Building Bridges to Student Growth: An LLM-Powered Feedback Generation System for Holistic Development
Abstract
. High-quality feedback is crucial for student learning, yet scaling it across both academic and extracurricular achievements remains difficult. This study introduces an LLM-powered system that generates professional feedback from structured extracurricular records. Using a dataset of 100 secondary students with 1,067 entries, five open-source models (DeepSeek-R1, Gemma3, GPT-OSS, Llama-3.1, Qwen3) produced 500 outputs evaluated across faithfulness, coverage, fluency, coherence, and relevance. Results show that GPT (4.48) and Gemma (4.45) achieved the highest overall means, with medium-sized models (20B-27B) offering the best trade-off between accuracy and inference stability. Strengths included consistently fluent and coherent prose (>4.6), while coverage remained the weakest dimension (3.48-4.25). These findings position LLMs as drafting aids rather than replacements. They can reduce teacher workload and accelerate feedback cycles, provide human oversight to ensure factual alignment, fairness, and contextual appropriateness. Keywords: large language models · generative AI in education · student development · feedback generation 1 Introduction High-quality feedback is widely recognized as a critical driver of student learning, particularly when it is timely, specific, and actionable (Hattie and Timperley, 2007). Brookhart (2017) provides a complementary discussion that producing individualized narrative comments at scale is labor-intensive, especially when schools wish to include not only academic performance but also extracurricular participation, leadership, and awards. The workload frequently leads to generic, low-utility remarks that offer limited guidance for improvement. Recent advances in large language models (LLMs) make it feasible to draft short, professional feedback for subsequent human review, potentially streamlining the drafting stage (Song et al., 2024). However, Van der Lee et al. (2021) noted that education facing Natural Language Generation (NLG) must be assessed by human raters and governed rigorously, as automatic metrics (e.g., BLEU) correlate imperfectly with human © The Author(s), under exclusive license to Springer Nature Singapore Pte