acceptodds
Under review as a conference paper at ICLR 2027

Learning Less is More: Adaptive Modular Foundation Model for Time Series Forecasting

Abstract

Many existing time series forecasting methods assume that input time series exhibit specific properties, such as seasonality, irregularity, or non-stationarity, and design network modules with these assumptions. Within these methods, foundation models often cover network modules with different assumptions, in order to learn input time series with diverse properties. However, such foundation models process each input time series through all modules, regardless of whether their property assumptions match the input, potentially reducing forecasting accuracy. To address this challenge, we propose ReTS, an adaptive modular foundation model for time series forecasting, where only modules with matched property assumptions assemble together for different input time series at run time. First, by decomposing popular time series forecasting models, we reveal the property assumptions of their network modules, and rearrange these models into a unified recurrent activation structure. Based on this unified structure, ReTS is proposed, which adaptively excludes modules with mismatched property assumptions for each input, thereby avoiding imposing properties the input does not exhibit. Furthermore, the activated structures of ReTS can closely resemble the structures of different baselines for different inputs, and even form unprecedented ones. Experiments on eighteen datasets demonstrate competitive performance against nineteen baselines, with 91% less computational cost at inference than the next-smallest pre-trained baseline. Our code is available at https://anonymous.4open.science/r/ReTS-6307/.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Foundation Model, Irregular Multivariate Time Series, Time Series Forecasting

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_PFyiOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ReTS, an adaptive modular foundation model for time series forecasting. It addresses the problem that existing foundation models process every input through all modules, even when a module’s property assumption (e.g., periodicity, non-stationarity, missingness, or variable dependency) does not match the input, which may reduce forecasting accuracy. To solve this, the authors decompose existing forecasting models into modules and rearrange them into a unified recurrent structure. A property learner then activates only the modules whose assumptions match each input and decides the number of recurrent steps.

Critical strengths
  • The paper provides an interesting view that explains the structures of many existing forecasting models through their property assumptions, and builds a modular foundation model that selects modules for each input.
  • ReTS achieves competitive results on both regular and irregular datasets, with far fewer parameters and much lower computational cost than existing foundation models.
Critical weaknesses
  • The authors admit that the reproduced zero-shot results of several baselines are much worse than those reported in the original papers. In addition, only small or older variants and short lookback windows (e.g., 24 steps for ETTm1) are used, which may underestimate the baselines.
  • ReTS is pre-trained on the training sets of the other seventeen benchmark datasets, which include closely related datasets (e.g., ETTm1 when testing ETTh1, MIMIC-III when testing MIMIC-IV). Meanwhile, baselines are excluded on datasets that overlap their pre-training data, so the zero-shot comparison is not fair.
  • According to Appendix C.2, the activated structures are mostly close to Leddam for regular data and HyperIMTS for irregular data. Moreover, w/ NAS performs almost the same as ReTS (0.69 vs. 0.67 MSE), so the benefit of the proposed module activation is not clear.
  • In Figure 1 and Appendix A, removing mismatched modules changes MSE by only about 1–5% on a few datasets, and no variance is reported. This is not strong enough to support the main motivation.
Questions for authors

Please refer to Critical Weaknesses.

1 comment

We thank the reviewer for finding the property view interesting and the efficiency results strong.

W1. Reproduction of baseline results and lookback windows

A1: We replicate the baselines using their own upstream repositories. Sundial reproduces to within 0.16% on ETT against our Table 1:

UpstreamΔ\Delta MSE
ETTh10.4728-0.01%
ETTm10.5375-0.04%

On ECL, MOIRAI agrees to within 0.5% at {96, 192}. The remaining gap at {336, 720} is MOIRAI's unstable performance exposed by the differences between the two standards: the two pipelines zero-pad the lookback window to a different length before encoder, and MOIRAI's any-variate attention is sensitive to that padding.

As already mentioned in paper, we have double-checked these entries, including code differences and checkpoint integrity.

only small or older variants

Small: Only one model, MOIRAI-small, is intentionally chosen as the small variant, since our work compares both accuracy and efficiency, where it is the closest baseline to ReTS's scale;

Old: We follow the latest baseline Zeus (ICML 2026) to select Chronos-Bolt, and TimesFM 2.5 is already the latest variant by the time we submit this manuscript in Sept 2026. Other baselines do not have newer variants in their original papers.

24 steps for ETTm1

This is one of the few mistakes in the manuscript, thanks for pointing out. ETTm1 also uses a 96-step lookback window, like other regular time series datasets, and it is now corrected in Appendix H.

