An LLM-based framework for personalized feedback generation and adaptive optimization in EFL writing

Abstract

Large language models, with their powerful natural language understanding and sequence generation capabilities, are gradually becoming a core component of automated educational feedback systems. However, their application in English as a Foreign Language (EFL) writing still faces challenges such as coarse intervention granularity and homogeneous feedback. This study proposes an adaptive EFL writing feedback generation framework that combines a dual-stream generative network, low-rank adaptation, and proximal policy optimization. Experiments were conducted on a longitudinal subset of the EF-Cambridge Open Language Database containing 17,859 essays written by 1,280 EFL learners. From this corpus, 2,500 expert-authored essay–feedback pairs were constructed and divided at the learner level into training, validation, and test sets in a ratio of 8:1:1, corresponding to 2,000, 250, and 250 pairs, respectively. The feedback generator was initialized from the Llama-2-7B-32K-Instruct checkpoint and trained on an eight-NVIDIA H100 GPU cluster. The system supported a maximum context length of 24,576 tokens. Under a batch size of 1 and a concurrency level of 1, the mean inter-token latency remained below 20 ms under the specified inference protocol. The normalized validation BLEU4 convergence ratio reached 99%, where this percentage represents convergence relative to the best validation checkpoint rather than an absolute BLEU-4 score. In a blind evaluation of all 250 test instances by three EFL experts, the pedagogical acceptability index reached 92.8%.