Writing study: Enhancing Generative Auto-bidding with Offline Reward Evaluation and Policy Search

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Zhiyu Mou, Yiqin Lv, Miao Xu, Qi Wang, Yixiu Mao, Jinghao Chen, Qichen Ye, Chao Li, Rongquan Bai, Chuan Yu, Jian Xu, Bo Zheng · 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: Learning & optimization algorithms (primary) · Theory & formal foundations.

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

32 physical pages

Before references / appendix

≈ 10.11 page-equivalents

Main figures / tables

5 / 5

Attached appendix figures / tables

8 / 8

Attached appendix span

≈ 15.89 page-equivalents

Before references / appendix: ≈ 31.59%References / intervening material: ≈ 18.75%Attached appendix: ≈ 49.66%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	≈ 0.97 pp.	3.03%
2 PRELIMINARIES	2	≈ 1.31 pp.	4.09%
3 METHOD	3	≈ 4.67 pp.	14.59%
4 EXPERIMENTS	8	≈ 2.32 pp.	7.25%
5 CONCLUSIONS	10	≈ 0.3 pp.	0.94%
REFERENCES	11	≈ 6 pp.	18.75%
A RELATED WORKS	17	≈ 0.69 pp.	2.16%
B AIGB METHOD DETAILS	17	≈ 1.61 pp.	5.03%
C THEORETICAL PROOFS	19	≈ 5.42 pp.	16.94%
D THEORETICAL PERFORMANCE RANGE TRADEOFF DISCUSSION	24	≈ 0.47 pp.	1.47%
E AIGB-PEARL ALGORITHM SUMMARY	25	≈ 1.55 pp.	4.84%
F ADDITIONAL EXPERIMENTS	26	≈ 5.37 pp.	16.78%
G EXTENDING AIGB-PEARL TO FIRST-PRICE AUCTIONS	32	≈ 0.23 pp.	0.72%
H EXTENDING AIGB-PEARL TO ONLINE SETTING DISCUSSION	32	≈ 0.25 pp.	0.78%
I LLM USAGE	32	≈ 0.09 pp.	0.28%
ACKNOWLEDGMENTS	32	≈ 0.22 pp.	0.69%

All detected figure and table locations

Main text

• Figure 1 · p. 5
• Figure 2 · p. 7
• Table 1 · p. 8
• Table 2 · p. 8
• Table 3 · p. 9
• Table 4 · p. 9
• Figure 3 · p. 10
• Table 5 · p. 10
• Figure 4 · p. 10
• Figure 5 · p. 10

Attached appendix

• Figure 6 · p. 19
• Figure 7 · p. 24
• Figure 8 · p. 26
• Table 6 · p. 27
• Table 7 · p. 28
• Figure 9 · p. 28
• Table 8 · p. 29
• Figure 10 · p. 29
• Table 10 · p. 29
• Figure 11 · p. 30
• Figure 12 · p. 30
• Table 9 · p. 30
• Figure 13 · p. 31
• Table 11 · p. 31
• Table 12 · p. 31
• Table 13 · p. 32

Argument and organization

1. Introduction

   Turns offline generative imitation's missing feedback into two explicit design challenges.

2. Preliminaries / Method

   Defines the auction MDP before deriving evaluator-bias bounds and their implementable regularizers.

3. Experiments

   Maps four research questions to overall performance, ablations, regularity and evaluator accuracy.

4. Appendices A–I

   Separates related work, proofs, implementation, expanded tests and scope extensions.

Problem, claim and contribution

The opening contrasts bootstrapped offline RL with conditional trajectory imitation, then identifies missing quality feedback and unprincipled extrapolation as distinct problems. Contributions answer those problems in the same order.

§1, pp.1–2 · Main paper

Argument and section order

Theorem 2 decomposes evaluator bias into sensitivity and imitation terms; Eq.8 converts that decomposition into two constraints before §3.2 presents the practical surrogate. This maintains a traceable theory-to-algorithm chain.

§§3.1–3.2, pp.4–7 · Main paper

Figures, captions and visual explanation

Figure 2 uses consistent colors for past/generated states and a shared-noise zoom-in to explain synchronous coupling. Its caption states the optimization role rather than merely naming modules.

Figure 2, p.7 · Main paper

Tables, comparisons and evidence

Four numbered research questions route readers to tests; Table 2 identifies 6,000 advertisers and 19 days, while Table 3 separately tests unseen advertisers. Table 1 explicitly warns that normalized absolute GMV has no standalone interpretation.

§4 and Tables 1–3, pp.8–9 · Main paper

Appendix and reproducibility

A contents list exposes proofs, executable algorithm structure, advertiser-feature prompts, seed-wise stability curves and hyperparameter sensitivity. Extensions to first-price and online settings explicitly require additional components outside the paper's scope.

pp.15–16; Appendices C, E–H, pp.19–32 · Main paper

Clarity and writing conventions

Sampling 8,000 trajectory pairs checks empirical regularity but does not certify a global Lipschitz bound. The text's guarantee language should be read alongside that finite-sample implementation distinction.

§4.4, p.10; Theorems 2–3 · Main paper

What to reuse—and what not to infer

Transferable lesson: Connect each proposed mechanism to a named failure, a mathematical or operational justification, and a dedicated falsification experiment; give readers an appendix map.

Boundary: Do not copy table deltas uncritically: Table 2 labels a 5.61-to-5.77 ROI change as +0.16%, inconsistent with a relative percentage. Algorithm 1 says maximize Eq.12 although the main text defines it as a loss. No standalone limitations section resolves these discrepancies.

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

PDF pp.1–10, abstract and §§1–5; references pp.11–14

Appendix / supplement

PDF pp.15–32, contents and Appendices A–I, including proofs C.1–C.6, Algorithm 1, supplementary experiments and LLM disclosure

Visually inspected PDF pages

7, 8

• official ICLR 2026 proceedings PDF · 32 pages