OPERA: Option-level Pedagogical Explanation via Reasoning for Augmented Knowledge Tracing

OPERA is a model-agnostic framework that injects option-level pedagogical signals into knowledge tracing models without any online LLM call. It guides an LLM to generate, for each option of a multiple-choice question, a structured explanation capturing either the correct reasoning pathway or the specific misconception behind a distractor. The generated content is validated by a multi-judge quality assurance mechanism, encoded into semantic vectors, and injected into the input layer of any KT backbone in place of the conventional ID embedding. All LLM computation is confined to one-time offline preprocessing, making OPERA model-agnostic and free from inference-time overhead. Experiments on three multiple-choice datasets show that OPERA consistently boosts lightweight ID-based KT backbones to state-of-the-art results that match or surpass much heavier process-aware models.

Original paper can be found at Huang Zhiwei et al. “OPERA: Option-level Pedagogical Explanation via Reasoning for Augmented Knowledge Tracing.” Proceedings of the CIKM 2026.