Quote of the day

A person who never made a mistake never tried anything new.

- Albert Einstein

Contact

Pedagogically Aligned LLM Tutors with Answer-Leakage Control for Engineering Mathematics

Open Education

Pedagogically Aligned LLM Tutors with Answer-Leakage Control for Engineering Mathematics

Pedagogically Aligned LLM Tutors with Answer-Leakage Control for Engineering Mathematics

Description

Design an LLM tutor that provides guided support without revealing full answers and evaluate its robustness against student strategies intended to extract direct solutions.

Details

Context and Problem Statement

A tutor that immediately gives complete solutions may reduce learning despite high short-term satisfaction. LLM tutors can be pressured into revealing answers through repeated reformulation or deceptive requests for verification.

Research Question

How can an LLM tutor be aligned with explicit pedagogical principles while remaining robust to answer-extraction strategies, and what effect does this constraint have on learner engagement?

Proposed Approach

Use preference-based alignment with a reward combining scaffolding quality, error diagnosis, and non-disclosure of final answers. Build an adversarial benchmark of answer-extraction strategies and evaluate the tutor with engineering-mathematics students.

Expected Contribution

A French-language benchmark for answer-leakage robustness and a quantified trade-off between pedagogical firmness and learner acceptance.

Expected Prototype

An LMS-integrated tutor providing progressive hints, error diagnosis, and teacher-facing analytics on recurring misconceptions.

Datasets

MathDial, academic tutor-evaluation datasets, and consent-based classroom dialogues.

Challenges

Human-evaluation cost, automated-judge bias, student-data protection, and learner frustration.

Research Question

Can a pedagogically aligned tutor resist sustained answer-extraction attempts without reducing learner engagement to an unacceptable degree?

Innovation

The project introduces adversarial evaluation of pedagogical behavior by treating answer extraction as a robustness problem.

Expected Deliverable

An aligned and evaluated tutor, a French answer-extraction benchmark, and a classroom-oriented user study.

Technologies

Large Language Models and Preference Alignment Reinforcement Learning from Human Feedback Automated Pedagogical Evaluation Adversarial Testing LMS Integration (LTI, Moodle)

Required Skills

  • Natural Language Processing
  • Model Alignment and Preference Learning
  • Human-Participant Evaluation Methodology
  • Web Development and LMS Integration

Datasets

  • MathDial
  • Academic tutor-evaluation datasets
  • Consent-based classroom dialogue corpus

Morocco & Africa Relevance

Large cohorts in Moroccan preparatory classes and engineering schools limit individualized support; a tutor that guides reasoning without replacing it can scale formative assistance.