Writing study: Vid-LLM: A Compact Video-based 3D Multimodal LLM with Reconstruction–Reasoning Synergy

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Haijier Chen, Bo Xu, Shoujian zhang, Haoze Liu, Jiaxuan Lin, Jingrong Wang · 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) · 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

20 physical pages

Before references / appendix

≈ 10.31 page-equivalents

Main figures / tables

5 / 6

Attached appendix figures / tables

2 / 3

Attached appendix span

≈ 3.89 page-equivalents

Before references / appendix: ≈ 51.55%References / intervening material: ≈ 29%Attached appendix: ≈ 19.45%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	≈ 0.76 pp.	3.8%
2 RELATED WORK	2	≈ 0.85 pp.	4.25%
3 METHOD	3	≈ 3.79 pp.	18.95%
4 EXPERIMENTS	7	≈ 3.27 pp.	16.35%
5 CONCLUSION	10	≈ 0.34 pp.	1.7%
6 REPRODUCIBILITY STATEMENT	11	≈ 0.09 pp.	0.45%
7 ACKNOWLEDGMENTS	11	≈ 0.11 pp.	0.55%
REFERENCES	11	≈ 5.8 pp.	29%
A APPENDIX	17	≈ 3.89 pp.	19.45%

All detected figure and table locations

Main text

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

Attached appendix

• Figure 6 · p. 17
• Figure 7 · p. 18
• Table 7 · p. 18
• Table 8 · p. 19
• Table 9 · p. 19

Argument and organization

1. Reconciling geometry and semantics

   Frames a representation-sharing problem rather than simply adding another encoder.

2. Shared encoder, cross-task adapter and staged training

   Separates architecture from supervision and optimization.

3. Task results, coupling ablations and failures

   Tests both branches and their interaction.

Problem, claim and contribution

The introduction distinguishes externally supplied 3D inputs from geometry inferred from video. This inference-time distinction is useful, but must not be rewritten as absence of geometry supervision: the training objective includes camera, depth and point-cloud targets.

Sections 1, 3.4, pp.1–2, 6 · Main paper

Argument and section order

Method subsections progress from shared features to metric scale, 3D token construction and losses. The explicit statement that language losses do not update the reconstruction branch is essential qualification for the broader language of mutually reinforcing tasks.

Sections 3.1–3.4, pp.4–6 · Main paper

Figures, captions and visual explanation

Figure 2 visually separates geometry and language streams while identifying the bridge tokens that exchange information. Figure 3 adds freeze/flame symbols and teacher branches, revealing the training schedule that a single deployment diagram would conceal.

Figures 2–3, pp.4–5 · Main paper

Tables, comparisons and evidence

Input modalities and reconstructed-geometry baselines are separated in tables. Nevertheless, Table 2 does not show best results on every caption metric: VGLLM has ROUGE 62.4 versus 61.8. Table 6 reports the full model's ScanRefer aggregate as 48.4, inconsistent with 59.8 in Table 5 under the stated averaging rule.

Tables 2, 5–6; evaluation details, pp.7–10 · Main paper

Appendix and reproducibility

Failure cases identify incomplete viewpoints and reflective surfaces as causes of missing chairs and undersized grounding boxes. Additional geometry tables distinguish ground-truth scale alignment from predicted metric depth, making their comparison more interpretable.

Appendix A.1–A.2, Figure 7 and Tables 7–9, pp.17–20 · Main paper

Clarity and writing conventions

The reproducibility promise is less informative than a complete configuration: learning rate, stage durations, training mixture and latency hardware are not clearly specified in this PDF. The label compact would benefit from a matched parameter and runtime accounting.

Sections 4.1, 6 and Table 5, pp.7, 9, 11 · Main paper

What to reuse—and what not to infer

Transferable lesson: Use complementary deployment and training diagrams, distinguish inference inputs from training supervision, and evaluate each branch separately from the joint task.

Boundary: Several numerical summaries and qualitative annotations need correction before reuse; Table 6's ScanRefer discrepancy cannot be resolved from the paper alone. Video-only input is not evidence of label-free or geometry-free learning.

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 1–11, Sections 1–7; references 11–16

Appendix / supplement

Pages 17–20, Appendix A.1–A.3 including failure cases, geometry evaluations and LLM-use statement

Visually inspected PDF pages

4, 5

• official ICLR 2026 proceedings PDF · 20 pages