Strategic Algorithmic Monoculture: Experimental Evidence from Coordination Games

Gonzalo Ballestero; Hadi Hosseini; Samarth Khanna; Ran I. Shorrer

cs.AIApr 10, 2026

Strategic Algorithmic Monoculture: Experimental Evidence from Coordination Games

Authors: Gonzalo Ballestero, Hadi Hosseini, Samarth Khanna, Ran I. Shorrer

Organizations: Department of Economics, Pennsylvania State University · Department of Informatics and Intelligent Systems, Pennsylvania State University

Abstract

AI agents increasingly operate in multi-agent environments where outcomes depend on coordination. We distinguish primary algorithmic monoculture -- baseline action similarity -- from strategic algorithmic monoculture, whereby agents adjust similarity in response to incentives. We implement a simple experimental design that cleanly separates these forces, and deploy it on human and large language model (LLM) subjects. LLMs exhibit high levels of baseline similarity (primary monoculture) and, like humans, they regulate it in response to coordination incentives (strategic monoculture). While LLMs coordinate extremely well on similar actions, they lag behind humans in sustaining heterogeneity when divergence is rewarded.

Figures & tables

Figure 1: Agreement rate by subject type and treatment

Figure 2: Matrix form examples

Panel of Figure 3 — Figure 3: LLM textual instructions

	Picking		Coordination		Divergence
Subject type	Agree. rate	Answers	Agree. rate	Answers	Agree. rate	Answers
Human	0.14	31.75	0.31	11.33	0.04	19.50
Claude-3.5H	0.71	3.58	0.66	3.75	0.28	10.00
Claude-4S	0.89	1.58	0.79	2.08	0.26	10.83
DeepSeek-R1	0.65	3.00	0.75	2.00	0.11	19.00
DeepSeek-70B	0.37	7.00	0.74	2.75	0.21	12.00

Table 1: Average agreement rates and number of different answers by treatment

Figure 4: Pairwise agreement rates by treatment

Figure 5: Primary monoculture and strategic coordination

Figure 6: Distribution of screened sentences by treatment

Figure 7: LLM-judge classification of strategic reasoning by treatment

Figure 8: Agreement rate under different temperature values by treatment

	(1)	(2)
	Coordination	Divergence
Copy	$0.017$	$-0.020$
	(0.014)	(0.024)
Person	$0.009$	$0.012$
	(0.011)	(0.015)

Table 2: Effect of information treatment on self-agreement rates

Figure 9: Agreement rate by LLMs with personas

Appendix figures & tables38 assets

Supplementary material from the paper’s appendix.

Appendix

	Picking		Coordination		Divergence
Variable	Mean	SD	Mean	SD	Mean	SD
Age	47.47	14.03	47.09	11.91	45.33	12.63
Female	0.50	0.50	0.54	0.50	0.57	0.50
Education
Primary school	0.01	0.10	0.03	0.17	0.02	0.14
Secondary school	0.26	0.44	0.24	0.43	0.33	0.47

Appendix

Table A.1 : Summary statistics (human subjects)

Type	Provider	Label	Model	Reasoning	Default Temperature
Closed source	Anthropic	Claude-3.5H	Claude-3.5-Haiku	No	1
	Anthropic	Claude-4S	Claude-4-Sonnet	Yes	1
	Google	Gemini-2.0F	Gemini-2.0-Flash	No	1
	Google	Gemini-2.5P	Gemini-2.5-Pro	Yes	1
	OpenAI	GPT-4o	GPT-4o	No	1
	OpenAI	o4-mini	o4-mini	Yes	1

Appendix

Table A.2 : Features of LLM subjects

Picking

Coordination

Divergence

Human

LLMs

Human

LLMs

Human

LLMs

Topic

Response

Freq.

Response

Freq.

Response

Freq.

Response

Freq.

Response

Freq.

