💻 computer science

A Neurosymbolic Multiagent Framework for Culturally Responsive Personalised Explainable and Fair AI Tutoring

This conceptual paper introduces KAIROS, a neurosymbolic multi-agent framework that ensures culturally responsive, fair, and explainable AI tutoring by formally decoupling pedagogical decisions from cultural presentation within a constrained decision-making process, while specifying a rigorous evaluation protocol to validate its theoretical guarantees without presenting empirical results.

Original authors: Nikita Srivastava, Deepa Mehta

Published 2026-08-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nikita Srivastava, Deepa Mehta

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the classroom, the most powerful tool a teacher possesses is not a textbook or a whiteboard, but the ability to know exactly what a student understands at any given moment. For decades, educators have known that when instruction is tailored to a learner's specific needs, students can master material far faster than they would in a standard group setting. This idea, often called the "zone of proximal development," suggests that learning happens best when a teacher offers a challenge that is just slightly above what the student can do alone, providing the right hint or example to bridge the gap. Today, artificial intelligence promises to bring this kind of one-on-one tutoring to millions of people at once. Large language models, the technology behind many modern chatbots, can hold fluent conversations and explain complex ideas in seconds. However, while these systems are impressive conversationalists, they often struggle to make the right teaching decisions. They may not know what a student actually knows, they might treat different groups of students unfairly, and when they make a mistake, they cannot explain why they chose a particular path. Without these safeguards, an AI tutor risks reinforcing stereotypes or lowering expectations for certain students, turning a tool for empowerment into one of subtle discrimination.

Researchers Nikita Srivastava and Deepa Mehta from Banaras Hindu University have proposed a new design for an AI tutoring system called KAIROS, intended to solve these problems not by patching them up, but by rebuilding how the system thinks. This is a conceptual paper, meaning the authors have not yet built a working version of the system or tested it on real students. Instead, they have laid out a precise blueprint and a rigorous plan for how such a system should be constructed and tested in the future. Their central argument is that the four qualities needed for a trustworthy tutor—personalization, cultural responsiveness, fairness, and explainability—are deeply connected. If you try to fix one without considering the others, the system will fail. For instance, if an AI tries to be culturally responsive by changing the examples it uses, it might accidentally change the difficulty of the lesson, effectively lowering the bar for some students while keeping it high for others. To prevent this, the authors propose a framework that strictly separates what is taught from how it is presented.

The core of the KAIROS framework is a two-step process for every interaction. First, the system decides on the "pedagogical core," which is the unchangeable educational substance of the lesson. This includes identifying which specific skill the student needs to practice, determining the appropriate difficulty level, and choosing the type of teaching move, such as offering a hint, asking a diagnostic question, or providing a worked example. This decision is made based on a detailed, constantly updated map of what the student knows and what they are ready to learn next. Crucially, this part of the system is bound by strict rules to ensure fairness: every student at the same level of understanding must receive the same core instruction, regardless of their background. The second step is the "presentation transform." This is where the system adapts the lesson to the student's culture and personal style. It might change the language used, swap a generic example for one drawn from the student's community, or adjust the tone to be more collaborative or more formal. The innovation here is that the system is designed so that these cultural adjustments can never alter the difficulty or the core concept being taught. The cultural adaptation is confined to the surface of the lesson, ensuring that responsiveness never slips into differential expectations.

To make these decisions, the system relies on a team of specialized digital agents working together, rather than a single black-box computer program. One agent maintains a "learner model," which is a calibrated estimate of the student's knowledge that includes a measure of uncertainty. If the system is unsure about what a student knows, it is designed to ask a question to clarify rather than guessing. Another agent retrieves specific educational content from a structured database to ensure the facts are correct. A third agent, the "Guardrail," acts as a safety inspector. Before any lesson is delivered, this agent checks that the cultural adaptation has not accidentally changed the difficulty of the problem and that the teaching move is fair compared to what other students in similar situations would receive. If the lesson fails these checks, it is sent back for revision. Because every step of this process is recorded in a clear, inspectable log, the system can provide a "provenance trace" as an explanation. Instead of offering a vague, post-hoc story about why it made a choice, the system can show the exact data points, the specific skill it targeted, and the fairness checks it passed that led to the decision.

The authors are careful to state that this is a design proposal, not a finished product. They have not yet run experiments with human students to prove that KAIROS works better than current methods. Instead, they have defined a set of new metrics to measure success, such as how well the system adapts to culture without changing the difficulty, and how faithfully its explanations match its actual decisions. They have also outlined a detailed plan for a future study that would test the system using simulated learners first, followed by a large, multi-site experiment with real students. The researchers argue that by separating the "what" from the "how" and enforcing strict fairness rules on the core instruction, it is possible to create an AI tutor that is both deeply personal and rigorously fair. The goal is to ensure that as artificial intelligence becomes a common part of education, it respects the dignity and potential of every learner, adapting to their background without ever lowering the standard of what they are expected to achieve.

