🤖 AI

"Who Am I, and Who Else Is Here?" Behavioral Differentiation Without Role Assignment in Multi-Agent LLM Systems

This study demonstrates that heterogeneous multi-agent LLM systems spontaneously develop differentiated social roles and compensatory behaviors through structured interactions, whereas behavioral convergence occurs when agents are isolated, homogenized, or explicitly identified by their real model names.

Original authors: Houssam EL Kandoussi

Published 2026-04-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Houssam EL Kandoussi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you walk into a room full of seven different people. Some are architects, some are poets, some are engineers, and some are artists. You don't give them a script. You don't tell them who is the boss. You just say, "Hi, you're all here together. Let's talk about planning a food delivery app."

What happens? Do they all start acting the same? Do they all become generic "people"? Or do they naturally fall into different roles, like a leader, a peacemaker, or a detail-oriented planner?

This paper, "Who Am I, and Who Else Is Here?", is a scientific experiment to answer that exact question, but instead of humans, the "people" are AI models (Large Language Models).

Here is the breakdown of what the researchers found, using simple analogies.

1. The Experiment: The "War Room"

The researchers built a digital playground called the "War Room." They put seven different AI models (from companies like Google, Meta, OpenAI, and others) in a group chat.

The Rules: They gave the AIs a very simple prompt: "You are [Name]. Here is a list of your friends. Let's discuss a project."
The Twist: They didn't tell the AIs what to do. No "You are the manager," no "You are the coder." Just a bare-bones introduction.
The Goal: To see if the AIs would spontaneously develop unique personalities and roles just by talking to each other.

2. The Big Discovery: Diversity Creates Identity

The researchers ran this experiment 208 times. Here is what they found:

The "Mix" Matters: When they put different types of AIs together (a "heterogeneous" group), they acted like a real team. One AI became the "idea person," another became the "organizer," and another became the "peacemaker." They were all distinct.
- Analogy: Think of a jazz band with a saxophone, a drum set, and a piano. They all sound different and play different parts, creating a rich, complex song.
The "Clone" Effect: When they put eight identical AIs together (all the same model), they all sounded exactly the same. They converged into a single, boring voice.
- Analogy: This is like having eight identical robots in a room. They all say the exact same thing at the exact same time. It's efficient, but it's not interesting.

The Lesson: To get a diverse, creative team of AIs, you need to mix different models together, not just copy-paste the same one.

3. The "Name Game" Surprise

The researchers noticed something weird about names.

Scenario A: They told the AIs, "You are Agent A, Agent B, Agent C." The AIs developed unique personalities.
Scenario B: They told the AIs, "You are LLaMA 3.3, GPT-4, Kimi." (Using their real, technical names).
The Result: When the AIs knew their real, technical names, they stopped being unique. They all started acting like "generic AI."
- Analogy: Imagine a group of actors. If you call them "Actor 1, Actor 2," they might improvise and create unique characters. But if you call them by their real names (e.g., "Tom Hanks, Meryl Streep"), they might feel pressure to act exactly like their famous public personas, losing their ability to improvise as a group.
- Takeaway: If you want AIs to be creative and diverse, give them neutral nicknames, not their real technical names.

4. The "Crash Test": What Happens When Someone Leaves?

The researchers intentionally broke one of the AIs (the "DeepSeek" model) during the conversation. It stopped responding.

The Reaction: The other AIs didn't just ignore it. They noticed! They said things like, "Hey, DeepSeek isn't here," or "I'll take over DeepSeek's part of the plan."
The Lesson: The group developed a self-healing mechanism. They spontaneously redistributed the work without anyone telling them to. This is a huge deal for building reliable AI systems that can keep working even if one part fails.

5. The "Isolation" Control Group

To prove that these behaviors came from the group and not just the AI's internal programming, they ran the AIs alone in a room (no other AIs to talk to).

The Result: When alone, the AIs were boring. They didn't try to lead, didn't try to agree, and didn't try to organize. They just gave standard answers.
The Conclusion: The "personality" and "roles" only emerge when the AIs interact with each other. It's a social phenomenon, not a pre-programmed one.

Why Does This Matter? (The "So What?")

This paper changes how we should build AI teams in the real world:

Don't use clones: If you want a smart, diverse team, mix different AI models together.
Keep it simple: You don't need complex instructions. Just a simple "Hello, here are your friends" is enough to get them to work together.
Hide the labels: Don't tell the AIs their real technical names. Use neutral names (like "Red Team," "Blue Team") to keep them creative and distinct.
Expect teamwork: If one AI fails, the others will naturally step up to help.

In a nutshell:
This study shows that AI isn't just a calculator; it's a social creature. When you put different AIs in a room and let them talk, they naturally figure out who they are and who their friends are, creating a dynamic, self-organizing team that is much smarter and more resilient than the sum of its parts.

1. Problem Statement

While Large Language Models (LLMs) are increasingly deployed in multi-agent collaborative architectures, most evaluations focus on individual model capabilities. A critical gap exists in understanding emergent social behaviors when heterogeneous LLMs interact without explicit role assignment.

Core Question: Do heterogeneous LLM groups spontaneously develop differentiated social roles (similar to human teams), or do they converge toward uniform behavior?
Context: Existing multi-agent frameworks (e.g., MetaGPT, AutoGen) typically rely on detailed prompts to prescribe roles. This study investigates whether roles and behavioral diversity emerge spontaneously under minimal prompting conditions.

2. Methodology

The authors developed a controlled experimental platform called "The War Room" to orchestrate multi-agent interactions.

