based on reused data, and which come from synthetic failure families? Which metric was declared primary before the results were seen (VV or EoffE_{\mathrm{off}}, Fig. 2 vs. Appendix F.1), and why is the candidate preferred over half-mix?
  4. Can the main text be made self-contained, with key terms defined where they first appear (e.g., "D3U", p. 7; "native fiber", p. 8; "gates", defined only in Appendix D) and a figure of the full pipeline? In addition, the later figures (Figs. 3, 6 and 7) need polishing: text reaches or crosses the box borders, a note box covers an axis label, there is much empty space, and the fonts are too small at print size. Finally, could the conclusion state the main findings concretely, i.e., what was achieved, on which data, and by how much?
Reviewer_Bt1GOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a “selective structural fidelity” evaluation contract that separates base-law quality, request execution and adoption evidence for joint time-series generators.

Critical strengths
  1. the overall motivation is clearly addressed and it is reasonable.
  2. the framework is easy to adapt and design principle is simple but effective.
Critical weaknesses
  1. the baseline is too weak. TimeGrad and Diffusion-TS only used 1-step diffusion, which does not make sense. Also, it's trained based on your simplified recipe.

  2. The evaluation data is repeatedly use and hard to find without reading the supplementary file.

  3. The EEG generator (Eq. 4) constructs each sample as a real training anchor plus a shared offset, so that Y_m − Ȳ = A_m − Ā holds exactly; the within-ensemble variability is therefore copied verbatim from training subjects rather than generated. Consistent with this, Table 15 shows the candidate’s nearest-training distance is 0.187, versus 1.026 for real test data and 0.946 for TimeGrad, and the zero-exact-copy result arises only because of the common offset.

Questions for authors

I think the paper still needs improvement to be more convincing.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.