Role-Based LLM Interaction Boosts Coding Performance While Cutting Tokens
Researchers propose PETITE, a multi-agent framework where two instances of the same LLM take on asymmetric tutor and student roles to solve coding problems more effectively than existing approaches. The student agent generates and refines solutions while the tutor provides structured feedback without access to ground truth, mimicking human cognitive development through role-based interaction. Tested on the APPS coding benchmark against Self-Consistency, Self-Refine, Multi-Agent Debate, and Multi-Agent Review, PETITE achieves comparable or better accuracy while using significantly fewer tokens, suggesting that role-differentiated interaction structures can enhance problem-solving efficiency without requiring stronger models or heterogeneous ensembles.
Researchers propose PETITE, a multi-agent framework where two instances of the same LLM take on asymmetric tutor and student roles to solve coding problems more effectively than existing approaches. The student agent generates and refines solutions while the tutor provides structured feedback without access to ground truth, mimicking human cognitive development through role-based interaction. Tested on the APPS coding benchmark against Self-Consistency, Self-Refine, Multi-Agent Debate, and Multi-Agent Review, PETITE achieves comparable or better accuracy while using significantly fewer tokens, suggesting that role-differentiated interaction structures can enhance problem-solving efficiency without requiring stronger models or heterogeneous ensembles.
- PETITE framework uses two instances of the same LLM in tutor-student roles to improve coding problem-solving performance
- Student agent iteratively refines solutions while tutor agent provides evaluative feedback without ground-truth access, creating complementary interaction
- Achieves similar or higher accuracy than state-of-the-art baselines on APPS benchmark while consuming significantly fewer tokens
- Demonstrates that structured role-based interaction can extract better performance from a single model without requiring stronger supervisory models or model ensembles
This work challenges the assumption that improving LLM performance requires either larger models or ensemble approaches. By showing that structured interaction patterns inspired by human learning can yield efficiency gains, it opens a new avenue for extracting more capability from existing models at lower computational cost, which has direct implications for cost-sensitive deployment and resource-constrained environments.
- Interaction structure and role assignment may be as important as model capacity for certain problem-solving tasks, potentially shifting focus from model scaling to interaction design
- Single-model multi-agent systems can compete with ensemble and debate-based approaches, reducing computational overhead and complexity in production systems
- The framework's reliance on developmental psychology principles suggests broader applicability beyond coding, potentially extending to other reasoning and planning domains
Our Briefing
Weekly signal. No noise. Built for founders, operators, and AI-curious professionals.
No spam. Unsubscribe any time.