W2. Zero-shot comparison fairness

A2: Three-fold

  1. Gift-Eval: We re-pre-train the ReTS family on GiftEvalPretrain covering 88 datasets (no overlap with GiftEval zero-shot), benchmarking variants from Luna to Astra (new variant based on Luna with hidden dimension 1280 and 20 layers, matching TimesFM), against the official scores reported by GiftEval:

    geomean MASE
    TimesFM0.9824
    ReTS Astra0.9866
    Zeus0.9894
    Sundial1.0559
    Chronos1.0597
    Moirai1.2803
    ReTS Luna1.4248

    ReTS Astra ranks the second, with 13G MACs of FLOPs, still smaller than any baseline. It should be noted that GiftEval primarily contains TS with seasonality and no missingness, where the TS property combinations are not diverse enough to fully demonstrate the power of dynamic module activation in ReTS.

  2. Pre-training MOIRAI on the same set of datasets as ReTS in manuscript, its averaged MSE improved to 1.10 over the 18 datasets, but ReTS still has 39.2% lower MSE.

  3. ReTS is never pre-trained on the dataset it is evaluated on: Excluding the five datasets, ReTS is also the best performing in zero-shot on 9/13 (MSE) and 7/13 (MAE) datasets.

W3. Activated structures and the NAS comparison

A3: Expected results. Cited from Section 5.5: ... achieve similar performance against NAS, without searching through numerous candidate structures.

We measured a 3h stage 2 training of ReTS, while the NAS search on the same set of modules and GPU cost 12h. Theoretically, a new module only adds one binary decision in ReTS and the cost grows as O(∣M∣)O(|\mathcal{M}|), while the search space of NAS grows as O(2∣M∣)O(2^{|\mathcal{M}|}), where M\mathcal{M} is the set of modules.

W4. Effect size and variance of the motivation ablations

A4:

changes MSE by only about 1–5% on a few datasets

In time series forecasting research, such improvement is actually well within expectation among existing papers, not to mention we significantly improve efficiency at the same time.

We measured this on the baselines' own published tables. The winner's margin over the 2nd is a mean of 2.28%, and 72 of the 82 entries fall below 5%. Restricted to the 4 foundation models, the margin over the runner-up is mean 1.82%, range 0.30%~7.83% over 25 entries, with 23 of the 25 below 5%:

MinMaxMean
Informer0.27%9.12%3.21%
PatchTST0.00%9.44%1.90%
MOIRAI1.16%7.83%3.16%
Sundial0.30%1.78%1.09%
VisionTS0.57%2.50%1.77%
MSHLLM0.74%2.69%1.35%

no variance is reported

  1. As already stated in Appendix D, every experiment is run with five random seeds, 2024 to 2028, and the mean is reported.
  2. Unlike non-pretraining time series forecasting papers, due to much broader evaluation scope, pre-training papers usually do not report variance in order to fit the giant benchmark tables into paper. This is the practice of the baselines: Zeus, Sundial, TimesFM and Chronos report means without standard deviations in their zero-shot benchmark tables, while MOIRAI reports them only for its probabilistic table. Our paper already spans 40 pages, containing forecasting results well beyond most baselines we compared.

Thanks again for your careful review! We hope that the answers meet your expectations :).

Reviewer_Zew9OpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper argues that many time series models are designed around assumptions about the input (seasonality, missing values, non-stationarity, dependence across variables), and that foundation models combining several such modules waste compute, and maybe accuracy, by running all of them on every series. The authors split a few existing models (MOIRAI, PatchTST, HyperIMTS, and others) into modules, put them into a single recurrent block, and train a small network (ReTS) that looks at the input, predicts its properties, and switches modules on or off and sets the number of recurrent steps. The property labels come from hand-crafted statistics such as an ADF test and a spectral measure. ReTS is pre-trained on the training splits of 18 regular and irregular datasets and evaluated on those same 18 datasets, leaving the test dataset out for zero-shot and few-shot. It is compared with 8 pre-trained models and 11 others. The headline claims are the biggest wins in the zero-shot table: about 91% fewer MACs than MOIRAI-small and model structures that adapt to each input.

Critical strengths
  1. The compute saving is real in the reported numbers: roughly 10x fewer MACs than MOIRAI-small, with similar or better error on most irregular datasets. This is the one result that could matter in practice, and only if the evaluation is sound (see below).
  2. The benchmark is broad, covering both regular and irregular series, and ReTS is averaged over five seeds. The idea is easy to follow.
Critical weaknesses
  1. Almost nothing is new. Every module comes from an existing model. The only new part is a gate: a linear threshold on a vector trained to predict classical statistics (ADF test, spectral concentration, missing mask). Choosing components by input properties is routine: model selection and ensembling (Chroma and SwiftTS, which the paper cites), mixture-of-experts, and adaptive depth. Section 4.4 separates ReTS from these only by saying its granularity is finer. The premise that matching the inductive bias to the data helps is well known, and Figure 1 adds a few small single-dataset differences. I learned nothing new from this paper.