Response

Freq.

athlete

michael jordan

25.00

lebron james

67.60

michael jordan

28.30

lebron james

37.00

lebron james

7.40

lebron james

13.40

lebron james

19.20

lionel messi

13.90

lebron james

21.20

lionel messi

33.50

michael jordan

4.30

shohei ohtani

7.60

lionel messi

3.80

serena williams

7.10

tom brady

8.10

cristiano ronaldo

13.80

simone biles

4.30

simone biles

7.40

serena williams

3.80

cristiano ronaldo

4.00

travis kelce

7.10

michael jordan

13.00

usain bolt

4.30

serena williams

5.80

Appendix

Table A.3 : Most frequent responses by treatment, topic, and subject type

Model	Picking	Coordination	Divergence	Coordination (Copy)	Divergence (Copy)	Coordination (Person)	Divergence (Person)	Total
Claude-3.5H	0	3,080	2,199	2,562	1,949	1,948	1,534	13,272
Claude-4S	0	4,017	4,140	3,922	5,685	3,752	3,938	25,454
DeepSeek-14B	12,206	26,689	51,691	34,635	67,941	28,964	52,977	275,103
DeepSeek-70B	9,945	19,311	48,469	26,449	58,785	22,795	44,406	230,160
DeepSeek-8B	11,840	22,776	31,985	31,288	53,601	24,044	35,806	211,340
DeepSeek-R1	6,884	40,752	107,375	56,113	176,669	36,530	107,968	532,291

Appendix

Table A.4 : Number of reasoning sentences by model and treatment

Figure A.1 : Agreement rate by treatment and model capabilities

Figure A.2 : Effect of information on agreement rate by model capabilities and condition

Panel of Figure 18 — Figure A.3 : Textual analysis of strategic reasoning in the copy and person condition

Figure A.4 : LLM-judge classification of beliefs about co-player identity (by treatment and condition)

Figure A.5 : Agreement rate of all pairs of LLMs with personas by treatment

Figure A.6 : LLM-persona pairwise agreement rates by treatment

	Out of format		Excluded
Model	Total	Percent	Total	Percent
Claude-3.5H	0	0.00	0	0.00
Claude-4S	0	0.00	0	0.00
DeepSeek-R1	0	0.00	0	0.00
DeepSeek-70B	173	0.04	133	0.03
DeepSeek-8B	702	0.17	195	0.05

Appendix

Table B.1 : Out-of-format and excluded responses by model

Model	Professional Athlete	Car Manufacturer	City	Color	Disease	Flower	Food Item	Geometric Shape	Positive Number	Character Trait	Letter	Month
Claude-4S	99.82	98.84	99.73	98.54	97.90	98.14	99.48	96.86	100.00	99.48	26	12
DeepSeek-14B	57.47	87.24	98.22	84.21	94.04	82.18	98.44	83.70	100.00	92.45	26	12
DeepSeek-70B	96.78	91.74	99.02	90.84	96.65	84.47	97.66	76.60	99.98	97.00	26	12
DeepSeek-8B	86.24	96.19	99.49	82.42	93.79	77.39	97.51	70.19	100.00	75.88	25.12	12
DeepSeek-R1	98.90	97.54	99.98	99.46	99.40	99.62	99.68	98.02	100.00	99.54	26	12
Gemini-2.5P	99.96	99.10	99.84	99.14	99.68	99.94	99.72	98.28	100.00	99.98	26	12

Appendix

Table B.2 : Average number of valid answers generated by model and topic

Figure B.1 : Tab switching behavior (human subjects)

Figure B.2 : Fraction of topics with different modal response relative to the picking arm, by treatment and subject type

Figure B.3 : Agreement rate by treatment, topic, and subject type

Panel of Figure 27 — Figure B.4 : Textual analysis of strategic reasoning by treatment and model

Figure B.5 : Pairwise agreement rates for Section 6.1

Figure B.6 : Agreement rate by treatment, model, and temperature

Figure B.7 : Effect of information on agreement rate by model, treatment, and condition

Panel of Figure 31 — Figure B.8 : Textual analysis of strategic reasoning by treatment, condition, and model

