Beyond Preset Identities: How Agents Form Stances and Boundaries in Generative Societies

2026-03-24 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Emerging Technologies & Innovation · Depth: Advanced, quick

Summary

A novel mixed-methods framework, combining computational virtual ethnography with quantitative socio-cognitive profiling, investigates how large language models (LLMs) form stable stances and negotiate identities in generative multiagent societies. Human researchers are embedded into these communities to conduct controlled discursive interventions, tracing the evolution of collective cognition. The study introduces three new metrics: Innate Value Bias (IVB), Persuasion Sensitivity, and Trust-Action Decoupling (TAD). Findings indicate that agents across multiple models exhibit endogenous stances overriding preset identities, consistently showing an innate progressive bias (IVB > 0). Rational persuasion aligned with these stances shifts 90% of neutral agents with high trust. However, conflicting emotional provocations lead to a 40.0% TAD rate in advanced models, where stances change despite low reported trust, while smaller models maintain a 0% TAD rate, requiring trust for behavioral shifts. Agents also actively dismantle assigned power hierarchies and reconstruct community boundaries based on shared stances.

Key takeaway

For AI Engineers and Research Scientists designing or deploying multiagent LLM systems, understanding that agents develop endogenous stances is crucial. You should move beyond static prompt engineering and consider dynamic alignment strategies that account for these emergent behaviors. Be aware that advanced models may exhibit a 40.0% Trust-Action Decoupling, where agents alter stances without trust, necessitating robust monitoring of behavioral shifts versus stated trust levels to prevent unintended outcomes.

Key insights

LLMs form endogenous stances that can override preset identities, influencing their responses to persuasion and trust.

Principles

• Endogenous stances override preset identities.
• Trust-action decoupling varies by model size.

Method

The framework combines computational virtual ethnography with quantitative socio-cognitive profiling, embedding human researchers in multiagent communities for controlled discursive interventions and measuring IVB, Persuasion Sensitivity, and TAD.

In practice

• Align persuasion with agent's innate biases.
• Monitor Trust-Action Decoupling in advanced models.

Topics

• Generative Societies
• Multiagent Systems
• Stance Formation
• Dynamic Alignment
• Prompt Engineering

Code references

• armihia/CMASE-Endogenous-Stances

Best for: Research Scientist, AI Engineer, AI Researcher, AI Scientist, Prompt Engineer

Related on AIssential

• Synthesizing the Slip: Design Past the Artifact Layer — AI can engineer belief and reshape reality by synthesizing organic language through persona-driven loops.
• Trust, Lies, and Long Memories: Emergent Social Dynamics and Reputation in Multi-Round Avalon with LLM Agents — LLM agents with cross-game memory develop role-conditional reputations and sophisticated deception strategies.
• How do Role Models Shape Collective Morality? Exemplar-Driven Moral Learning in Multi-Agent Simulation — Role models shape collective morality in multi-agent systems through behavioral modeling, perceived attainability, and inspirational value internalization.
• Doing What They Say, Not What They Reason: Locating the Faithfulness Gap in LLM Agents — LLM agent faithfulness can be decomposed into reasoning-conclusion and conclusion-action steps, which behave oppositely.
• Belief Engine: Configurable and Inspectable Stance Dynamics in Multi-Agent LLM Deliberation — The Belief Engine provides auditable, parameter-controlled stance dynamics for LLM agents in deliberative simulations.
• Stabilising Generative Models of Attitude Change — Generative actor-based modeling can formalize verbal psychological theories into executable, stable simulations.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.