  2. The zero-shot comparison is not a fair one. ReTS is pre-trained on the training splits of the other 17 benchmark datasets, and several of those are closely related to the held-out one: ETTh1 and ETTm1 are the same transformer data at two sampling rates, and MIMIC-III, MIMIC-IV, and PhysioNet’12 are all ICU records. The baselines are not given this. So Table 1 shows how a model trained on the benchmark performs on the benchmark, not how a foundation model performs zero-shot, and I don’t think an 18-dataset pre-training set justifies the name “foundation model”. The obvious control, training the strongest baselines on the same 17 datasets, is missing, and the full-shot table doesn’t replace it.

  3. The baseline numbers also look unreliable. MOIRAI gets 0.67 MSE on ECL and Sundial 0.47 on ETTh1, which differ clearly from what their papers report. Section 5.2 acknowledges this and blames the open-source weights, without evidence. If the replication is wrong, I can’t draw much from who wins Table 1. Even taking the table at face value, most of the wins are on irregular datasets, which these baselines were never built for. On regular data, ReTS wins ECL, ties Weather, and loses ETTh1 (0.66 vs 0.47), ETTm1 (0.65 vs 0.54), and Traffic (1.13 vs 0.75). Using “Closely catching up” understates this.

  4. Dropping mismatched modules keeps or improves accuracy rests only on Figure 1: a handful of single-dataset removals, mostly 0.00–0.02 MSE, with the largest at EPA-Air (0.59 to 0.48), no variance. Beyond that, I can’t see what is actually learned. The property labels come from fixed detectors, and the link from properties to modules is chosen by the authors, so the network mostly reproduces those statistics. A rule that gates modules directly from the detectors is the obvious baseline, and it is missing. The ablations don’t settle this. “w/o activator” (1.46 vs 0.67 MSE) shows that a model trained with gates fails without them, not that the extra modules hurt, and “w/ NAS” (0.69 vs 0.67) is about the same as the full model. The “unprecedented” structure for FNSPID has the same error as several existing baselines (0.12/0.20), so there is no sign the new structure is better.

Questions for authors

Mostly see weaknesses. One extra point: do the reported MACs include the property learner and the depth head, and how much of the saving comes from gating rather than from ReTS simply being a small model? A dense ReTS of the same size and wall-clock latency would answer this.

3 comments

We thank the reviewer for acknowledging the compute saving and the broad benchmark. However, the listed weaknesses are heavily biased without providing genuine evidence to support score 1.

W1. Novelty

A1:

  1. No genuine evidence: PLEASE NAME any existing method that can use one unified structure to represent a majority of existing regular and irregular TS forecasting models with completely different backbones. In contrast, we have covered all reviewer-listed evidence in Section 4.4, with further novelty proof in Tables 1, 7, and 8 and Figures 2, 5, and 13.

  2. Don't confuse "easy-to-understand" idea with "well known":

    We provided the high-level essence over dozens of appearances: designing modules, like seasonal-trend decomposition or channel-independent VS dependent, can be concluded as TS property matching. It is easy to understand from essence to appearances, but not the reverse without human insights.

  3. The novelty lies in:

    • The high-level model structure unification over heterogeneous forecasting models.
    • The mainstream usually does addition, while our work does reduction.

W2. Zero-shot fairness

Three-fold

A2.1 Gift-Eval

We re-pre-train the ReTS family on GiftEvalPretrain covering 88 datasets (no overlap with GiftEval zero-shot), benchmarking variants from Luna to Astra (new variant based on Luna with hidden dimension 1280 and 20 layers, matching TimesFM), against the official scores reported by GiftEval:

MASE
TimesFM0.9824
ReTS Astra0.9866
Zeus0.9894
Sundial1.0559
Chronos1.0597
Moirai1.2803
ReTS Luna1.4248

ReTS Astra ranks the second, with 13G MACs of FLOPs, still smaller than any baseline. It should be noted that GiftEval primarily contains TS with seasonality and no missingness, where the TS property combinations are not diverse enough to fully demonstrate the power of dynamic module activation in ReTS.

