ScioMind Framework Introduces Cognitively Grounded Multi-Agent Simulation with Memory-Anchored Belief Updates

ScioMind Framework Introduces Cognitively Grounded Multi-Agent Simulation with Memory-Anchored Belief Updates

Researchers have introduced ScioMind, a framework for multi-agent social simulation that bridges structured opinion dynamics with large language model (LLM)-based reasoning, according to a paper published on arxiv.org.

According to the paper, ScioMind integrates three key components: a memory-anchored belief update rule that modulates susceptibility to influence through personality-conditioned anchoring strength; a hierarchical memory architecture supporting persistent, experience-driven belief formation; and dynamic agent profiles derived from a corpus-grounded retrieval pipeline.

The framework was evaluated on real-world policy debate scenarios. According to arxiv.org, “across metrics including polarisation, diversity, extremization, and trajectory stability, the proposed components consistently yield improvements in behavioural realism.” The research found that dynamic profiles increased opinion diversity, while memory and reflection reduced unstable oscillation. Anchoring induced persistent belief trajectories that “better align with patterns reported in political psychology,” the paper states.

This work appears alongside broader research on LLM-based multi-agent systems. A separate arxiv.org paper accepted at ACM FAccT 2026 introduced the Mechanism Plausibility Scale, operationalizing a four-level scale for evaluating LLM-based agent-based models by separating generative sufficiency from mechanistic plausibility. Another arxiv.org study on embodied multi-agent coordination found that dialogue reduced action conflicts by 40 to 83 percentage points but degraded task success relative to silent coordination.