Empirical validation of a generative AI framework for personalized education assessment

The tension between personalized learning demands and standardized evaluation mechanisms presents a persistent challenge in contemporary education. This study proposes a comprehensive personalized education assessment framework driven by generative artificial intelligence technologies. The framework adopts a five-layer hierarchical architecture integrating data collection, processing, intelligent analysis, assessment generation, and feedback optimization components. ChatGLM3-6B, fine-tuned on 50,000 expert-curated programming feedback instances assembled through a human-in-the-loop process combining authentic instructor records, newly authored examples, and AI-assisted human-verified content, enables contextually responsive feedback generation, while dynamic learner profiling and knowledge graph modeling support precise diagnostic assessment. Empirical validation involving 449 undergraduate students in introductory Python programming courses demonstrated that the framework achieved assessment accuracy correlating at 0.847 with expert consensus (Fleiss’ κ = 0.74 for inter-rater reliability) while reducing generation time by over 99% compared to manual evaluation. Ablation experiments confirmed that knowledge graph integration contributed most substantially to accuracy improvements, with removal of this component reducing correlation by 0.055. Experimental participants exhibited significantly higher learning gains (Cohen’s d = 0.56), with particularly pronounced effects among initially lower-performing students. The framework also enhanced learner engagement and satisfaction compared to conventional assessment approaches. These findings suggest that generative AI can effectively operationalize personalized assessment at scale while maintaining pedagogical quality and transparency.