A Cognitive Evaluation of Instruction Generation Agents tl;dr They Need Better Theory-of-Mind Capabilities
share
We mathematically characterize the cognitive capabilities that enable humans to effectively guide others through natural language. We show that neural-network-based instruction generation agents possess similar cognitive capabilities, and design an evaluation scheme for probing those capabilities. Our results indicate that these agents, while capable of effectively narrowing the search space, poorly predict the listener's interpretations of their instructions and thus often fail to select the best instructions even from a small candidate set. We augment the agents with better theory-of-mind models of the listener and obtain significant performance boost in guiding real humans. Yet, there remains a considerable gap between our best agent and human guides. We discuss the challenges in closing this gap, emphasizing the need to construct better models of human behavior when interacting with AI-based agents.
READ FULL TEXT