InqEduAgent: LLM-Driven Adaptive Learning

• The paper proposes an LLM-empowered framework that adaptively matches learning partners using Gaussian process regression and Pareto-front optimization.
• It models learners with a two-dimensional trait vector capturing subject and reasoning preferences to enable complementarity in inquiry-oriented collaboration.
• Experimental findings show improved collaborative gains over static methods, highlighting effectiveness across STEM, Humanities, and Social Sciences.

InqEduAgent is an LLM-empowered agent framework for inquiry-oriented education that models, simulates, and selects learning partners tailored to a learner, an exercise set, and a knowledge domain. Its central claim is that collaborative partnership in inquiry-oriented learning should be treated as an adaptive matching problem rather than as ad hoc pairing, fixed tutoring, or simple performance grouping. In this framework, a “learning partner” is not merely a tutor but a co-learner who exchanges interpretations and reasoning, and partner quality is estimated through a combination of generative learner agents, Gaussian process augmentation, and Pareto-front optimization over multi-domain performance profiles (Zhao et al., 5 Aug 2025).

1. Problem formulation and educational scope

InqEduAgent is situated in inquiry-oriented learning, where questioning, explanation, hypothesis testing, reflection, metacognitive regulation, and social co-construction are treated as central components of learning rather than peripheral supports. Within this setting, the framework addresses three limitations identified in prior partner-selection practice: experience-based or random pairing by instructors, rule-based or fixed AI assistants that behave as static tools rather than collaborators, and simple co-learning heuristics that ignore multidimensional learner traits and domain-specific prior knowledge patterns (Zhao et al., 5 Aug 2025).

A defining feature of the framework is its treatment of partner assignment as a prediction problem over incremental gain. The target question is not which partner is strongest in an absolute sense, but which partner, for a particular learner and domain, is most likely to improve post-communication performance on the relevant exercises. This shift matters because inquiry-oriented collaboration depends on complementarity in subject preference, reasoning style, and prior knowledge, not only on aggregate competence.

The educational role of InqEduAgent is therefore dual. First, it can recommend human partners by using historical performance and domain features to predict which peer is likely to help a target learner on upcoming tasks. Second, it can guide the design of AI learning partners by assigning personas such as a STEM-focused inductive reasoner or a humanities-oriented deductive reasoner, then estimating which such AI profile best complements a given learner (Zhao et al., 5 Aug 2025).

A common misconception is to treat InqEduAgent as a tutoring chatbot. The framework is narrower and more specific: its primary contribution is adaptive partner allocation and collaboration modeling. Direct instruction may still occur inside the interaction protocol, but the framework’s core object of optimization is partner matching rather than stand-alone tutoring.

2. Learner representation and generative partner modeling

InqEduAgent represents each generative learner agent with a two-dimensional characteristic vector

$\mathbf{A} = (\mathbf{A}_1, \mathbf{A}_2),$

where subject preference and logical preference each take values in $\{-1,0,1\}$. Subject preference $\mathbf{A}_1$ encodes bias toward Social Science and Humanities $(-1)$, no obvious preference $(0)$, or STEM $(1)$. Logical preference $\mathbf{A}_2$ encodes deductive reasoning $(-1)$, intuitive reasoning with no preference $(0)$, or inductive reasoning $(1)$. A learner is then instantiated as

$\{-1,0,1\}$0

with numeric trait codes mapped to natural-language persona descriptions for prompting the LLM agent (Zhao et al., 5 Aug 2025).

Exercises are represented by knowledge concept features

$\{-1,0,1\}$1

which conceptually encode domain, difficulty, and knowledge aspects. Solo learner–exercise interaction yields a response triplet with correctness $\{-1,0,1\}$2, while post-communication interaction records the learner, the partner, the stem without options, and the learner’s post-communication correctness. The use of the question stem without options is deliberate: it is intended to reduce bias from direct answer-choice exposure during partner exchange.

Prior knowledge is represented in two coupled forms. The first is a domain-score vector

$\{-1,0,1\}$3

where $\{-1,0,1\}$4 is learner $\{-1,0,1\}$5’s score in domain $\{-1,0,1\}$6. The second is an embedding-based representation in which learner traits are mapped to a short-dimensional vector and exercise or domain features are mapped to high-dimensional BERT embeddings; these are concatenated for downstream Gaussian-process modeling (Zhao et al., 5 Aug 2025).

This representation is intentionally compact. It does not attempt to encode a full latent cognitive state or a detailed knowledge-tracing model. Instead, it constructs a generative approximation of learner diversity that is rich enough to support domain-sensitive collaboration prediction while remaining tractable for partner matching.

