Effective Explanations Support Planning Under Uncertainty

Short summary

Research proposes a computational model that uses LLMs to convert natural-language explanations into executable action plans (policy + value maps), tested via planning agents under partial observability. Preregistered human study across 1,200 explanations shows high-quality explanations significantly improve navigation performance, demonstrating that procedural explanation is shaped by how language grounds into action.

• •LLM-based system converts explanations into policy priors and value maps executed by planning agents
• •Preregistered experiments with 1,200 explanations across 24 maps show quality-correlated improvements in human navigation
• •Demonstrates procedural explanation as utility-guided communication under uncertainty

Generated with AI, which can make mistakes.