MERIT反馈促进大型语言模型谈判能力的提升

Bargaining is often perceived as a logical arena rather than an art or intuition. However, Large Language Models (LLMs) struggle in this domain due to limited strategic depth and challenges in adapting to complex human factors. Current benchmarks rarely capture these limitations. To address this gap, a utility feedback-centric framework is proposed. Key contributions include the development of AgoraBench, a new benchmark encompassing nine challenging scenarios (e.g., deception, monopoly) that supports diverse strategy modeling, and the introduction of human-aligned, economically grounded metrics derived from utility theory. These metrics are operationalized through agent utility, negotiation power, and acquisition ratio, which implicitly measure the alignment of negotiations with human preferences.