New PETITE framework shows LLMs solve harder problems when assigned opposing tutor-student roles, mimicking how human learning benefits from structured social interaction.

Researchers tested whether two instances of the same LLM can achieve better problem-solving outcomes by assigning them asymmetric roles rather than using stronger models

The tutor-student multi-agent system was evaluated on autonomous coding problems, where the student generates solutions and the tutor provides evaluative feedback without access to correct answers

The approach is inspired by human cognitive development, which shows that structured role-based exchanges between tutors and learners enable solutions neither could achieve independently

The PETITE framework demonstrates that complementary role assignments can create synergistic effects that push LLMs beyond their performance in existing single-agent frameworks