Writing study: FINER: MLLMs Hallucinate under Fine-grained Negative Queries

CVPR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← CVPR 2026 oral writing corpus

Rui Xiao, Sanghwan Kim, Yongqin Xian, Zeynep Akata, Stephan Alaniz · Official CVPR 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: Datasets, benchmarks & measurement (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

10 physical pages

Before references / appendix

≈ 8.22 page-equivalents

Main figures / tables

5 / 5

Attached appendix figures / tables

0 / 0

Attached appendix span

Separate file recorded below; not in this PDF total

Before references / appendix: ≈ 82.2%References / intervening material: ≈ 17.8%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	≈ 0.99 pp.	9.9%
2. FINER Benchmarks	2	≈ 1.49 pp.	14.9%
3. Training with FINER (FINER-Tuning)	3	≈ 0.98 pp.	9.8%
4. Experiments	4	≈ 3.05 pp.	30.5%
5. Related Works	7	≈ 0.88 pp.	8.8%
6. Conclusion and Limitation	8	≈ 0.49 pp.	4.9%
References	9	≈ 1.78 pp.	17.8%

All detected figure and table locations

Main text

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

Attached appendix

No recognized caption labels.

Separate source files—excluded from canonical-PDF totals

A separate author version may repeat main text and references. Whole-file lengths and counts are not extra appendix material and are never added to the canonical PDF totals. Read the study's source scope for which sections received substantive reading versus layout inspection.

Additional source 1

Separate PDF (or ZIP containing it): whole-file layout of 27 physical pages; 20 detected figures and 12 tables. These may include overlapping main-text material.

Located appendix only: starts on physical p.1; ≈ 26.94 page-equivalents, 20 figure labels and 12 table labels. This interval is distinct from the whole file.

Detected heading	PDF page	Estimated span
Supplementary Material	1	≈ 0.02 pp.
A. Extended Related Works	1	≈ 1.1 pp.
B. FINER Benchmark Details	2	≈ 3.66 pp.
C. No, but I can see {Z2} in this image.	5	≈ 0.01 pp.
D. No, but I can see {Z3} in this image.	5	≈ 0.01 pp.
E. No, but I can see {Z4} in this image.	5	≈ 8.4 pp.
F. Human Study	14	≈ 0.67 pp.
G. Templates	14	≈ 10.12 pp.
References	25	≈ 2.96 pp.

• Figure 6 · p. 3
• Figure 7 · p. 3
• Figure 8 · p. 4
• Table 6 · p. 4
• Table 7 · p. 5
• Table 8 · p. 5
• Figure 9 · p. 6
• Table 9 · p. 8
• Table 10 · p. 8
• Figure 10 · p. 9
• Table 11 · p. 10
• Table 12 · p. 10
• Figure 11 · p. 12
• Table 13 · p. 13
• Table 14 · p. 13
• Table 15 · p. 13
• Table 16 · p. 13
• Table 17 · p. 14
• Figure 12 · p. 14
• Figure 15 · p. 14
• Figure 13 · p. 15
• Figure 14 · p. 17
• Figure 16 · p. 19
• Figure 17 · p. 19
• Figure 18 · p. 20
• Figure 19 · p. 20
• Figure 20 · p. 21
• Figure 21 · p. 21
• Figure 22 · p. 22
• Figure 23 · p. 23
• Figure 24 · p. 24
• Figure 25 · p. 24

Argument and organization

1. Failure probe and benchmark construction

   Begin with granularity-sensitive hallucinations, then define scene graphs, negative queries and paired accuracy.

2. Remedy and retained capabilities

   Introduce FINER-Tuning, measure target and transfer benchmarks, then check general skills and training choices.

Problem, claim and contribution

The question is not whether MLLMs hallucinate in general but whether they can reject one subtle false component inside a mostly true description.

Introduction and Figure 1, PDF pp. 1–2 · Main paper · Additional source 2

Argument and section order

The benchmark and its paired metric are defined before tuning; related work is deferred until after results, prioritizing the empirical problem and measurement contract.

Sections 2–5, PDF pp. 2–7 · Main paper · Additional source 2

Figures, captions and visual explanation

The data-construction diagram combines positive/negative scene graphs, verification stages and actual multiple-choice examples; the page-6 small multiples then expose accuracy as object, attribute or relation count increases.

Figure 2, PDF p. 3; Figure 4, PDF p. 6; visual inspection · Main paper · Additional source 2

Tables, comparisons and evidence

Before/after tuning rows retain the base model and show metric deltas. Separate general-capability tests address over-rejection rather than treating lower hallucination alone as sufficient.

Tables 1–3 and Sections 4.3–4.4, PDF pp. 5–7 · Main paper · Additional source 2

Appendix and reproducibility

The long supplement separates question generation, training/evaluation, bias checks, human study and full prompt templates; its human study splits positive and negative partners across different survey versions to avoid remembered answers.

Supplement Sections B–G, PDF pp. 3–24; especially F, p. 14 · Main paper · Additional source 2

Clarity and writing conventions

Subset abbreviations and metric directions are explained in captions, and Gemini's partial evaluation is disclosed instead of being presented as identical coverage.

Tables 1–2, PDF pp. 5–6 · Main paper · Additional source 2

What to reuse—and what not to infer

Transferable lesson: For a failure benchmark, specify the minimal counterfactual change, use a metric that prevents trivial response bias, and test whether the proposed remedy damages unrelated capabilities.

Boundary: Machine-generated negatives still require validation. Relative comparisons between different tuned MLLMs are not controlled causal comparisons of tuning methods.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-09-26

Reading status

Writing-focused manuscript scan + selected visuals inspected

Main text

Manuscript-wide writing scan of PDF pp. 1–10: framing, argument sequence, section organization, figure/table captions, evaluation interpretation and limitations; reference sections identified. This is a writing-focused scan, not an exhaustive technical proof or reproduction audit. Main PDF pages 3 and 6 were rendered and visually inspected.

Appendix / supplement

Writing-focused scan of the separate supplementary PDF, pp. 1–27, including its organization, captions, protocol/evidence detail and limitation material where present; proofs and code listings were inspected for exposition, not independently verified or executed. Supplement Sections B–G, PDF pp. 3–24; especially F, p. 14. Linked video/audio demonstrations and non-PDF archive contents were not assessed.

Visually inspected PDF pages

3, 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 CVPR 2026 CVF open-access accepted manuscript; retrieved 2026-09-26 · 10 pages
• Official CVPR 2026 CVF supplementary PDF; retrieved 2026-09-26 · 27 pages