REFINE: Real-world Exploration of Interactive Feedback and Student Behaviour
agents
| Source: ArXiv | Original article
A team of researchers from the University of Copenhagen and the Norwegian University of Science and Technology has released a new arXiv pre‑print, REFINE: Real‑world Exploration of Interactive Feedback and Student Behaviour (arXiv:2603.29142v1). The paper introduces REFINE, a hybrid system that pairs a pedagogically‑grounded feedback‑generation agent with an “LLM‑as‑a‑judge” regeneration loop and a self‑reflective tool‑calling interactive agent. The judge, trained on human‑aligned data, evaluates the quality of generated feedback and prompts the generator to revise until the response meets educational criteria. The interactive agent then fields follow‑up questions from students, drawing on tool‑calling capabilities to supply context‑aware, actionable advice.
The authors argue that the architecture tackles a long‑standing bottleneck in digital learning: delivering timely, individualized formative feedback at scale. In pilot deployments across two Nordic high schools, REFINE reduced the average feedback latency from hours to under two minutes while maintaining rubric‑aligned quality scores comparable to teacher‑generated comments. Student surveys reported higher perceived relevance and increased willingness to ask clarification questions, suggesting the system may improve engagement beyond static auto‑graded quizzes.
The development builds on recent advances in LLM‑driven educational tools, such as the ToolTree planning framework reported earlier this month, and signals a shift from one‑shot feedback generators toward iterative, judge‑guided loops that can adapt to learner input. Industry observers will watch whether platforms like Nearpod or ThingLink integrate REFINE’s API to enrich their formative‑assessment suites. Equally important will be longitudinal studies measuring learning gains and the system’s ability to mitigate bias in feedback. If the early results hold, REFINE could become a cornerstone of next‑generation AI‑assisted instruction, prompting schools and ed‑tech firms to accelerate trials and standard‑setting discussions.
Sources
Back to AIPULSEN