Ken Utilization Layer: Hebbian Replay Within a Student's Ken for Adaptive Exercise Recommendation

TL;DR

This work addresses the challenge of real-time, few-shot exercise recommendation under sparse supervision by introducing KUL-Rec, a memory-augmented architecture grounded in Complementary Learning Systems. It combines a fast Hebbian memory (Notebook) with a slow-learning Student network and a Loss-aligned Internal Target to enable continual personalization without storing raw data or relying on cohort training, and it supports both tabular and open-ended inputs through a unified embedding space. The approach is evaluated via bidirectional, rank-sensitive metrics on ten tabular KT datasets and a live 13-week short-answer deployment, demonstrating superior ranking performance, reduced memory and latency, and meaningful improvements in learner experience. The findings suggest that Hebbian replay with bounded consolidation offers a scalable, interpretable path to adaptive educational technology across modalities, while highlighting embedding robustness as a critical consideration for deployment.

Abstract

Adaptive exercise recommendation (ER) aims to choose the next activity that matches a learner's evolving Zone of Proximal Development (ZPD). We present KUL-Rec, a biologically inspired ER system that couples a fast Hebbian memory with slow replay-based consolidation to enable continual, few-shot personalization from sparse interactions. The model operates in an embedding space, allowing a single architecture to handle both tabular knowledge-tracing logs and open-ended short-answer text. We align evaluation with tutoring needs using bidirectional ranking and rank-sensitive metrics (nDCG, Recall@K). Across ten public datasets, KUL-Rec improves macro nDCG (0.316 vs. 0.265 for the strongest baseline) and Recall@10 (0.305 vs. 0.211), while achieving low inference latency and an $\approx99$\% reduction in peak GPU memory relative to a competitive graph-based model. In a 13-week graduate course, KUL-Rec personalized weekly short-answer quizzes generated by a retrieval-augmented pipeline and the personalized quizzes were associated with lower perceived difficulty and higher helpfulness (p < .05). An embedding robustness audit highlights that encoder choice affects semantic alignment, motivating routine audits when deploying open-response assessment. Together, these results indicate that Hebbian replay with bounded consolidation offers a practical path to real-time, interpretable ER that scales across data modalities and classroom settings.

Figures (3)

• Figure 1: Overview of KUL-Rec Modeling: A) A learner's longitudinal interactions with quiz questions generate sequential data typically analyzed using recurrent neural networks (RNNs) to output KC graphs of the likelihood the student will get KCs correct given the current learning state. Our proposed model augments this framework by incorporating biologically inspired dual-process memory mechanisms, enabling continual adaptation, rapid instance learning, and generalized knowledge consolidation. B) Detailed view of the Learning and Consolidation phases. During Learning, the model rapidly stores interactions using a Hebbian update with weight decay. Each mini-batch $B$ consists of questions $q_B$ and responses $r_B$, which are encoded into the Hebbian memory module $N$, forming an update $N_B$ at time step $t$. During Consolidation, stored memories are recalled and used to refine representations in the Student network $S$.
• Figure 2: Tabular dataset results.A) Model architecture of KUL-Rec used in this experiment. B) Histogram of nDCG scores across all ten datasets, showing consistent performance gains over recurrent, transformer, and graph-based baselines. Plasticity, retention, and consolidation in continual learning.C) Heatmaps of nDCG and Recall@10 across combinations of encoding rate ($\eta$) and forgetting rate ($\rho$) on ASSIST2009 and EdNet-Sm. On math tasks, lower $\eta$ and moderate forgetting ($\rho$) perform best, while language tasks benefit from higher values. Performance is relatively stable across settings. D) Effect of replay consolidation iterations. Fewer replay steps yield better generalization, supporting the view that lighter rehearsal avoids overfitting. Small improvements in nDCG result from increased patience and extended consolidation epochs.
• Figure 3: Short-answer adaptation pipeline and outcomes.A: Weekly personalized quiz generation. A Retrieval-Augmented Generation (RAG) pipeline constructs a candidate bank from learning objectives, readings, and Bloom's levels. Each student's KUL-Rec model ranks questions by predicted answer similarity, and one per objective is selected at the 66th percentile. B: Sankey diagram showing the distribution of assigned Bloom's Taxonomy (BT) levels per student over 10 weeks, highlighting divergence in personalized trajectories. C–D: Mean ± s.e.m. survey ratings of perceived difficulty and helpfulness, showing significant improvement in the adaptive quiz condition.