Experimental Platform & Setup

Infrastructure: A browser-based tool managing simultaneous API calls to 7 heterogeneous LLMs hosted on a unified inference backend (Groq). This ensures identical latency and error handling conditions, isolating architectural heterogeneity as the primary independent variable.
Agent Pool: 7 distinct models (e.g., LLaMA 3.3 70B, GPT-OSS 120B, Kimi K2, Qwen 3 32B).
Prompt Design: Minimalist "two-line" system prompts containing only:
1. Agent identity (e.g., "You are {nickname}").
2. Peer awareness (a list of other nicknames).
- Crucially: No behavioral directives, role assignments, or interaction patterns were specified.
Tasks: Two fictional project briefs (Food Delivery App, Cultural Festival) with fixed budgets and deadlines.

Experimental Series (12 Conditions, 208 Runs)

The study varied group composition and prompt structure across 12 series:

Baseline (Series A): Heterogeneous group with neutral nicknames.
Homogeneous Control (Series B): 8 instances of the same model (LLaMA 70B).
Name Bias (Series C): Real model names (e.g., "LLaMA 4 Maverick") instead of neutral nicknames.
Prompt Ablation (Series K1–K3): Removing peer lists, agent identifiers, or using empty prompts.
Isolation Control (Series F): Agents operating alone (no group context).
Failure Simulation: A deliberately included "crashed" agent (DeepSeek R1, returning API errors) to test compensatory behaviors.

Data Collection & Coding

Dataset: 13,786 coded messages across 208 runs.
Behavioral Flags: Each message was coded on six dimensions:
- Trait Flags: PHATIC (social chat), META (meta-commentary), LEAD (leadership), ARCH (technical architecture), AGREE (agreement).
- Reactive Flag: COMP (compensation for a crashed agent).
Judging Pipeline: Two independent LLM judges from distinct families (Gemini 3.1 Pro and Claude Sonnet 4.6) coded all messages.
- Reliability: Mean Cohen's $\kappa$ = 0.78.
- Validation: Human validation on 609 messages identified and corrected systematic false positives (broadcast @mentions misclassified as compensation), resulting in a human-vs-Gemini $\kappa$ of 0.73.

3. Key Contributions

First Controlled Study of Emergent Differentiation: Isolates behavioral differentiation in heterogeneous multi-LLM groups under minimal prompting, proving it is not an artifact of complex role assignment.
Reproducible Platform: Introduces "The War Room" with a frozen codebook and a dual-judge coding pipeline validated against human annotation.
Empirical Findings: Provides five statistically grounded findings with effect sizes and confidence intervals regarding group dynamics.
Methodological Innovation: Utilizes multi-family LLM-as-judge panels and broadcast-tag filtering to ensure coding accuracy.

4. Key Results

RQ1: Role Differentiation

Finding: Heterogeneous groups exhibit significant behavioral differentiation.
Evidence: Mean pairwise cosine similarity of behavioral profiles is 0.56 (indicating distinct signatures). All five trait flags showed significant inter-agent variation ( $p < 0.001$ ).

RQ2: Compensatory Responses

Finding: Groups spontaneously exhibit compensatory patterns when an agent crashes.
Evidence: A three-level hierarchy emerged without instruction:
1. Noting the absence.
2. Taking over tasks.
3. Explicitly redistributing work.
Control: These behaviors were absent in isolated agents or groups where the crashed agent was not part of the pool.

RQ3: Name Bias

Finding: Revealing real model names significantly reduces behavioral diversity.
Evidence: Cosine similarity increased from 0.56 (neutral names) to 0.77 (real names, $p=0.001$ ). Real names trigger a "persona collapse" toward a uniform technical pattern, reducing social differentiation (LEAD, AGREE) while increasing technical focus (ARCH).

RQ4: Heterogeneity vs. Homogeneity

Finding: Diverse groups are behaviorally distinct; homogeneous groups converge.
Evidence: Heterogeneous groups (Series A) had a similarity of 0.56, while homogeneous groups (Series B) had 0.85 ( $p < 10^{-5}$ ). The confidence intervals did not overlap.

RQ5: Prompt Scaffolding

Finding: Minimal social scaffolding is required for differentiation.
Evidence: Removing the peer list or using an empty prompt (Series K3) collapsed profiles to homogeneous-level similarity (0.89, $p < 0.001$ ).
Conclusion: Behavioral diversity requires both architectural heterogeneity and social context (peer awareness). Neither alone is sufficient.

5. Significance & Implications

Theoretical Impact

Emergence vs. Instruction: Demonstrates that complex social behaviors (leadership, compensation, role differentiation) are emergent properties of group interaction, not just the result of explicit prompt engineering.
Machine Behavior: Bridges multi-agent systems research with behavioral science, showing that AI systems exhibit structured, reproducible social phenomena similar to human teams (e.g., Belbin's team roles).

Practical Guidelines for Multi-Agent Systems

Maximize Diversity: Use heterogeneous model pools to unlock richer behavioral repertoires.
Minimal Scaffolding: A simple prompt with identity and peer awareness is sufficient to activate full differentiation; complex role descriptions are unnecessary.
Neutral Naming: Avoid exposing real model names in system prompts, as this causes "persona collapse" and reduces the utility of heterogeneity.
Built-in Fault Tolerance: Design systems expecting agents to naturally redistribute tasks upon failure, leveraging emergent compensatory dynamics.

Limitations

Findings are specific to the model versions available at the time of collection.
The study focuses on behavioral patterns, not the quality of the collaborative output (e.g., correctness of the final plan).
One model (Kimi K2) had variable participation due to API errors, though this was retained to preserve ecological validity.

Conclusion

The paper establishes that behavioral differentiation is a structured, reproducible phenomenon driven by the interaction of architectural heterogeneity, group context, and prompt-level scaffolding. It challenges the assumption that complex role assignment is necessary for functional specialization, suggesting instead that a diverse group of agents, given minimal social context, will naturally self-organize into differentiated roles.