Writing study: World Knowledge in the Weights: Reading Concept Circuits of Vision Transformers
← ECCV 2026 oral writing corpus
Yanlin Chen, Tang Li, Xi Peng · Official ECCV 2026 oral record
Figure, table and section ledger
Editorial contribution type: Empirical analysis, diagnosis & auditing (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
- 19 physical pages
- Before references / appendix
- ≈ 14.35 page-equivalents
- Main figures / tables
- 7 / 2
- Attached appendix figures / tables
- 0 / 0
- Attached appendix span
- No boundary detected; not assumed zero
Extraction notes: No attached appendix boundary detected. Zero means no recognized caption, not independently verified absence.
Source-bound correction: All top-level main headings and the reference boundary checked against source PDF text; source-bound positions use physical single-column y/page-height, not a two-column reading-order surrogate. Running page-number/author titles, affiliations, numbered contribution bullets, footnotes and bibliography continuations are excluded. This linked manuscript contains no appended supplementary section; separate supplement links were not scanned.
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 | ≈ 2 pp. | 10.53% |
| 2 Preliminary | 3 | ≈ 0.97 pp. | 5.11% |
| 3 Method | 4 | ≈ 5.91 pp. | 31.11% |
| 4 Experiments | 10 | ≈ 3.2 pp. | 16.84% |
| 5 Related Work | 13 | ≈ 0.89 pp. | 4.68% |
| 6 Conclusion | 14 | ≈ 0.56 pp. | 2.95% |
| Acknowledgements | 15 | ≈ 0.13 pp. | 0.68% |
| References | 15 | ≈ 4.65 pp. | 24.47% |
All detected figure and table locations
Main text
Attached appendix
No recognized caption labels.
Argument and organization
- 1 Introduction; 2 Preliminary
Contrasts internal concept interactions with input saliency, motivating an inspectable graph rather than treating a heat map as the model's complete explanation.
- 3 Method: CLT adaptation and global/instance circuits
Defines sparse dictionaries before separating global, input-invariant circuits from instance-specific circuits; experiments test discovery, removal and model differences.
- 4 Experiments; 5 Related Work; 6 Conclusion
The study uses mean ablations, faithfulness/completeness curves and intervention tests rather than relying only on visually plausible concept labels.
Problem, claim and contribution
Contrasts internal concept interactions with input saliency, motivating an inspectable graph rather than treating a heat map as the model's complete explanation.
Sections 1–2, pp.1–4 · Main paper
Argument and section order
Defines sparse dictionaries before separating global, input-invariant circuits from instance-specific circuits; experiments test discovery, removal and model differences.
Sections 2–4, pp.3–12 · Main paper
Figures, captions and visual explanation
The instance-circuit diagram links concept example patches through weighted edges; another panel groups spurious, non-discriminative and selected discriminative concepts around a common diagnostic plane.
Figure 4, p.8; Figure 6, p.10 · Main paper
Tables, comparisons and evidence
The study uses mean ablations, faithfulness/completeness curves and intervention tests rather than relying only on visually plausible concept labels.
Table 1, p.7; Figure 5, p.9; Section 4.2, p.11 · Main paper
Appendix and reproducibility
The CLT adaptation section sends additional implementation details to supplementary material; only this main-text pointer, not the separate supplement, was inspected.
Section 3.1, p.5 · Main paper
Clarity and writing conventions
Short takeaway boxes precede experiment explanations, making each test's intended inference explicit; the final limitations qualify the graph's causal meaning.
Sections 4.1–4.3, pp.10–12; Limitations, p.15 · Main paper
What to reuse—and what not to infer
Transferable lesson: Interpretability writing needs examples of concepts plus interventions that could falsify the explanation, and a clear distinction between a diagram and a causal model.
Boundary: The authors explicitly note that global graphs do not fully capture causal relations and that CLTs primarily capture MLP pathways. Do not repeat the title as literal complete world knowledge.
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-oriented scan across the manuscript main argument, physical PDF pp.1–15: framing and section roadmaps, method/argument explanations, figure and table captions, comparison protocols, conclusions and limitations. Selected full-page rasters were inspected. This is not exhaustive line-by-line technical reading, proof verification or experimental reproduction.
- Appendix / supplement
- Main-manuscript pointers to appendix/supplement and the reference boundary were checked. Separate supplementary PDFs were not retrieved or inspected; their contents are not included in inspected-page counts.
- Visually inspected PDF pages
- 8, 10
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.