Enterprise innovation has become a core engine for cultivating new quality productive forces and achieving high-quality economic development. To effectively implement the innovation-driven development strategy, the Chinese government has utilized tax preferential policies as a crucial measure to incentivize enterprise R&D and innovation, with annual increases in tax reductions. However, issues such as an imbalanced R&D investment structure and inadequate policy implementation have hindered the full realization of China’s enterprise innovation potential. Therefore, based on data from Chinese listed companies, this paper innovatively introduced a machine learning-based heterogeneity analysis method, building upon the traditional PSM-DID method, to deeply explore the heterogeneous impact of tax preferential policies on enterprise innovation. The research found that, on one hand, tax preferential policies demonstrate a certain overall positive incentive effect on stimulating enterprise innovation investment activities and serve as an important policy tool for promoting the development of new quality productive forces in China; but on the other hand, some issues remain to be optimized in policy implementation. Specifically, many enterprises exploit tax incentives through policy arbitrage, significantly weakening the positive effects of the policy and even causing adverse consequences in some cases. Through machine learning analysis to identify the heterogeneous features of policy effects, it is found that the positive impact of tax preferential policies on patent quantity and quality is mainly concentrated in approximately 15% of enterprises. This indicated that many enterprises have not yet been able to fully and effectively utilize policy incentives to unleash their innovation potential. These under-benefited enterprises generally exhibit characteristics such as high R&D risk, R&D manipulation, and strong entrepreneurial control, warranting attention in subsequent policy optimizations and adjustments. Furthermore, the machine learning-based heterogeneity analysis method introduced in this study overcomes the limitations of traditional linear models, enabling a more accurate identification of heterogeneous policy effects, and will provide a new theoretical perspective and methodological reference for precise policy implementation and promoting the vigorous development of new quality productive forces.