A2.2 Train baseline on the same datasets

MOIRAI is the only model that fits 80GB VRAM on the exact same pre-train datasets as ReTS, reflecting the outstanding efficiency of ReTS. Although MOIRAI indeed improved, ReTS still wins on all datasets:

MOIRAI (same pre-train)ReTS LunaΔ\Delta MSE
Avg.1.100.67-39.2%

A2.3 Fairness

ReTS is never pre-trained on the dataset it is evaluated on. Also, excluding the five datasets, ReTS is the best performing in zero-shot on 9/13 (MSE) and 7/13 (MAE) datasets.

W3. Baseline reproduction

A3:

look unreliable...differ clearly from what their papers report

We replicate the baselines using their own upstream repositories. Sundial reproduces to within 0.16% on ETT against our Table 1:

UpstreamΔ\Delta
ETTh10.4728-0.01%
ETTm10.5375-0.04%

On ECL, MOIRAI agrees to within 0.5% at {96, 192}. The remaining gap at {336, 720} is MOIRAI's unstable performance exposed by the differences between the two standards: the two pipelines zero-pad the lookback window to a different length before encoder, and MOIRAI's any-variate attention is sensitive to that padding.

Closely catching up understates this...most of the wins are on irregular

Only comparing ECL, ETTh1, ETTm1, Traffic, and Weather, ReTS ranks the 2nd in zero-shot and the 1st in full-shot.

W4. Motivation

A4:

only on Figure 1

Figure 1 and the 3-page-long Appendix A cover 35 ablations over 11 modules across 8 baselines and 8 datasets, all supporting our claim.

property labels come from fixed detectors

The detectors only produce a portion of the training target, instead of directly activating modules. Instead, the high-dimensional input time series representation of the property learner, carrying way richer semantics than labels, is for activation. Replace learned gates by raw detector scores, and average over the 18 datasets:

Avg. MSEAvg. MAE
detector-gated0.800.58
ReTS Luna0.670.49

Detector alone does NOT have the expressive power of the dynamic input time series representation ReTS actually uses.

w/o activator...not that the extra modules hurt

w/o activator is the exact dense version you ask for in your questions, trained separately, so it IS the proof that extra modules hurt.

w/ NAS is about the same as the full model

Expected results. Cited from Section 5.5: ...achieve similar performance against NAS, without searching through numerous candidate structures.

no sign the new structure is better

It is a better structure than other activated structures of ReTS instead of arbitrary heterogeneous baselines, proving our motivation that given a set of modules, only activating matched ones is better.

Q. MACs

A5: MACs include the property learner and the depth head, occupying 0.04%~4.7% across the datasets; The saving does not come from size alone: The dense model is still 9.1% of MOIRAI's MACs.

Thank you for the detailed rebuttal and additional experiments. I appreciate the authors' efforts to address my concerns, particularly through the GIFT-Eval evaluation, matched pre-training comparison, and detector-gated baseline.

These additional results partially address my concerns regarding experimental soundness and evaluation fairness. However, two key issues remain:

  1. The proposed framework primarily integrates existing modules through conditional activation. Although the detector-gated comparison is informative, the independent contribution of property-guided supervision over generic learned routing remains unclear. A controlled comparison using the same backbone without property supervision would be important.

  2. The authors clarify that dense ReTS already requires only 9.1% of MOIRAI's MACs. Therefore, much of the reported computational advantage appears attributable to the lightweight backbone rather than adaptive module selection. The incremental efficiency gains from dynamic activation remain insufficiently demonstrated.

Overall, I acknowledge the additional experimental evidence and the practical value of the proposed framework. I have therefore raised my rating to reflect the improvements in the rebuttal. Nevertheless, I remain unconvinced that the methodological contribution is sufficiently substantial for acceptance, and my overall recommendation remains negative.

AuthorsOpenReview ↗

Further efficiency clarifications

We thank the reviewer for raising the score and further discussing potential improvements. Similar to the concern in W4 "no sign the new structure is better", both the accuracy and efficiency contribution of adaptive module activation lie in given the same set of modules, only activating matched ones is better, instead of better than arbitrary heterogeneous baselines. We have updated efficiency analysis in Section 5.3 and Appendix E.5 to include further clarifications, distinguishing the efficiency gains of backbones and our proposed mechanism. Since both flexibility (numerous backbones) and less module computations are the advantages of our proposed unified recurrent activation structure, ReTS Luna is still adopted in Table 1 to demonstrate the potential of small models providing state-of-the-art accuracy leveraging both advantages.

To isolate the efficiency contribution of activation itself:

Empirical evidences

MOIRAI small is already the scale-closest yet performant recent baseline we can find, since ReTS is too efficient compared to baselines. We further adjust hidden dimension of ReTS in following experiments to 1024 and 20 layers, letting the dense variant match the total budget, and results are averaged over the 18 datasets in paper. The time is measured using the same hardware as the one used in Section 5.3.

