A Risk-Adaptive and Evidence-Constrained Framework for Generative AI Feedback in Programming Education

This paper addresses the challenge of providing timely, personalized feedback in introductory programming courses using generative AI. The author proposes a framework that integrates learning analytics, predictive modeling, and evidence-constrained generation. By treating feedback as a sequential decision-making process, the system determines when to intervene, what evidence to use, and the appropriate level of assistance based on a student's risk of persistent failure.

Using 2,993 failed programming submission states from 215 students, the author developed models to predict persistent failure and other behavioral outcomes. The framework employs a student-disjoint evaluation to prevent data leakage and uses logistic regression as a transparent baseline against more complex deep learning models. A key innovation is the evidence-gated progressive assistance strategy, which maps calibrated risk scores to four levels of feedback intensity—ranging from simple self-checks to localized hints—while enforcing strict source-traceability to prevent AI hallucinations.

Predictive models achieved moderate discrimination, with logistic regression performing competitively against deep learning architectures. The study demonstrates that broader student history improves the prediction of specific behaviors like unmodified resubmission. The generation pipeline successfully produced structurally complete feedback, with a source-verification gate ensuring that only verifiable code references were included. Furthermore, the risk-adaptive policy effectively prioritized interventions, allowing instructors to manage limited support capacity while capturing a significant proportion of persistent failures.

This research provides a reproducible, auditable pipeline for deploying generative AI in education. By separating the prediction of student risk from the generation of feedback, the framework allows educators to maintain oversight and ensure that automated support remains grounded in actual student performance. This approach moves beyond simple fluency metrics, focusing instead on the educational validity and operational feasibility of AI-driven interventions.