Systematizing Research on AI-Based Automated Assessment: A Two-Axis Review Framework for Open-Ended Responses

Automated assessment of open-ended responses has evolved from rule-based and statistical approaches to large language model (LLM)-based and hybrid architectures. Despite this progress, the literature remains fragmented across scoring, formative feedback, contextual grounding, reliability, and system integration. This review proposes a two-axis classification framework that organizes existing research by research focus and level of implementation, from offline benchmark studies to learning management system (LMS)-integrated systems. The framework identifies six analytical categories and reveals an uneven pattern of development: scoring-oriented studies are methodologically more mature, while grounding, formative feedback, and institutional deployment often remain at the pilot stage. The review also highlights key gaps, including the lack of end-to-end architectures, limited reactive LMS integration, weakly structured outputs, and underdeveloped workflow-oriented perspectives. The central limitation of current research is not the absence of promising components, but their insufficient integration into coherent and reproducible assessment systems.