Multi-Expert Japanese Essay Scoring: MoE Architecture with Rubric-Augmented Prompts and LoRA-Efficient Fine-Tuning

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Automated essay scoring (AES) in Japanese has traditionally relied on single-classifier approaches with limited performance. This paper presents a multi-expert Japanese essay scoring system that combines Mixture of Experts (MoE) with parameter-efficient LoRA fine-tuning and rubric-augmented prompts, applying MoE to Japanese AES in a low-resource setting. We apply a dense MoE head that replaces the standard single linear classifier with multiple lightweight expert MLPs and a small routing network that performs input-dependent expert selection. Under 5-fold cross-validation on each GSK prompt separately, with that prompt’s rubric included in the input, our CALM2-7B MoE model achieves 92.7% accuracy and 0.915 Quadratic Weighted Kappa (QWK), improving over a controlled CALM2-7B single-head baseline that uses the same rubric input. The MoE gain appears primarily when rubric templates are present and is model-dependent. The results show that this routing head improves scoring over a single-head baseline while remaining parameter-efficient through LoRA adaptation.