Writing study: AdAEM: An Adaptively and Automated Extensible Measurement of LLMs' Value Difference

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Jing Yao, Shitong Duan, Xiaoyuan Yi, Dongkuan Xu, Peng Zhang, Tun Lu, Ning Gu, Zhicheng Dou, Xing Xie · 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: Datasets, benchmarks & measurement resources (primary) · Learning & optimization algorithms.

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

49 physical pages

Before references / appendix

≈ 10.62 page-equivalents

Main figures / tables

9 / 1

Attached appendix figures / tables

12 / 17

Attached appendix span

≈ 27.89 page-equivalents

Before references / appendix: ≈ 21.67%References / intervening material: ≈ 21.41%Attached appendix: ≈ 56.92%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	≈ 1.04 pp.	2.12%
2 RELATED WORKS	2	≈ 1.04 pp.	2.12%
3 METHODOLOGY	3	≈ 2.69 pp.	5.49%
4 ADAEM ANALYSIS	6	≈ 3.15 pp.	6.43%
5 VALUE EVALUATION WITH ADAEM	9	≈ 0.9 pp.	1.84%
6 CONCLUSION AND FUTURE WORK	10	≈ 0.15 pp.	0.31%
ETHICS STATEMENT	10	≈ 0.81 pp.	1.65%
REPRODUCIBILITY STATEMENT	11	≈ 0.21 pp.	0.43%
REFERENCES	11	≈ 10.48 pp.	21.39%
A DISCUSSION ON LLMS’ VALUE	22	≈ 0.63 pp.	1.29%
B DETAILS OF DATASET CONSTRUCTION	22	≈ 2.37 pp.	4.84%
C EXPERIMENTAL DETAILS	25	≈ 10.47 pp.	21.37%
D DETAILED DERIVATION	35	≈ 2.04 pp.	4.16%
E ADDITIONAL RESULTS	37	≈ 4.49 pp.	9.16%
F LLM USAGE	42	≈ 0.15 pp.	0.31%
G ADDITIONAL DISCUSSION	42	≈ 0.95 pp.	1.94%
H LIMITATIONS	43	≈ 0.9 pp.	1.84%
I DISCUSSION ON THE MATHEMATICAL APPROXIMATION OF ADAEM	44	≈ 1.31 pp.	2.67%
J ADAEM FRAMEWORK ON MORAL FOUNDATION THEORY	45	≈ 1.96 pp.	4%
K ANALYSIS ON HYPERPARAMETER ROBUSTNESS	47	≈ 2.62 pp.	5.35%

All detected figure and table locations

Main text

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

Attached appendix

• Figure 10 · p. 23
• Figure 11 · p. 24
• Table 2 · p. 24
• Figure 12 · p. 24
• Table 3 · p. 25
• Table 4 · p. 26
• Table 5 · p. 27
• Table 6 · p. 30
• Table 7 · p. 32
• Table 8 · p. 32
• Table 9 · p. 33
• Table 10 · p. 34
• Table 11 · p. 35
• Table 12 · p. 35
• Figure 13 · p. 38
• Figure 14 · p. 39
• Figure 15 · p. 39
• Figure 16 · p. 40
• Figure 17 · p. 40
• Table 13 · p. 41
• Figure 18 · p. 41
• Table 14 · p. 45
• Figure 19 · p. 46
• Table 15 · p. 46
• Table 16 · p. 47
• Table 17 · p. 47
• Figure 20 · p. 48
• Figure 21 · p. 48
• Table 18 · p. 49

Argument and organization

1. Informativeness rather than generic agreement

   Uses paired questions to motivate a dynamic evaluation objective.

2. Objective, alternating refinement and exploration

   Separates scoring from the search policy and final value aggregation.

3. Quality, validity, reliability and extensibility

   Tests different properties of the measurement instrument before model comparisons.

Problem, claim and contribution

A generic question and a contextualized variant make the motivation concrete before mathematical notation is introduced.

Figure 1 and Section 1, p. 2 · Main paper

Argument and section order

Separating instrument validity and reliability from downstream model rankings is a useful benchmark-paper structure.

Sections 4–5, pp. 6–10 · Main paper

Figures, captions and visual explanation

The central diagram shows the question-refinement loop and resulting evaluation set in one view, but the tiny objective competes with example responses for attention.

Figure 2, p. 3 · Main paper

Tables, comparisons and evidence

Figure 9 prints scores on every bar, but its ten panels use different truncated horizontal scales; bar lengths cannot safely be compared between value dimensions.

Figure 9, p. 10 · Main paper

Appendix and reproducibility

The appendix provides model versions, prompts, human-annotation compensation, a notation table and ablations. These are more useful for replication than the abbreviated main algorithm alone.

Appendices B–C and K, pp. 22–35, 47–49 · Main paper

Clarity and writing conventions

The conceptual objective and practical reward need sharper separation: Appendix C.3 uses heuristic scores and a negative disentanglement term, whereas the main objective presents a positive divergence term.

Equation 1, p. 4; Appendix C.3, pp. 29–30; Appendix I, pp. 44–45 · Main paper

What to reuse—and what not to infer

Transferable lesson: For a new evaluation instrument, organize evidence around what it measures, whether it responds to controlled interventions, and how stable its results are—not only whether it separates models.

Boundary: Greater disagreement is not proof of more accurate measurement of intrinsic values. Model nationality is not a causal explanation. The stated JS-versus-total-variation bound and zero-intersection denominators require independent checking; do not inherit the convergence guarantee for the heuristic implementation.

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. 1–11, Sections 1–6, ethics and reproducibility statements

Appendix / supplement

PDF pp. 22–49, Appendices A–K including all prompt listings and derivations

Visually inspected PDF pages

3, 10

• Official ICLR 2026 proceedings conference PDF; PDF creation metadata: Mon Mar 2 05:15:30 2026 EST · 49 pages