Effective Explanations Support Planning Under Uncertainty
Researchers from Stanford University and collaborators have proposed a new computational model to understand how procedural explanations facilitate planning under uncertainty. Published on arXiv, the study addresses the challenge of explaining navigation from point A to B by mentally simulating listener actions. The model utilizes a large language model to translate natural language explanations into program-like guidance, specifically creating a policy prior and value map. A planning agent then executes these instructions under conditions of partial observability. The effectiveness of explanations is scored based on path efficiency and reliability, with penalties for replanning. Through four preregistered experiments involving 1,200 explanations across 24 maps, the team collected data on helpfulness judgments and navigation performance. Results indicate that higher-scoring explanations are perceived as more helpful and significantly improve navigation outcomes compared to baseline or low-quality explanations. This research highlights procedural explanation as utility-guided communication, demonstrating how language can be effectively grounded into action despite uncertainty.
Wire timeline
Effective Explanations Support Planning Under Uncertainty
Researchers from Stanford University and collaborators have proposed a new computational model to understand how procedural explanations facilitate planning under uncertainty. Published on arXiv, the study addresses the challenge of explaining navigation from point A to B by mentally simulating listener actions. The model utilizes a large language model to translate natural language explanations into program-like guidance, specifically creating a policy prior and value map. A planning agent then executes these instructions under conditions of partial observability. The effectiveness of explanations is scored based on path efficiency and reliability, with penalties for replanning. Through four preregistered experiments involving 1,200 explanations across 24 maps, the team collected data on helpfulness judgments and navigation performance. Results indicate that higher-scoring explanations are perceived as more helpful and significantly improve navigation outcomes compared to baseline or low-quality explanations. This research highlights procedural explanation as utility-guided communication, demonstrating how language can be effectively grounded into action despite uncertainty.
cs.AI updates on arXiv.org