Enhancing coherence in descriptive writing tasks using LLMs for private university EFL learners in Dhaka

Abstract

This study examines whether short-term use of Large Language Models (LLMs) can enhance coherence in descriptive writing among private university EFL learners in Dhaka compared with peer review. In a two-week intervention, a total of eighteen undergraduates participated and were divided into two groups: an experimental group (n=9) using LLM tools for paragraph-level scaffolding and a control group (n=9) using traditional peer review. Both groups completed pre-tests and post-tests on descriptive paragraph writing, which were scored using a ten-point analytic rubric. In addition, six students participated in semi-structured interviews. The experimental group started with higher baseline scores (8.78 versus 6.89). Results showed that both groups made gains, but distributional analysis revealed different patterns. The experimental group's score range narrowed from four to two points, with the standard deviation dropping from 1.30 to 0.87, and four students reaching the top scores. On the other hand, the control group's range widened from three to five points, showing a more uneven distribution. Interview data showed that AI was used as scaffolding for paragraph organization and transitions by experimental participants. In contrast, control participants described peer feedback as being primarily word-level, focusing mostly on vocabulary. These findings suggest that LLMs can be effectively used as a scaffold for coherence when students use them for paragraph-level support. The study is limited by its short timeframe, single site, and small sample, but it contributes a Bangladeshi perspective on AI integration in EFL writing instruction.