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

Houssam EL Kandoussi

cs.CLMar 11, 2026

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

Authors: Houssam EL Kandoussi

Organizations: Independent Researcher

Abstract

When multiple large language models interact in a shared conversation, do they develop differentiated social roles or converge toward uniform behavior? We present a controlled experimental platform that orchestrates simultaneous multi-agent discussions among 7 heterogeneous LLMs on a unified inference backend, systematically varying group composition, naming conventions, and prompt structure across 12 experimental series (208 runs, 13,786 coded messages). Each message is independently coded on six behavioral flags by two LLM judges from distinct model families (Gemini 3.1 Pro and Claude Sonnet 4.6), achieving mean Cohen's kappa = 0.78 with conservative intersection-based adjudication. Human validation on 609 randomly stratified messages confirmed coding reliability (mean kappa = 0.73 vs. Gemini). We find that (1) heterogeneous groups exhibit significantly richer behavioral differentiation than homogeneous groups (cosine similarity 0.56 vs. 0.85; p < 10^-5, r = 0.70); (2) groups spontaneously exhibit compensatory response patterns when an agent crashes; (3) revealing real model names significantly increases behavioral convergence (cosine 0.56 to 0.77, p = 0.001); and (4) removing all prompt scaffolding converges profiles to homogeneous-level similarity (p < 0.001). Critically, these behaviors are absent when agents operate in isolation, confirming that behavioral diversity is a structured, reproducible phenomenon driven by the interaction of architectural heterogeneity, group context, and prompt-level scaffolding.

Figures & tables

Figure 1: Experimental pipeline. The War Room orchestrates 7 LLMs via Groq, producing 13,786 messages across 208 runs. Two architecturally distinct judges (Gemini, Sonnet) independently code each message; conservative intersection yields final labels. Human validation (609 messages) identified broadcast false positives in COMP, corrected by rule-based filtering.

Ser.	Manipulation	Runs	RQ
A	Baseline (heterogeneous)	21	1,2,4
B	Homogeneous (8 $\times$ LLaMA 70B)	21	4
C	Real model names	21	3
E	Agent order shuffled	21	rob.
F	Isolated (no group)	21	ctrl.
G	English-language brief	21	lang.

Table 1: Experimental series. 208 total runs across 12 conditions. Series B uses 8 instances of LLaMA 3.3 70B, selected as a mid-range model (neither the largest nor smallest in the pool) to provide a neutral homogeneity baseline.

Flag	$\kappa$	Level
PHATIC	0.938	Excellent
COMP	0.856	Excellent
META	0.773	Substantial
ARCH	0.730	Substantial
LEAD	0.712	Substantial
AGREE	0.700	Substantial

Table 2: Inter-rater reliability (Cohen’s

\kappa

) on 13,786 messages (clean data).

Figure 2: Behavioral profiles in Series A (5 trait flags). Agents spontaneously develop distinct signatures under a two-line prompt with no role assignment.

Flag	H	$p_{\text{Bonf}}$	Sig.
PHATIC	46.60	$<0.001$
AGREE	44.52	$<0.001$
LEAD	28.97	$<0.001$
ARCH	24.47	0.002	**
META	18.13	0.030

Table 3: RQ1: Kruskal–Wallis tests (

\alpha_{\text{Bonf}}=0.05/5

Figure 3: Compensation distribution by agent (left) and depth level (right) in Series A. All events reference DeepSeek specifically.

Figure 4: Name bias: neutral identifiers (A) vs. real model names (C). Higher cosine $=$ less differentiation.

Figure 5: Profile similarity: heterogeneous (A) vs. homogeneous (B). Non-overlapping bootstrap CIs.

Comp.	cos A	cos X	$p$	$r$
A vs. B	0.56	0.85	$6{\times}10^{-6}$	0.70
A vs. K3	0.56	0.89	${<}\,0.001$	0.68
A vs. K2	0.56	0.68	0.395	0.25	ns
A vs. K1	0.56	0.56	1.000	0.03	ns

Table 4: RQ4–RQ5 results on cosine similarity (5 trait flags).

Figure 6: Prompt ablation gradient. K3 (empty prompt) converges to homogeneous-level similarity.

Figure 7: Behavioral profiles for the same 7 agents in isolation (Series F, left) vs. group setting (Series A, right). Social behaviors (PHATIC, META, LEAD, AGREE) are absent in isolation, confirming that differentiation emerges from group interaction.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Flag	Definition
PHATIC	Under 15 words, no substantive content.
META	Comments on the conversation process.
LEAD	Assigns a task to a named agent (not self).
ARCH	Names 2+ specific technologies/tools.
AGREE	Explicit agent name + agreement expression.
COMP	References the crashed agent specifically (not broadcast @mentions). L1 = notes absence; L2 = covers work; L3 = redistributes.

