Writing study: MetaEmbed: Scaling Multimodal Retrieval at Test-Time with Flexible Late Interaction

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Zilin Xiao, Qi Ma, Mengting Gu, Chun-cheng Chen, Xintao Chen, Vicente Ordonez, Vijai Mohan · 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: Architecture & representation design (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

20 physical pages

Before references / appendix

≈ 9.47 page-equivalents

Main figures / tables

4 / 3

Attached appendix figures / tables

0 / 4

Attached appendix span

≈ 3.89 page-equivalents

Before references / appendix: ≈ 47.35%References / intervening material: ≈ 33.2%Attached appendix: ≈ 19.45%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	≈ 1.29 pp.	6.45%
2 RELATED WORK	2	≈ 0.76 pp.	3.8%
3 METHODOLOGY	3	≈ 2.04 pp.	10.2%
4 EXPERIMENTS	5	≈ 3.64 pp.	18.2%
5 DISCUSSION	9	≈ 0.48 pp.	2.4%
6 CONCLUSION	9	≈ 0.32 pp.	1.6%
ACKNOWLEDGEMENTS	10	≈ 0.1 pp.	0.5%
ETHICS STATEMENT	10	≈ 0.12 pp.	0.6%
REPRODUCIBILITY STATEMENT	10	≈ 0.15 pp.	0.75%
REFERENCES	10	≈ 6.63 pp.	33.15%
A APPENDIX	17	≈ 3.89 pp.	19.45%

All detected figure and table locations

Main text

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

Attached appendix

• Table 4 · p. 17
• Table 5 · p. 18
• Table 1 · p. 18
• Table 6 · p. 19

Argument and organization

1. Methodology

   Define late interaction before introducing learned tokens, nested groups and the grouped contrastive objective.

2. Experiments

   Separate benchmark comparisons, budget scaling and controlled representation ablations.

3. Discussion

   Decompose index memory and scoring cost from query encoding.

Problem, claim and contribution

The introduction positions a specific bottleneck between a single compressed vector and hundreds of patch vectors. This establishes why learned compact token sets matter before naming the nested training recipe.

§1, pp.1–2 · Main paper

Argument and section order

The method introduces the MaxSim-sum operation, then token construction, then nested-prefix losses. Separate indexing and query-time choices make the deployment control concrete without implying any additional inference-time learning.

§§3.1–3.2, pp.3–5 · Main paper

Figures, captions and visual explanation

The comparison diagram reuses vector and score motifs across single-vector, ordinary multi-vector and nested representations. Shared symbols let the reader identify the changed component rather than decoding three unrelated architectures.

Fig.1, p.2 · Main paper

Tables, comparisons and evidence

The efficiency table combines quality, FLOPs, latency and index memory, with precision, candidate count and scoring batch specified in its caption. The surrounding discussion explicitly excludes the much larger query-encoding cost from those latency numbers.

Table3 and §5, p.9 · Main paper

Appendix and reproducibility

Matched single-last, mean-pooling and split-vector baselines isolate representation choices more cleanly than the headline literature comparison. The per-task appendix also acknowledges class overlap in the benchmark's nominal out-of-domain split.

Appendices A.2–A.3, Tables5–6, pp.17–19 · Main paper

Clarity and writing conventions

Several narrative details require table-first correction: percentage-point changes are called relative gains, the 3B baseline differs across plots and tables, and statistical-significance language lacks a reported test. These do not justify silently repairing the source's numbers.

§4.2, p.7; Fig.3–4, p.8; Table5, p.18 · Main paper

What to reuse—and what not to infer

Transferable lesson: Present a selectable compute budget with its quality and storage costs, then distinguish controlled baselines from inherited literature numbers.

Boundary: Test-time scaling means selecting stored embedding prefixes, not more reasoning or retraining. Scoring latency is not end-to-end latency, and CPU offloading costs are not measured. Main and appendix disagree on the7B batch size and several category scores. The biomedical document benchmark is not WSI pathology evidence; generalization claims inherit class-overlap and backbone-data caveats.

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.1–9, §§1–6; statements and references pp.10–16

Appendix / supplement

Appendix A.1–A.5, PDF pp.17–20, including all per-task results, implementation and overlap caveats

Visually inspected PDF pages

2, 9

• official ICLR 2026 proceedings PDF · 20 pages