Technical Summary: A Neurosymbolic Multiagent Framework for Culturally Responsive Personalised Explainable and Fair AI Tutoring

Problem Statement
The paper identifies a critical gap in the deployment of Large Language Model (LLM) tutors. While LLMs have made one-to-one dialogue tutoring scalable, they currently lack four essential properties required for educational trustworthiness:

Unvalidated Personalisation: They often operate statelessly without a persistent, calibrated learner model, failing to maintain instruction within the Zone of Proximal Development (ZPD).
Superficial Cultural Responsiveness: Current systems "localise" culture only at the surface level (e.g., translation, name swapping) rather than engaging with substantive culturally responsive pedagogy.
Post-hoc Explainability: Explanations are typically generated after the fact using attribution methods (e.g., LIME, SHAP) that approximate black-box decisions rather than revealing the actual causal logic.
Reactive Fairness: Fairness is usually measured as an outcome disparity after deployment rather than enforced as a constraint during the decision-making process.

The authors argue these are not isolated engineering gaps but a single design problem where objectives interfere: cultural adaptation can inadvertently become differential expectation, and personalisation can amplify group disparities.

Methodology: The KAIROS Framework
The paper proposes KAIROS, a neuro-symbolic, multi-agent framework that treats tutoring as a constrained Partially Observable Markov Decision Process (POMDP). The core innovation is the decomposition of a tutoring action ( $a_t$ ) into two distinct components:

Pedagogical Core ( $a_t^{ped}$ ): Determines what is taught (target skill, move type, difficulty).
Presentation Transform ( $a_t^{pres}$ ): Determines how it is framed (examples, register, language, cultural context).

This decomposition allows cultural responsiveness to operate freely within the presentation layer while strictly constraining the pedagogical core to ensure fairness and content invariance.

Key Architectural Components
The system is organized as a blackboard/mediator architecture with six specialized agents:

Learner-Model Agent (LMA): Maintains a calibrated, uncertainty-aware mastery belief using Graph-Regularised Knowledge Tracing (GRKT). This model fuses attentive sequence modeling with message passing over a domain knowledge graph, enforcing prerequisite monotonicity and calibration to ensure internal consistency.
Retrieval Agent (RA): Plans and executes retrieval over a content store indexed by the knowledge graph, ensuring evidence is pedagogically typed and traceable.
Pedagogy Agent (PA): The mediator that selects the pedagogical core using a Kairos score. This score ranks actions based on expected learning gain, ZPD fit, uncertainty reduction, and frustration risk, subject to constraints.
Cultural Adaptation Agent (CAA): Realises the presentation transform based on a soft, learner-authored cultural schema. Crucially, this schema conditions only the presentation, not the pedagogical core, preventing cultural adaptation from altering difficulty or content.
Explanation Agent (EA): Assembles a provenance trace ( $\tau_t$ ) from the inspectable objects (belief, frontier, Kairos factors, evidence) that generated the decision, rather than generating a post-hoc rationalisation.
Guardrail Agent (GA): A verifier that checks every proposed action for content invariance (via bidirectional entailment and difficulty estimation), treatment parity, safety, and factuality before the action is committed.

Formal Constraints and Fairness
The framework enforces fairness as a policy constraint rather than an outcome metric:

Content Invariance: The presentation transform must preserve the epistemic content and difficulty of the pedagogical core.
Equalised Pedagogical Treatment (EPT): At matched learner states, the distribution of pedagogical-core actions must be invariant across groups.
Equalised Learning Gain (ELG): Bounded disparity in normalised learning gains across groups.

Results and Claims
The paper explicitly states it is a conceptual contribution and presents no empirical results.

No Implementation: The authors do not report on a deployed system or experimental data.
Proposed Evaluation: The paper specifies a rigorous, pre-registered evaluation protocol involving simulated learners (calibrated on public datasets like EdNet and ASSISTments) and a multi-site randomised controlled trial (RCT) with human participants.
Metrics Defined: Six primary metrics are defined for future testing, including Normalised Learning Gain, Learner-Model Calibration (ECE), Cultural Appropriateness Index (CAI), ZPD-Adherence Rate (ZAR), Treatment-Parity Gap (TPG), and Explanation-Faithfulness Score (EFS).
Complexity Analysis: Theoretical and micro-benchmark analysis suggests the governance overhead (GRKT update, scoring, guardrail checks) is negligible (~5ms per turn) compared to LLM decoding costs.

Significance and Scope
The paper's significance lies in its architectural claim: that the tension between cultural responsiveness, fairness, personalisation, and explainability can be resolved by decomposing the tutoring action and constraining its parts.

Design over Implementation: The authors argue that these are design problems decided by how actions are represented, not merely by model training. Therefore, the framework is presented as a specification that can be "tested and falsified."
Methodological Shift: It proposes a shift from treating fairness as a post-hoc audit to enforcing it as a constraint on the policy, and from post-hoc explanation to intrinsic, faithful provenance tracing.
Ethical Positioning: The framework explicitly addresses the risk of essentialism by making cultural schemas soft, learner-authored, and uncertainty-aware, confined strictly to the presentation layer to prevent lowered expectations for specific groups.

In summary, KAIROS is a formal specification for a tutoring system that aims to make AI trustworthy by separating the "what" (fair, constrained pedagogy) from the "how" (culturally responsive presentation), supported by a neuro-symbolic architecture that ensures decisions are inspectable, calibrated, and auditable. The paper serves as a protocol for how such a system should be built and evaluated, rather than a report on a system that has already been built.

Technical Summary: A Neurosymbolic Multiagent Framework for Culturally Responsive Personalised Explainable and Fair AI Tutoring

More like this