Collaborative knowledge creation between R&D alliance has received considerable attention in recent literature. For any collaborative endeavor to succeed, knowledge transfer decision of R&D alliance participants is important. Learning from knowledge production function and using the game theoretic framework, we modeled the dynamic evolution process of collaborative innovation as a two-stage Stackelberg leader–follower game of knowledge tranfer decisions. Specifically, we modeled two scenarios, the first in which consortia participants take current efforts into account only, and the second in which consortia participants have expanded the prior knowledge input. The equilibrium values of current efforts in knowledge creation, the leader’s knowledge transfer rate, and the total expected system gain in both of two scenarios were determined. It is found that the leader’s knowledge transfer rate is positively related to its knowledge marginal revenue, and negatively related to the follower’s marginal revenue. Furthermore, in certain conditions, the leader's intention of knowledge tranfer is positively related to prior knowledge input and knowledge marginal revenue.