LearnLens: LLM-Enabled Personalised, Curriculum-Grounded Feedback with Educators in the Loop

TL;DR

LearnLens addresses the challenge of delivering timely, high-quality feedback at scale in science education by introducing a modular LLM-based system that combines error-aware assessment, curriculum-grounded feedback via a Chain-of-Concept memory, and educator-in-the-loop interfaces. The approach overcomes prior limitations by grounding feedback in curriculum topics rather than surface similarity and by enabling teacher control through verifiers and interactive revisions. Key contributions include an error-aware assessment module, a topic-linked memory for coherent retrieval, a verification-loop that ensures feedback quality, and dual teacher/student interfaces that support scalable classroom deployment, validated by local deployments across multiple LLMs. Overall, LearnLens demonstrates potential to reduce teacher workload while delivering personalised, curriculum-aligned feedback, offering a practical path toward scalable AI-assisted learning in real classrooms, with future work addressing student evaluation and broader generalization.

Abstract

Effective feedback is essential for student learning but is time-intensive for teachers. We present LearnLens, a modular, LLM-based system that generates personalised, curriculum-aligned feedback in science education. LearnLens comprises three components: (1) an error-aware assessment module that captures nuanced reasoning errors; (2) a curriculum-grounded generation module that uses a structured, topic-linked memory chain rather than traditional similarity-based retrieval, improving relevance and reducing noise; and (3) an educator-in-the-loop interface for customisation and oversight. LearnLens addresses key challenges in existing systems, offering scalable, high-quality feedback that empowers both teachers and students.

Figures (5)

• Figure 1: LearnLens: A modular LLM-based system delivering personalised, curriculum-aligned feedback through error analysis, topic memory, and educator oversight.
• Figure 2: LearnLens: Overall framework.
• Figure 3: Comparison between traditional RAG and the proposed Chain-of-Concept approach. In RAG (top), past assessment records are stored sequentially without explicit logical structure, forcing the retriever to rely solely on surface-level similarity. This often introduces significant noise, especially as the database grows. In contrast, our approach (bottom) organises past assessments by topic-level relationships, enabling the model to retrieve more contextually relevant information and generate more personalised, coherent feedback.
• Figure 4: Full Questionnaire Response Distribution (N = 30 per Question). Shades of blue denote more favourable evaluations of the tool, while shades of red indicate dissatisfaction. Grey segments represent neutral factual responses that are not direct indicators of user sentiment toward the tool.
• Figure 5: Grouped feedback by participants’ prior methods (manual, prompt-based, or automated). Left: perceived improvements across five dimensions. Right: future preferences for task completion. Blue = positive toward our tool; red = preference for previous methods; orange = combine both; grey = neutral or other.