Writing study: Quantized Reasoning Models Think They Need to Think Longer, but They Do Not

NeurIPS overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← NeurIPS 2026 oral writing corpus

Sanae Lotfi, Polina Kirichenko, Steven Li, Zechun Liu · Official NeurIPS 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: Empirical analysis, diagnosis & auditing (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

≈ 9.47 page-equivalents

Main figures / tables

5 / 1

Attached appendix figures / tables

5 / 11

Attached appendix span

≈ 9.93 page-equivalents

Before references / appendix: ≈ 43.05%References / intervening material: ≈ 11.82%Attached appendix: ≈ 45.14%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Source-bound correction: Top-level heading and reference/appendix boundaries checked against source PDF text. Bibliography titles/author initials, numbered contribution/property bullets, diagram steps, equations and TOC entries are not substantive headings. Section spans are page-coordinate estimates, not ink area or prose-only length. This is an author manuscript mirror, not a verified accepted-version template. Source layout uses physical single-column y/page-height, including equation-heavy pages. Complete numbered-caption inventory cross-checked against source text; restored real punctuation-free captions and excluded inline references. Selected pages visually inspected, not every image/table.

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	≈ 1.05 pp.	4.77%
2 Related Work	2	≈ 0.76 pp.	3.45%
3 Experimental Setup	3	≈ 0.89 pp.	4.05%
4 Quantization Exacerbates Overthinking in Reasoning Models	4	≈ 2.25 pp.	10.23%
5 Penalizing Overthinking Markers	6	≈ 2.73 pp.	12.41%
6 Why Does Quantization Exacerbate Overthinking?	9	≈ 0.6 pp.	2.73%
7 Conclusion, Limitations, and Future Work	10	≈ 0.39 pp.	1.77%
References	10	≈ 2.61 pp.	11.86%
Appendix	13	≈ 0.05 pp.	0.23%
A Experimental Details	13	≈ 1.3 pp.	5.91%
B Additional Results	14	≈ 4.19 pp.	19.05%
C Token Lists Used in Experiments	18	≈ 1.08 pp.	4.91%
D GPT-5 Prompt for Error Categorization	19	≈ 3.3 pp.	15%

All detected figure and table locations

Main text

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

Attached appendix

• Figure 6 · p. 15
• Figure 7 · p. 15
• Figure 8 · p. 16
• Figure 9 · p. 16
• Figure 10 · p. 17
• Table 2 · p. 18
• Table 3 · p. 19
• Table 4 · p. 19
• Table 5 · p. 20
• Table 6 · p. 20
• Table 7 · p. 20
• Table 8 · p. 21
• Table 9 · p. 21
• Table 10 · p. 21
• Table 11 · p. 21
• Table 12 · p. 22

Argument and organization

1. 1 Introduction

   Motivates quantization-induced reasoning overhead.

2. 2 Related Work

   Positions quantization and overthinking studies.

3. 3 Experimental Setup

   Specifies models, quantization and reasoning benchmarks.

4. 4 Quantization Exacerbates Overthinking in Reasoning Models

   Localizes behavioral and token-level failure patterns.

5. 5 Penalizing Overthinking Markers

   Tests targeted penalties against alternative token lists.

6. 6 Why Does Quantization Exacerbate Overthinking?

   Discusses a mechanism for the observed behavioral change.

7. 7 Conclusion, Limitations, and Future Work

   States the efficiency lesson and model/task limits.

Problem, claim and contribution

The introduction turns the apparent compute advantage of quantization into a testable counter-effect: longer reasoning can erase part of its benefit. It states a behavioral problem before introducing the decoding intervention.

PDF pp.1–2, Introduction and Figure 1 · Main paper

Argument and section order

The evidence order is diagnosis, error taxonomy, token-level localization, then an intervention with alternative token controls. This gives the remedy an explanatory motivation rather than presenting only an accuracy table.

PDF pp.4–9, Sections 4–5 · Main paper

Figures, captions and visual explanation

Figure 3 uses the same four error classes in four quantization panels, with control token lists encoded consistently across panels. Table 1 pairs Base/+Pen rows and accuracy/length values for each benchmark, making both quality and output cost visible.

PDF p.5, Figure 3; p.8, Table 1 · Main paper

Tables, comparisons and evidence

Equal-sized high-KL, low-KL, and random-token controls challenge whether any penalty would help. The large comparison table retains cells where accuracy falls. Its caption specifies best-lambda results and variation across lambda, so these deviations should not be described as repeated-seed uncertainty.

PDF pp.7–9, Section 5.2 and Table 1 · Main paper

Appendix and reproducibility

The appendix separates implementation/calibration details, extra model and benchmark results, literal token lists, and the error-labeling prompt. This supports reproducibility of an intervention whose precise word list and annotation procedure are part of the method.

PDF pp.13–22, Appendices A–D · Main paper

Clarity and writing conventions

The error classes receive concrete definitions, including reaching a correct intermediate answer but subsequently abandoning it. The paper states that a manual annotation subset was used to calibrate an LLM judge rather than implying that all scaled labels were human annotations.

PDF p.5, Section 4.2; p.10, limitations · Main paper

What to reuse—and what not to infer

Transferable lesson: For an efficiency diagnosis, pair mechanism-specific controls with both quality and cost, and define exactly what each uncertainty or variation statistic measures.

Boundary: Best-penalty selection and variation across penalty strengths are not an unbiased deployment estimate or seed-based error bars. Marker lists, model families, and language/task coverage limit the generalization. The scanned arXiv version is not verified identical to the accepted revision.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-10-02

Reading status

Writing-focused manuscript scan + selected visuals inspected

Main text

Writing-oriented source scan of framing, argument roadmap, selected substantive passages, method/theorem statements, main figure/table captions, evidence and limits across PDF pp.1–10. Two selected main-body pages were actually inspected as rendered images. This is not exhaustive line-by-line technical verification or experiment reproduction. Title and all author names compared against official program metadata and manuscript first page; matching authors, with capitalization, middle initials or expanded given names where applicable.

Appendix / supplement

Appendix A–D headings, implementation openings, selected extended-result captions, literal token-list and annotation-prompt locations were checked; the full experiment suite was not reproduced.

Visually inspected PDF pages

5, 8

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.

• Public author manuscript arXiv:2606.00206v1; not verified identical to the accepted OpenReview revision · 22 pages