Writing study: VALUEFLOW: Toward Pluralistic and Steerable Value-based Alignment in Large Language Models

ICML overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICML 2026 oral writing corpus

Woojin Kim, Sieun Hyeon, Jusang Oh, Jaeyoung Do · 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: Datasets, benchmarks & measurement (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

63 physical pages

Before references / appendix

≈ 9.54 page-equivalents

Main figures / tables

8 / 4

Attached appendix figures / tables

34 / 11

Attached appendix span

≈ 49.91 page-equivalents

Before references / appendix: ≈ 15.14%References / intervening material: ≈ 5.63%Attached appendix: ≈ 79.22%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	≈ 1.72 pp.	2.73%
2. Related Work	3	≈ 0.21 pp.	0.33%
3. Preliminaries	3	≈ 0.87 pp.	1.38%
4. Hierarchical Value Embedding Space	4	≈ 1.23 pp.	1.95%
5. Value Evaluation Framework	5	≈ 0.98 pp.	1.56%
6. Experiments	6	≈ 2.32 pp.	3.68%
7. Analysis	8	≈ 1.02 pp.	1.62%
8. Conclusion	9	≈ 0.32 pp.	0.51%
Impact Statement	10	≈ 0.15 pp.	0.24%
Acknowledgments	10	≈ 0.35 pp.	0.56%
References	10	≈ 3.54 pp.	5.62%
A. Related Works	14	≈ 3 pp.	4.76%
B. Hierarchical Value Embedding Space Construction	17	≈ 14 pp.	22.22%
C. Value Intensity DB	31	≈ 4 pp.	6.35%
D. Steerability Experiment	35	≈ 22 pp.	34.92%
E. Demographic Alignment	57	≈ 2.27 pp.	3.6%
F. Framework Extension	59	≈ 1.03 pp.	1.63%
G. Human Evaluation	60	≈ 2.7 pp.	4.29%
H. Limitation	63	≈ 0.21 pp.	0.33%
I. LLM Usage	63	≈ 0.08 pp.	0.13%
J. License	63	≈ 0.63 pp.	1%

All detected figure and table locations

Main text

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

Attached appendix

• Table 5 · p. 17
• Figure 9 · p. 17
• Figure 10 · p. 18
• Figure 11 · p. 18
• Table 6 · p. 18
• Table 7 · p. 19
• Figure 12 · p. 22
• Figure 13 · p. 22
• Table 8 · p. 23
• Table 9 · p. 24
• Table 10 · p. 24
• Table 11 · p. 24
• Figure 14 · p. 28
• Figure 15 · p. 28
• Figure 16 · p. 29
• Figure 17 · p. 30
• Figure 18 · p. 34
• Figure 19 · p. 34
• Table 12 · p. 35
• Table 13 · p. 38
• Figure 20 · p. 39
• Figure 21 · p. 40
• Figure 22 · p. 41
• Figure 23 · p. 42
• Figure 24 · p. 43
• Figure 25 · p. 43
• Figure 26 · p. 44
• Figure 27 · p. 44
• Figure 28 · p. 45
• Figure 29 · p. 50
• Figure 30 · p. 50
• Figure 31 · p. 51
• Figure 32 · p. 52
• Figure 33 · p. 53
• Figure 34 · p. 54
• Figure 35 · p. 54
• Figure 36 · p. 54
• Figure 37 · p. 55
• Figure 38 · p. 58
• Figure 39 · p. 59
• Figure 40 · p. 60
• Table 14 · p. 60
• Table 15 · p. 61
• Figure 41 · p. 62
• Figure 42 · p. 62

Argument and organization

1. Problem and contribution

   Distinguishes value direction from graded intensity and connects extraction, controllable generation and evaluation through the same explicitly named infrastructure.

2. Argument sequence

   The paper first motivates rating instability, then constructs the embedding and ranking-based reference database before evaluating single- and multi-value control.

3. Evidence and verification

   The ranking model is fitted separately for each value, so normalized scores are explicitly value-conditional rather than one globally comparable intensity scale.

Problem, claim and contribution

Distinguishes value direction from graded intensity and connects extraction, controllable generation and evaluation through the same explicitly named infrastructure.

Introduction and Figure 1, p. 2 · Main paper

Argument and section order

The paper first motivates rating instability, then constructs the embedding and ranking-based reference database before evaluating single- and multi-value control.

Sections 3-7, pp. 3-9 · Main paper

Figures, captions and visual explanation

An end-to-end worked example carries profiles through steering and anchor comparison; signed, default-marked small multiples retain hard-to-steer and polarity-asymmetric cases.

Figure 1, p. 2; Figure 6, p. 8; both PNGs inspected · Main paper

Tables, comparisons and evidence

The ranking model is fitted separately for each value, so normalized scores are explicitly value-conditional rather than one globally comparable intensity scale.

Section 5.1, p. 5 · Main paper

Appendix and reproducibility

Consensus, human escalation and exclusion rules document label construction; the final pages distinguish imperfect dose-response control, limited personalization and heterogeneous upstream data terms.

Appendix B.2, p. 20; H-J, p. 63 · Main paper

Clarity and writing conventions

The text groups steering behavior into interpretable patterns but acknowledges ceiling effects and failed negative or composite targets instead of implying uniform controllability.

Sections 6.2-7.1, pp. 7-8; Appendix H, p. 63 · Main paper

What to reuse—and what not to infer

Transferable lesson: For a new construct-measurement resource, make calibration, human review and non-comparability boundaries as visible as the end-to-end use case.

Boundary: Demographic profiles are task-derived aggregates, not individual truths; shared numerical ranges do not license comparing absolute intensities across different values.

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-focused distributed manuscript scan, not exhaustive technical/proof validation. Sections 1-8, pp. 1-9: value intensity motivation, hierarchical representation, anchor/ranking evaluation, steering, demographic profiles and composition analysis.

Appendix / supplement

Writing-focused appendix scan; code, data and external assets not executed. Appendices A-J, pp. 14-63: hierarchy construction, training, intensity database, steering prompts/results, demographic evaluation, human evaluation, limitations and licenses.

Visually inspected PDF pages

2, 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 ICML 2026 proceedings PDF, PMLR volume 306; retrieved 2026-10-07. Public OpenReview decision notes were not independently reverified. · 63 pages