Writing study: Plug-and-Play Compositionality for Boosting Continual Learning with Foundation Models

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Weiduo Liao, Fei Han, Hisao Ishibuchi, Qingfu Zhang, Ying Wei · 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: Learning & optimization algorithms (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

34 physical pages

Before references / appendix

≈ 10.4 page-equivalents

Main figures / tables

4 / 2

Attached appendix figures / tables

10 / 11

Attached appendix span

≈ 17.89 page-equivalents

Before references / appendix: ≈ 30.59%References / intervening material: ≈ 16.79%Attached appendix: ≈ 52.62%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	≈ 1.93 pp.	5.68%
2 RELATED WORKS	3	≈ 0.9 pp.	2.65%
3 PRELIMINARIES	4	≈ 0.39 pp.	1.15%
4 METHODS	4	≈ 2.62 pp.	7.71%
5 EXPERIMENTS	7	≈ 3.21 pp.	9.44%
6 CONCLUSION	10	≈ 0.22 pp.	0.65%
7 ETHICS STATEMENT	10	≈ 0.34 pp.	1%
8 REPRODUCIBILITY STATEMENT	11	≈ 0.1 pp.	0.29%
ACKNOWLEDGMENTS	11	≈ 0.11 pp.	0.32%
REFERENCES	11	≈ 5.71 pp.	16.79%
A APPENDIX CONTENTS	17	≈ 1.16 pp.	3.41%
B DISCUSSIONS	18	≈ 0.84 pp.	2.47%
C ADDITIONAL RELATED WORKS	19	≈ 0.27 pp.	0.79%
D THEOREM	19	≈ 1.2 pp.	3.53%
E HYPERPARAMETERS AND EXPERIMENTAL SETTINGS	20	≈ 0.29 pp.	0.85%
F PSEUDO CODE	20	≈ 1.86 pp.	5.47%
G VISUALIZATION	22	≈ 3.86 pp.	11.35%
H ADDITIONAL RESULTS	26	≈ 7.22 pp.	21.24%
I USE OF LARGE LANGUAGE MODELS	33	≈ 1.19 pp.	3.5%

All detected figure and table locations

Main text

• Figure 1 · p. 2
• Figure 2 · p. 6
• Table 1 · p. 8
• Figure 3 · p. 9
• Table 2 · p. 9
• Figure 4 · p. 10

Attached appendix

• Table 3 · p. 21
• Table 4 · p. 21
• Figure 5 · p. 23
• Figure 6 · p. 24
• Figure 7 · p. 24
• Figure 8 · p. 25
• Figure 9 · p. 25
• Figure 10 · p. 26
• Table 5 · p. 27
• Table 6 · p. 28
• Figure 11 · p. 28
• Figure 12 · p. 29
• Figure 13 · p. 29
• Table 7 · p. 30
• Table 8 · p. 31
• Table 9 · p. 31
• Table 10 · p. 31
• Table 11 · p. 32
• Table 12 · p. 32
• Figure 14 · p. 33
• Table 13 · p. 33

Argument and organization

1. Compositional motivation → concept extraction → relational distillation

   Connects class-relevant slots to a learner-independent regularizer.

2. Matched baseline pairs → component controls → visual mechanism

   Tests usefulness across different continual-learning families.

Argument and section order

Concepts, primitives, selection, and primitive-logit alignment are introduced in dependency order; the worked Chihuahua example keeps the intermediate representation interpretable.

Sections 4.1–4.2, pages 5–7 · Main paper

Figures, captions and visual explanation

Figure 1 separates the general learner/logit branch from a magnified slot-learning branch and labels every loss at its point of application.

Figure 1, page 2 · Main paper

Tables, comparisons and evidence

Table 1 pairs each baseline with its augmented version and labels rehearsal methods separately, which is clearer than pooling heterogeneous learners into one ranking.

Table 1, page 8 · Main paper

Capacity-matched controls and direct-primitive-classification failures help distinguish relational distillation from simply adding parameters or replacing the classifier.

Table 2, page 9; Table 11, page 32 · Main paper

Appendix and reproducibility

A table of contents leads to stage-specific hyperparameters and pseudocode; the latter makes the no-gradient concept branch during learner training explicit.

Appendices A, E–F, pages 17–22 · Main paper

Clarity and writing conventions

Claims of consistent forgetting reduction and free application exceed the tables: some forgetting values worsen, and the cost table separates 5.5 hours of slot learning from the downstream 9.1-to-10.1-hour comparison.

Table 1, page 8; Table 13, page 33 · Main paper

What to reuse—and what not to infer

Transferable lesson: For an add-on method, present matched with/without pairs across host algorithms, isolate capacity and representation-use controls, and disclose both reusable preparation cost and per-host overhead.

Boundary: No concept annotations are used for training, but class labels supervise primitive learning; this is not wholly unsupervised. Slot stability examples and compositional definitions are not a guarantee of semantic disentanglement or universal forgetting prevention.

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

Pages 1–11, Sections 1–8; references through page 16

Appendix / supplement

Pages 17–34, Appendices A–I: definitions/proofs, settings, both algorithms, visualizations, complete additional experiments, costs, and disclosure

Visually inspected PDF pages

2, 8, 33

• ICLR 2026 final proceedings · 34 pages