Reproduce “Belief Engine: Configurable and Inspectable Stance Dynamics in Multi-Agent LLM Deliberation” on OpenResearch

About this paper

Joshua C. Yang, Maurice Flechtner, Damian Dailisan, Michiel A. Bakker

Researchers at ETH Zurich and MIT developed the Belief Engine (BE), an auditable simulation-control layer for LLM agents that enables explicit, configurable log-odds belief updating. This system ensures agent stance changes are tied to an explicit, evidence-level update trail, allowing control over properties like evidence uptake and prior anchoring, and proving useful for diagnosing human deliberative behaviors.

Key insights

• Decoupling belief maintenance from generative reasoning provides an auditable and configurable framework for explicit stance dynamics in LLM agents.
• A Bayesian-inspired log-odds update rule allows for additive accumulation of argument weights, leading to bounded stance and diminishing returns near certainty, reflecting a more stable and predictable belief evolution.
• The framework can diagnose human deliberative behavior by identifying which observed stance movements are reconstructable by extracted evidence and specific uptake/anchoring profiles, or indicate uncaptured factors.

Summary by alphaXiv.