Distilling Bayesian Belief States into Language Models for Auditable Negotiation
Researchers have developed a new framework called BOND (Bayesian Opponent-belief Negotiation Distillation) to make negotiation agents more auditable. This system uses a large language model to infer and update beliefs about an opponent's values during dialogue. A smaller, distilled model then uses these beliefs to make decisions and can output normalized posterior beliefs, outperforming existing state-of-the-art methods on the CaSiNo negotiation dataset. AI
IMPACT Introduces a method for making LLM-based negotiation agents more transparent and auditable.