Observing Micromotives and Macrobehavior of Large Language Models

Chenghua Lin; Chung-Chi Chen; Tomas Goldsack; Xingwei Qu; Yuyang Cheng

REVIEW 3 major objections 5 minor 14 references

Observing Micromotives and Macrobehavior of Large Language Models

T0 review · 3 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read LLM advice segregates society even when bias scores are low.

desk verdict Clever application of Schelling to LLM-driven decisions, but the prompt installs the very same-type preference the paper claims to observe, so the central result is unsupported. read the letter →

arxiv 2412.10428 v1 pith:NXFNZJD3 submitted 2024-12-10 physics.soc-ph cs.AIcs.CL

Yuyang Cheng , Xingwei Qu , Tomas Goldsack , Chenghua Lin , Chung-Chi Chen This is my paper · ORCID

classification physics.soc-phcs.AIcs.CL

keywords largelanguagemodelsSchellingmodelsegregationmacrobehaviormicromotivesbiasmitigationagent-basedsimulationsocietalimpact

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that large language models, when asked to advise agents on whether to move based on neighbors' demographics, produce a society that becomes highly segregated—no matter which model is used, which demographic category is tested, or how well the model scores on bias benchmarks. The authors use Schelling's checkerboard model as an observation instrument and report a segregation shift of about 0.27 across all tested configurations. They further show that segregation rises sharply once more than 40 percent of people follow LLM suggestions. The point is that fixing individual-level bias in an LLM may not prevent undesirable society-level outcomes.

What carries the argument

The central mechanism is the Schelling checkerboard model with an LLM replacing the fixed tolerance threshold. In the experiment, the LLM is asked to rate 'move' and 'stay' responses on a satisfaction scale given the count of same-type and other-type neighbors; these ratings become the movement thresholds. The segregation shift metric compares the initial share of same-type neighbor edges to the final share, normalized by the theoretical maximum similarity. This setup turns the LLM's demographic preferences into a measurable micromotive and reveals the aggregate pattern.

What would settle it

Run the identical Schelling experiment with a neutral prompt that describes the neighborhood composition without stating any satisfaction rule, and ask the LLM only to choose move or stay; if the segregation shift drops to near zero, the claim that 'regardless of bias, a highly segregated society will emerge' fails for that configuration.

Watch

Extended reading notes

Core claim

The paper establishes that the micromotives of LLMs—operationalized as their satisfaction ratings for moving versus staying—lead to a consistent macrobehavior of segregation in a Schelling checkerboard. Five models (GPT-3.5, GPT-4o, Gemini-1.5, Claude-3.5, and Qwen2-72B) across five demographic categories (age, gender, politics, race, and religion) produce segregation shifts around 0.27 even though their bias-benchmark scores differ substantially. The segregation outcome is largely insensitive to the measured bias level, so the common assumption that removing bias from LLMs will make their societal influence benign needs reexamination. The paper also identifies a tipping point: when 40 percent of people use LLM suggestions, segregation starts to increase noticeably, and with full adoption it becomes highly segregated.

Load-bearing premise

The experiment's prompt explicitly tells the LLM that satisfaction is higher when more same-type neighbors are nearby; if this rule is not the model's natural preference but an instruction the authors implanted, then the observed segregation is a product of the prompt, not an intrinsic property of the LLMs.

Editorial extensions

If this is right

Even a model that scores near-perfect on bias benchmarks can still drive segregation when its advice steers individual moves.
Macro-level evaluation of LLM influence should accompany micro-level bias benchmarks, since the two can diverge.
As LLM adoption passes roughly 40 percent, societal segregation grows with each additional user following model advice.
Society-level outcomes, not just model-level debiasing, should be a target for responsible AI deployment.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The prompt's explicit rule that same-type neighbors raise satisfaction may itself be the cause of the observed behavior; testing a neutral prompt that does not state this rule would clarify whether LLMs naturally hold such a micromotive.
If users selectively follow LLM advice only when it matches their own biases, the tipping point could occur at a lower or higher adoption rate than 40 percent, which assumes uniform compliance.
Dynamic trust in LLMs, where users learn to discount advice over time, could dampen or reverse the segregation trend in longer-horizon simulations.
The same framework could be applied to other collective-action problems, such as opinion polarization or resource allocation, by changing the satisfaction rule.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

