AdvisingWise: Supporting Academic Advising in Higher Education Settings Through a Human-in-the-Loop Multi-Agent Framework¶
🕒 Published (v1): 2025-11-07 20:55 UTC · Source: Arxiv · link
Ask a follow-up
Open an assistant pre-loaded with this paper's context.
💬 Ask ChatGPT✦ Ask Claude
TL;DR¶
AdvisingWise is a human-in-the-loop multi-agent system for academic advising that automates query processing, information retrieval, and response drafting while requiring advisor review before any response reaches students. Built on LangGraph with a three-phase architecture (query preprocessing → ReAct-style information collection → quality-controlled generation), it achieves 84.2% high-accuracy responses on expert-evaluated queries and improves advisor willingness to adopt AI tooling after direct use.
Problem¶
High student-to-advisor ratios in higher education create bottlenecks during peak advising periods. Existing LLM-based advising tools either over-automate (cutting out human judgment and trust) or under-deliver on reliability and personalization, failing to handle multi-source inference queries or maintain student-specific context across sessions.
Method¶
A three-phase multi-agent pipeline implemented in LangGraph, using a tiered model strategy (GPT-4o-mini for lightweight preprocessing agents, GPT-4o for answer generation and quality control):
- Query Preprocessing: Three agents—Academic Info Extractor (stores student profile as key-value pairs), Query Rewriter (contextualizes queries using conversation history), and Query Classifier (filters off-topic queries).
- Information Collection: A ReAct-style loop between a Thought Agent and an Action Agent, running up to 4 iterations. Actions are:
search_knowledge_base(MongoDB vector store with AzureOpenAI embeddings),search_course_db(exact lookup),search_web(Tavily + GPT-4o relevance filter),request_student_info(LangGraph human-in-the-loop pause), orprovide_answer. Institutional sources are prioritized for policy queries; student prompting is used for personalization gaps. - Response Generation: Answer Generator produces a five-component structured response (completeness assessment, confidence score, evidence-based answer, source references, identified gaps). Quality Controller validates factual accuracy against retrieved documents, checks citations, and flags hallucinations.
All draft responses—including advisor-only uncertainty notes—route through a Rocket.Chat interface for one-click approval or editing before delivery to students.
Key Contributions¶
- Human-in-the-loop system design that keeps both students (context provisioning) and advisors (output validation) in the workflow loop, not as optional overseers.
- Three-phase multi-agent architecture combining profile-aware query preprocessing, iterative ReAct retrieval, and hallucination-checked response generation.
- Tiered LLM cost strategy: cheaper model (GPT-4o-mini) for retrieval/classification agents, stronger model (GPT-4o) for final generation.
- Mixed-methods evaluation combining expert accuracy assessment, LLM-as-a-judge pairwise retrieval comparison, and a qualitative user study with 8 advisors.
Results¶
- Expert evaluation on 20 sampled queries (of 100 synthesized): 84.2% (16/19 evaluable) achieved high accuracy; lower-scoring cases reflect conservative refusal-to-answer under uncertainty rather than hallucination.
- LLM-as-a-judge pairwise comparison: ReAct-style retrieval outperformed standard RAG baseline, with the largest margin on Handbook-Implicit queries (multi-source inference required).
- User study (\(n=8\) advisors): advisors' initial concerns about reliability and personalization diminished after direct system use; willingness to adopt increased post-interaction.
Limitations¶
- Deployed and evaluated only for CS graduate advising at a single private U.S. university; generalizability to other programs/institutions is untested.
- Expert evaluation covers only 20 of 100 benchmark queries, and 1 query was unevaluable, leaving coverage gaps.
- ReAct loop is capped at 4 iterations, which may be insufficient for highly ambiguous or multi-step queries.
- No longitudinal evaluation of advisor trust or student outcomes; user study reflects short-term lab perceptions.
- Relies on GPT-4o/GPT-4o-mini and external services (Tavily, MongoDB/Azure), introducing cost, latency, and vendor-dependency concerns not analyzed.
- Advisor review adds latency; the paper does not measure response time relative to unassisted advising.
Relevance to Harnesses / Meta-Harnesses¶
AdvisingWise is a domain-specific multi-agent harness that pipelines distinct specialist agents—preprocessor, rewriter, classifier, reasoner, retriever, generator, quality controller—under a coordinating orchestration layer (LangGraph) with explicit human checkpoints. The tiered model routing (cheap agents for classification/retrieval, capable models for synthesis/QC) is a canonical harness pattern for managing cost-quality tradeoffs across agent roles. The ReAct-style inner loop with bounded iterations and action-type heuristics illustrates how domain constraints can be baked into a meta-harness's control flow rather than left to unconstrained LLM planning. The advisor-in-the-loop gate demonstrates a mandatory human checkpoint pattern that any safety-critical harness must support.