A relevance model of human sparse communication in cooperation

TL;DR

The paper introduces a model explaining how humans choose what to communicate during cooperative tasks, showing it outperforms AI like GPT-4.

Contribution

A novel relevance model based on decision theory and Theory of Mind that captures human communication choices in cooperative settings.

Findings

The relevance model accurately predicts which information humans choose to communicate in a navigation task.

Humans perform better and rate AI higher when it uses the relevance model compared to a heuristic model.

The model outperforms GPT-4 in simulating human communication patterns.

Abstract

Human real-time communication creates a limitation on the flow of information, which requires the transfer of carefully chosen and condensed data in various situations. We introduce a model that explains how humans choose information for communication by utilizing the concept of “relevance” derived from decision-making theory and Theory of Mind (ToM). We evaluated the model by conducting experiments where human participants and an artificial intelligence (AI) agent assist each other to avoid multiple traps in a simulated navigation task. The relevance model accurately depicts how humans choose which trap to communicate. It also outperforms GPT-4, which participates in the same task by responding to prompts that describe the game settings and rules. Furthermore, we demonstrated that when humans received assisting information from an AI agent, they achieved a much higher performance and gave higher ratings to the AI when it utilized the relevance model compared to a heuristic model. Together, these findings provide compelling evidence that a relevance model rooted in decision theory and ToM can effectively capture the sparse and spontaneous nature of human communication.

Figures