---
source_url: https://arxiv.org/abs/2507.04295v4
ingested: 2026-05-07
sha256: 2010c58618ace03a5cd5ca24a9d192f58601a1d3f49d81398c52f1628b0bec9d
---

EMNLP 2025. Modular LLM-based system for science education feedback. Three components: (1) error-aware assessment module capturing nuanced reasoning errors; (2) curriculum-grounded generation module using structured topic-linked memory chain rather than similarity-based retrieval, improving relevance and reducing noise; (3) educator-in-the-loop interface for customisation and oversight. Addresses challenges of generic/non-personalised feedback and noisy retrieval in RAG systems.