A Qualitative Study on the Impact of Personalized Language Learning Based on Large Language Models on Students’ Academic Self-Efficacy

Abstract

Academic self-efficacy is a core non-cognitive predictor of language learning outcomes, yet traditional large-class teaching fails to address individual differences, leading to widespread low self-efficacy among middle school students. Large Language Models (LLMs) enable scalable personalized instruction, but existing research lacks multi-method qualitative evidence to unpack the underlying mechanisms. Grounded in Bandura's Self-Efficacy Theory and Bloom's Mastery Learning Theory, this study conducted an 8-week teaching experiment with 160 Grade 8 and Grade 10 students in Sichuan, China, covering both Chinese and English subjects. Adopting a qualitative-led mixed-methods design, it collected data through semi-structured interviews (14 participants, 60,000 words), longitudinal learning logs (320 valid entries, 40,000 words) and quantitative classroom observations (64 sessions). NVivo 12 was used for three-level coding of qualitative texts, with triangulation and member checking ensuring validity. Results show that LLM-based personalized learning improves academic self-efficacy via four synergistic pathways: reshaping success experiences, regulating emotional arousal, strengthening positive verbal persuasion, and empowering autonomous learning. This study provides localized empirical evidence for digital language education and non-cognitive ability cultivation.