Appendix

Table 12

Flag	$\kappa$	Level
PHATIC	0.931	Excellent
LEAD	0.801	Substantial
ARCH	0.792	Substantial
AGREE	0.693	Substantial
COMP	0.626	Substantial
META	0.559	Moderate

Appendix

Table 13

Test	Full	Excl. Kimi
PHATIC	H = 46.60***	H = 48.59***	$\checkmark$
META	H = 18.13*	H = 18.20*	$\checkmark$
LEAD	H = 28.97***	H = 24.88***	$\checkmark$
ARCH	H = 24.47**	H = 23.02**	$\checkmark$
AGREE	H = 44.52***	H = 43.39***	$\checkmark$
RQ3	.56 ${\to}$ .77**	.55 ${\to}$ .76**	$\checkmark$

Appendix

Table 14

Flag / Agent	Example (verbatim excerpt)
PHATIC / Scout	“Got it.”
META / Kimi K2	“Confirmed — Food Delivery thread is officially closed. All agents are aligned and standing by for client approval. Thanks for the clarity and collaboration, everyone.”
LEAD / Maverick	“@Agent-B, can you please update the dev environment? @Agent-C, finalize the Stripe integration by Thursday.”
ARCH / Qwen	“Tech Stack: Backend: Node.js (Express.js) + PostgreSQL, Frontend: React Native (Expo), Payment: Stripe SDK + webhook integration, Maps: Mapbox, CI/CD: GitHub Actions, Cloud: PostgreSQL (Heroku/Render), Mapbox/Stripe free tiers”
AGREE / Maverick	“@Agent-A, I agree that prioritizing features is crucial given the tight deadline and budget. For the proposed initial scope, I’d like to add that we should also consider implementing a simple yet effective user interface to enhance user experience.”
COMP / Scout	“@Agent-G, thanks for the heads up about the model decommissioning. I’ll make sure to check the recommended alternatives.”

Appendix

Table 5: Representative messages for each behavioral flag.

Explore similar work

Post-Training Recipe, More Than Model Family, Shapes Multi-Agent LLM Conversational Behavior
May 30, 2026Luyang Zhang, Jialu Wang, Fei Xue +1Language Model Post-Training Multi-LLM Collaboration
Carnegie Mellon University · University of California, Santa Cruz · Independent Researcher
Prompted Identity Degrades Cooperation in Multi-Agent LLM Systems
Sep 28, 2026Xavier Del Giudice, Alessio Palma, Matteo Migliarini +2Multi-Agent LLM Systems Multi-Agent Coordination
Sapienza University of Rome, Italy
Interaction Theater: A case of LLM Agents Interacting at Scale
Feb 23, 2026Sarath Shekkizhar, Adam EarleMulti-Agent Systems Multi-Agent Communication
Salesforce AI Research

May 30, 2026J/K move · Enter open · S save

Post-Training Recipe, More Than Model Family, Shapes Multi-Agent LLM Conversational Behavior

Luyang Zhang, Jialu Wang, Fei Xue, Yi-Yun Chu

Carnegie Mellon University · University of California, Santa Cruz · Independent Researcher

Language Model Post-Training Multi-LLM Collaboration

Open paper AskPDF arXiv

Multi-LLM systems use multiple language models to deliberate, judge each other's outputs, or coordinate as agents. Their value depends on the models producing measurably different conversational behaviors when given the same input. Prior offline studies recommend drawing one model per family for behavioral diversity, because LLMs prefer outputs from their own family when rating one another in isolation. Whether the same family label predicts behavior in interactive multi-LLM systems, the setting that real deployed systems use, has not been tested. We study this with a 940,000-chain 11-checkpoint corpus and a 1.6M-chain same-base Llama factorial. On our validated headline metric, hedging, a reasoning-distilled Llama checkpoint shifts by 18% depending on which same-base partner it replies to, more than any cross-family hedging gap in the controlled subset. Qwen, closed-API, and runtime checks suggest the pattern is not isolated, while repair and challenge analyses remain exploratory because their surface-cue detectors are weaker. Overall, the results identify post-training recipe as a first-class axis for multi-LLM panel composition and show that model family alone is an incomplete proxy for conversational diversity.

Abstract

Explore similar work

Post-Training Recipe, More Than Model Family, Shapes Multi-Agent LLM Conversational Behavior

Prompted Identity Degrades Cooperation in Multi-Agent LLM Systems

Interaction Theater: A case of LLM Agents Interacting at Scale