DiagRe: Diagnosis-Guided Response Replay for Continual Cognitive State Modeling
Fangzhou Yao , Qi Liu , Linan Yue , Weibo Gao
Modeling users’ cognitive states has become an important task for intelligent systems, as it enables adaptive and personalized services based on users’ behavioral records. Cognitive diagnostic models (CDMs) provide a principled approach to estimating users’ knowledge states from their response logs, thereby supporting fine-grained cognitive state modeling and personalized interventions. Most existing studies on CDMs assume that the collected response data are independent and identically distributed (i.i.d.). However, this assumption does not align with practical learning processes, where response samples arrive continuously and are intrinsically non-stationary due to evolving learner states and changing interaction contexts. To address this gap between real-world scenarios and current research, we first formulate the continual cognitive state modeling problem. To mitigate the catastrophic forgetting problem, we propose a model-agnostic framework called Diagnosis-guided Response Replay (DiagRe). DiagRe utilizes a response replay buffer to cache important previous response logs when new response data arrives. In each training iteration, new response logs are combined with cached responses for the current stage of training. Simultaneously, to prevent losing predictive knowledge when retraining on previous logs, we introduce a prediction regularization term that encourages the prediction module to replicate its original inferences at each training stage. In this way, DiagRe can effectively combat the forgetting problem during continual learning from incoming response logs. Extensive experiments on real-world data validate DiagRe’s superiority and robustness for continual cognitive state modeling.
Cognitive diagnosis / Catastrophic forgetting / Response replay
Higher Education Press 2026
/
| 〈 |
|
〉 |
PDF
(8573KB)