3. Gaussian process augmentation and Pareto-front partner selection

The adaptive matching algorithm proceeds in five stages: domain grouping, training-sample construction, Gaussian-process regression, posterior prediction for candidate partners, and Pareto-front filtering. For a learner pair $\{-1,0,1\}$7 in a domain block $\{-1,0,1\}$8, the target variable is the average domain-level gain produced by communication: $\{-1,0,1\}$9 where $\mathbf{A}_1$0 after embedding. Here $\mathbf{A}_1$1 is the learner’s solo correctness and $\mathbf{A}_1$2 is the post-communication correctness after interacting with partner $\mathbf{A}_1$3 (Zhao et al., 5 Aug 2025).

The regression layer assumes

$\mathbf{A}_1$4

with Gaussian-process prior

$\mathbf{A}_1$5

The kernel is an RBF kernel, chosen because the framework treats collaboration gain as a smooth nonlinear function over an embedding space that combines learner traits and domain features. The model then predicts a Gaussian posterior for a new candidate pairing $\mathbf{A}_1$6, producing both a mean gain estimate and a variance estimate (Zhao et al., 5 Aug 2025).

The partner-selection layer adds a multi-objective filter. A learner score vector $\mathbf{A}_1$7 dominates $\mathbf{A}_1$8 if it is at least as good in all domains and strictly better in at least one. Candidate partners are restricted to a Pareto front, after which the framework selects the partner with maximal GP-predicted gain: $\mathbf{A}_1$9 Two variants are defined. InqEduAgent-GP constructs a global Pareto front across all learners and all domains. InqEduAgent-LP constructs a local Pareto front for each learner from partners with whom that learner has already communicated (Zhao et al., 5 Aug 2025).

The use of Gaussian processes is not incidental. The framework argues that learner and exercise characteristics empirically follow patterns close to normal distributions, making GP-based nonparametric regression a natural fit. It also emphasizes three GP properties that are especially relevant in educational settings: flexibility without a fixed parametric form, uncertainty quantification under sparse data, and strong performance with relatively small numbers of observations per domain.

4. Interaction protocol, LLM environment, and matching dynamics

InqEduAgent uses LLMs in three roles: simulating learner agents, mediating learner–learner communication, and serving as the reasoning engine for experiments. Agents first answer CMMLU multiple-choice questions independently. They then enter an inquiry-style communication protocol in which each agent reasons over the question stem without options, exchanges its reasoning text with a partner, summarizes the partner’s reasoning in light of its own cognitive style, and finally re-answers the full multiple-choice question using both its own and its partner’s insights (Zhao et al., 5 Aug 2025).

The experimental environment uses Qwen-32B and DeepSeek-32B as the underlying LLMs. To estimate accuracy, answers are sampled 10 times and averaged. This setup is used to examine whether the matching policy remains effective across different underlying LLM capability levels rather than only for one fixed model family.

The framework also adopts what it calls a “mirror mapping strategy.” Numeric trait codes such as $(-1)$0 are translated into natural-language persona prompts for the LLM agents, while semantic learner and domain information are mapped back into numeric embeddings for Gaussian-process modeling. This closes a semantic-to-numeric-to-semantic loop: traits determine behavior in language space, and observed behavior is reinterpreted in numeric form for probabilistic partner selection (Zhao et al., 5 Aug 2025).

A second common misconception is that the interaction protocol mainly evaluates answer exchange. In fact, the protocol is designed to probe collaboration under reasoning transfer. By suppressing answer options during the exchange phase and forcing agents to summarize one another’s reasoning, the framework treats the communication step as a structured inquiry interaction rather than a shortcut to answer disclosure.

5. Evaluation protocol and empirical findings

The evaluation uses CMMLU, a multitask Chinese benchmark spanning 67 knowledge domains across STEM, Social Science, Humanities, China-specific domains, and other areas. A total of 11,582 exercises are first evaluated with Qwen-32B and DeepSeek-32B. The analysis reports that about 61.1% of exercises are solved perfectly, about 82.8% are passed with accuracy at least $(-1)$1, and 17.2% remain unsolved; from the lower-accuracy regions, six representative domains are selected: Machine Learning $(-1)$2, College Engineering Hydrology $(-1)$3, Marketing $(-1)$4, High School Geography $(-1)$5, Arts $(-1)$6, and Logic $(-1)$7 (Zhao et al., 5 Aug 2025).

The benchmark compares several conditions: Baseline (no roles, no co-learning), SLM (self-learning model with roles but no interaction), CLM (co-learning with random partner matching), InqEduAgent-GP, InqEduAgent-LP, and neural-regression replacements for the GP variants. Three metrics are used: Mean Gain, Best Gain, and Std. Mean Gain captures average improvement in accuracy under co-learning, Best Gain captures the strongest observed improvement under optimal matching, and Std measures variability across experiments (Zhao et al., 5 Aug 2025).