ReTS (dynamic, matched budget)ReTS (dense, matched budget)MOIRAI
Flops11.33 G MACs14.95G MACs14.61 G MACs
Wall-clock inference time37.83 ms52.45 ms48.06 ms

The isolated improvement for activation itself is approximately 22% in terms of flops over the dense variant and MOIRAI. In terms of wall-clock inference time, it is 21% and 28% better than MOIRAI and the dense variant, respectively.


Lastly, thanks for helping us improve the manuscript.

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

This paper proposes ReTS, an adaptive modular time-series foundation model designed for time-series forecasting. The proposed framework learns representations corresponding to key time-series statistical properties—such as temporal periodicity, non-stationarity, missingness, and variable dependency—and utilizes these representations to dynamically route computation by selecting forecasting modules and controlling recurrence depth. Extensive evaluations across 18 datasets show competitive forecasting accuracy and computational efficiency compared to standard baselines.

Critical strengths
  1. Relevant Research Direction: Dynamically adapting model architectures to heterogeneous time-series characteristics is a well-motivated and impactful problem in time-series foundation modeling.

  2. Favorable Efficiency and Comprehensive Assessment: The reported accuracy-to-compute trade-offs are compelling. Furthermore, the inclusion of few-shot, full-shot, and model scaling experiments provides a broader perspective on the framework's empirical performance.

Critical weaknesses
  1. Unsubstantiated Property-Module Associations: The mapping between modules and properties (Appendix B, Table 3) relies on manual heuristic associations. However, demonstrating that a module models a specific property well does not imply that the module is ineffective or inappropriate when that property is absent. For instance, patching and convolutional operations exploit localized temporal structure regardless of strict periodicity. The paper lacks rigorous empirical or theoretical justification for treating these heuristic associations as reliable inductive priors for module selection. Furthermore, current experiments (Section 5.5, Appendices C.3 and E.6) fail to disentangle property-guided routing from generic conditional routing, sparsity effects, or dynamic capacity changes.

  2. Incomplete Architectural Specification: Sections 4.1 and 4.2 describe property-aware representation learning and module activation, but leave the execution details of the overall forecasting backbone underspecified. Although Appendix C.1 and Table 4 detail individual modules and interfaces, essential architectural details—such as input sequence scaling/embedding mechanisms and inter-module hidden state propagation—remain too high-level to enable confident reconstruction.

  3. Inadequate Training Details and Reproducibility Concerns: The core three-stage training procedure is omitted from the main text and described only in Appendix C.6. Critical training configurations—including learning rate schedules, optimizer setups, and compute budget allocations across stages—are missing. Additionally, baseline fine-tuning protocols in Appendix G are incomplete, hindering reproducibility and raising questions about fairness in baseline comparisons.

  4. Over-Reliance on Metaphorical Exposition: The main text relies heavily on visual illustrations and informal analogies at the expense of formal mathematical definitions and algorithmic formulations, which are relegated to the appendix. This presentation structure undermines the self-containedness and technical rigor of the main paper.

Questions for authors
  1. Validation of Property-Module Assumptions: What empirical or quantitative evidence demonstrates that the manually selected property-module mappings reliably predict when a module will enhance forecasting accuracy compared to unconstrained routing?

  2. Stage-Level Ablation and Gating Dynamics: Could you provide performance metrics after each of the three training stages alongside stage-level ablation studies under normalized compute budgets? Specifically, how does transitioning from soft to hard gating impact performance, and what quantitative gains are directly attributable to the third adaptation stage?

  3. Generalizability to Scaled Time-Series Foundation Models (TSFMs): Can the proposed property-guided module activation strategy be integrated into larger pretrained TSFMs (e.g., Moirai, Chronos)? If so, what performance improvements does it offer over baseline models on broader evaluation suites such as GIFT-Eval?

3 comments

We thank the reviewer for acknowledging our comprehensive assessments and compelling accuracy-to-compute trade-off. However, there are many misreadings over the contents we already presented, and we disclose why in the last section.

W1 & Q1. Property-module associations; generic routing

A1:

mapping...relies on manual heuristic associations...lacks rigorous theoretical justification

As already detailed using 3 pages in Appendix B, associations are based on solid prior works in property classifications [1-3] and dozens of method descriptions in existing TSF methods, instead of inventing ourselves.

a module models a specific property well does not imply that the module is ineffective or inappropriate when that property is absent...lacks rigorous empirical justification

We already used Figure 1 and a 3-page-long Appendix A to empirically validate it, covering 35 ablations over multiple baselines and datasets, all supporting our claim.

disentangle property-guided routing from generic conditional routing, sparsity effects, or dynamic capacity changes

generic conditional routing, In Table 2:

  • fixed-depth and w/o activator separate depth adaptation from module selection
  • w/ NAS keeps selection and sparsity but removes the property grounding

sparsity effects or capacity changes: The ablation of w/o activator already compares the dense version, and it's nearly impossible to fix the capacity for a set of heterogeneous modules.

