Data Valuation for Generative Models

The rapid advancement of generative AI has fundamentally shifted machine learning from a discriminative paradigm to a generative one. This shift has rendered traditional data valuation methods, predominantly designed for discriminative models, inadequate for assessing the value of data in generative tasks. The field now faces four fundamental challenges: (1) Comprehensive data valuation measures customized for generative AI tasks, (2) Scalable data valuation adapted to the complexity of generative models, (3) Coherent data valuation aligned across all value propagation stages, and (4) Practical data valuation mechanisms for fair and trustworthy rewards. To address these challenges, this tutorial presents the first systematic synthesis of emerging solutions, organized around a four-step framework: defining value through generative utility functions, computing value with scalable algorithms, refining value via temporal and multi-stage attribution, and distributing value through trustworthy incentive mechanisms. For each step, we present the key design ideas, review cutting-edge approaches, and outline future directions. Our goal is to equip participants with a structured understanding of this rapidly evolving landscape. By the end of the tutorial, attendees will gain a clear grasp of the four fundamental challenges, a practical understanding of the major solution families, and a vision for future research toward a unified data valuation framework for the generative AI economy.