Writing study: Is Your LLM Overcharging You? Tokenization, Transparency, and Incentives

ICML overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICML 2026 oral writing corpus

Ander Artola Velasco, Stratis Tsirtsis, Nastaran Okati, Manuel Gomez Rodriguez · 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: Theory & formal analysis (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

25 physical pages

Before references / appendix

≈ 9.28 page-equivalents

Main figures / tables

3 / 0

Attached appendix figures / tables

8 / 1

Attached appendix span

≈ 11.91 page-equivalents

Before references / appendix: ≈ 37.12%References / intervening material: ≈ 15.24%Attached appendix: ≈ 47.64%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Source-bound correction: Join wrapped printed heading text using adjacent source lines; original heading start coordinates are unchanged.

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	≈ 1.65 pp.	6.6%
2. A Principal-Agent Model of Delegated Autoregressive Generation	3	≈ 1.33 pp.	5.32%
3. Provider Incentives Under the Pay-Per-Token Pricing Mechanism	4	≈ 3.1 pp.	12.4%
4. Incentive-Compatible Pricing Mechanisms	7	≈ 1.16 pp.	4.64%
5. Discussion and Limitations	8	≈ 0.95 pp.	3.8%
6. Conclusions	9	≈ 0.24 pp.	0.96%
Acknowledgements	10	≈ 0.1 pp.	0.4%
Impact Statement	10	≈ 0.14 pp.	0.56%
References	10	≈ 3.81 pp.	15.24%
A. Misreporting policies πm	14	≈ 1 pp.	4%
B. Additional Experimental Details	15	≈ 1 pp.	4%
C. Proofs	16	≈ 3 pp.	12%
D. Additional Experimental Results	19	≈ 6.91 pp.	27.64%

All detected figure and table locations

Main text

• Figure 1 · p. 5
• Figure 2 · p. 7
• Figure 3 · p. 8

Attached appendix

• Figure 4 · p. 19
• Figure 5 · p. 20
• Figure 6 · p. 21
• Figure 7 · p. 22
• Figure 8 · p. 22
• Figure 9 · p. 23
• Figure 10 · p. 24
• Table 1 · p. 24
• Figure 11 · p. 25

Argument and organization

1. 1 Introduction; 2 Principal-Agent Model

   Turns tokenization ambiguity into a formal information-asymmetry and utility model.

2. 3 Pay-Per-Token Incentives

   Separates misreporting incentives, worst-case optimization hardness and a feasible proof-of-concept vulnerability.

3. 4 Incentive-Compatible Pricing; 5 Discussion; 6 Conclusion

   Characterizes a restricted mechanism family and discusses operational and modeling limits.

Problem, claim and contribution

The opening asks what incentives a pricing rule creates rather than claiming evidence that named commercial providers have committed billing fraud.

Abstract and Introduction, p.1 · Main paper

Argument and section order

The paper distinguishes worst-case hardness of optimal plausible misreporting from the existence of a profitable heuristic, then presents the alternative pricing characterization.

Sections 3.2-4, pp.5-8 · Main paper

Figures, captions and visual explanation

Paired plots separate extra billed tokens from plausibility; the utility plot marks the break-even margin and spells out sampling and prompt conditions in its caption.

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

Tables, comparisons and evidence

The principal-agent definition specifies additive pricing and GPU-energy utility accounting, making the scope of the later characterization visible before the theorem.

Section 2, p.3; Theorem 4.3, pp.7-8 · Main paper

Appendix and reproducibility

Experimental details state model versions, prompt length/language filters, five generation seeds, 90 percent intervals and isolated GPU energy measurement.

Appendix B, p.15 · Main paper

Clarity and writing conventions

The discussion explicitly notes that character-based pricing does not prevent verbosity inflation and that single-provider interactions do not cover broader market dynamics.

Section 5, p.9 · Main paper

What to reuse—and what not to infer

Transferable lesson: Distinguish an incentive vulnerability from observed misconduct, and keep theorem restrictions and cost accounting attached to a proposed remedy.

Boundary: This is not an audit of proprietary billing behavior or a complete market model; the characterization concerns additive mechanisms and leaves other forms of misreporting outside its remedy.

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, 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 15. Remaining appendix proofs and details were not exhaustively read. No separately linked supplement inspected.

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.

• Published proceedings PDF, PMLR volume 306, retrieved 2026-10-07 · 25 pages