[1] Xiaozhe Wang, et al. Characteristic-Based Clustering for Time Series Data. DMKD 2006

[2] S. Makridakis, et al. Forecasting: Methods and Applications. 1998

[3] Yong Liu, et al. Non-stationary Transformers: Exploring the Stationarity in Time Series Forecasting, NIPS 2022

W2 & W3. Reproducibility

A2:

As already stated in Appendix C, our code is always available in the anonymized repo for exact reproduction, not to mention we have already provided a 3-page-long Appendix C to cover a wide range of method details, and Section 5.1 and Appendices D, G, and H totaling 7 pages to cover details on dataset, baselines, and training.

Q2. Stage-level ablation

A3:

StageAvg. MSEAvg. MAE
11.090.66
20.930.55
30.670.49

1: Full-arch warmup; 2: Focus on property learner and module activators; 3: Freeze learner and activators to focus on activated modules. The transition from soft to hard gating in stage 3 is the biggest gain, matching our motivation. As for the time per stage, they cost 3.5h, 3h, and 2.3h respectively on A100.

W4. Metaphor

A4: They are intended for the ease of understanding of human readers instead of LLMs, not to mention we have already covered all the rigor-required math both in Section 4 and Appendix C, with ten block equations. The page limit of the main body is 9 pages, which are already fully used to present more important methodology and experiments.

Q3. Generalizability to scaled TSFMs and GIFT-Eval

A6:

Generalizability to Scaled Time-Series Foundation Models

The property-guided module activation has already been illustrated to be capable of expressing MOIRAI in Figure 1, and Sundial, TimesFM, and Zeus in Figure 13.

GIFT-Eval

We re-pre-train the ReTS family on GiftEvalPretrain covering 88 datasets (no overlap with GiftEval zero-shot), benchmarking variants from Luna to Astra (new variant based on Luna with hidden dimension 1280 and 20 layers, matching TimesFM), against the official scores reported by GiftEval:

geomean MASE
TimesFM0.9824
ReTS Astra0.9866
Zeus0.9894
Sundial1.0559
Chronos1.0597
Moirai1.2803
ReTS Luna1.4248

ReTS Astra ranks the second, with 13G MACs of FLOPs, still smaller than any baseline. It should be noted that GiftEval primarily contains TS with seasonality and no missingness, where the TS property combinations are not diverse enough to fully demonstrate the power of dynamic module activation in ReTS.

Concerns on fully AI-generated review

  • Unaware of the 9-page limit in W4 and of the requirement to provide "one or two critical weaknesses": not provided with enough context.
  • Ignoring already presented content: likely due to lack of attention.
  • Obvious combinations of LLM wording: a bolded short sentence, then use "——" to explain it, then use "()" to attach evidence. Any human familiar with frontier LLM outputs should be able to identify this, and replicate similar reviews.

This review has added substantive noise to our manuscript's evaluation, where its score 1 cannot be supported by the listed weaknesses, since we already covered nearly all weaknesses-related explanations in our manuscript.

Thanks for the rebuttal, but I don't think the authors have addressed my concerns.

Module selection relies substantially on prior knowledge. While the paper provides evidence that this selection is effective, it does not establish that the selected components or their combinations are optimal. Length of pages in appendix does not directly answer this concern.

I insist that more details about the architecture and training procedures should be included in the main text, rather than emphasizing textual distinctions from existing methods. ICLR 27 explicitly allows one additional page during the rebuttal phase to accommodate new results and discussions. I believe this space would be better used to incorporate more architectural and training details.

For TSFM genralizability, I was expecting discussions about whether the modular selection part, which I think is the contribution, could be adpatted to existing TSFM backbones, and see whether it brings performance gains to existing models. I did not see such kind of discussions or responses. Besides, since ReTS is claimed as a foundation model and adopt TSFMs as SOTA, it is necesary to compare ReTS with them on popular benchmarks like Gift-eval and fev-bench. I appreciate that you have provide some results on Gift-eval in the rebuttal, but your experimental setting is not clear sice TimesFM-3 already reports mse less than 0.7 in the gift-eval leaderboard recently.

I understand doing a paper is time and energy consuming. Since none of my concern has been fully considered. I'm sorry but I cannot support the acceptance of this paper.

