Writing study: CauKer: Classification Time Series Foundation Models Can Be Pretrained on Synthetic Data

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Shifeng Xie, Vasilii Feofanov, Jianfeng Zhang, Themis Palpanas, Ievgen Redko · Official ICLR 2026 oral record

Observed writing choices, not acceptance rules. This study examines the inspected manuscript version. It does not establish why the paper was accepted or that its presentation choices should be copied unchanged.

Figure, table and section ledger

Editorial contribution type: Datasets, benchmarks & measurement resources (primary) · Learning & optimization algorithms.

Layout extraction, not a required paper template. Counts are distinct caption labels; a multi-panel figure counts once. Section spans are estimated from heading positions, including floats and whitespace. Method and limitations.

Canonical PDF

29 physical pages

Before references / appendix

≈ 10.23 page-equivalents

Main figures / tables

7 / 2

Attached appendix figures / tables

6 / 11

Attached appendix span

≈ 14.74 page-equivalents

Before references / appendix: ≈ 35.28%References / intervening material: ≈ 13.9%Attached appendix: ≈ 50.83%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Section-by-section lengths and source pages

The title/abstract/front matter precede the first numbered heading. A span ends at the next detected top-level heading, not at the last sentence; end-page whitespace is included. References and statements are listed separately. These are not measurements of text-only or artwork area.

Top-level heading	Starts on PDF page	Estimated span	Share of entire PDF
1 INTRODUCTION	1	≈ 0.87 pp.	3%
2 RELATED WORK	2	≈ 0.81 pp.	2.79%
3 OUR CONTRIBUTIONS	3	≈ 2.26 pp.	7.79%
4 EXPERIMENTAL RESULTS	5	≈ 5.06 pp.	17.45%
5 CONCLUSION	10	≈ 0.52 pp.	1.79%
ACKNOWLEDGEMENTS	11	≈ 0.12 pp.	0.41%
REFERENCES	11	≈ 4.03 pp.	13.9%
APPENDIX	15	≈ 0.22 pp.	0.76%
A OVERVIEW OF PRE-TRAINING DATASETS FOR TIME SERIES FOUNDATION	15	≈ 0.33 pp.	1.14%
B LOSS AND ARCHITECTURE OF MANTIS AND MOMENT	15	≈ 0.78 pp.	2.69%
C DETAILS OF CAUKER	16	≈ 3.21 pp.	11.07%
D ADDITIONAL QUALITATIVE ANALYSIS	19	≈ 1.44 pp.	4.97%
E EXPERIMENTAL DETAILS OF SECTION 4.2.1	21	≈ 0.87 pp.	3%
F EXPERIMENTAL DETAILS OF SECTION 4.2.2	22	≈ 0.68 pp.	2.34%
G EXPERIMENTAL DETAILS OF SECTION 4.4	22	≈ 0.57 pp.	1.97%
H DOMAIN-WISE ANALYSIS ON UCR	23	≈ 1.17 pp.	4.03%
I SUPPLEMENTARY EVALUATION ON WOODS	24	≈ 0.68 pp.	2.34%
J PRE-TRAINING CHRONOS ON CAUKER AND ZERO-SHOT EVALUATION	25	≈ 1.49 pp.	5.14%
K VISUALIZATION OF EMBEDDINGS	26	≈ 1.16 pp.	4%
L ATTENTION ROLLOUT ANALYSIS.	27	≈ 0.89 pp.	3.07%
M ADDITIONAL EXPERIMENTS ON DOWNSTREAM FINE-TUNING	28	≈ 0.71 pp.	2.45%
N SUPPLEMENTARY EVALUATION ON IRREGULAR TIME SERIES	29	≈ 0.21 pp.	0.72%
O USE OF LARGE LANGUAGE MODELS	29	≈ 0.34 pp.	1.17%

All detected figure and table locations

Main text

• Figure 1 · p. 4
• Table 1 · p. 6
• Table 2 · p. 6
• Figure 2 · p. 6
• Figure 3 · p. 8
• Figure 4 · p. 8
• Figure 5 · p. 8
• Figure 6 · p. 9
• Figure 7 · p. 9

Attached appendix

• Table 3 · p. 15
• Figure 8 · p. 18
• Figure 9 · p. 19
• Table 4 · p. 20
• Table 5 · p. 20
• Table 6 · p. 20
• Table 7 · p. 21
• Table 8 · p. 22
• Figure 10 · p. 23
• Table 9 · p. 23
• Table 10 · p. 24
• Table 11 · p. 25
• Figure 11 · p. 26
• Figure 12 · p. 27
• Figure 13 · p. 28
• Table 12 · p. 29
• Table 13 · p. 29

Argument and organization

1. Classification needs both temporal patterns and cluster structure

   Motivates combining Gaussian-process kernels with structural causal models.

2. Five-step synthetic generator

   Turns the motivation into a numbered recipe and an aligned pipeline figure.

3. Q1–Q3 and extended evaluations

   Separates generator quality, scaling and sample efficiency before domain-specific appendix checks.

Problem, claim and contribution

The gap is expressed as two data properties, not simply a claim that more synthetic data is useful; each motivates a distinct generator component.

Sections 1–3, pp. 1–5 · Main paper

Argument and section order

Explicit research questions become experiment headings, making it straightforward to trace the contribution list to evidence.

Section 4, pp. 5–10 · Main paper

Figures, captions and visual explanation

Figure 1 aligns kernel matrices, sampled trajectories and a causal graph with the five numbered method steps; the caption explains the represented transformations.

Figure 1 and Section 3, p. 4 · Main paper

Tables, comparisons and evidence

Figure 7 places corpus size and UCR-training-data inclusion beside accuracy and learning curves; this exposes a consequential overlap difference that a headline score would hide.

Figure 7, p. 9 · Main paper

Appendix and reproducibility

The appendices document generator pseudocode, kernel sensitivity, out-of-domain classification, forecasting, fine-tuning and clinical time series, while also exposing protocol differences that must accompany comparisons.

Appendices C, G–N, pp. 16–29 · Main paper

Clarity and writing conventions

The term zero-shot needs a local definition: encoders are frozen but downstream random-forest or SVM classifiers use labeled training splits. Synthetic-data efficiency is not automatically matched-compute efficiency.

Section 4.1 and Appendix G, pp. 5–6, 20–21 · Main paper

What to reuse—and what not to infer

Transferable lesson: Pair each motivating data property with a pipeline component, then organize experiments around explicit questions and display overlap, data volume and evaluation protocol beside the results.

Boundary: Do not repeat scaling-law or causal-generalization language as a theorem. The 100K Mantis comparison differs by 0.11 percentage points, not less than 0.1; larger-model trends are not uniformly monotone, and some baseline training budgets and model sizes differ.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-09-19

Reading status

Full text + selected visuals inspected

Main text

PDF pp. 1–11, Sections 1–5 including all equations and experiments

Appendix / supplement

PDF pp. 15–29, Appendices A–O including implementation, additional benchmarks and limitations

Visually inspected PDF pages

4, 9

• Official ICLR 2026 proceedings conference PDF; PDF creation metadata: Wed May 6 05:37:47 2026 EDT · 29 pages