Writing study: Compact SO(3) Equivariant Atomistic Foundation Models via Structural Pruning

NeurIPS overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← NeurIPS 2026 oral writing corpus

Chen Wang, Siyu Hu, Guangming Tan, Weile Jia · 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: Efficiency, inference & systems (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

27 physical pages

Before references / appendix

≈ 9.09 page-equivalents

Main figures / tables

3 / 8

Attached appendix figures / tables

2 / 16

Attached appendix span

≈ 13.91 page-equivalents

Before references / appendix: ≈ 33.67%References / intervening material: ≈ 14.81%Attached appendix: ≈ 51.52%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.

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	≈ 1.44 pp.	5.33%
2 Related Work	3	≈ 0.42 pp.	1.56%
3 Method	3	≈ 2.3 pp.	8.52%
4 Experiments	5	≈ 3.85 pp.	14.26%
5 Conclusion	9	≈ 0.43 pp.	1.59%
References	10	≈ 4 pp.	14.81%
A Background	14	≈ 1.11 pp.	4.11%
B Motivation and Theory	15	≈ 2.65 pp.	9.81%
C Comparison of Depth and Width Pruning	17	≈ 1.44 pp.	5.33%
D Experimental Details	19	≈ 5.54 pp.	20.52%
E Applicability of the Proposed Method	24	≈ 1.72 pp.	6.37%
F Licenses for Datasets and Models	26	≈ 0.55 pp.	2.04%
G Broader Impacts	27	≈ 0.91 pp.	3.37%

All detected figure and table locations

Main text

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

Attached appendix

• Figure 4 · p. 17
• Table 9 · p. 18
• Table 10 · p. 19
• Table 11 · p. 20
• Table 12 · p. 20
• Table 13 · p. 21
• Table 14 · p. 21
• Table 15 · p. 22
• Table 16 · p. 22
• Table 17 · p. 22
• Figure 5 · p. 23
• Table 18 · p. 23
• Table 19 · p. 23
• Table 20 · p. 24
• Table 21 · p. 25
• Table 22 · p. 25
• Table 23 · p. 26
• Table 24 · p. 26

Argument and organization

1. Compression gap and alternatives

   Distinguishes structural pruning from distillation and quantization before proposing a symmetry-preserving unit.

2. Pruning mechanism

   Defines block granularity, importance and structural alignment before the full workflow.

3. Accuracy, efficiency and transfer

   Separates benchmark quality, resource cost, downstream applications and combinations with other compression methods.

Problem, claim and contribution

The paper identifies a symmetry constraint on compression and explicitly distinguishes pruning, knowledge distillation and quantization. This establishes the intervention without conflating all small-model methods.

Introduction; Table 1 · Main paper

Argument and section order

The method decomposes into granularity, importance estimation, structural alignment and workflow. Experiments separately examine accuracy, efficiency, application and combinations, mirroring the claims rather than listing datasets alone.

Sections 3.1–3.4 and 4.2–4.5 · Main paper

Figures, captions and visual explanation

A three-part diagram connects the equivariant block being removed, the weight/tensor restructuring, and the four-stage training pipeline. The inset structure makes the changed computation more inspectable than a row of generic model boxes.

Figure 2, PDF p. 4 · Main paper

Tables, comparisons and evidence

Tables retain parameter/data budgets and metric directions. The Matbench table visibly preserves metrics where the baseline wins; the learning curve uses training samples, not an unlabeled speed axis.

Tables 3–5 and Figure 3, PDF p. 7 · Main paper

Appendix and reproducibility

The attached material covers symmetry arguments, pruning alternatives, calibration/retraining details, pipeline GPU-hours, latency versus throughput, and licenses. These distinguish a smaller checkpoint from a cheaper end-to-end procedure.

Appendices B–F · Main paper

Clarity and writing conventions

Cost claims specify what is being compared, while the distillation combination is named as a separate condition. In a compression manuscript, avoid letting the headline parameter reduction stand in for measured inference savings.

Section 4.3; Table 8 · Main paper

What to reuse—and what not to infer

Transferable lesson: For an efficiency contribution, draw the legal unit of compression and report accuracy, training cost, deployment cost and compatibility as separate evidence.

Boundary: The source is a public arXiv version, not a verified accepted revision. A favorable subset of metrics does not imply dominance on every benchmark or hardware setting.

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-focused scan across framing, method organization, captions, result presentation and conclusion; title and author overlap checked against the official program. Selected physical pages inspected as rendered images.

Appendix / supplement

Attached appendix headings, reproduction routes and representative supporting prose/captions scanned. Long tables, every proof and every appendix page were not individually validated.

Visually inspected PDF pages

4, 7

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 manuscript arXiv:2605.08885v1; not verified identical to the accepted OpenReview revision · 27 pages