Writing study: ReQAT: Achieving Full-Precision Reasoning Accuracy with 4-bit Floating-Point Quantization-Aware Training

ICML overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICML 2026 oral writing corpus

Janghwan Lee, Sihwa Lee, Jinseok Kim, Yongjik Kim, Jieun Lim, Jinwook Oh, Jungwook Choi · 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: Training, objectives & optimization (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

22 physical pages

Before references / appendix

≈ 9.21 page-equivalents

Main figures / tables

7 / 7

Attached appendix figures / tables

4 / 10

Attached appendix span

≈ 7.91 page-equivalents

Before references / appendix: ≈ 41.86%References / intervening material: ≈ 22.18%Attached appendix: ≈ 35.95%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	≈ 1.27 pp.	5.77%
2. Background	2	≈ 0.79 pp.	3.59%
3. Empirical Analysis	3	≈ 1.38 pp.	6.27%
4. Methods	4	≈ 2.11 pp.	9.59%
5. Experiments	6	≈ 1.99 pp.	9.05%
6. Discussions	8	≈ 0.75 pp.	3.41%
7. Related Works	9	≈ 0.22 pp.	1%
8. Conclusion	9	≈ 0.16 pp.	0.73%
Acknowledgements	10	≈ 0.11 pp.	0.5%
Impact Statement	10	≈ 0.05 pp.	0.23%
References	10	≈ 4.88 pp.	22.18%
A. Microscaling Formats	15	≈ 0.37 pp.	1.68%
B. Details for Token Entropy Analysis in Sec. 3	15	≈ 2.33 pp.	10.59%
C. ReQAT Details	17	≈ 2.3 pp.	10.45%
D. Experimental Details and More Results	20	≈ 2.91 pp.	13.23%

All detected figure and table locations

Main text

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

Attached appendix

• Table 8 · p. 15
• Figure 8 · p. 16
• Figure 9 · p. 17
• Table 9 · p. 18
• Figure 10 · p. 19
• Figure 11 · p. 19
• Table 10 · p. 20
• Table 11 · p. 21
• Table 12 · p. 21
• Table 13 · p. 22
• Table 14 · p. 22
• Table 15 · p. 22
• Table 16 · p. 22
• Table 17 · p. 22

Argument and organization

1. 1 Introduction; 2 Background; 3 Empirical Analysis

   Diagnoses low-entropy sampling errors before proposing a training change.

2. 4 Methods

   Maps the diagnosis to trace alignment, selective entropy loss and calibration initialization.

3. 5 Experiments; 6 Discussion; 7 Related Work; 8 Conclusion

   Separates budget-matched accuracy, serving throughput, ablations and supervision limits.

Problem, claim and contribution

The opening connects compression to a particular reasoning failure, not simply a global increase in reconstruction error.

Abstract, p.1; Section 3, pp.3-4 · Main paper

Argument and section order

Controlled mixed-precision routing is separated from the more scalable logit-noise proxy before the method is introduced.

Sections 3.2-4, p.4 · Main paper

Figures, captions and visual explanation

The diagnostic composite links token examples, probability changes, routing and perturbation curves; the results tables organize precision and training budgets rather than merging all configurations.

Figure 2, p.3; Tables 1-3, p.7 · Main paper

Tables, comparisons and evidence

The text matches fine-tuning token budgets, but the tables show that adding every component is not monotonically beneficial and that the 8B model remains below its BF16 fine-tuned baseline.

Tables 1-2 and accompanying claims, p.7 · Main paper

Appendix and reproducibility

Evaluation details distinguish AIME-120 from single-year reporting, state eight seeds and separate calibration from training and serving settings.

Appendix D.1-D.3, p.20 · Main paper

Clarity and writing conventions

The limitations identify dependence on trace quality, while the broad full-precision headline requires model-, budget- and component-specific qualification.

Limitations and Conclusion, p.9 · Main paper

What to reuse—and what not to infer

Transferable lesson: Make the diagnostic-to-intervention link explicit, then audit broad recovery and monotonic-improvement language against every relevant table row.

Boundary: The title is not a universal full-precision guarantee; hardware throughput reflects batching and memory as well as arithmetic, and logit noise is a proxy intervention.

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, 4, 5, 7, 8, 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 20. Remaining appendix proofs and details were not exhaustively read. No separately linked supplement inspected.

Visually inspected PDF pages

3, 7

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 · 22 pages