Abstract: Compared to ChatGPT, the domestic large-scale model DeepSeek has achieved significant innovation in model architecture, training paradigm, and inference mechanism. This means that the current large-scale model has achieved a qualitative leap from a mechanical perception feedback mechanism to a human like subject cognitive decision-making paradigm, which has also led to risks of data input disorder, algorithm running out of control, and content output disorder. The core of value alignment lies in promoting the transformation of big models from instrumental rationality to value rationality, which has certain necessity and feasibility. However, it also faces technical challenges and normative difficulties, and the reason lies in the lack of a sound value alignment system. In view of this, it is urgent to adopt a collaborative paradigm of technical regulation, ethical adjustment, and legal governance, that is, to build a data compliance framework based on the risk stratification concept, strengthen the full cycle traceability supervision of algorithm decision-making, and optimize the content governance accountability framework; improve the mechanism for extracting ethical consensus, enhance the ethical review response system, and shape the ethical subject responsibility network; establish a clear hierarchy of value alignment standards, clarify the nature of value alignment with the unity of power and responsibility, and design a value alignment evaluation method for dynamic evolution, so as to accelerate the formation of new quality productivity to promote the construction of Chinese path to modernization.

Key words: large-scale artificial intelligence model, value alignment, collaborative governance, alignment standards, alignment evaluation