Jie Zhou

Large language models (LLMs) offer new opportunities for lesson planning, remediation, and personalized practice, but deploying them in real educational settings remains challenging for two reasons. First, direct use of learner interaction traces can expose sensitive information about student knowledge states and behavioral patterns. Second, unconstrained generation can produce recommendations that are pedagogically inconsistent with the adopted curriculum, such as skipping prerequisite concepts or drifting outside the prescribed textbook scope. We propose PrivPath, a stage–subject–textbook-aligned framework for privacy-preserving curriculum planning that explicitly separates on-device learner modeling from server-side content generation. Its core module, Tri-Index Private Path Planning (TIPP), first restricts planning to a scoped curriculum graph defined by the target educational context, then privatizes learner mastery summaries under local differential privacy before transmission, and finally computes a curriculum-feasible instructional path that is coupled with evidence-grounded constrained generation. This design preserves personalization while reducing reliance on raw student traces and improving controllability of the generated teaching sequence. Experiments on three educational interaction datasets paired with open-textbook curriculum graphs show that PrivPath substantially improves structural validity and privacy robustness. In particular, it raises graph feasibility from 82.5% to 99.2% relative to TriIndex-RAG, improves the offline pedagogical utility proxy $\Delta$AUC from 0.013 to 0.018, and lowers membership inference AUC from 0.74 to 0.52 at ${\epsilon }_{\mathrm{loc}}=1.0$. These results suggest that curriculum-aligned path planning, privacy-preserving learner adaptation, and practically useful LLM-based educational generation can be achieved within a unified framework.

Educational platforms increasingly rely on vector databases to store and retrieve embedding representations of large-scale learning corpora (e.g., lecture notes, assignments, feedback, and student Q&A) for retrieval-augmented generation and analytics. However, directly indexing educational text embeddings raises privacy risks (student identities, sensitive performance signals, and protected attributes) and creates a management challenge: embeddings drift as curricula evolve, access policies change, and new content arrives continuously. This paper studies privacy-preserving intelligent management of educational-corpus vector libraries and proposes a novel, end-to-end algorithmic framework that jointly optimizes (i) privacy leakage control, (ii) retrieval quality, and (iii) operational efficiency under streaming updates. We introduce a hierarchical policy-aware vector lifecycle model, a privacy budget scheduler for adaptive re-embedding and re-indexing, and a secure-aware clustering-and-routing mechanism that supports fast query-time filtering with minimal accuracy loss. The resulting system, PP-EDUVec, enables compliant similarity search across multi-tenant educational data while automatically maintaining index health (freshness, redundancy, and utility) over time. On the EDU-Mix benchmark, PP-EDUVec achieves Recall@10 $=0.835$ while reducing representation leakage (LeakRep) from $0.215$ to $0.136$ ($-36.7\%$) and access-pattern leakage (LeakAP) from $0.398$ to $0.255$ ($-35.9\%$), and lowering mean latency from $42.1$ ms to $33.4$ ms ($-20.7\%$) and weekly maintenance time from $55.0$ to $35.8$ min/week ($-34.9\%$) compared with PostFilter.