Writing study: DOUBT: Decoupled Object-level Understanding and Bridging via vMF-based Trustworthiness for Hallucination Detection in MLLMs

ICML overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICML 2026 oral writing corpus

Kaiqi Chen, Yang Qin, Changhao He, Xi Peng, Peng Hu · Official ICML 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: Domain-grounded methods & physical 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

24 physical pages

Before references / appendix

≈ 9.23 page-equivalents

Main figures / tables

4 / 3

Attached appendix figures / tables

0 / 8

Attached appendix span

≈ 10.91 page-equivalents

Before references / appendix: ≈ 38.46%References / intervening material: ≈ 16.08%Attached appendix: ≈ 45.46%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	≈ 1.42 pp.	5.92%
2. Related Work	3	≈ 0.6 pp.	2.5%
3. Method	3	≈ 2.24 pp.	9.33%
4. Experiment	5	≈ 2.49 pp.	10.38%
5. Computational Cost Analysis	8	≈ 1.39 pp.	5.79%
6. Conclusion	9	≈ 0.28 pp.	1.17%
Impact Statement	10	≈ 0.08 pp.	0.33%
Acknowledgments	10	≈ 0.11 pp.	0.46%
References	10	≈ 3.85 pp.	16.04%
A. Hallucination Mitigation through OUB	14	≈ 0.43 pp.	1.79%
B. Performance under Different vMF Parameters	14	≈ 0.29 pp.	1.21%
C. Effect of Introducing a More Capable External Model	14	≈ 0.5 pp.	2.08%
D. Impact of Different Uncertainty Estimation Metrics	15	≈ 0.44 pp.	1.83%
E. Ablation on the Number of Recognized Objects	15	≈ 0.86 pp.	3.58%
F. Hard Case Analysis	16	≈ 0.22 pp.	0.92%
G. Robustness to Stage-2 Prompt Formats	16	≈ 0.9 pp.	3.75%
H. Robustness to Different Semantic Embedding Models	17	≈ 0.59 pp.	2.46%
I. Case Study	18	≈ 6.67 pp.	27.79%

All detected figure and table locations

Main text

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

Attached appendix

• Table 4 · p. 14
• Table 5 · p. 14
• Table 6 · p. 15
• Table 7 · p. 15
• Table 8 · p. 16
• Table 9 · p. 17
• Table 10 · p. 17
• Table 11 · p. 18

Argument and organization

1. Problem and contribution

   Motivates the method with confidently consistent wrong answers, separating diversity induced by a grounding prompt from the reliability statistic used to score the resulting responses.

2. Argument sequence

   Prompt construction precedes a geometric consistency measure and an explicit binary evaluation rule, followed by component ablations and a separate computational-cost section.

3. Evidence and verification

   Detection accuracy checks whether the predicted reliability flag agrees with a reference answer's correctness; the separate mitigation appendix measures answer improvement, while hard-case recovery uses a deliberately selected failure subset.

Problem, claim and contribution

Motivates the method with confidently consistent wrong answers, separating diversity induced by a grounding prompt from the reliability statistic used to score the resulting responses.

Introduction and Figure 1, pp. 1-2; Section 3, pp. 3-5 · Main paper

Argument and section order

Prompt construction precedes a geometric consistency measure and an explicit binary evaluation rule, followed by component ablations and a separate computational-cost section.

Sections 3-5, pp. 3-8 · Main paper

Figures, captions and visual explanation

The ablation table retains configurations where removing a component wins locally, and the cost table exposes 21 forward passes plus latency and generated tokens rather than describing training-free operation as cost-free.

Table 2, Figure 3 and Table 3, p. 8; p. 5 metric-definition and p. 8 PNGs inspected · Main paper

Tables, comparisons and evidence

Detection accuracy checks whether the predicted reliability flag agrees with a reference answer's correctness; the separate mitigation appendix measures answer improvement, while hard-case recovery uses a deliberately selected failure subset.

Section 3.4, p. 5; Appendix A, p. 14; F, pp. 16-17 · Main paper

Appendix and reproducibility

Prompt, embedding-model and object-count comparisons expose deployment choices, while selected full response sets let readers inspect how bridging changes variation rather than merely seeing aggregate scores.

Appendices B-H, pp. 14-17; I, pp. 18-24 · Main paper

Clarity and writing conventions

The equation defines flag=1 as non-hallucinatory despite the phrase hallucination flag, so a reusable account must preserve the polarity; the trustworthiness statistic is internal consistency, not a calibrated correctness probability.

Equations 4-6 and Section 3.4, p. 5 · Main paper

What to reuse—and what not to infer

Transferable lesson: Define the target, flag polarity and denominator before presenting detection results, and report inference overhead separately from the absence of model retraining.

Boundary: Consistency can remain wrong, the 39.10% recovery figure concerns a selected semantic-entropy failure subset, and illustrative medical imagery is not clinical validation; black-box prompting also uses an external sentence-embedding model.

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-focused distributed manuscript scan, not exhaustive technical/proof validation. Sections 1-6, pp. 1-9: false-consistency motivation, object-bridged prompting, directional embedding concentration, detection definition, evaluations, ablations, cost and qualitative mitigation.

Appendix / supplement

Writing-focused appendix scan; code, data and external assets not executed. Appendices A-I, pp. 14-24: answer mitigation, stronger object-list generators, prompt and hyperparameter robustness, hard-case subsets, embedding choices and extended examples.

Visually inspected PDF pages

5, 8

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.

• Published ICML 2026 proceedings PDF, PMLR volume 306; retrieved 2026-10-07. Public OpenReview decision notes were not independently reverified. · 24 pages