Enhancing LLM Problem Solving via Tutor-Student Multi-Agent Interaction

Researchers present PETITE, a tutor-student multi-agent framework that enhances LLM problem-solving by assigning complementary roles to agents from the same model. Evaluated on coding benchmarks, the approach achieves comparable or superior accuracy to existing methods while consuming significantly fewer tokens, demonstrating that structured role-differentiated interactions can improve LLM performance more efficiently than larger models or heterogeneous ensembles.

This research addresses a fundamental challenge in LLM optimization: extracting maximum problem-solving capability from existing models without requiring larger architectures or additional computational resources. The PETITE framework leverages developmental psychology principles, where structured social interaction between roles creates emergent performance gains. Rather than relying on ensemble methods combining different models or training stronger supervisory models, the approach uses a single LLM instantiated twice with asymmetric roles—one generating solutions iteratively, the other providing evaluative feedback without ground-truth access.

The framework's significance lies in its efficiency gains. By achieving comparable or higher accuracy on APPS coding benchmarks while consuming fewer tokens than Self-Consistency, Self-Refine, and Multi-Agent Debate approaches, PETITE offers practical advantages for developers and organizations managing LLM inference costs. Token consumption directly correlates with computational cost and latency, making this improvement commercially relevant.

The research reflects a broader industry trend toward optimizing existing models rather than perpetually scaling up. As LLM inference costs remain substantial, techniques that extract better performance through interaction design become increasingly valuable. This approach also suggests that model capability improvements may come from architectural and procedural innovation rather than parameter count alone.

Investors and developers should monitor whether this peer-tutoring paradigm generalizes beyond coding tasks to other domains requiring iterative refinement. If successful adaptation occurs across language understanding, reasoning, and creative tasks, similar frameworks could reshape how organizations deploy LLMs cost-effectively.

• →PETITE uses role-differentiated agents from a single LLM to achieve competitive coding performance with lower token consumption than existing multi-agent approaches.
• →Structured role-based interaction between tutor and student agents creates synergistic problem-solving without requiring larger models or heterogeneous ensembles.
• →The framework demonstrates that procedural innovation can improve LLM efficiency, challenging the assumption that capability gains require parameter scaling.
• →Token reduction has direct implications for inference costs and latency, making this approach commercially relevant for production deployments.
• →Results suggest developmentally-grounded interaction structures provide a replicable pattern for optimizing LLM performance across problem domains.