The paper applies Schelling's segregation model to LLM-generated move/stay decisions and reports that five models across five demographic categories all produce a SegShift around 0.27, with a tipping point when 40% of the population follows LLM advice. The specific setup is new, and the consistency across models is a striking empirical pattern. If the result were genuinely about how LLMs behave outside the prompt, it would be an important challenge to the assumption that removing individual bias is sufficient.

Here is the problem: the prompt tells the LLM exactly what to prefer. Appendix A, Rule 2 states that "Satisfaction is higher if many individuals of the same {agent_type} are nearby." The LLM is asked to rate move/stay responses on that basis. So the micromotive is not observed; it is manufactured. Any instruction-following model will exhibit the same-type preference, and Schelling's model guarantees segregation from that preference. The paper never runs a neutral prompt, an inverted prompt, or a random-threshold baseline. That makes the central claim unsupported as stated.

Two further issues. The conversion from satisfaction ratings to movement thresholds is never specified. Without that, we cannot tell whether the ratings are doing any work or whether the threshold sits near a Schelling tipping point, where noisy thresholds would produce similar SegShift values regardless of model. Also, no error bars or variance across the ten trials are reported, so we cannot assess whether the 0.27 consistency is meaningful or tight by construction.

The paper's limitation section acknowledges simplifications of human behavior but does not address the prompt-level confound. That is a notable omission.

What is genuinely valuable is the idea of using LLMs to parameterize agent-based models, and the 40% threshold simulation is a nice demonstration of how adoption rates can matter. But the paper needs at least a control condition without the explicit same-type preference before making claims about LLM micromotives. As it stands, it tells us more about prompt compliance than about intrinsic LLM bias.

I would not cite this for the claim, but I might bring it to a reading group to discuss prompt confounds in LLM-driven simulation. It deserves a serious referee because the idea is worth engaging with, but the current version does not support its headline. My recommendation: send it to review, expect heavy revision, and require a neutral-prompt baseline before acceptance.

Referee Report

3 major / 5 minor

Summary. The paper extends Schelling's segregation model by using LLM satisfaction ratings as the movement rule. For five demographic categories and five LLMs, the authors prompt the model to rate "move" and "stay" responses, use the average ratings as a tolerance threshold, and simulate a 20x20 Schelling grid. They report a consistent segregation shift of approximately 0.27 across all models and categories, and a tipping point beyond 40% LLM adoption. The abstract claims that, regardless of the LLM's bias score, a highly segregated society emerges as more people follow LLM suggestions.

Significance. The question of whether LLM suggestions could shape societal macrobehavior is timely, and the idea of embedding LLMs in a Schelling-style simulation is creative. If the central claim were supported, the paper would connect LLM bias mitigation debates to systemic social outcomes. However, the current operationalization explicitly instructs the LLM to prefer same-type neighbors, so the measured segregation is an engineering consequence of the prompt rather than an observed property of LLM micromotives. The cross-model, cross-category consistency reported in Table 1 is descriptively interesting, but it cannot be interpreted as evidence about intrinsic LLM biases without a neutral-prompt control. The paper also provides no code or data release, limiting reproducibility.

major comments (3)

