ParaTutor: LLM Mediated Parent Child Tutoring through Role Separated Scaffolding Interface in Real Time

Synthesis: - Effective parent–child tutoring depends on preserving distinct, asymmetric roles: parents guide the learning process while children remain actively engaged in understanding and reasoning.

• Recurring challenges include parents struggling to understand problem structure, lacking sufficient knowledge to provide support, and encountering communication difficulties that disrupt shared understanding.
• Generic Large Language Models (LLMs) assistance tends to reduce the parent's tutoring role; ParaTutor better preserves parent-led support and sustains children's participation in reasoning through role-separated Scaffolding.
• In multi-user learning, the value of LLM support depends not only on model capability but on how support is distributed across users with different roles.
• Emotionally supportive language — gentle reminders, positive reinforcement, neutral phrasing such as "let's try a different way" — enhances persistence, reduces frustration, and fosters learner confidence.

Key Findings

Study Design & Method

The work began with a formative study in the Chinese home tutoring context that identified locally common math word Problem Solving procedures, common challenges in parental tutoring, and seven strategies commonly adopted by parents. ParaTutor was then developed using multi-agent chatbots, providing parents with guidance for tutoring and children with visual grounding for problem solving. The system was evaluated with 23 parent–child dyads (children aged 10–12) under four tutoring conditions that varied how LLM assistance was delivered. Design features for emotional buffering include sentence starters such as "You're on the right track," "let's think it through together," and "That's okay, let's try another way," which help parents maintain a constructive tone when children struggle or make mistakes.

What this means for practice

• Educators. Split AI support by role rather than handing the child a general chatbot. In this evaluation with 23 parent–child dyads, ParaTutor's role-separated Scaffolding gave parents strategy, language, repair, and phase scaffolds while children received visual grounding; generic LLM assistance offered useful explanations but did not consistently support parent-led tutoring or the children's active reasoning.
• EdTech designers. Time support to breakdown moments instead of offering it continuously: the 312 coded parent-facing scaffolding moments — strategy (104, 33.3%), language (96, 30.8%), repair (58, 18.6%), and phase (54, 17.3%) — clustered where parents needed help deciding how to continue.
• Educators. Treat a tutoring breakdown as a signal to change the next move, not to repeat the same explanation. In the condition without role-separated support, dyads failed to reach a justified solution within the 10-minute limit in 11 cases across five families, and parents described embarrassment and rising impatience after restating an explanation the child did not follow.
• EdTech designers. Budget for parental interpretation. Keeping parents in the guiding role still required them to read system guidance, decide when to intervene, rephrase prompts in their own words, and regulate the emotional tone of the exchange, which the authors note may be less workable for families with limited time, confidence, or patience.

Limitations

• Role separation left the tutoring labor with parents: they still had to interpret system guidance, choose when to intervene, translate prompts, and manage tone, and the authors state that parent-facing scaffolding may be less effective for families with limited time, low confidence, or limited patience.
• The evaluation captured short-term interaction with 23 parent–child dyads (children aged 10–12) using a research prototype; it did not establish longer-term effects on children's mathematical reasoning, strategy transfer, or sustained changes in family tutoring practice.
• The prototype's multi-agent pipeline, built on the DeepSeek API, responded noticeably more slowly than direct interaction with the base model, and its LLM-generated scaffolds and visual representations were occasionally incomplete or inconsistent enough to require parental verification.
• The 23 dyads were recruited through Xiaohongshu and local after-school institutions between August 2024 and May 2025 within the Chinese home tutoring context, so the interactional patterns reflect that setting and its math word-problem curriculum.

Citation

Lan Luo, Anqi Wang, Muzhi Zhou, Junhua Zhu, Jie Cai, Ao Yu, Hui Pan (2026). ParaTutor: LLM Mediated Parent Child Tutoring through Role Separated Scaffolding Interface in Real Time. arXiv cs.HC.