Agent Sociodrama

Social-Dynamics-Guided Persona Recasting for LLM Agent Teams

Yirui Qi^1,2 Xiaoming Zhang^1† Jaehong Yoon^2†

¹Beihang University ²Nanyang Technological University, Singapore

^†Corresponding authors

Motivation and overview of Agent Sociodrama — **Motivation and overview.** Agent Sociodrama models multi-agent interaction through stable personas, evolving emotions, and directed relationships. We assess the social fidelity of these dynamics against social-theory patterns and human judgments, and use social trajectories to guide team composition optimization for better collaboration.

Abstract

Large language model (LLM) agents are increasingly deployed as teams, yet they fail in characteristically social ways: information holders stay silent, dominant speakers anchor groups on wrong answers, and norm violations go unsanctioned. We present Agent Sociodrama, a framework for modeling and validating evolving social dynamics in LLM agent teams. Under fixed personas, agents' emotional states and directed, asymmetric relationships are updated every round by a unified external LLM judge and shape their subsequent interactions. Across six LLM backbones and five social scenarios, we assess social fidelity through classical social-theory patterns and agreement with human judgments. The resulting trajectories exhibit structured social patterns and model-specific differences.

We then introduce Trajectory-Guided Persona Recasting (TPR), which diagnoses group-dynamic failures from interaction transcripts and social trajectories and recasts agent personas before rerunning the task. Across four collaborative benchmarks, TPR improves performance from both heterogeneous and randomly initialized agent persona compositions, yielding up to 83.3%p of accuracy improvement over the initial casting. On representative distributed-information cooperation and mixed-motive bargaining tasks, we outperform SOTOPIA and AgentVerse in three of four comparisons, by up to 37.5%p in Pareto rate. These results show that validated social trajectories provide actionable signals for repairing LLM agent teams.

Agent Sociodrama

Each agent has a fixed persona, an emotional state S, and a directed relationship R toward every other agent. In every round, agents decide whether to speak publicly, message someone privately, or stay silent, and speak in order of urgency. An external LLM judge then updates each agent's emotional state and relationships from the events it could observe.

Social Fidelity across LLMs

We run five scenarios of increasing social pressure, from a hospital ethics review to a demolition standoff, with six LLM backbones and five nine-round episodes per condition. From Hospital to Demolition, mean final favorability drops from 0.75 to 0.23, structural balance from 100.0% to 71.7%, and repair from 58.5% to 18.9%, while relationships emerge gradually over rounds rather than being fixed at initialization.

**Team-level favorability trajectories across nine rounds.** Relationships warm steadily under low social pressure and flatten or decline as conflict intensifies.

Model variation across social-pressure settings

LLMs show distinct social dynamics

Final favorability differs significantly across the six models in every scenario (Kruskal–Wallis, all p < 0.01). Model-specific social dynamics are multidimensional: positive affect and high reciprocity do not necessarily imply favorable or repairable relationships.

Agreement with human judgments

On items where at least seven of thirteen human raters agree, the LLM judge matches the human majority on 76.2% of relationship updates and 88.2% of emotional updates.

Trajectory-Guided Persona Recasting

TPR treats team composition as a searchable design variable. It evaluates a casting over repeated episodes, diagnoses team failures from the transcripts and the emotional and relational trajectories, and recasts the personas of the agents involved by majority vote across episodes, for at most three recasts.

Personas are drawn from five archetypes: balanced, dominant, analytical, agreeable, and reserved. Each agent's private knowledge remains fixed under recasting.

4 / 4benchmarks improved from both heterogeneous and random initial castings

+83.3%plargest accuracy gain over the initial casting

3 / 4comparisons won against SOTOPIA and AgentVerse

**Performance across recasting steps.** Means over six backbones and two initializations, compared with the variant without social dynamics.

Comparison with collaborative baselines (final means over six backbones)

	NASA L₁↓ (Het.)	NASA L₁↓ (Rand.)	Pareto↑ (Het.)	Pareto↑ (Rand.)
Ours	12.03	12.92	48.6	44.4
SOTOPIA	13.69	14.08	18.1	6.9
AgentVerse	14.06	12.89	22.2	40.3

Case Study

On Lost at Sea (DeepSeek V4-Flash, Random), the diagnosis finds that Blake's argument for using the fuel can for signaling is undervalued. After three recasts, the team contains four distinct personas: Blake dominant, Casey analytical, Drew agreeable, and Avery balanced. Blake now argues for ranking the fuel can high.

Mean L₁ falls from 21.83 to 18.00, 14.33, and finally 11.33. Each value averages 12 episodes; dialogue is paraphrased from logs. The full diagnoses and transcripts of this run are in the repository.

BibTeX

@article{qi2026agentsociodrama,
  title   = {Agent Sociodrama: Social-Dynamics-Guided Persona Recasting for LLM Agent Teams},
  author  = {Qi, Yirui and Zhang, Xiaoming and Yoon, Jaehong},
  journal = {arXiv preprint arXiv:2610.15788},
  year    = {2026}
}