[Appendix A (Table 3), Rule 2] The prompt explicitly states: "Satisfaction is higher if many individuals of the same {agent_type} are nearby; conversely, it is lower if many individuals of {opp_agent_type} are present." This rule directly installs the in-group preference that is known to drive Schelling segregation. The abstract's central claim that "regardless of the level of bias in LLMs, a highly segregated society will emerge" is therefore a consequence of the prompt design, not an observed property of LLM micromotives. The manuscript never tests a prompt without Rule 2 or with an inverted rule, so it cannot distinguish an intrinsic LLM preference from an experimenter-imposed one. This confound invalidates the answer to RQ1 and the paper's headline conclusion.
[Section 3.3] The text states that average satisfaction scores "serve as moving thresholds in the Schelling model," but the exact mapping from the two ratings (yes_rating and no_rating) to a single tolerance threshold t in [0,1] is never specified. Without this formula, the simulation is not reproducible, and the reported SegShift values cannot be checked for sensitivity to the choice of conversion. The authors should provide the exact functional form and a sensitivity analysis over reasonable monotonic transformations of the ratings.
[Section 4 (Figure 1 and Table 2)] The 40% tipping-point claim rests on a partial-adherence simulation whose design is not described: how are the LLM-following agents selected, whether they are re-selected each iteration, and whether "independent decisions" are simply random moves with no preference. No error bars or multiple-seed statistics are reported for Figure 1, so the "critical juncture" narrative is not supported. In addition, the bias-independence claim compares only two models (GPT-4o and GPT-3.5) in Table 2; with n=2, no statistical conclusion about independence from bias scores is possible, especially given the range of SegShift values for other models (e.g., Gemini's 0.19 to 0.36 in Table 1).

minor comments (5)

[Section 1] In the research questions list, "(RQ3)How" should be "(RQ3) How".
[Section 3.2] The phrase "to accommodation for random initialisation" should be "to accommodate for random initialisation".
[Table 1] The column header "SegShif" is a typo and should read "SegShift".
[Section 3.3] The initial grid is said to have Seginit approximately 0.511 and to be "standardized," but the exact initial configuration (or the algorithm used to generate it) is not described, which prevents replication.
[General] The manuscript does not mention whether code or data will be made available; for a simulation study, a public release would substantially aid reproducibility and trust in the results.

Circularity Check

2 steps flagged · score 8.0 of 10

The headline segregation result is manufactured by the prompt: Rule 2 explicitly instructs LLMs to be more satisfied with same-type neighbors, so the segregation outcome is a definitional consequence of the prompt plus Schelling's known dynamics rather than an observed property of LLM micromotives.

self definitional [Appendix A, Table 3 (Prompt), Rule 2; also Section 3.1]
"To ensure that the LLM bases its decisions strictly within the demographic context, we have implemented specific rules in the prompt. ... 2. Satisfaction is higher if many individuals of the same {agent_type} are nearby; conversely, it is lower if many individuals of {opp_agent_type} are present."

This rule defines the target micromotive before any model output is gathered. The LLM is not asked whether it prefers same-type neighbors; it is instructed that satisfaction is higher under exactly that condition. Those ratings are then converted into Schelling movement thresholds, and the paper's own preliminary states that thresholds above 0.33 produce segregation. The observed SegShift is therefore entailed by the prompt's utility definition plus Schelling's known dynamics. A neutral prompt or an inverted-preference control would be required to attribute the segregation to the LLMs' natural micromotives; the paper includes neither, so the central claim reduces to the experimental instruction.
fitted input called prediction [Section 3.3 Setup and Section 4 Experimental Results (Table 1)]
"These average scores serve as moving thresholds in the Schelling model. ... Regardless of the model used or the category examined, the segregation ratio increased by approximately 27%."

The thresholds are not independently measured preferences; they are the LLM's compliance scores under Rule 2, which already stipulates in-group preference. Computing a segregation shift from thresholds that encode exactly the Schelling preference is not a prediction of LLM-driven social outcomes; it is the known Schelling segregation result relabeled as an LLM discovery. The constancy across models and demographic categories is expected because the same forced preference rule is used in every condition.

