Writing study: CATP: Contextually Adaptive Token Pruning for Efficient and Enhanced Multimodal In-Context Learning

AAAI overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← AAAI 2026 oral writing corpus

Yanshu Li, Jianjiang Yang, Zhennan Shen, Ligong Han, Haoyan Xu, Ruixiang Tang · Official AAAI 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

9 physical pages

Before references / appendix

≈ 7.07 page-equivalents

Main figures / tables

6 / 3

Attached appendix figures / tables

0 / 0

Attached appendix span

No boundary detected; not assumed zero

Before references / appendix: ≈ 78.56%References / intervening material: ≈ 21.44%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Extraction notes: No attached appendix boundary detected. Zero means no recognized caption, not independently verified absence.

Source-bound correction: Checked 2026-10-25 against the authored writing scan and source bounding boxes. Main headings are unnumbered and use a slightly smaller extracted glyph height than some other AAAI PDFs; chart labels and the Method table-column title are not headings. Captions match Figures 1–6 and Tables 1–3. The linked extended version is not part of these measurements.

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
Introduction	1	≈ 1.52 pp.	16.89%
Related Work	2	≈ 0.52 pp.	5.78%
Method	3	≈ 2.24 pp.	24.89%
Experiments	5	≈ 2.23 pp.	24.78%
Conclusion	7	≈ 0.18 pp.	2%
Acknowledgments	8	≈ 0.03 pp.	0.33%
References	8	≈ 1.93 pp.	21.44%

All detected figure and table locations

Main text

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

Attached appendix

No recognized caption labels.

Argument and organization

1. Introduction → Related Work

   Expose the single-image to interleaved-context transfer gap with qualitative examples and a controlled comparison before defining the proposed pruning mechanism.

2. Method: Preliminary and Motivation → Contextually Adaptive Token Pruning

   Define the sequence and attention objects, diagnose existing criteria, then place the pre-decoder and shallow-decoder stages at their actual computation boundaries.

3. Experiments: Setup → Main Results → Efficiency Analysis → Ablation Study

   Name workload, hardware and token budget, then separate task scores from measured latency, FLOPs, memory and sensitivity to context configuration.

Problem, claim and contribution

The opening narrows a broad efficiency problem to a missing setting: image-token pruning inside multimodal in-context sequences. Figures 1–3 make the failure case visible before the new method, linking the contribution to cross-image/text interactions rather than merely claiming fewer tokens.

Introduction pp.1–2; Figures 1–3 pp.1–3 · Main paper

Argument and section order

Definitions of demonstrations, query and token sequence precede the selection equations. The method overview specifies two distinct execution locations; experiments subsequently separate accuracy, resource use and sensitivity. This sequence lets the reader identify what is removed and where its overhead occurs.

Method pp.3–5; Experiments pp.5–7 · Main paper

Figures, captions and visual explanation

Figure 4 follows token movement vertically through projector and decoder layers, using token colors and pruning marks to expose the stage boundary. It gives architecture placement, while Figure 5 examines context configuration and Figure 6 examines hyperparameters. The very short architecture caption leaves important interpretation in adjacent prose; a self-contained caption would help reuse.

Figure 4 physical p.4; Figures 5–6 pp.6–7 · Main paper

Tables, comparisons and evidence

Table 3 keeps latency, FLOPs, KV-cache memory and relative task performance separate. In its LLaVA-Next-7B block, DivPrune reduces FLOPs more than CATP but has higher latency than vanilla, illustrating why a proxy cost cannot substitute for a measured endpoint. The caption says four models, but this published table visibly contains three model blocks; preserve that source mismatch.

Efficiency Analysis p.6; Table 3 physical p.7; setup p.5 · Main paper

Appendix and reproducibility

The accepted PDF carries the core selection equations and settings in its main body, and links an extended version separately. These observations and counts use only the nine-page publisher file. Do not infer that the extended version lacks proofs, extra experiments or pseudocode merely because they are absent here.

Extended-version link p.1; Eqs.1–10 pp.3–5; end pp.8–9 · Main paper

Clarity and writing conventions

Terms for demonstrations and the query make the input contract concrete, while bold local signposts help the reader distinguish design decisions. Calibrate the conclusion to the tested setting: the relative aggregate is not absolute accuracy or a win on every benchmark, and “first” language still requires a defensible novelty search.

Stage 1/Stage 2 pp.4–5; Tables 1–3 pp.6–7; Conclusion p.7 · Main paper

What to reuse—and what not to infer

Transferable lesson: When proposing an efficient inference method, show where the operation occurs and pair task quality with measured end-to-end cost; report the workload and denominator so token removal is not mistaken for proportional speedup.

Boundary: No independent timing, proof audit or extended-version reading was performed. Table 3 has a caption/model-count mismatch, and source-reported aggregate improvements must not be rewritten as every-task gains.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-10-25

Reading status

Writing-focused manuscript scan + selected visuals inspected

Main text

Writing-oriented scan of the problem, method, evaluation and conclusion on physical pp.1–7; references/end checked for structural scope, not an independent proof or runtime reproduction.

Appendix / supplement

The inspected publisher PDF ends with references and has no attached appendix. It links an extended arXiv version on p.1; that separate version was not inspected or included in these measurements.

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.

• Official AAAI-26 published proceedings PDF retrieved 2026-10-25, printed pp.6619–6627 · 9 pages