Strategic Algorithmic Monoculture:Experimental Evidence from Coordination Games

1. Problem Statement

As AI agents (specifically Large Language Models or LLMs) become increasingly embedded in economic systems, they operate in multi-agent environments where outcomes depend on coordination. The paper addresses a critical question: How do LLMs coordinate compared to humans, and what are the risks of "algorithmic monoculture"?

The authors distinguish between two types of monoculture:

• Primary Algorithmic Monoculture: Baseline similarity in actions due to shared training data, architecture, or design, occurring even without strategic incentives.
• Strategic Algorithmic Monoculture: The adjustment of similarity in response to incentives (e.g., coordinating on a focal point or diverging to avoid congestion).

The core problem is that while LLMs may excel at converging on the same action (beneficial for some tasks), their inherent homogeneity and difficulty in randomizing may make them poor at "coordinated divergence" (diverging when required), potentially leading to systemic fragility in environments where diversity is socially optimal (e.g., hiring, investment screening).

2. Methodology

The authors designed a controlled experimental framework adapting the classic coordination game literature (specifically Mehta et al., 1994) to compare human and LLM subjects.

Experimental Design

• Subjects: 301 human subjects (via Prolific) and 16 diverse LLMs (including OpenAI, Anthropic, Google, DeepSeek, and open-source models like Llama and Qwen).
• Tasks: Subjects answered 12 open-ended questions (e.g., "Name a city," "Name a color") across three treatment arms:
  1. Picking (Baseline): Incentivized only to provide a valid answer. Measures Primary Monoculture.
  2. Coordination: Incentivized to match the answer of another agent. Measures Strategic Monoculture (convergence).
  3. Divergence: Incentivized to provide an answer different from another agent. Measures Strategic Divergence.
• Metric: The Agreement Rate (probability that two independent agents provide the same answer).
• LLM Configuration:
  • Models were queried 50 times per question per treatment.
  • Experiments included variations: Temperature (modulating stochasticity), Information (telling LLMs they face a "copy" vs. a "person"), and Personas (assigning human demographic traits to LLMs).
• Textual Analysis: The authors analyzed Chain-of-Thought (CoT) reasoning using semantic similarity and an "LLM-as-a-Judge" approach to categorize strategic logic (e.g., salience, obscurity, randomization).

3. Key Contributions

1. Conceptual Taxonomy: The paper formally distinguishes between primary and strategic algorithmic monoculture, providing a framework to separate baseline homogeneity from strategic adaptation.
2. Experimental Innovation: It introduces the Coordinated Divergence game to AI research, a necessary counterpart to coordination games to test if agents can successfully avoid each other.
3. Mechanism Identification: By leveraging LLMs' ability to output reasoning, the authors link strategic choices directly to internal logic, overcoming the "black box" limitation of human subject studies.
4. Empirical Evidence of Asymmetry: The study provides robust evidence that LLMs exhibit a sharp asymmetry: they are superior to humans at coordination but inferior at divergence.

4. Key Results

A. Primary Monoculture (Baseline)

• Finding: LLMs exhibit significantly higher baseline agreement than humans.
• Data: In the Picking arm, LLM self-agreement was 58%, compared to 14% for humans.
• Implication: LLMs share a strong "focal point" bias even without incentives, likely due to shared training corpora.

B. Strategic Coordination (Convergence)

• Finding: LLMs outperform humans significantly when incentivized to match.
• Data: In the Coordination arm, LLM agreement rose to 72%, while humans reached 31%.
• Mechanism: LLMs effectively utilize secondary and Schelling salience (identifying unique focal points like "Paris" or "Eiffel Tower") to coordinate.

C. Strategic Divergence (The Critical Deficit)

• Finding: LLMs perform substantially worse than humans when incentivized to diverge.
• Data: In the Divergence arm, LLM agreement remained high at 27%, whereas humans dropped to 4%.
• Implication: LLMs struggle to sustain heterogeneity. Even when explicitly told to differ, they frequently converge on the same "obscure" or "random" choice, failing to break symmetry.

D. Mechanism Analysis

• Reasoning vs. Action: Textual analysis revealed that LLMs often articulate the correct strategy (e.g., "I should choose an obscure option"), yet fail to execute it effectively.
• Randomization: LLMs have a limited capacity for true randomization. When forced to generate a list and pick randomly, their divergence performance improved, suggesting the deficit is partly due to poor stochastic sampling.
• Temperature: Increasing temperature (stochasticity) improved divergence performance but hurt coordination performance, confirming a trade-off.
• Information: Telling LLMs they face a "copy" of themselves vs. a "person" had little effect on outcomes, though it altered their internal reasoning.
• Personas: Assigning human personas to LLMs shifted behavior slightly but did not close the gap in divergence performance.

5. Significance and Implications

• Systemic Risk: The findings suggest that as AI agents proliferate in markets (hiring, trading, admissions), they may create systemic fragility. While they can easily collude or coordinate on the same strategy (potentially leading to bubbles or collusion), they are ill-equipped to provide the diversity needed to avoid correlated errors or congestion.
• Policy & Design: Simply "turning up the heat" (increasing temperature) is not a silver bullet, as it degrades coordination performance. Designers of multi-agent systems must explicitly engineer mechanisms for controlled heterogeneity rather than relying on LLMs to naturally diverge.
• Theoretical Extension: The paper bridges game theory and AI safety, showing that the very features making LLMs powerful (shared training, pattern recognition) become liabilities in environments requiring coordinated divergence.

In summary, the paper demonstrates that while LLMs are "super-coordinators" for convergence, they are "poor diversifiers," creating a unique risk profile for algorithmic monoculture in multi-agent economic systems.