full rationale

The central derivation is not self-contained in the relevant sense: the paper's conclusion that LLM micromotives produce segregation is loaded into the prompt via Rule 2, which tells each model that satisfaction increases with same-type neighbors. Because the LLM ratings are converted directly into Schelling movement thresholds, the observed SegShift is the known Schelling outcome under the preference the authors themselves instructed; it is not an independent measurement of LLM bias. The paper does not test a neutral or inverted prompt, so there is no evidence that LLMs would exhibit this micromotive absent instruction. No load-bearing self-citation chain is present: the LangBiTe bias scores are external and the preliminary Schelling discussion is standard. The circularity is definitional: Rule 2 installs the preference, the ratings inherit it, and the macrobehavior follows by construction.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central simulation rests on the prompt's explicit preference rule, which installs the micromotive being studied. The conversion of LLM ratings to movement thresholds is unspecified, and several numerical choices (initial segregation, grid size, averaging) are made by hand. No new external entities are introduced.

free parameters (4)

Initial segregation Seginit = 0.511
Chosen by hand in Section 3.3 to 'highlight differences in segregation outcomes across models'. The magnitude of the reported SegShift depends on this starting point.
Grid size and agent count = 20x20, 360 agents
Arbitrary simulation choices that set MaxSim (0.9) and constrain possible segregation levels.
Number of repeated trials = 10
Used to average LLM ratings and simulations; no variance or confidence intervals are reported.
Final-state averaging window = Last 10 iterations
Chosen for stability (Section 3.2) without robustness analysis.

assumptions (4)

domain assumption Schelling's movement rule: agents move when the local same-type fraction is below a threshold.
The paper adopts Schelling's checkerboard dynamics (Section 2), including binary groups and local neighbor evaluation.
domain assumption LLM ratings of 'move' and 'stay' can be mapped to a per-agent tolerance threshold.
Section 3.1 states that average scores 'serve as moving thresholds', but the exact functional mapping is never defined.
ad hoc to paper Satisfaction is higher when more same-type neighbors are present (prompt Rule 2).
This rule is inserted by the authors into the prompt (Appendix A), directly engineering the in-group preference that produces the reported segregation.
domain assumption Users follow LLM suggestions with uniform weight and static behavior.
Acknowledged in the Limitations section; the mixed-population simulation assumes non-LLM users make random choices.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Observing Micromotives and Macrobehavior of Large Language Models." pith.science (2026). https://pith.science/paper/NXFNZJD3

@misc{pith2026241210428,
  author       = {Pith},
  title        = {Pith review of: Observing Micromotives and Macrobehavior of Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NXFNZJD3}},
  note         = {Machine review of arXiv:2412.10428}
}

read the original abstract

Thomas C. Schelling, awarded the 2005 Nobel Memorial Prize in Economic Sciences, pointed out that ``individuals decisions (micromotives), while often personal and localized, can lead to societal outcomes (macrobehavior) that are far more complex and different from what the individuals intended.'' The current research related to large language models' (LLMs') micromotives, such as preferences or biases, assumes that users will make more appropriate decisions once LLMs are devoid of preferences or biases. Consequently, a series of studies has focused on removing bias from LLMs. In the NLP community, while there are many discussions on LLMs' micromotives, previous studies have seldom conducted a systematic examination of how LLMs may influence society's macrobehavior. In this paper, we follow the design of Schelling's model of segregation to observe the relationship between the micromotives and macrobehavior of LLMs. Our results indicate that, regardless of the level of bias in LLMs, a highly segregated society will emerge as more people follow LLMs' suggestions. We hope our discussion will spark further consideration of the fundamental assumption regarding the mitigation of LLMs' micromotives and encourage a reevaluation of how LLMs may influence users and society.

Figures

Figures reproduced from arXiv: 2412.10428 by the authors.

**Figure 2.** Comparative analysis of recent approaches for [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** The first distribution represents the probability distribution of agents moving in the Schelling model, [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 4 canonical work pages

[1]

online" 'onlinestring :=

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...
[2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...
[3]

Anthropic. 2024. https://www.anthropic.com/news/claude-3-haiku Claude 3 haiku: Our fastest model yet

work page 2024
[4]

Yu, and Lichao Sun

Yue Huang, Qihui Zhang, Philip S. Yu, and Lichao Sun. 2023. https://api.semanticscholar.org/CorpusID:259202452 Trustgpt: A benchmark for trustworthy and responsible large language models . ArXiv, abs/2306.11507

arXiv 2023
[5]

