Belief Engine makes AI agent stance changes auditable and configurable

A team of researchers including Joshua C. Yang and collaborators have introduced the Belief Engine (BE), a novel system that brings transparency and controllability to stance dynamics in multi-agent LLM deliberations. The system treats an agent's belief as an evidential state over a proposition, exposed as a scalar stance value. BE extracts arguments made during deliberation into structured memory, then updates stance using a log-odds rule governed by two key parameters: evidence uptake (u) and prior anchoring (a). This allows researchers to configure how strongly agents integrate new evidence versus stick to initial positions.

The Belief Engine was validated across multiple base LLMs, with parameter sweeps showing that u and a reliably shape stance dynamics while preserving a complete evidence-level update trail. Testing on the DEBATE dataset (human deliberation with pre/post opinions) showed BE best reconstructs participants whose final stance follows the evidence presented. Cases where participants held stable or evidence-opposing stances point to anchoring effects or factors outside the extracted evidence stream. This provides infrastructure for studying evidence-grounded deliberation where openness, commitment, convergence and disagreement can be traced to explicit assumptions rather than hidden prompt effects.