Writing study: Agent0-VL: Exploring Self-Evolving Agent for Tool-Integrated Vision-Language Reasoning

ICML overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICML 2026 oral writing corpus

Jiaqi Liu, Kaiwen Xiong, Peng Xia, Yiyang Zhou, Haonian Ji, Lu Feng, Siwei Han, Mingyu Ding, Huaxiu Yao · 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: Training, objectives & optimization (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

18 physical pages

Before references / appendix

≈ 8.1 page-equivalents

Main figures / tables

4 / 4

Attached appendix figures / tables

8 / 1

Attached appendix span

≈ 5.91 page-equivalents

Before references / appendix: ≈ 45%References / intervening material: ≈ 22.17%Attached appendix: ≈ 32.83%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	≈ 1.02 pp.	5.67%
2. Related Work	2	≈ 0.63 pp.	3.5%
3. Preliminaries	3	≈ 0.35 pp.	1.94%
4. Methodology	3	≈ 2.12 pp.	11.78%
5. Experiments	5	≈ 3.09 pp.	17.17%
6. Conclusion	8	≈ 0.18 pp.	1%
Impact Statement	9	≈ 0.05 pp.	0.28%
References	9	≈ 3.99 pp.	22.17%
A. Notation	13	≈ 0.03 pp.	0.17%
B. Implementation Details	13	≈ 0.52 pp.	2.89%
C. Prompting and Templates	13	≈ 0.08 pp.	0.44%
D. Training Data Construction Pipeline	13	≈ 1.94 pp.	10.78%
E. Case Studies	15	≈ 3.35 pp.	18.61%

All detected figure and table locations

Main text

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

Attached appendix

• Figure 5 · p. 13
• Table 5 · p. 14
• Figure 6 · p. 15
• Figure 7 · p. 15
• Figure 8 · p. 16
• Figure 9 · p. 17
• Figure 10 · p. 17
• Figure 11 · p. 18
• Figure 12 · p. 18

Argument and organization

1. Problem and contribution

   The introduction links tool-grounded verification to self-improvement, but its zero-external-reward language requires the implementation's explicit phase boundary.

2. Argument sequence

   A shared policy and step-level feedback definition lead to an algorithm separating experience generation from policy updates; evaluation asks four explicit questions before reporting results.

3. Evidence and verification

   Iteration tables, retrained component removals and Best-of-8 critic comparisons test different uses of the shared model; the illustrated successful repair is an example rather than an error-rate estimate.

Problem, claim and contribution

The introduction links tool-grounded verification to self-improvement, but its zero-external-reward language requires the implementation's explicit phase boundary.

Introduction, p. 2; Appendix B, p. 13 · Main paper

Argument and section order

A shared policy and step-level feedback definition lead to an algorithm separating experience generation from policy updates; evaluation asks four explicit questions before reporting results.

Sections 3-5.1 and Algorithm 1, pp. 3-5 · Main paper

Figures, captions and visual explanation

The algorithm visually separates inner and outer loops; a four-phase case diagram shows initial answer, verifier feedback, patch and corrected reasoning next to paired PRM result rows.

Algorithm 1, p. 5; Figure 4 and Table 4, p. 8; both PNGs inspected · Main paper

Tables, comparisons and evidence

Iteration tables, retrained component removals and Best-of-8 critic comparisons test different uses of the shared model; the illustrated successful repair is an example rather than an error-rate estimate.

Sections 5.2-5.5, pp. 6-8 · Main paper

Appendix and reproducibility

Implementation discloses a two-epoch external-correctness RL warm-up before internal-reward self-evolution, plus teacher-generated SFT and manual spot-checks, materially qualifying the headline training description.

Appendix B, p. 13; D.2-D.4, p. 14 · Main paper

Clarity and writing conventions

Some cross-references and conditions need reconciliation: the p. 8 case points to Appendix 5.5 rather than Appendix E, and the p. 16 case triggers high-confidence error repair whereas Algorithm 1 uses low confidence.

Algorithm 1, p. 5; Section 5.5, p. 8; Appendix E, p. 16 · Main paper

What to reuse—and what not to infer

Transferable lesson: Separate phase-specific supervision claims from whole-pipeline claims, and audit pseudocode, example triggers and appendix links against the same operational definition.

Boundary: Zero external rewards describes the later self-evolving phase, not the entire training pipeline; successful selected cases do not establish verifier calibration, and the source's repair-gate descriptions are not fully consistent.

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. 1-8: unified solver/verifier motivation, POMDP formulation, rewards and selective repair, inner/outer optimization loops, iterative evaluation, ablations and PRM transfer.

Appendix / supplement

Writing-focused appendix scan; code, data and external assets not executed. Appendices A-E, pp. 13-18: notation, implementation and warm-up, role prompts, teacher/data construction, manual checks and full case studies.

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. · 18 pages