ScioMind: Cognitively Grounded Multi-Agent Social Simulation with Anchoring-Based Belief Dynamics and Dynamic Profiles

Problem
This paper addresses the limitations of existing multi-agent social simulation frameworks that either employ fixed update rules with minimal cognitive grounding or rely heavily on unconstrained interactions among large language models (LLMs). The authors propose ScioMind, a cognitively grounded simulation framework that aims to enhance the realism of social opinion dynamics by integrating structured belief updates with LLM-based reasoning. This work is presented as a preprint and has not yet undergone peer review.

Method
ScioMind introduces three core components to achieve its objectives:

1. Memory-Anchored Belief Update Rule: This rule modulates agents’ susceptibility to influence based on personality-conditioned anchoring strength, allowing for more nuanced belief updates.
2. Hierarchical Memory Architecture: This architecture supports persistent and experience-driven belief formation, enabling agents to retain and reflect on past experiences, which influences their current beliefs.
3. Dynamic Agent Profiles: These profiles are generated through a corpus-grounded retrieval pipeline, allowing for heterogeneous personalities and rationales that evolve over time. This dynamic aspect enhances the agents’ internal states and interactions.

The training and evaluation of ScioMind are conducted through multiple case studies in a real-world policy debate scenario, although specific training compute details are not disclosed.

Results
ScioMind demonstrates significant improvements across various metrics when compared to baseline models. Key findings include:

• Polarization: Enhanced by the memory and reflection components, leading to more stable belief trajectories.
• Diversity: Dynamic profiles contribute to increased opinion diversity among agents.
• Extremization: The anchoring mechanism induces persistent belief trajectories that align with established patterns in political psychology.
• Trajectory Stability: The combination of memory and belief updates reduces unstable oscillations in agent opinions.

These results indicate that ScioMind outperforms traditional models in behavioral realism, with specific effect sizes not quantified in the abstract but implied to be substantial.

Limitations
The authors acknowledge that while ScioMind improves upon existing frameworks, it may still be limited by the inherent biases of the LLMs used for agent reasoning. Additionally, the reliance on a corpus for dynamic profile generation may restrict the diversity of agent personalities to the content of the corpus. The paper does not address potential scalability issues when simulating larger populations of agents or the computational overhead introduced by the hierarchical memory architecture.

Why it matters
The development of ScioMind has significant implications for the field of social simulation and opinion dynamics. By providing a framework that integrates cognitive principles with LLM capabilities, it opens avenues for more realistic modeling of social interactions and belief formation. This could enhance the understanding of social phenomena in various domains, including political science, sociology, and behavioral economics. Furthermore, the insights gained from ScioMind could inform the design of more effective interventions in public policy and communication strategies.

Authors: Yitian Yang, Yiqun Duan, Linghan Huang, Yiqi Zhu, Francesco Bailo, Chunmeizi Su, Huaming Chen
Source: arXiv:2605.13725
URL: https://arxiv.org/abs/2605.13725v1