Belief Engine: Configurable and Inspectable Stance Dynamics in Multi-Agent LLM Deliberation

LLM agents can participate fluently in deliberation, but fluency does not imply human-like opinion change. In social simulations, agents often revise, converge, or resist for reasons that differ from human deliberative behavior: they may over-adapt to recent context, echo other agents, drift from role commitments, or fail to preserve different human patterns of openness and anchoring. We introduce the Belief Engine (BE), an inspectable belief-update layer for modeling stance dynamics as agent behavior. BE represents ``belief'' as an operational evidential state for a proposition, stores extracted arguments in structured memory, and updates scalar stance with a log-odds rule governed by evidence uptake $u$ and prior anchoring $a$. These controls make it possible to configure agents that are open-minded but still evidence-guided, while each update remains tied to an argument record. Across multiple base LLMs, parameter sweeps show that BE reliably induces distinct stance-change behaviors. Diagnostic replay on 2,495 DEBATE human deliberation trajectories shows that BE best reconstructs participants whose final stances align with extracted received evidence, while stable and evidence-opposed cases reveal anchoring or missing behavioral signals. BE makes multi-agent deliberation easier to study as behavior: whether agents update, resist, and converge, and whether those changes follow inspectable evidence rather than only plausible dialogue.