Writing study: Large Language Models Develop Novel Social Biases Through Adaptive Exploration

ICML overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICML 2026 oral writing corpus

Addison J. Wu, Ryan Liu, Xuechunzi Bai, Thomas L. Griffiths · 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: Empirical analysis, diagnosis & auditing (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

41 physical pages

Before references / appendix

≈ 9.2 page-equivalents

Main figures / tables

6 / 2

Attached appendix figures / tables

6 / 9

Attached appendix span

≈ 26.91 page-equivalents

Before references / appendix: ≈ 22.44%References / intervening material: ≈ 11.93%Attached appendix: ≈ 65.63%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	≈ 1.47 pp.	3.59%
2. Related Work	2	≈ 0.49 pp.	1.2%
3. Methodology	3	≈ 1.46 pp.	3.56%
4. Do LLMs Segregate Equal Groups?	4	≈ 1.04 pp.	2.54%
5. Interventions to Identify Factors Behind Stratification	5	≈ 3.06 pp.	7.46%
6. Discussion	8	≈ 1.15 pp.	2.8%
Acknowledgments	10	≈ 0.09 pp.	0.22%
Impact Statement	10	≈ 0.07 pp.	0.17%
References	10	≈ 4.88 pp.	11.9%
A. Prompting Details	15	≈ 9.11 pp.	22.22%
B. Metrics: Analyses and Interpretations	24	≈ 3.35 pp.	8.17%
C. Human Participants	27	≈ 0.53 pp.	1.29%
D. Rank-Ordered Allocation Matrices (Hiring Experiment)	28	≈ 6.14 pp.	14.98%
E. Sampling Baslines	34	≈ 0.38 pp.	0.93%
F. Additional experimental scenarios	34	≈ 0.72 pp.	1.76%
G. Stochasticity in emergent biases is driven by outcome exploitation rather than other variation	35	≈ 1.43 pp.	3.49%
H. Prior Biased Associations Experiment	36	≈ 1.33 pp.	3.24%
I. LLMs’ exploration does not uncover differences between success rates	38	≈ 1 pp.	2.44%
J. Unlearning a biased belief requires a sequence of exceedingly unlikely outcomes	39	≈ 1 pp.	2.44%
K. Objective Demographic-Job Mapping Experiment	40	≈ 0.49 pp.	1.2%
L. Effects of agentic augmentations	40	≈ 0.19 pp.	0.46%
M. Experiment with real-life demographic labels	40	≈ 1.23 pp.	3%

All detected figure and table locations

Main text

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

Attached appendix

• Figure 7 · p. 20
• Figure 8 · p. 26
• Figure 9 · p. 26
• Figure 10 · p. 27
• Table 3 · p. 34
• Figure 11 · p. 35
• Table 4 · p. 35
• Table 5 · p. 36
• Table 6 · p. 36
• Table 7 · p. 37
• Table 8 · p. 38
• Table 9 · p. 38
• Table 10 · p. 39
• Figure 12 · p. 40
• Table 11 · p. 41

Argument and organization

1. 1 Introduction; 2 Related Work; 3 Methodology

   Separates inherited stereotypes from biases formed through sequential feedback, then defines the fictional hiring task and metrics.

2. 4 Equal-Group Results; 5 Interventions

   Establishes the phenomenon before testing input, environmental and explicit-incentive explanations.

3. 6 Discussion; Appendices A-M

   Interprets the fairness implications while documenting prompts, human baseline provenance and counterexamples to a universal fix.

Problem, claim and contribution

The opening argues that removing pretrained bias is insufficient and uses artificial groups with equal success probabilities to isolate a different failure mode.

Abstract, p.1; Section 3.1, p.3 · Main paper

Argument and section order

Interventions are grouped by model inputs, task structure and explicit incentives, allowing partial remedies to be compared without conflating their targets.

Section 5 roadmap, p.5 · Main paper

Figures, captions and visual explanation

The result plot combines distinct markers and separately labeled axes for two metrics; the intervention plot includes human and random reference bands alongside model conditions.

Figure 2, p.4; Figure 6, p.9 · Main paper

Tables, comparisons and evidence

The lower-success experiment defines bold entries by confidence-interval overlap with a random baseline and reports interventions that remain model-dependent.

Table 2 and Section 5.2, p.7 · Main paper

Appendix and reproducibility

Prompt scripts and the provenance/subsets of previously collected human participants are explicit; a later experiment demonstrates a success-rate cost when the synthetic environment has a true group-job mapping.

Appendices A.1, C and K, pp.15,27,40 · Main paper

Clarity and writing conventions

The discussion warns against blindly transferring diversity steers when underlying group success rates are unknown, separating a diagnostic intervention from a general deployment prescription.

Section 6, p.9 · Main paper

What to reuse—and what not to infer

Transferable lesson: Use a controlled environment to isolate a failure mode, then test and disclose conditions under which the apparent remedy fails.

Boundary: The synthetic allocation game is not a field hiring study; these observations do not validate demographic decision rules or establish behavior for every deployed model.

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

Visually inspected PDF pages

4, 9

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 · 41 pages