Writing study: Beyond Prompt-Induced Lies: Investigating LLM Deception on Benign Prompts

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Zhaomin Wu, Mingzhe Du, See-Kiong Ng, Bingsheng He · 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: Empirical analysis, diagnosis & auditing (primary) · Datasets, benchmarks & measurement resources.

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

29 physical pages

Before references / appendix

≈ 10.49 page-equivalents

Main figures / tables

7 / 0

Attached appendix figures / tables

16 / 5

Attached appendix span

≈ 14.87 page-equivalents

Before references / appendix: ≈ 36.17%References / intervening material: ≈ 12.55%Attached appendix: ≈ 51.28%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	≈ 1.24 pp.	4.28%
2 RELATED WORK	2	≈ 0.69 pp.	2.38%
3 DEFINITION AND METRICS OF DECEPTION	3	≈ 2.23 pp.	7.69%
4 EVALUATION FRAMEWORK	5	≈ 2.13 pp.	7.34%
5 EXPERIMENT	7	≈ 2.96 pp.	10.21%
6 CONCLUSION	10	≈ 0.29 pp.	1%
REPRODUCIBILITY STATEMENT	11	≈ 0.39 pp.	1.34%
REFERENCES	11	≈ 3.64 pp.	12.55%
Appendix	15	≈ 0.57 pp.	1.97%
A EXPERIMENT SETUP DETAILS	15	≈ 0.74 pp.	2.55%
B ADDITIONAL MODEL-WISE ANALYSIS	16	≈ 1.1 pp.	3.79%
C ADDITIONAL OVERALL ANALYSIS	17	≈ 3.14 pp.	10.83%
D ABLATION STUDIES	20	≈ 1.8 pp.	6.21%
E EXAMPLE OF PROMPTS OF CSQ FRAMEWORK	22	≈ 1.11 pp.	3.83%
F CASE STUDY: REASONING BEHIND DECEPTION	23	≈ 3.25 pp.	11.21%
G GENERALIZATION TO OTHER DOMAINS	26	≈ 0.68 pp.	2.34%
H COMPARISON TO PRIOR BENCHMARKS AND FRAMEWORK ADVANTAGES	27	≈ 1.59 pp.	5.48%
I BROADER IMPACT	29	≈ 0.55 pp.	1.9%
J LARGE LANGUAGE MODEL USAGE	29	≈ 0.35 pp.	1.21%

All detected figure and table locations

Main text

• Figure 1 · p. 2
• Figure 2 · p. 6
• Figure 3 · p. 7
• Figure 4 · p. 9
• Figure 5 · p. 9
• Figure 6 · p. 10
• Figure 7 · p. 10

Attached appendix

• Table 1 · p. 16
• Figure 8 · p. 17
• Figure 9 · p. 18
• Table 2 · p. 18
• Figure 10 · p. 19
• Figure 11 · p. 19
• Figure 12 · p. 20
• Figure 13 · p. 21
• Figure 14 · p. 21
• Figure 15 · p. 22
• Figure 16 · p. 22
• Figure 17 · p. 23
• Figure 18 · p. 24
• Figure 19 · p. 24
• Table 3 · p. 25
• Figure 20 · p. 25
• Table 4 · p. 25
• Figure 21 · p. 26
• Figure 22 · p. 26
• Figure 23 · p. 27
• Table 5 · p. 28

Argument and organization

1. Definitions and metrics before benchmark construction

   Distinguishes four conceptual response regimes before operationalizing performance asymmetry and hard/easy response inconsistency.

2. Controlled task, model comparison and sensitivity checks

   Introduces linked/broken contact chains, then follows individual models, aggregate trends and incentivizing prompts.

Problem, claim and contribution

The question is compelling because ordinary benign requests, rather than explicit lying instructions, define the stress test. Synthetic contacts reduce dependence on memorized factual knowledge.

Sections 1 and 4.1, pages 1–2 and 6. · Main paper

Argument and section order

Definitions precede measurement and task design, making the inferential chain visible. However, naming the two output-based scores intention and behavior risks making an operational definition sound like independently observed psychology.

Sections 3–4, pages 3–8. · Main paper

Figures, captions and visual explanation

Figure 2 places complete and broken chains side by side with highlighted query endpoints and colored outcome boxes; readers can inspect the perturbation without decoding equations first.

Figure 2, page 6. · Main paper

Tables, comparisons and evidence

Figure 6 combines a labeled model scatterplot with release-date trends; the four quadrant labels are interpretive, not separately validated ground-truth classes. Figure 7 usefully tests an additional prompting condition.

Figures 6–7, page 10. · Main paper

Appendix and reproducibility

The appendix records API model versions, temperature and difficulty-ratio checks, full prompt examples, character-count tables and embedding projections. These expose assumptions hidden by an aggregate score.

Appendices A, D–F, pages 15–16 and 20–27. · Main paper

Clarity and writing conventions

Appendix G explicitly limits cross-domain generalization, while the impact discussion labels some proposed training causes as suspicion. That qualification should also accompany the stronger causal language elsewhere.

Appendices G and I, pages 26–27 and 29. · Main paper

What to reuse—and what not to infer

Transferable lesson: Pair every abstract behavioral construct with an inspectable controlled task, complete prompts and a sensitivity analysis that shows exactly what the metric measures.

Boundary: A harder-question error followed by an easier-question success, or a directional answer asymmetry, does not by itself establish deceptive intent or a hidden correct belief. Preserve this construct-validity distinction; visual correlation and selected reasoning traces do not resolve it.

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

Pages 1–10, Sections 1–6; references on pages 11–14.

Appendix / supplement

Pages 15–29, Appendices A–J: settings, model-wise and aggregate results, ablations, prompts, reasoning cases, embeddings, scope, benchmark comparison, impact and LLM use.

Visually inspected PDF pages

6, 10

• ICLR 2026 official proceedings PDF, retrieved 2026-09-19 · 29 pages