Writing study: InfoTok: Adaptive Discrete Video Tokenizer via Information-Theoretic Compression

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Haotian Ye, Qiyuan He, Jiaqi Han, Puheng Li, Jiaojiao Fan, Zekun Hao, Fitsum Reda, Yogesh Balaji, Huayu Chen, Sheng Liu, Angela Yao, James Y Zou, Stefano Ermon, Haoxiang Wang, Ming-Yu Liu · 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) · Theory & formal foundations · Inference, algorithms & systems efficiency.

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

23 physical pages

Before references / appendix

≈ 10.29 page-equivalents

Main figures / tables

4 / 3

Attached appendix figures / tables

2 / 3

Attached appendix span

≈ 9.9 page-equivalents

Before references / appendix: ≈ 44.74%References / intervening material: ≈ 12.22%Attached appendix: ≈ 43.04%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	≈ 1.7 pp.	7.39%
2 ADAPTIVE TOKENIZATION: THE FRAMEWORK	3	≈ 1.92 pp.	8.35%
3 INFOTOK: INFORMATION-THEORETIC ADAPTIVE TOKENIZATION	5	≈ 1.46 pp.	6.35%
4 EXPERIMENT	6	≈ 3.13 pp.	13.61%
5 RELATED WORKS	9	≈ 0.56 pp.	2.43%
6 DISCUSSIONS & LIMITATIONS	10	≈ 0.4 pp.	1.74%
7 CONCLUSION	10	≈ 0.31 pp.	1.35%
ETHICS STATEMENT	11	≈ 0.11 pp.	0.48%
REPRODUCIBILITY STATEMENT	11	≈ 0.08 pp.	0.35%
REFERENCES	11	≈ 2.81 pp.	12.22%
Supplementary Material	14	≈ 1 pp.	4.35%
A ILLUSTRATION OF ADAPTIVE TOKENIZATION	15	≈ 2 pp.	8.7%
B PROOFS	17	≈ 4.28 pp.	18.61%
C EXPERIMENTAL DETAILS	21	≈ 1.94 pp.	8.43%
D ADDITIONAL EXPERIMENT RESULTS	23	≈ 0.17 pp.	0.74%
E USE OF LARGE LANGUAGE MODELS	23	≈ 0.5 pp.	2.17%

All detected figure and table locations

Main text

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

Attached appendix

• Figure 5 · p. 15
• Figure 6 · p. 16
• Table 4 · p. 22
• Table 5 · p. 23
• Table 6 · p. 23

Argument and organization

1. Define tokenization as a coding problem

   Introduces encoder, quantizer and decoder before explaining why content-independent lengths can waste tokens.

2. Separate allocation from compression

   Uses a routing theorem to motivate length selection, then describes how a transformer compressor stores information and token positions.

3. Test reconstruction and overhead separately

   Compares matched average bit budgets, adaptive strategies and visual reconstructions; later gives measured latency and untested downstream scope.

Problem, claim and contribution

The paper turns a practical video-token bottleneck into a precise question about representation length and reconstruction. Contribution bullets then mirror theory, router/compressor design and empirical evaluation. The framing is strongest when near-optimality remains attached to the idealized coding assumptions, rather than implying that the practical lossy tokenizer realizes Shannon-optimal video representation.

Sections 1–3; Theorems 2.1, 2.2 and 3.1 · Main paper

Argument and section order

A common adaptive-tokenizer interface is presented before criticizing the uniform router. Algorithm 1 maps each operation to the relevant subsection. The implementation then separates length allocation from spatial token selection, preventing two uses of adaptivity from being conflated. Main-text limitations explicitly state that generation and action-understanding experiments were not performed.

Sections 2.2–3.2; Algorithm 1; Section 6 · Main paper

Figures, captions and visual explanation

Figure 1 uses a closed encode/compress/quantize/decompress/decode loop, with a separate router control arrow and two video examples showing different budgets. This makes the output contract much clearer than a list of modules. However, the drawing does not expose the extra decoder pass needed by the router; a reproduction-oriented figure should show that cost as well.

Figure 1, PDF page 2; Section 3.1 · Main paper

Tables, comparisons and evidence

The main table states processed 256×256 evaluation sets and matched average BPP16 budgets. Table 3 splits compressor ablation from architecture/mechanism comparison. Table 6 then shows a key trade-off: 1.23 seconds for Cosmos plus InfoTok versus 0.61 for fixed Cosmos, despite beating search-based adaptive alternatives. Fewer tokens and fewer search evaluations are not automatically faster tokenization than the fixed base.

Section 4.1; Tables 1–3; Appendix D, Table 6, PDF page 23 · Main paper

Appendix and reproducibility

The supplement has an explicit contents page and separates illustrative token masks, coding-tree proofs, architecture/hyperparameters and extra results. Appendix C.3 records both evaluation-dataset normalization and a training-EMA alternative, an important deployment detail not obvious from the overview. Model size, mask-bit overhead, minimum retained length and 33-frame chunking are all part of the reproducibility contract.

Supplement contents; Appendices A–D; Table 4; Section C.3 · Main paper

Clarity and writing conventions

Several statements need reconciliation before being reused: the abstract emphasizes 20% token savings while the introduction says approximately 50%; Table 2 labels its search result optimal even though some LPIPS/FVD entries are worse, so optimality is criterion-specific. The 0.81 DAVIS InfoTok-Flex SSIM/LPIPS values also differ between Tables 1 and 2. These are reasons to audit repeated numbers and name the optimized metric, not to copy a headline.

Abstract; end of Section 1; Tables 1–2, PDF pages 7 and 9 · Main paper

What to reuse—and what not to infer

Transferable lesson: Present a compression method as three contracts: what information is retained, how the rate is allocated, and what compute/side-information is needed to encode and decode. Pair the rate–quality plot with actual latency and clearly untested downstream tasks.

Boundary: The theory is an idealized lossless coding argument with assumptions on routing, optimization and likelihood approximation. Practical reconstruction-error routing and lossy video reconstruction need separate validation; the paper does not establish downstream VLM, action-recognition or WSI performance. A searched optimum for one criterion is not a universal upper bound on every reported metric.

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

All Sections 1–7 and ethics/reproducibility statements, PDF pages 1–11: coding argument, router, compressor, all main comparisons and limitations.

Appendix / supplement

All Appendices A–E, PDF pages 15–23: adaptive visual examples, full coding-tree proof and router bound, training/inference configurations, higher-resolution and latency results, LLM-use disclosure.

Visually inspected PDF pages

2, 9, 23

• Official ICLR 2026 proceedings manuscript including appendices · 23 pages