A. Palomino, D. Cervantes
Retrieval-Augmented Generation (RAG) systems promise to reduce large language model hallucinations by grounding responses in verified sources, which makes them candidates for supporting students at academic risk; however, evidence on their use for institutional risk monitoring remains scattered. This systematic review, conducted following PRISMA guidelines across eight databases, identified 135 studies published between 2020 and 2026, of which 100 met the inclusion criteria, complemented by a scientometric analysis of the scholarly output. Results reveal a recently formed corpus —two thirds of the studies belong to the 2025-2026 period— and a bifurcation between two bodies of literature that rarely cite each other: predictive academic risk analytics, active since 2020, and conversational artificial intelligence, whose RAG implementations do not appear before 2024 yet reach 38% of the corpus. Both traditions rely on disjoint metric families: no study jointly evaluates risk prediction accuracy together with chatbot reliability, and the two mitigation repertoires —class imbalance treatment and generative text control— never coexist within a single design. The review concludes that no hybrid architecture coupling risk prediction with retrieval-augmented generation exists to date, and proposes a research agenda focused on designing such architectures, building joint evaluation frameworks, and combining both mitigation repertoires under teacher oversight.
Keywords:
Published on 21/07/26
Licence: CC BY-NC-SA license
Views 12Recommendations 0
