Can Language Models Teach Weaker Agents? Teacher Explanations Improve Students via Personalization

Language models can effectively teach weaker AI agents through natural language explanations, demonstrating that the reasoning abilities of large language models translate into teaching capabilities that improve student agent performance on tasks.

Selective intervention is more efficient than explaining everything: The research shows teachers should strategically choose which data points to explain rather than providing explanations for all examples, allowing students to perform well independently after receiving targeted guidance within a communication budget.

Personalized explanations based on student understanding are key to effective teaching: The study proposes using "mental models" of the student to personalize explanations, suggesting that effective AI teachers—like human teachers—should adapt their communication style to match what individual learners need rather than using one-size-fits-all explanations.

Teacher interventions help students generalize to new, unexplained tasks: The research demonstrates that explanations don't just help with immediate performance but enable students to apply learned knowledge to future problems they encounter without additional teacher guidance, suggesting genuine learning transfer occurs.

This work has practical implications for AI system design: By establishing frameworks for when and how AI systems should communicate with each other, the research provides actionable guidance for building more efficient multi-agent AI systems that use communication resources responsibly rather than wastefully.