Writing study: FlashVID: Efficient Video Large Language Models via Training-free Tree-based Spatiotemporal Token Merging

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Ziyang Fan, Keyu Chen, Ruilong Xing, Yulin Li, Li Jiang, Zhuotao Tian · 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: Inference, algorithms & systems efficiency (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

30 physical pages

Before references / appendix

≈ 10.11 page-equivalents

Main figures / tables

4 / 6

Attached appendix figures / tables

6 / 8

Attached appendix span

≈ 13.9 page-equivalents

Before references / appendix: ≈ 33.7%References / intervening material: ≈ 19.97%Attached appendix: ≈ 46.33%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 heading	Starts on PDF page	Estimated span	Share of entire PDF
1 INTRODUCTION	1	≈ 2.08 pp.	6.93%
2 BACKGROUND AND MOTIVATION	3	≈ 1.83 pp.	6.1%
3 METHODOLOGY	5	≈ 1.81 pp.	6.03%
4 EXPERIMENTS	7	≈ 3.35 pp.	11.17%
5 CONCLUSION	10	≈ 0.42 pp.	1.4%
REFERENCES	11	≈ 6 pp.	20%
Supplementary Material	17	≈ 1.45 pp.	4.83%
A MORE EXPERIMENTAL RESULTS	18	≈ 2.1 pp.	7%
B EVALUATION BENCHMARKS	20	≈ 1.09 pp.	3.63%
C IMPLEMENTATION DETAILS	21	≈ 2.52 pp.	8.4%
D RELATED WORK	24	≈ 0.96 pp.	3.2%
E MORE VISUALIZATIONS	25	≈ 4.14 pp.	13.8%
F USAGE OF LARGE LANGUAGE MODELS	29	≈ 1.63 pp.	5.43%

All detected figure and table locations

Main text

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

Attached appendix

• Table 7 · p. 18
• Table 8 · p. 18
• Table 9 · p. 19
• Table 10 · p. 19
• Table 12 · p. 20
• Table 11 · p. 20
• Table 13 · p. 20
• Table 14 · p. 20
• Figure 5 · p. 26
• Figure 6 · p. 27
• Figure 7 · p. 28
• Figure 8 · p. 28
• Figure 9 · p. 29
• Figure 10 · p. 30

Argument and organization

1. Observation before mechanism

   Explains why fixed spatial correspondences miss video motion, then compares inference-compression paradigms before introducing the tree representation.

2. Complementary modules

   Presents representative-token selection and residual-token merging as different responsibilities, joined by an explicit algorithm.

3. Accuracy, capacity and runtime

   Separates fixed-retention accuracy, more frames at a token budget, component ablations and measured inference timing.

Problem, claim and contribution

The motivation identifies a concrete failed assumption—semantic correspondence need not remain at the same image coordinate. That observation supports the tree-merging design more persuasively than a general statement that videos contain redundant tokens.

Section 1 and Section 2.2; Figures 1 and 3 · Main paper

Figures, captions and visual explanation

The overview gives token selection and merging parallel panels, then shows where their outputs enter the existing video-language pipeline. Distinct fills distinguish representative from merged tokens. This separates the contribution from the unchanged backbone, although the full hybrid pruning step is explained later in the appendix.

Figure 4, page 5; Appendix C.2 · Main paper

Tables, comparisons and evidence

Accuracy tables group retention ratios and backbone models, while Table 6 separately itemizes vision encoding, compression, LLM forward time and time to first token. Reporting both prefill and end-to-end timing makes clear why a 6.3× prefill gain becomes only about 2.1× at first token.

Tables 1–3 and 6; Section 4.4 · Main paper

The ablations retain unsuccessful depth/breadth constraints rather than implying every plausible structural constraint helps. Failure visualizations admit that feature similarity can merge different entities. These are useful boundaries for a method whose headline emphasizes negligible degradation.

Appendix A.3, Tables 11–12; Appendix E.5, Figure 9 · Main paper

Appendix and reproducibility

A supplementary contents page routes readers to extra results, baseline reproduction, implementation, budget alignment and visual failures. The budget derivation and baseline modifications are particularly important: without them the apparent same-budget comparison is underspecified.

Supplementary contents, page 17; Appendix C.1–C.3 · Main paper

Clarity and writing conventions

Budget terminology needs tighter scope than the headline: matching average tokens per transformer layer does not prove equal total video-encoding cost or wall-clock time when input frames increase tenfold. The paper's detailed timing table is the stronger model for reporting deployment efficiency.

Table 3; Section 4.4; Appendix C.3, Equations 12–15 · Main paper

What to reuse—and what not to infer

Transferable lesson: Turn an efficiency paper into three distinct tests: quality at a fixed compression level, quality when spending saved capacity, and actual measured runtime including the compression overhead. State exactly which resource is held fixed.

Boundary: Token similarity is not verified object correspondence. The main overview is not the complete deployed pipeline: segmentation, budget-filling compression and inner-LLM pruning also matter. Average layer-token alignment is only a proxy for equal compute, and more source frames can increase costs outside the LLM.

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

Complete Sections 1–5, pages 1–10: motivation, efficiency definitions, method, experiments and conclusion.

Appendix / supplement

Complete supplementary contents and Appendices A–F, pages 17–30, including algorithms, baseline adaptation, token-budget algebra, failure cases and language-model disclosure.

Visually inspected PDF pages

5, 10

• Official ICLR 2026 proceedings manuscript with supplementary material A–F · 30 pages