Introduction
Inquiry-oriented education, which prioritizes critical thinking, active questioning, and self-regulated learning, is fundamentally collaborative. However, identifying an ideal study partner remains a significant challenge. Traditional classroom pairings often rely on heuristics or random assignment, while digital AI "tutors" typically function as rigid, one-sided information sources rather than dynamic peers.
This research addresses the gap in Human-AI Co-Learning. Current systems often treat AI as a "supertool" or an authoritative teacher, neglecting its potential as a teammate with distinct cognitive biases and strengths. The InqEduAgent framework aims to simulate human social behaviors and facilitate two-way knowledge exchange by matching human learners with AI "co-learners" optimized for specific academic tasks.
Methods
The InqEduAgent framework shifts the paradigm from "AI-as-teacher" to "AI-as-peer" through three primary components:
1. Agentic Personas
Agents are assigned specific "personalities" defined by two dimensions: Subject Preference (e.g., STEM, Social Sciences, Humanities) and Logical Preference (e.g., Deductive vs. Inductive reasoning). This allows the system to simulate a diverse classroom environment.
2. Inquiry-Based Interaction
The framework employs a four-step communication protocol to foster deep learning:
- Independent Thinking: The agent processes the problem individually.
- Exchange: Thinking contents are shared between partners.
- Summarization: A joint perspective is synthesized.
- Re-evaluation: The final answer is derived from the collaborative process.
3. Gaussian Process (GP) Augmentation and Selection
Gaussian Process (GP) modeling serves as the system's mathematical engine, mapping learner characteristics and exercise difficulty into a high-dimensional space to predict partner effectiveness. To select the optimal partner, the system utilizes a Pareto-based selection algorithm, identifying "non-dominated" candidates who provide the best trade-offs across multiple knowledge domains.
Datasets
The study utilizes the CMMLU (Chinese Massive Multitask Language Understanding) benchmark, covering 67 knowledge domains. For controlled experimentation, the authors curated the RKD (Representative Knowledge Domains) set, focusing on 41 domains across STEM, Social Science, and Humanities, including high-difficulty areas like Machine Learning and College Engineering Hydrology.
Key Results and Insights
- Performance Gains: The InqEduAgent-LP (Local Pareto) model achieved an overall accuracy of 30.47%, a +4.47% improvement over the baseline of non-collaborative agents.
- STEM Efficacy: The model was notably effective in STEM subjects, showing performance increases exceeding 10%.
- GP vs. Neural Networks: The Gaussian Process approach outperformed standard Neural Network variants, particularly in handling uncertainty and identifying patterns within limited domain knowledge.
- Stability: While the Local Pareto version offered higher peak performance by adapting to individual history, the Global Pareto (GP) version demonstrated greater stability (lowest standard deviation of 0.0319), making it a reliable choice for general applications.
Limitations and Future Work
Despite promising results, challenges remain regarding data scarcity in specific Humanities sub-domains and the inherent complexity of human learners (e.g., emotional intelligence and cultural background). Future research will explore expanding persona dimensions and integrating AI partners into physical classrooms to support "triangular" human-human-AI collaboration.
Summary updated