Skip to content
Computational Pathology Reading Roadmap

Writing study: daVinci-Dev: Agent-native Mid-training for Software Engineering

← ICML 2026 oral writing corpus

Ji Zeng, Dayuan Fu, Tiantian Mi, Zhuang Yumin, Yaxing Huang, Xuefeng Li, Lyumanshan Ye, Muhang Xie, Qishuo Hua, Zhen Huang, Mohan Jiang, Hanning Wang, Jifan Lin, Yang Xiao, Jie Sun, Yunze Wu, Pengfei Liu · Official ICML 2026 oral record

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
≈ 9.34 page-equivalents
Main figures / tables
4 / 6
Attached appendix figures / tables
3 / 0
Attached appendix span
≈ 4.56 page-equivalents
Before references / appendix: ≈ 51.89%References / intervening material: ≈ 22.78%Attached appendix: ≈ 25.33%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 headingStarts on PDF pageEstimated spanShare of entire PDF
1. Introduction1≈ 1.43 pp.7.94%
2. Background and Problem Setup3≈ 0.48 pp.2.67%
3. Agent-Native Data: Design and Synthesis3≈ 1.78 pp.9.89%
4. Experiments5≈ 2.03 pp.11.28%
5. Analysis7≈ 1.82 pp.10.11%
6. Related Work9≈ 0.58 pp.3.22%
7. Conclusion9≈ 0.13 pp.0.72%
Impact Statement9≈ 0.36 pp.2%
Acknowledgments10≈ 0.09 pp.0.5%
References10≈ 4.1 pp.22.78%
A. Training Details14≈ 0.65 pp.3.61%
B. Evaluation-Infrastructure Fixes for SWE-Bench Verified15≈ 0.59 pp.3.28%
C. LLM prompts used for PR rendering15≈ 0.09 pp.0.5%
D. Dataset Formats15≈ 2.92 pp.16.22%
E. Benchmark decontamination18≈ 0.31 pp.1.72%
All detected figure and table locations

Main text

Attached appendix

Argument and organization

  1. Problem and contribution

    Distinguishes contextually native trajectories, which preserve process structure, from environmentally native trajectories, which require real execution feedback.

  2. Argument sequence

    Data definitions precede training comparisons; robustness is then divided into post-training regime and model scale, before broader transfer and scaling discussion.

  3. Evidence and verification

    The table distinguishes author-rerun results from a value estimated from another figure; adjacent prose qualifies the 72B 0.3-point increment as within per-entry standard error.

Problem, claim and contribution

Distinguishes contextually native trajectories, which preserve process structure, from environmentally native trajectories, which require real execution feedback.

Introduction and Figure 1, p. 2 · Main paper

Argument and section order

Data definitions precede training comparisons; robustness is then divided into post-training regime and model scale, before broader transfer and scaling discussion.

Sections 3-5, pp. 3-9 · Main paper

Figures, captions and visual explanation

The overview uses two parallel data paths with distinct coverage/authenticity roles. Table 1 exposes mid-training data, post-training data and method rather than collapsing them into a model-name leaderboard.

Figure 1, p. 2; Table 1, p. 6; both PNGs inspected · Main paper

Tables, comparisons and evidence

The table distinguishes author-rerun results from a value estimated from another figure; adjacent prose qualifies the 72B 0.3-point increment as within per-entry standard error.

Table 1 and Section 4.3, p. 6 · Main paper

Appendix and reproducibility

Prompt and rendered-PR examples show the exact supervision format. Decontamination defines the n-gram overlap denominator and reports removed evaluation instances.

Appendices C-E, pp. 15-18 · Main paper

Clarity and writing conventions

Exploratory SWE-Bench Pro/Multilingual results are explicitly withheld from comparative claims because repository sources overlap training data.

Section 4.4, p. 6 · Main paper

What to reuse—and what not to infer

Transferable lesson: Expose the whole training/evaluation recipe in comparison columns and place contamination or uncertainty qualifications next to the gain.

Boundary: The overview's best-baseline bar comparison is less controlled than the full table; retain table annotations and do not treat small increments as established improvements.

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 1-7, pp. 1-9: deployment mismatch, two trajectory types, data construction, training/evaluation comparisons, scaling and conclusion.
Appendix / supplement
Writing-focused appendix scan; code, data and external assets not executed. Appendices A-E, pp. 14-18: stage and hyperparameter details, evaluation-infrastructure fixes, prompt/data-format examples and benchmark decontamination.
Visually inspected PDF pages
2, 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.