Writing study: TiRex-2: Generalizing TiRex to Multivariate Data and Streaming

NeurIPS overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← NeurIPS 2026 oral writing corpus

Patrick Podest, Marco Pichler, Elias Bürger, Levente Zólyomi, Bernhard Voggenberger, Wilhelm F Berghammer, Daniel Klotz, Sebastian Böck, Günter Klambauer, Sepp Hochreiter · Official NeurIPS 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: Architecture & representation (primary).

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

31 physical pages

Before references / appendix

≈ 11.07 page-equivalents

Main figures / tables

6 / 1

Attached appendix figures / tables

3 / 6

Attached appendix span

≈ 14.93 page-equivalents

Before references / appendix: ≈ 35.71%References / intervening material: ≈ 16.13%Attached appendix: ≈ 48.16%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Source-bound correction: Top-level heading and reference/appendix boundaries checked against source PDF text. Bibliography titles/author initials, numbered contribution/property bullets, diagram steps, equations and TOC entries are not substantive headings. Section spans are page-coordinate estimates, not ink area or prose-only length. This is an author manuscript mirror, not a verified accepted-version template. Source layout uses physical single-column y/page-height, including equation-heavy pages. Appendix span includes its explicit cover/contents/list-of-figures frontmatter; actual lettered subsection boundaries use their substantive starts, not TOC links. Rendered pp.18–19 reveal invisible duplicate older caption text: visible appendix figures are 7, 8, 9, not the overlapping extracted 6, 7, 8, 9. Retain visible labels only. Restore split Figure 6 and Table 1 captions p.10.

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	≈ 1.23 pp.	3.97%
2 Related work	2	≈ 0.93 pp.	3%
3 A TiRex architecture for multivariate forecasting with covariates	3	≈ 4.34 pp.	14%
4 Experiments	8	≈ 2.63 pp.	8.48%
5 Conclusion	10	≈ 0.72 pp.	2.32%
6 Acknowledgments and Disclosure of Funding	11	≈ 0.65 pp.	2.1%
References	12	≈ 5 pp.	16.13%
Appendix	17	≈ 1 pp.	3.23%
A Extended results	18	≈ 2 pp.	6.45%
B Asymmetric group attention: leakage derivation and details	20	≈ 2.75 pp.	8.87%
C Binary-aware tail-compressing scaler	22	≈ 0.99 pp.	3.19%
D Training Setup	23	≈ 1.86 pp.	6%
E Pre-Training Corpus	25	≈ 3.84 pp.	12.39%
F Synthetic Multivariate Coupling: Background and Design	29	≈ 1.32 pp.	4.26%
G Evaluation Metrics	30	≈ 1.16 pp.	3.74%

All detected figure and table locations

Main text

• Figure 1 · p. 1
• Figure 2 · p. 4
• Figure 3 · p. 7
• Figure 4 · p. 9
• Figure 5 · p. 9
• Figure 6 · p. 10
• Table 1 · p. 10

Attached appendix

• Figure 7 · p. 18
• Figure 8 · p. 18
• Figure 9 · p. 19
• Table 2 · p. 22
• Table 3 · p. 25
• Table 4 · p. 26
• Table 5 · p. 28
• Table 6 · p. 29
• Table 7 · p. 29

Argument and organization

1. Problem and contribution

   Defines the desired combination of multivariate forecasting, future-known covariates and bounded-state streaming before introducing the architecture, making the information-access constraint part of the contribution.

2. Argument sequence

   Problem setup precedes the time/variate mixing design, streaming update rules and synthetic coupling; evaluation then separates forecasting accuracy, streaming behavior and component ablations.

3. Evidence and verification

   Table 1 explicitly distinguishes the full model's absolute MASE from ablation deltas and states variability over seeds. The covariate-only subset remains separate from the complete benchmark.

Problem, claim and contribution

Defines the desired combination of multivariate forecasting, future-known covariates and bounded-state streaming before introducing the architecture, making the information-access constraint part of the contribution.

Introduction and Section 3.1, pp. 1–3 · Main paper

Argument and section order

Problem setup precedes the time/variate mixing design, streaming update rules and synthetic coupling; evaluation then separates forecasting accuracy, streaming behavior and component ablations.

Sections 3–4, pp. 3–10 · Main paper

Figures, captions and visual explanation

Figure 6 aligns streaming length, prediction horizon and covariate shift in three panels with a shared model legend. Dashed context boundaries distinguish trained context lengths from extrapolated streaming lengths. The central stack in Figure 2 is flanked by detailed time-mixer and variate-mixer insets. Persistent colors distinguish targets, past covariates and future-known covariates; tensor reshapes, shared weights and the asymmetric attention mask show how the method controls information flow instead of leaving future availability implicit.

Figure 6, p. 10 (render inspected); Figure 2, p. 4 (additional rendered inspection) · Main paper

Tables, comparisons and evidence

Table 1 explicitly distinguishes the full model's absolute MASE from ablation deltas and states variability over seeds. The covariate-only subset remains separate from the complete benchmark.

Table 1 and Section 4.3, p. 10 · Main paper

Appendix and reproducibility

A leakage derivation and counterexample explain the asymmetric masking design; the training-corpus tables document benchmark-related inclusions rather than hiding overlap behind a single dataset total.

Appendices B and E, pp. 20–22 and 25–29 · Main paper

Clarity and writing conventions

The limitations qualify when changed future covariates require recomputation, preventing constant-memory streaming from being misread as constant cost for every possible update.

Section 5, p. 11 · Main paper

What to reuse—and what not to infer

Transferable lesson: Pair a capability claim with explicit information-flow constraints, extrapolation boundaries and an ablation table that labels absolute values versus deltas.

Boundary: The extraction layer repeats some captions, although the inspected rendered page is clean. Do not double-count captions or generalize a conditional streaming guarantee to all covariate updates.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-10-02

Reading status

Writing-focused manuscript scan + selected visuals inspected

Main text

Writing-oriented scan across manuscript framing, section sequence, figure/table captions, selected argument/evaluation prose and conclusion; not exhaustive technical or proof verification. Sections 1–5, pp. 1–11; architecture, streaming semantics, synthetic coupling, evaluation and limitations were scanned. Source is the exact public author version listed below, not a verified accepted/camera-ready revision.

Appendix / supplement

Writing-oriented appendix/documentation scan: Appendices A–G, pp. 18–31: additional plots, leakage analysis, scaling, training, corpus composition, synthetic data and metrics. Source code, data, model outputs and external video assets were not executed or independently validated.

Visually inspected PDF pages

4, 10

This is a writing-focused scan of the manuscript and the stated supplementary scope: framing, organization, visual explanation, evidence presentation and wording. It is not a line-by-line proof audit, reproduction of results or exhaustive scientific review.

• Public author arXiv manuscript 2607.01204v1; not verified identical to the accepted or camera-ready OpenReview revision · 31 pages