Writing study: MC-Search: Evaluating and Enhancing Multimodal Agentic Search with Structured Long Reasoning Chains

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Xuying Ning, Dongqi Fu, Tianxin Wei, Mengting Ai, Jiaru Zou, Ting-Wei Li, Hanghang Tong, Yada Zhu, Hendrik Hamann, Jingrui He · Official ICLR 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 resources (primary) · Learning & optimization algorithms.

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

35 physical pages

Before references / appendix

≈ 10.11 page-equivalents

Main figures / tables

5 / 4

Attached appendix figures / tables

5 / 10

Attached appendix span

≈ 21.9 page-equivalents

Before references / appendix: ≈ 28.89%References / intervening material: ≈ 8.54%Attached appendix: ≈ 62.57%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	≈ 1.94 pp.	5.54%
2 MC-SEARCH BENCHMARK	3	≈ 2.14 pp.	6.11%
3 AGENTIC MM-RAG PIPELINE AND PROCESS-LEVEL ALIGNMENT	5	≈ 0.84 pp.	2.4%
4 EXPERIMENTS	6	≈ 3.04 pp.	8.69%
5 RELATED WORK	9	≈ 0.23 pp.	0.66%
6 CONCLUSION	9	≈ 0.32 pp.	0.91%
ACKNOWLEDGEMENTS	10	≈ 1 pp.	2.86%
REFERENCES	11	≈ 3 pp.	8.57%
Appendix	14	≈ 1 pp.	2.86%
A USE OF LARGE LANGUAGE MODELS	15	≈ 0.14 pp.	0.4%
B RELATED WORK	15	≈ 0.86 pp.	2.46%
C DATA EXAMPLE	16	≈ 4 pp.	11.43%
D BASELINE RAG PERFORMANCE	20	≈ 2 pp.	5.71%
E DATASET STATISTICS	22	≈ 0.58 pp.	1.66%
F DATASET QUALITY VERIFICATION	22	≈ 0.42 pp.	1.2%
G CROSS-MODEL AND HUMAN CONSISTENCY ANALYSIS	23	≈ 1.67 pp.	4.77%
H FINE-GRAINED IMPROVEMENT ANALYSIS OF SEARCH-ALIGN	24	≈ 0.72 pp.	2.06%
I SEARCH-ALIGN FINE-TUNING DETAILS	25	≈ 0.61 pp.	1.74%
J SOFT HPS EVALUATION WITH THRESHOLDED SEMANTIC MATCHING	26	≈ 0.37 pp.	1.06%
K EVALUATION UNDER TOP-k RETRIEVAL	26	≈ 1.63 pp.	4.66%
L AGENTIC MM-RAG CASE STUDY	28	≈ 2 pp.	5.71%
M THEORETICAL JUSTIFICATION OF TOPOLOGY DESIGNS	30	≈ 1 pp.	2.86%
N PROMPT FOR DATA CONSTRUCTION	31	≈ 3 pp.	8.57%
O PROMPT TEMPLATE FOR AGENTIC MM-RAG PIPELINE	34	≈ 1 pp.	2.86%
P PROMPT TEMPLATE FOR LLM-AS-JUDGE	35	≈ 0.89 pp.	2.54%

All detected figure and table locations

Main text

• Figure 1 · p. 1
• Table 1 · p. 2
• Figure 2 · p. 3
• Table 2 · p. 4
• Table 3 · p. 7
• Figure 3 · p. 8
• Figure 4 · p. 8
• Table 4 · p. 9
• Figure 5 · p. 9

Attached appendix

• Table 5 · p. 20
• Table 6 · p. 21
• Table 7 · p. 22
• Table 8 · p. 22
• Table 9 · p. 23
• Figure 6 · p. 23
• Table 10 · p. 24
• Table 11 · p. 24
• Figure 7 · p. 25
• Table 12 · p. 25
• Table 13 · p. 26
• Figure 8 · p. 26
• Table 14 · p. 27
• Figure 9 · p. 28
• Figure 10 · p. 29

Argument and organization

1. Gap and benchmark construction

   Explains why final-answer accuracy misses failures in long multimodal search, then defines topologies, filtering and process metrics.

2. Unified pipeline and process supervision

   Separates the evaluated retrieval loop from training on verified trajectories.

3. Research-question-led evaluation

   Organizes results by performance, length, retrieval depth, modality coverage and error types.

Problem, claim and contribution

The opening turns a vague reasoning goal into five annotated retrieval structures. Sub-question, modality, evidence and intermediate answer become explicit units of evaluation.

Sections 1–2.1, pages 1–4. · Main paper

Argument and section order

Five named research questions map directly onto the results subsections, separating benchmark ranking from diagnosis of chain length, retrieval and modality errors.

Section 4, pages 6–9. · Main paper

Figures, captions and visual explanation

Figure 2 pairs data construction with the evaluated agent loop; repeated image/text icons connect the taxonomy to chain matching. The miniature introductory conversation is motivating but needs enlargement for readable details.

Figure 1, page 1; Figure 2, page 3. · Main paper

Tables, comparisons and evidence

Table 3 groups models by topology and distinguishes answer metrics from process metrics. Shaded groups, tiny topology diagrams and red fine-tuning rows aid navigation, while the dense full-page layout needs careful reading.

Table 3, page 7. · Main paper

Appendix and reproducibility

The indexed supplement includes complete data examples, prompts, human/judge agreement, soft matching and top-k retrieval tests. These support auditability beyond a single final-answer score.

Contents page 14; Appendices C and G–P, pages 16–35. · Main paper

Clarity and writing conventions

Cross-table reconciliation is still necessary: aligned Qwen's Multi-Images Fork HPS is 38.01 in Table 3 but 41.20 for Retrieval@1 in Table 14; Table 4's InternVL overall image count also does not equal its subgroup sum.

Tables 3, 4 and 14, pages 7, 9 and 27. · Main paper

What to reuse—and what not to infer

Transferable lesson: For a benchmark paper, let the annotation units determine the metrics and organize the experiments around diagnostic questions; expose prompts and concrete success/failure traces alongside aggregate scores.

Boundary: Model-generated references and judges do not establish independent truth. The curated local, usually top-1 retrieval setting and restricted topology family are not unrestricted web search; cross-table inconsistencies and exact train/evaluation split boundaries need resolution before numerical reuse.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-09-19

Reading status

Full text + selected visuals inspected

Main text

Pages 1–10, Sections 1–6 and ethics/reproducibility statements; references on pages 11–13.

Appendix / supplement

Pages 14–35, contents and Appendices A–P, including all examples, comparison tables, judge checks, topology assumptions and prompt templates.

Visually inspected PDF pages

1, 3, 7, 9

• ICLR 2026 official proceedings PDF, retrieved 2026-09-19 · 35 pages