Writing study: Token Activation Map to Visually Explain Multimodal LLMs

ICCV overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICCV 2025 oral writing corpus

Yi Li, Hualiang Wang, Xinpeng Ding, Haonan Wang, Xiaomeng Li · Official ICCV 2025 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

11 physical pages

Before references / appendix

≈ 8.05 page-equivalents

Main figures / tables

8 / 4

Attached appendix figures / tables

0 / 0

Attached appendix span

No boundary detected; not assumed zero

Before references / appendix: ≈ 73.18%References / intervening material: ≈ 26.82%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.

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	≈ 1.35 pp.	12.27%
2. Related Work	2	≈ 0.67 pp.	6.09%
3. Method	3	≈ 2 pp.	18.18%
4. Experiments	5	≈ 3.31 pp.	30.09%
5. Conclusion	8	≈ 0.16 pp.	1.45%
Acknowledgment	8	≈ 0.17 pp.	1.55%
References	9	≈ 2.95 pp.	26.82%

All detected figure and table locations

Main text

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

Attached appendix

No recognized caption labels.

Argument and organization

1. 1 Introduction; 2 Related Work

   Motivates token-dependent interference using progressive language generation and a concrete fork/plate example.

2. 3 Method

   Defines token maps, context correction and ranked filtering, then introduces plausibility metrics before results.

3. 4 Experiments; 5 Conclusion

   Separates module controls, baseline families, model-scale transfer, qualitative attributes and failure cases.

Problem, claim and contribution

The motivation is the difference between a single class output and sequential token outputs: earlier context can contaminate a later token’s visual map.

Figure 1 and Sections 1–2, physical pp.1–3 · Main paper

Argument and section order

Evaluation definitions sit in the method rather than being buried after results, explicitly distinguishing object words and function words.

Sections 3.1–3.4, pp.3–5 · Main paper

Figures, captions and visual explanation

The method figure assigns equation ranges to overview, interference removal, filtering and metric panels, letting readers move between a worked activation example and the formal definition.

Figure 2, p.4 · Main paper

Tables, comparisons and evidence

Main comparisons group gradient, attention, combined and logit methods, repeating object/function/combined IoU across datasets; added-method rows use signed gains rather than disguising them as absolute scores.

Tables 3–4, p.6 · Main paper

Appendix and reproducibility

The main text routes metric interpretation and tagging details to Supplement E, baseline implementations to D, and extended model/failure examples to named later sections. These references, not the supplement, were inspected.

Sections 3.4–4.3, pp.5–8 · Main paper

Clarity and writing conventions

The evaluation states that its IoU metrics measure plausibility rather than faithfulness, an important qualification beside the causal terminology.

Section 3.4, p.5 · Main paper

What to reuse—and what not to infer

Transferable lesson: Define what an explanation metric measures before claiming improvement; align a method overview with its equations and preserve signed non-improvements.

Boundary: Cleaner activation maps and mask overlap are not proof of faithful causal explanations or better model reasoning. The inspected method caption labels the filtering panel as d even though the drawing uses c; panel references need proofreading.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-10-07

Reading status

Writing-focused manuscript scan + selected visuals inspected

Main text

Writing-oriented scan across the main manuscript, physical PDF pp.1–8, covering framing, section progression, method explanations, captions, comparison protocols and concluding scope; reference boundary checked. Selected full-page rasters were inspected. Not exhaustive line-by-line reading, proof verification or experimental reproduction.

Appendix / supplement

Main-paper appendix/supplement references were checked. A separate supplementary source is linked in the inventory but was not retrieved or inspected; its contents are not included in the reading or visual-page counts.

Visually inspected PDF pages

4, 6

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 CVF open-access ICCV 2025 accepted-paper PDF, retrieved 7 October 2026; CVF states identical to accepted version except watermark; not assumed identical to initial submission · 11 pages