Writing study: CAT-Q: Cost-efficient and Accurate Ternary Quantization for LLMs

ICML overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICML 2026 oral writing corpus

Shigeng Wang, Chao Li, Yangyuxuan Kang, Jiawei Fan, Anbang Yao · Official ICML 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: Efficiency, inference & systems (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

19 physical pages

Before references / appendix

≈ 9.04 page-equivalents

Main figures / tables

3 / 13

Attached appendix figures / tables

11 / 2

Attached appendix span

≈ 6.91 page-equivalents

Before references / appendix: ≈ 47.58%References / intervening material: ≈ 16.05%Attached appendix: ≈ 36.37%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Source-bound correction: Restore the source’s letter-only appendix captions: Tables A-B and Figures A-K on physical pp14-19. These are caption labels, not numbered main sections.

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	≈ 1.55 pp.	8.16%
2. Method	2	≈ 2.76 pp.	14.53%
3. Experiments	5	≈ 3.63 pp.	19.11%
4. Related Work	9	≈ 0.5 pp.	2.63%
5. Conclusion	9	≈ 0.1 pp.	0.53%
Impact Statement	9	≈ 0.14 pp.	0.74%
References	10	≈ 3.04 pp.	16%
Appendix	13	≈ 0.2 pp.	1.05%
A. Datasets Used in Experiments	13	≈ 0.8 pp.	4.21%
B. Implementation Details of CAT-Q	14	≈ 0.69 pp.	3.63%
C. Illustrations of the Smooth Transition Function f (·) with Varying Sharpness	14	≈ 0.71 pp.	3.74%
D. Comparison of Loss Curves of CAT-Q with vs. without the ST Component	15	≈ 0.17 pp.	0.89%
E. A Pilot Study of CAT-Q on Challenging Mathematics and Coding Tasks	15	≈ 0.43 pp.	2.26%
F. A More Comprehensive Comparison of Different Strategies for Determining α and ∆.	16	≈ 0.24 pp.	1.26%
G. Discussion of Limitations	16	≈ 3.67 pp.	19.32%

All detected figure and table locations

Main text

• Figure 1 · p. 3
• Figure 2 · p. 3
• Figure 3 · p. 4
• Table 1 · p. 6
• Table 2 · p. 6
• Table 3 · p. 6
• Table 4 · p. 6
• Table 5 · p. 7
• Table 6 · p. 7
• Table 7 · p. 7
• Table 8 · p. 8
• Table 9 · p. 8
• Table 10 · p. 8
• Table 11 · p. 8
• Table 12 · p. 8
• Table 13 · p. 9

Attached appendix

• Table A · p. 14
• Figure A · p. 14
• Figure B · p. 15
• Table B · p. 15
• Figure C · p. 17
• Figure D · p. 17
• Figure E · p. 17
• Figure F · p. 18
• Figure G · p. 18
• Figure H · p. 18
• Figure I · p. 19
• Figure J · p. 19
• Figure K · p. 19

Argument and organization

1. 1 Introduction; 2 Method

   Separates ternary post-training quantization from training low-bit models from scratch.

2. 2.2-2.4 Components

   Builds modulation, soft-to-hard conversion and the sliding-layer reconstruction objective.

3. 3 Experiments; 4 Related Work; 5 Conclusion

   Separates model quality, component ablations and measured deployment performance.

Problem, claim and contribution

The low-data headline concerns calibration of pretrained models, making inherited pretraining an essential part of the comparison context.

Abstract, p.1; Tables 2-3, p.6 · Main paper

Argument and section order

The method distinguishes a learnable distribution adjustment from the transition schedule and the multi-layer objective that uses both.

Section 2, pp.3-5 · Main paper

Figures, captions and visual explanation

The overview places learned factors and statistical priors beneath the conversion flow; deployment results put hardware, batch size, length and actual storage formats in the caption.

Figure 1, p.3; Table 13, p.9 · Main paper

Tables, comparisons and evidence

Full-precision rows retain the accuracy loss after ternarization, while comparisons with QAT models explicitly count quantization-stage tokens.

Tables 1-3, p.6 · Main paper

Appendix and reproducibility

The supplement begins with an itemized roadmap and ends by acknowledging remaining accuracy gaps and limited optimized kernel support.

Appendix opening, p.13; Appendix G, p.16 · Main paper

Clarity and writing conventions

The broad efficiency claim is made more interpretable by separate quality, calibration-cost and hardware-throughput reports rather than bit width alone.

Sections 3.3-3.4, pp.6-9 · Main paper

What to reuse—and what not to infer

Transferable lesson: Separate inherited model cost, calibration cost, representation size and measured deployment speed instead of treating one as a proxy for all four.

Boundary: The large token reduction is not an end-to-end pretraining saving; accuracy remains below full precision and deployment gains are hardware/format-specific.

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-oriented scan of the main argument through physical p.9, with focused physical-page text inspection on 1, 3, 5, 6, 7, 9. Framing, section progression, mechanism/proof explanation, displays, comparison conditions and conclusions were inspected. This is not exhaustive technical reading, proof verification or reproduction.

Appendix / supplement

Appendix organization and main-text pointers were checked; focused text inspection of physical pages 13, 16. Remaining appendix proofs and details were not exhaustively read. No separately linked supplement inspected.

Visually inspected PDF pages

3, 9

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.

• Published proceedings PDF, PMLR volume 306, retrieved 2026-10-07 · 19 pages