Skip to content

ExpertAgent: Enhancing Personalized Education through Dynamic Planning and Retrieval-Augmented Long-Chain Reasoning

🕒 Published (v1): 2025-10-08 19:03 UTC · Source: Arxiv · Venue: NEURIPS 2025 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

ExpertAgent is an LLM-based educational agent that combines RAG over a curated curriculum repository, chain-of-thought (CoT) reasoning, and a dynamically updated student model to deliver personalized, hallucination-reduced tutoring. The system proactively adapts lesson content, difficulty, and exercise recommendations after each student interaction. A small user-acceptance study (Likert-scale) reports favorable scores on usability and performance expectancy.

Problem

Existing AI-in-education (AIED) systems apply RAG, CoT, and dynamic planning in isolation: RAG lacks curricular grounding, CoT is narrowly scoped to math/logic, and dynamic planning is borrowed from robotics with little student-modeling adaptation. Together these gaps produce systems with poor long-term context, error propagation across pipeline stages, and superficial personalization.

Method

ExpertAgent follows a two-sided architecture:

  1. Teaching side — Instructional materials are chunked, embedded into a vector database, and retrieved via RAG on each student query. Retrieved context is injected into an LLM prompt that generates a CoT-structured response with source attribution.
  2. Student side — A structured student model records per-topic mastery state, exercise outcomes, and misconceptions. After each interaction the model is updated and drives subsequent planning: difficulty adjustment, targeted exercise selection, and knowledge-map visualization (states: Untouched / Learning / Mastered / Weak).

The bidirectional loop is: student query → RAG retrieval → CoT generation → student-model update → next instructional plan.

Key Contributions

  • Unified framework integrating RAG, long-chain CoT reasoning, and dynamic student modeling within a single agent.
  • Proactive instructional planning driven by a continuously updated student model rather than reactive QA.
  • Knowledge map visualization mapping topic-level mastery states for learner self-monitoring.
  • Modular architecture with document management, multi-level feedback collection, and structured trajectory logging (CSV/JSON).

Results

  • User Acceptance Model (internal volunteer study, Likert scale):
  • Performance Expectancy: 4.33 / 5
  • Effort Expectancy: 4.22 / 5
  • Facilitating Conditions: 4.22 / 5
  • Social Influence: 2.78 / 5
  • No controlled comparison against baselines; no learning-outcome metrics (accuracy, retention) are reported.

Limitations

  • No quantitative learning-outcome evaluation; only a subjective user-acceptance survey on an unreported number of volunteers.
  • No comparison against baseline systems (e.g., vanilla RAG, GPT tutoring without student model).
  • Social influence score (2.78) indicates low peer/organizational adoption momentum.
  • Retrieval latency and long-term dialogue coherence issues acknowledged but not measured.
  • Reinforcement learning for mathematical/symbolic reasoning is deferred to future work.
  • Study is single-subject/domain; generalizability across subjects and learner populations untested.

Relevance to Harnesses / Meta-Harnesses

ExpertAgent exemplifies a task-specific agent harness: it wraps an LLM with a persistent state layer (student model), a retrieval pipeline (RAG), and an orchestration loop that sequences retrieval → generation → state-update → replanning across turns. The bidirectional teaching–learning feedback loop is architecturally analogous to meta-harness patterns where a controller module maintains world-state and issues adaptive sub-task prompts. For researchers building harnesses, the student-model-as-planner pattern—where accumulated interaction history drives next-step selection rather than a fixed prompt chain—is a concrete example of stateful orchestration beyond single-turn RAG. The modular design (document management, feedback collection, trajectory logging) also mirrors harness infrastructure concerns around observability and continuous improvement.