Belief Engine: Auditable Stance Dynamics for Multi-Agent LLMs
The Belief Engine (BE) has been unveiled by researchers, designed to allow for the auditing of stance changes in LLM-based agents by conceptualizing belief as an evidential state related to a proposition. BE organizes arguments into a structured memory and modifies stances using a log-odds rule influenced by evidence uptake and prior anchoring. Parameter sweeps conducted across various base LLMs demonstrate that these controls consistently influence stance dynamics while maintaining a trace of evidence-level updates. On the DEBATE dataset, BE excels in reconstructing participants whose final stance aligns with the evidence extracted.
Key facts
- Belief Engine (BE) is an auditable belief-update layer for LLM agents.
- BE treats belief as an evidential state over a proposition exposed as scalar stance.
- BE extracts arguments into structured memory.
- Stance is updated with a log-odds rule controlled by evidence uptake u and prior anchoring a.
- Parameter sweeps were performed across multiple base LLMs.
- BE preserves an evidence-level update trail.
- DEBATE is a human deliberation dataset with pre/post opinions.
- BE best reconstructs participants whose final stance follows extracted evidence.
Entities
Institutions
- arXiv