Writing study: DepthLM: Metric Depth from Vision Language Models

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

zhipeng cai, Ching-Feng Yeh, Hu Xu, Zhuang Liu, Gregory P. Meyer, Xinjie Lei, Changsheng Zhao, Shang-Wen Li, Vikas Chandra, Yangyang Shi · 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: Architecture & representation design (primary) · Empirical analysis, diagnosis & auditing.

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

15 physical pages

Before references / appendix

≈ 9.65 page-equivalents

Main figures / tables

8 / 3

Attached appendix figures / tables

2 / 3

Attached appendix span

≈ 1.89 page-equivalents

Before references / appendix: ≈ 64.33%References / intervening material: ≈ 23.07%Attached appendix: ≈ 12.6%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	2	≈ 0.51 pp.	3.4%
2 RELATED WORK	2	≈ 0.6 pp.	4%
3 DEPTHLM	3	≈ 4.2 pp.	28%
4 RESULTS	7	≈ 3.03 pp.	20.2%
5 CONCLUSION	10	≈ 0.21 pp.	1.4%
REFERENCES	10	≈ 3.46 pp.	23.07%
A APPENDIX	14	≈ 1.89 pp.	12.6%

All detected figure and table locations

Main text

• Figure 1 · p. 1
• Figure 2 · p. 4
• Figure 3 · p. 4
• Figure 4 · p. 5
• Figure 5 · p. 6
• Figure 6 · p. 6
• Figure 7 · p. 7
• Table 1 · p. 8
• Table 2 · p. 8
• Figure 8 · p. 9
• Table 3 · p. 10

Attached appendix

• Figure 9 · p. 14
• Table 4 · p. 14
• Figure 10 · p. 14
• Table 5 · p. 15
• Table 6 · p. 15

Argument and organization

1. Analysis of the performance gap

   Pose separate questions about pixel reference, objective and camera ambiguity.

2. Findings assembled into DepthLM

   Map each diagnostic result to one component of the final pipeline.

3. Depth and multi-task evaluation

   Compare sparse point estimates and extend the same text interface to related geometric questions.

Problem, claim and contribution

The central question asks whether a standard VLM needs a new dense head or regression loss. The paper instead traces failure to input reference and camera scale, making a simple intervention scientifically legible.

Introduction and Section 3, pages 2–3 · Main paper

Argument and section order

Four numbered findings organize controlled comparisons before the final method. Section 3.4 explicitly connects each design choice back to its supporting finding, a clear diagnosis-to-design narrative.

Sections 3.1–3.4, pages 3–7 · Main paper

Figures, captions and visual explanation

Figure 6 uses a short pipeline with intrinsics and markers entering at different stages, while the adjacent fixed-budget label-density plot explains why image diversity matters. The visual and causal roles are distinct.

Figures 5–6, page 6 · Main paper

Tables, comparisons and evidence

Table 1 groups general, spatial and depth-trained VLMs, including prompt-adjusted baselines. Table 2 separately labels values taken from prior pure-vision papers, which are not a uniform rerun of the same evaluation pipeline.

Tables 1–2, page 8 · Main paper

Appendix and reproducibility

The compact appendix gives actual marker shapes, available-versus-used dataset counts, prompts and per-model settings. Its 35M/60M training-sample counts clarify that the final models differ from the one-point-per-image diagnostic.

Appendices A.1–A.7, pages 14–15 · Main paper

Clarity and writing conventions

Some claims need narrower wording: the pose task predicts camera travel distance rather than a full six-degree-of-freedom pose, and querying 10,000 pixels separately is not equivalent to efficient dense prediction.

Section 3.4, page 7; visualization discussion, page 9; Table 3, page 10 · Main paper

What to reuse—and what not to infer

Transferable lesson: Organize an ablation-driven method as explicit questions and numbered findings, then show exactly how the final pipeline follows from those findings.

Boundary: Known intrinsics and undistorted pinhole geometry are assumed. Sparse query accuracy, borrowed dense-model numbers and selected point-cloud visualizations do not establish matched-compute dense reconstruction or universal 3D reasoning.

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 1–5

Appendix / supplement

Pages 14–15, Appendices A.1–A.7, markers, dataset statistics, GRPO prompts and training settings

Visually inspected PDF pages

6, 8, 13

• ICLR 2026 final proceedings · 15 pages