Robust knowledge tracing via bidirectional context and conditional diffusion

Abstract

Knowledge Tracing (KT) is essential in online education for modeling learners’ evolving knowledge states to support personalized instruction. However, conventional KT approaches often exhibit instability when confronted with abrupt shifts in learner performance—such as sudden errors amid consistent success or unexpected recoveries after prolonged failure. These non-stationary response patterns challenge models that rely on local sequential updates, leading to overreaction, delayed adaptation, or inconsistent predictions. To address this, we propose DiffKT , a knowledge tracing framework that enhances robustness through holistic sequence modeling. DiffKT first introduces a Bidirectional Global Context-aware State Estimator (GCSE), which leverages the full interaction sequence—past and future—to construct stable, context-rich representations of latent knowledge states, thereby mitigating myopic bias. Building on these representations, DiffKT employs a conditional diffusion process to iteratively refine knowledge state estimates by denoising corrupted sequence-level proxies. This enables the model to distinguish transient anomalies from genuine learning signals and maintain temporal consistency across dynamic response patterns. Extensive experiments on four benchmark datasets show that DiffKT achieves competitive or superior performance compared to state-of-the-art methods in next-response prediction. Further case studies and ablation analyses confirm that the integration of bidirectional context and diffusion-based refinement jointly improves robustness to abrupt performance shifts and generalization across diverse learning trajectories.