Enhancing LLM Problem Solving via Tutor-Student Multi-Agent Interaction
Summary: arXiv:2604.08931v1 Announce Type: new
Abstract: Human cognitive development is shaped not only by individual effort but by structured social interaction, where role-based exchanges such as those between a tutor and a learner enable solutions that neither could achieve alone. Inspired by these developmental principles, we ask the question whether a tutor-student multi-agent system can create a synergistic effect by pushing Large Language Model (LLM) beyond what it can do within existing frameworks.
Introduction
The advancement of artificial intelligence, particularly in the domain of Large Language Models (LLMs), has opened new avenues for problem-solving capabilities. Researchers are exploring innovative frameworks that not only utilize the inherent strengths of LLMs but also enhance their functionality through structured interactions. One such framework is the PETITE model, which enables a tutor-student interaction between two agents instantiated from the same LLM.
The PETITE Framework
In the PETITE framework, two agents are assigned asymmetric roles in the context of an autonomous coding problem domain:
- Student Agent: This agent generates and iteratively refines solutions to coding problems.
- Tutor Agent: This agent provides structured evaluative feedback without access to ground-truth answers.
The aim of this approach is to extract better problem-solving performance from a single model by structuring interactions through these complementary roles. This method diverges from traditional approaches that rely on stronger supervisory models or heterogeneous ensembles.
Evaluation and Results
The model’s effectiveness was evaluated against the APPS coding benchmark, comparing its performance to state-of-the-art approaches such as:
- Self-Consistency
- Self-Refine
- Multi-Agent Debate
- Multi-Agent Review
Results indicate that the PETITE model achieves similar or even higher accuracy while utilizing significantly fewer tokens. This efficiency suggests that developmentally grounded, role-differentiated interaction structures can provide a principled and resource-efficient paradigm for enhancing LLM problem-solving through structured peer-like interactions.
Conclusion
The findings from the PETITE framework underscore the potential of leveraging structured social interactions between artificial agents to enhance problem-solving capabilities. By mimicking human cognitive development processes, the tutor-student agent dynamic offers a promising avenue for future research and application in the field of artificial intelligence. As LLMs continue to evolve, exploring collaborative frameworks like PETITE may yield significant advancements in their performance and efficiency.
Index Terms
- Peer Tutoring
- Scaffolding
- Large Language Models
- Multi-Agent Systems
- Code Generation