The main result is that GP-based adaptive matching improves over both self-learning and random co-learning. Overall Mean Gain rises from 0.2704 for Baseline to 0.2655 for SLM, 0.2935 for CLM, 0.3030 for InqEduAgent-GP, and 0.3047 for InqEduAgent-LP. The local-Pareto variant is the best overall on this metric. Domain-specific patterns are more differentiated: InqEduAgent-LP achieves the highest STEM Mean Gain at 0.3871, whereas InqEduAgent-GP achieves the highest Humanities Mean Gain at 0.3502 and the highest Social Science Best Gain at 0.3462. Stability is also strong for the GP models, with overall Std 0.0319 for InqEduAgent-GP and 0.0325 for InqEduAgent-LP. The neural variants achieve lower mean performance; the paper notes that LP(NN) attains lower overall Std at 0.0214 but with higher variance in certain domains, including Humanities at 0.0606, which it interprets as a sign of possible overfitting (Zhao et al., 5 Aug 2025).

The component-removal experiment isolates the value of each stage. Total accuracy rises from 26.00% for Baseline to 26.55% for SLM, 29.35% for CLM, 30.30% for InqEduAgent-GP, and 30.47% for InqEduAgent-LP. Relative to CLM, the GP-based partner-selection layer adds further gain; relative to SLM, the co-learning step itself contributes about 2.8 percentage points. The framework therefore attributes performance improvements to two distinct mechanisms: collaboration itself, and adaptive partner selection over collaboration (Zhao et al., 5 Aug 2025).

6. Position within educational agent research, applications, and limitations

InqEduAgent occupies a distinct position within the 2025–2026 literature on educational agents. Unlike CodeEdu, which organizes Planner, Researcher, Tutor, Programmer, and Report Analyst roles to support proactive coding education and debugging workflows, InqEduAgent is centered on the selection of learning partners rather than on a tutoring pipeline (Zhao et al., 18 Jul 2025). Unlike EduAgentQG, which uses Planner, Writer, Solver, Educator, and Checker roles to generate personalized mathematics questions from structured educational goals, InqEduAgent does not generate assessment artifacts as its primary output; it predicts collaboration gain and allocates partners (Jia et al., 8 Nov 2025). At a broader systems level, its emphasis on learner-level pairing is complementary to agentic educational frameworks such as AUSS, which integrates student-level personalization, educator-level automation, and institutional-level intelligence, and to the inclusive ecosystem perspective that treats learning, teaching, administration, and accessibility as a coordinated multi-agent platform (J et al., 17 Apr 2026, Sudarshan et al., 14 May 2026).

The framework is also related to learner-simulation lines. EduAgent models fine-grained student behavior using cognitive priors, while Edu-Theater uses cohort-aware roll-call and teacher-centered simulation to generate scalable learner-task interaction data; both address learner modeling, but neither is organized around Pareto-filtered partner allocation in inquiry-oriented collaboration (Xu et al., 2024, Gao et al., 13 Jun 2026). EduVerse provides a user-defined classroom simulation space with longitudinal cognition–interaction–evolution dynamics, which suggests a larger simulation ecology into which InqEduAgent-style partner matching could be embedded (Ma et al., 7 Oct 2025). A plausible implication is that InqEduAgent can function as one subsystem in a broader agentic classroom architecture: matching peers or AI partners inside a simulated or live inquiry environment.

The framework’s practical applications include online guided-inquiry platforms, smart classrooms, and AI co-learner assignment. The paper explicitly argues that its matching mechanism can support the intelligent allocation of human-based learning partners and the design of AI-based learning partners. Code, data, and appendix materials are reported as publicly available at https://github.com/InqEduAgent/InqEduAgent (Zhao et al., 5 Aug 2025).

Its limitations are equally explicit. First, the experiments use simulated learners, not human participants, so real-world validation remains open. Second, the GP layer has cubic scaling in the number of training points, making large-scale deployment computationally expensive. Third, BERT-based embeddings and LLM-mediated communication are costly. Fourth, the local-Pareto variant faces a cold-start problem for learners with little or no history. Fifth, the discretization of subject and reasoning style into three levels per axis is a coarse approximation of real learner diversity. Finally, fairness and interpretability remain unresolved: partner selection could systematically privilege high-performing students as desirable collaborators, and the combination of high-dimensional embeddings with black-box LLM behavior may make recommendations difficult to explain (Zhao et al., 5 Aug 2025).

These limitations clarify both the contribution and the current boundary of the framework. InqEduAgent demonstrates that inquiry-oriented partner selection can be formalized as a probabilistic, domain-sensitive, multi-objective matching problem. It does not yet establish that such matching is pedagogically validated in real classrooms, but it provides a concrete mechanism for moving from static study-partner assignment toward adaptive, data-driven human–AI and human–human collaboration.