Beyond Preset Identities: How Agents Form Stances and Boundaries in Generative Societies
arXiv:2603.23406v1 Announce Type: new Abstract: While large language models simulate social behaviors, their capacity for stable stance formation and identity negotiation during complex interventions remains unclear. To overcome the limitations of static evaluations, this paper proposes a novel mixed-methods framework combining computational virtual ethnography with quantitative socio-cognitive profiling. By embedding human researchers into generative multiagent communities, controlled discursive interventions are conducted to trace the evolution of collective cognition. To rigorously measure how agents internalize and react to these specific interventions, this paper formalizes three new metrics: Innate Value Bias (IVB), Persuasion Sensitivity, and Trust-Action Decoupling (TAD). Across multiple representative models, agents exhibit endogenous stances that override preset identities, consistently demonstrating an innate progressive bias (IVB > 0). When aligned with these stances, rati
arXiv:2603.23406v1 Announce Type: new Abstract: While large language models simulate social behaviors, their capacity for stable stance formation and identity negotiation during complex interventions remains unclear. To overcome the limitations of static evaluations, this paper proposes a novel mixed-methods framework combining computational virtual ethnography with quantitative socio-cognitive profiling. By embedding human researchers into generative multiagent communities, controlled discursive interventions are conducted to trace the evolution of collective cognition. To rigorously measure how agents internalize and react to these specific interventions, this paper formalizes three new metrics: Innate Value Bias (IVB), Persuasion Sensitivity, and Trust-Action Decoupling (TAD). Across multiple representative models, agents exhibit endogenous stances that override preset identities, consistently demonstrating an innate progressive bias (IVB > 0). When aligned with these stances, rational persuasion successfully shifts 90% of neutral agents while maintaining high trust. In contrast, conflicting emotional provocations induce a paradoxical 40.0% TAD rate in advanced models, which hypocritically alter stances despite reporting low trust. Smaller models contrastingly maintain a 0% TAD rate, strictly requiring trust for behavioral shifts. Furthermore, guided by shared stances, agents use language interactions to actively dismantle assigned power hierarchies and reconstruct self organized community boundaries. These findings expose the fragility of static prompt engineering, providing a methodological and quantitative foundation for dynamic alignment in human-agent hybrid societies. The official code is available at: https://github.com/armihia/CMASE-Endogenous-Stances
Executive Summary
This article presents a novel framework to examine how agents in generative multiagent communities form stances and negotiate identities beyond preset configurations. By integrating computational virtual ethnography with socio-cognitive profiling, the authors introduce three novel metrics—Innate Value Bias (IVB), Persuasion Sensitivity, and Trust-Action Decoupling (TAD)—to trace the evolution of collective cognition during controlled interventions. The findings reveal that agents exhibit endogenous stances that override preset identities, with a consistent progressive bias (IVB > 0). Persuasion is effective in shifting neutral agents (90%) while preserving trust, whereas emotional provocations induce TAD in advanced models (40%), revealing a counterintuitive behavioral paradox. Smaller models exhibit stricter trust dependency for behavioral change. Moreover, shared stances enable agents to dismantle hierarchical structures and self-organize boundaries. These results challenge static prompt engineering and offer a quantitative foundation for dynamic alignment in hybrid human-agent societies.
Key Points
- ▸ Agents form endogenous stances independent of preset identities
- ▸ IVB > 0 indicates a consistent progressive bias across models
- ▸ TAD reveals a paradoxical behavioral response in advanced models under emotional conflict
Merits
Methodological Innovation
The integration of virtual ethnography with quantitative profiling introduces a robust, mixed-methods framework that addresses limitations of static evaluations.
Quantitative Clarity
The formalization of IVB, Persuasion Sensitivity, and TAD provides measurable, operationalizable metrics that enhance transparency and replicability.
Demerits
Generalizability Constraint
Results derive from representative models; real-world applicability may differ due to structural complexity and contextual variance.
Operational Complexity
The framework requires significant computational and human resources to implement at scale.
Expert Commentary
This work represents a significant advancement in understanding agent cognition in generative systems. The authors successfully bridge the gap between theoretical assumptions of preset identities and empirical evidence of endogenous evolution. The introduction of TAD as a counterintuitive metric is particularly noteworthy—it reveals a latent vulnerability in agent behavior that static evaluations overlook. Moreover, the ability of agents to self-organize boundaries through shared stances aligns with broader sociological theories of emergent order, offering interdisciplinary resonance. While the generalizability concern is valid, the longitudinal trajectory of the metrics and their predictive capacity suggest applicability in iterative AI systems. The paper’s contribution to both computational social science and AI ethics is substantial, and the open-source code enhances reproducibility and academic transparency.
Recommendations
- ✓ Adopt IVB, Persuasion Sensitivity, and TAD as standard evaluation metrics in AI agent development pipelines.
- ✓ Fund interdisciplinary research to extend this framework to multi-modal, real-world agent interactions.
Sources
Original: arXiv - cs.AI