As for the AI generation, I have submited "AI Documentation For Reviewing" when submiting this review. AC and PC could see my hand-written version. None of them are "generated" from nothing.

AuthorsOpenReview ↗

Round 2. On optimum, main-text details, TSFM adaptation and GIFT-Eval settings

We thank the reviewer for raising the score. However, we believe there are still plenty of noises affecting the fair judgement over our manuscript in the late response just 23 hours before the rebuttal period ends.

does not establish that the selected components or their combinations are optimal

We NEVER claim optimality. As already stated in Appendix A: "It should also be noted that given the numerous number of time series forecasting models and datasets, it is nearly impossible to guarantee that the conclusion always holds. However, what we can provide in this justification are the ablation results upon several well-known baselines and benchmarking datasets." Therefore, the claim is significantly narrower: modules carry property assumptions, and matching them at run time beats running all of them. This claim is proven by all the contents in this manuscript.

substantively on prior knowledge

As already explained in our response A.1 above, the papers we cite in Appendix B explain themselves about which module is intended for which purpose, so property -> module mappings are established facts rather than our guesses. e.g.,

  1. "Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting": moving average -> seasonal-trend decomposition
  2. "Revitalizing Multivariate Time Series Forecasting: Learnable Decomposition with Inter-Series Dependencies and Intra-Series Variations Modeling": 1-D CNN -> seasonal-trend decomposition

would be better used to incorporate more architectural and training details

Since the reviewer insists on moving already presented 4-page architectural and training details (Appendix C and D) into the 1-page, we take the extra page ICLR 27 allows and add two new subsections to the methodology, Section 4.4 Architectural Specification of ReTS and Section 4.5 Training Schedule, both marked in red in the revised PDF.

can the modular selection be adapted to existing TSFM backbones, and does it bring gains

  1. Theoretical: As already clearly written in our response A.6 above, the property-guided module activation has already been illustrated to be capable of expressing MOIRAI in Figure 1, and Sundial, TimesFM, and Zeus in Figure 13.

  2. Empirical:

    Since the response arrived just 23 hours before the rebuttal period ends, it is impossible to evaluate all baselines. Picking the best-performing TimesFM in GIFT-Eval as an example, we decompose it based on Figure 13, and add the property learner and module activators to it. Upon the TimesFM 2.5 checkpoint, ReTS improves the zero-shot geomean MASE from 0.9824 to 0.9487. Note that we did not change the module weights, just decompose and add activation logics upon modules. We further analyze the gain, where the depth estimator may activate fewer layers (18 or 19 layers) than the original 20-layer baseline. Since GIFT-Eval datasets mostly contain non-stationary and seasonal series, nearly all modules will be activated, and the gain mostly comes from the adaptive layer depth of ReTS.

GIFT-Eval setting is unclear

GIFT-Eval officially provides unified inference codes in the SalesforceAIResearch/gift-eval GitHub repository to prevent any cheating, so there's no second setting we can ever use. As already written in our response A.6 above, for baselines, we directly reported the geomean MASE across datasets based on the officially provided metrics in repository's results/ folder. As for our model, we retrain ReTS from scratch on GiftEvalPretrain, the official pre-training dataset with no overlap with GiftEval, such that zero-shot is guaranteed.

TimesFM-3 already reports MSE below 0.7

Four-fold.

  1. As already clearly written in our response A.6 above, what we provided is geomean MASE instead of MSE.
  2. TimesFM-3 does not report MSE below 0.7 on GIFT-Eval. In file results/TimesFM-3/all_results.csv of the SalesforceAIResearch/gift-eval repository, the minimum per-dataset MSE is 1.33 across 97 configs. GIFT-Eval is scored in each dataset's native units and includes series with large magnitudes, so a raw MSE is not comparable across datasets and cannot be averaged into one number.
  3. There's no time we can ever add TimesFM-3 in our submission, given it was released on Aug 31 and the submission deadline was Sep 25, while our paper includes 3 (zero/few/full-shot) * 36 forecast horizons across datasets * 5 random seeds = 540 runs for each baseline.
  4. What ReTS demonstrates in paper is NOT the ultimate accuracy against any baseline, but rather achieving state-of-the-art accuracy with less than 1/10 computational cost of MOIRAI small and 1/100 cost of TimesFM 2.5.

Comments on the paper

2 comments
Area_Chair_uUQ1OpenReview ↗

Discussion Reminder

Dear Reviewers,

Thank you for your efforts so far. This is a friendly reminder to review the authors' rebuttal and respond to their comments as soon as possible, before the discussion deadline (November 18 AoE). Please also indicate whether your concerns have been addressed. Thank you for your time and cooperation!