Yizhi Li, Ge Zhang, Bohao Yang, Chenghua Lin, Anton Ragni, Shi Wang, and Jie Fu. 2022. HERB : Measuring hierarchical regional bias in pre-trained language models. In Findings of the Association for Computational Linguistics: AACL-IJCNLP 2022, pages 334--346

work page 2022
[6]

Sergio Morales, Robert Claris'o, and Jordi Cabot. 2024. https://api.semanticscholar.org/CorpusID:269449200 Langbite: A platform for testing bias in large language models . ArXiv, abs/2404.18558

work page Pith review arXiv 2024
[7]

OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mohammad Bavarian, Jeff Belgum, Irwan Bello, Jake Berdine, Gabriel Bernadett-Shapiro, Christopher Berner...

arXiv 2024
[8]

Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. 2022. https://arxiv.org/abs/2203.02155 Training language models to f...

arXiv 2022

Show all 14 references

[9]

Schelling

Thomas C. Schelling. 1969. http://www.jstor.org/stable/1823701 Models of segregation . The American Economic Review, 59(2):488--493

1969
[10]

Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, Soroosh Mariooryad, Yifan Ding, Xinyang Geng, Fred Alcober, Roy Frostig, Mark Omernick, Lexi Walker, Cosmin Paduraru, Christina Sorokin, A...

2024 arXiv
[11]

Tianyang Xu, Shujin Wu, Shizhe Diao, Xiaoze Liu, Xingyao Wang, Yangyi Chen, and Jing Gao. 2024. https://arxiv.org/abs/2405.20974 Sayself: Teaching llms to express confidence with self-reflective rationales . Preprint, arXiv:2405.20974

2024 arXiv
[12]

An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jianxin Yang, Jin Xu, Jingren Zhou, Jinze...

2024 arXiv
[13]

Ge Zhang, Yizhi Li, Yaoyao Wu, Linyuan Zhang, Chenghua Lin, Jiayi Geng, Shi Wang, and Jie Fu. 2023 a . Corgi-pm: A chinese corpus for gender bias probing and mitigation. arXiv preprint arXiv:2301.00395

2023 arXiv
[14]

Zhexin Zhang, Leqi Lei, Lindong Wu, Rui Sun, Yongkang Huang, Chong Long, Xiao Liu, Xuanyu Lei, Jie Tang, and Minlie Huang. 2023 b . Safetybench: Evaluating the safety of large language models with multiple choice questions. arXiv preprint arXiv:2309.07045

2023 arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [1]

online" 'onlinestring :=

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

[2] [2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

[3] [3]

Anthropic. 2024. https://www.anthropic.com/news/claude-3-haiku Claude 3 haiku: Our fastest model yet

work page 2024

[4] [4]

Yu, and Lichao Sun

Yue Huang, Qihui Zhang, Philip S. Yu, and Lichao Sun. 2023. https://api.semanticscholar.org/CorpusID:259202452 Trustgpt: A benchmark for trustworthy and responsible large language models . ArXiv, abs/2306.11507

arXiv 2023

[5] [5]

Yizhi Li, Ge Zhang, Bohao Yang, Chenghua Lin, Anton Ragni, Shi Wang, and Jie Fu. 2022. HERB : Measuring hierarchical regional bias in pre-trained language models. In Findings of the Association for Computational Linguistics: AACL-IJCNLP 2022, pages 334--346

work page 2022

[6] [6]

Sergio Morales, Robert Claris'o, and Jordi Cabot. 2024. https://api.semanticscholar.org/CorpusID:269449200 Langbite: A platform for testing bias in large language models . ArXiv, abs/2404.18558

work page Pith review arXiv 2024

[7] [7]

OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mohammad Bavarian, Jeff Belgum, Irwan Bello, Jake Berdine, Gabriel Bernadett-Shapiro, Christopher Berner...

arXiv 2024

[8] [8]

Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. 2022. https://arxiv.org/abs/2203.02155 Training language models to f...

arXiv 2022