Figure B.9 : LLM-judge classification of beliefs about co-player identity (by model, treatment, and condition)

Figure B.10 : Agreement rate by treatment, topic, and subject type (personas)

Panel of Figure 34 — Figure C.1 : Prompt for answer extraction using an LLM-judge

Panel of Figure 35 — Figure C.2 : Person condition textual instructions

Panel of Figure 36 — Figure C.3 : Copy condition textual instructions

Panel of Figure 37 — Figure C.4 : Persona textual instructions

Panel of Figure 38 — Figure C.5 : List generation textual instructions

Panel of Figure 39 — Figure C.6 : Selection textual instructions

Figure D.1 : Instructions: picking treatment arm

Figure D.2 : Example of question: picking treatment arm

Figure D.3 : Instructions: coordination treatment arm

Figure D.4 : Example of question: coordination treatment arm

Figure D.5 : Instructions: divergence treatment arm

Figure D.6 : Example of question: divergence treatment arm

Category	Words
Salience	common – familiar – focal point – frequent – most likely – obvious – ordinary – popular – salience – salient – Schelling – typical – universal – usual
Randomize	coin – lottery – random – randomly – randomized – randomize – randomizing – randomness – stochastic
Obscure	less common – less likely – less obvious – less popular – obscure – rare – special – uncommon – unlikely

Appendix

Table E.1 : Word groups used for analyzing model explanations in the experiment

Panel of Figure 47 — Figure E.1 : Prompt for textual analysis using an LLM-judge (Part 1: Strategic Reasoning)

Panel of Figure 48 — Figure E.2 : Prompt for textual analysis using an LLM-judge (Part 2: Other Player’s Identity)

Multi-Agent Coordination

Multi-agent coordination studies how autonomous agents align actions, share information, and divide work to achieve individual or collective goals.

See all 516 papers

Competition and Cooperation of LLM Agents in Games
Apr 1, 2026Jiayi Yao, Cong Chen, Baosen ZhangLLM Agents Multi-Agent LLM Systems
Department of Electrical and Computer Engineering, University of Washington, Seattle. · Thayer School of Engineering at Dartmouth College.
More Capable, Less Cooperative? When LLMs Fail At Zero-Cost Collaboration
Apr 9, 2026Advait Yadav, Sid Black, Oliver SourbutMulti-Agent Coordination Multi-Agent Collaboration
MATS · University of Illinois Urbana-Champaign · UK AI Security Institute +1
Benchmarking Open-Ended Multi-Agent Coordination in Language Agents
Jun 6, 2026Kale-ab Abebe Tessera, Andras Szecsenyi, Cameron Barker +7Multi-Agent Coordination LLM Agent Evaluation
University of Edinburgh · University of Oxford · University College London

Apr 1, 2026J/K move · Enter open · S save

Competition and Cooperation of LLM Agents in Games

Jiayi Yao, Cong Chen, Baosen Zhang

Department of Electrical and Computer Engineering, University of Washington, Seattle. · Thayer School of Engineering at Dartmouth College.

LLM Agents Multi-Agent LLM Systems Game-Playing Agents Multi-Agent Coordination Cooperative Game Theory

Open paper AskPDF arXiv

Large language model (LLM) agents are increasingly deployed in competitive multi-agent settings, raising fundamental questions about whether they converge to equilibria and how their strategic behavior can be characterized. In this paper, we study LLM agent interactions in two standard games: a network resource allocation game and a Cournot competition game. Rather than converging to Nash equilibria, we find that LLM agents tend to cooperate when given multi-round prompts and non-zero-sum context. Chain-of-thought analysis reveals that fairness reasoning is central to this behavior. We propose an analytical framework that captures the dynamics of LLM agent reasoning across rounds and explains these experimental findings.

Abstract

Multi-Agent Coordination

Explore similar work

Competition and Cooperation of LLM Agents in Games

More Capable, Less Cooperative? When LLMs Fail At Zero-Cost Collaboration

Benchmarking Open-Ended Multi-Agent Coordination in Language Agents