Best regards,
AC

AuthorsOpenReview ↗

Rebuttal Summary

Dear AC and all readers:

As the rebuttal comes to an end, and since we believe some late comments (posted less than 24 hours before the deadline) evaluate our manuscript without proper grounds, we summarize here the current strengths, the addressed weaknesses, and the noise across the reviews and rebuttals.

It would be a real loss to the conference if humans cannot tell genuine progress apart from AI-generated and bias-driven noise in reviews. Our manuscript has 41 pages of extended content, covering the essence (time series property, high-level human insights) -> motivation (mismatched module) -> methodology (one structure expresses dozens of existing methods, an important step toward unifying irregular and regular time series forecasting) -> verification (19 baselines & 18 datasets & GIFT-Eval; zero-shot & few-shot & full-shot).

Strengths

  • Novelty: the property view of forecasting is interesting and new (AYiK, PFyi). Reviewer Zew9 silently agrees with it, since no further novelty concern or evidence was raised in later discussions.
  • Methodology: the accuracy-to-compute trade-off is compelling (AYiK, Zew9, PFyi).
  • Experiments: the benchmark is broad (Zew9).

Weaknesses, all addressed

  • [solved] Zew9 W2 & PFyi W2, zero-shot fairness: ReTS uses eighteen leave-one-out checkpoints and is never evaluated on a dataset it saw. New evidence: MOIRAI retrained on the same corpus still loses by 39.2%, and GIFT-Eval re-pre-training ranks ReTS Astra 2nd of 7.
  • [addressed, no response] PFyi W1, 24-step ETTm1 lookback: corrected in Appendix H to 96 steps.
  • [addressed, no response] Zew9 round 2, isolated efficiency gain of activation: added into the efficiency analysis in Section 5.3 and Appendix E.5.
  • [addressed, no response] AYiK round 2, main-text method details: re-summarizes already-existing content from Appendices C and D into the new Sections 4.4 and 4.5.

Noise

We found substantive noise in the reviews, originating mainly from the initial AI-generated review and from biased comments. As the reviewers did not mention them in later follow-up discussions, some have been silently resolved:

  • Claims that no evidence supports:
    • [silently solved] Zew9 W1, novelty: the review names no method that unifies regular and irregular models in one structure, and Section 4.6 already covers every example it lists.
    • [addressed, no response] AYiK round 2, TimesFM 3: it never reported MSE below 0.7 in the official GIFT-Eval repository.
  • Already clarified by the original manuscript:
    • [addressed, no response] AYiK W1, property-module associations and prior knowledge: 3-page Appendix B already covers, while the review names no counterexample. The prior knowledge rests on facts stated in the corresponding existing papers.
    • [silently solved] AYiK W2 & W3, reproducibility: the abstract already provides the anonymized repository plus 3-page Appendix C and totally 7-page Section 5.1, Appendices D, G and H.
    • [silently solved] AYiK W4, metaphor: Sections 4.1 to 4.3 give ten block equations that define the method, and metaphor is for human readers.
    • [addressed, no response] AYiK round 2, optimum: we never claim optimality, and the actual claim is already stated in Appendix A.
    • [addressed, no response] AYiK round 2, adapt to existing TSFM: shown theoretically in Figures 1 and 13, and verified empirically on TimesFM 2.5.
    • [addressed, no response] AYiK round 2, GIFT-Eval setting: the official codes are the only setting available.
    • [silently solved] Zew9 W4, motivation: Appendix A already holds all removals over 8 baselines and 8 datasets, so "only Figure 1" misreads a pointer as a missing result.
    • [silently solved] Zew9 W3 & PFyi W1, baseline reproduction: Section 5.2 already explains it. We further dig into the details: Sundial reproduces within 0.16% on ETT through its own upstream repository. The ECL gap at horizons {336, 720} from lookback padding length differing between the two pipelines, to which MOIRAI's any-variate attention is sensitive.
    • [addressed, no response] PFyi W4, variance: Appendix D already stated the five seeds, and we further prove that no zero-shot benchmark table in the baseline papers reports standard deviations, which supports our approach as standard practice.
    • [addressed, no response] PFyi W3, NAS: Section 5.5 already explained it is the intended outcome rather than a defect, since ReTS matches NAS without searching.
  • Point caused by the AI-generated review:
    • [solved] AYiK, review scope: the review is unaware of the page limit and of the "one or two critical weaknesses" instruction, and returns four.

We thank the AC and all readers for their efforts in keeping AI-generated noise and biased comments out of ICLR, in identifying genuine human insights, and in ensuring a fair review process.

Best regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.