关键词: 质量信息不对称, 大模型, 信号博弈, 开源程度

Abstract:

As public concern about the quality of large language models （LLMs） continues to grow， vendors are increasingly adopting open-source strategy as an indirect means to signal quality， aiming to enhance consumers’ perception of model performance and alleviate security concerns. A signaling game model is constructed in which the open-source degree of LLMs serves as a quality signal， and how a vendor’s quality disclosure strategy affects the open-source degree， market demand， and profits is examined. It is found that when the quality difference between the vendor’s potential high and low-type states is small， the vendor conceals its quality type under a non-disclosure strategy （pooling equilibrium）. Compared with active disclosure， in this case a high-type vendor chooses a lower open-source degree and experiences lower first-period market demand but higher second-period demand， while the opposite pattern holds for a low-type vendor. When the quality difference is large， the vendor reveals its type under non-disclosure （separating equilibrium）， where a high-type vendor adopts a higher open-source degree and achieves higher first-period demand but lower second-period demand compared to active disclosure， while a low-type vendor’s open-source degree and market demand remain unaffected. Active quality disclosure always enhances the profit of a high-type vendor， but reduces the profit of a low-type vendor when the quality difference is substantial.

Key words: quality information asymmetry? large language models? signaling game? open-source degree