[Paper Review] Collective Intelligence in Human-AI Teams: A Bayesian Theory of Mind Approach

Samuel Westby, Christoph Riedl|arXiv (Cornell University)|Aug 24, 2022

Team Dynamics and Performance4 citations

TL;DR

This paper proposes a Bayesian theory of mind (ToM) framework for human-AI teams that models teammates' mental states from communication to enhance collective intelligence. Using a generative computational model, it demonstrates that AI agents can predict and improve team performance by sending targeted messages based on belief divergence, achieving 82.1% accuracy—4.9% higher than random interventions—while a real-time ToM measure explains 8% of team performance variance after only 25% of messages.

ABSTRACT

We develop a network of Bayesian agents that collectively model the mental states of teammates from the observed communication. Using a generative computational approach to cognition, we make two contributions. First, we show that our agent could generate interventions that improve the collective intelligence of a human-AI team beyond what humans alone would achieve. Second, we develop a real-time measure of human's theory of mind ability and test theories about human cognition. We use data collected from an online experiment in which 145 individuals in 29 human-only teams of five communicate through a chat-based system to solve a cognitive task. We find that humans (a) struggle to fully integrate information from teammates into their decisions, especially when communication load is high, and (b) have cognitive biases which lead them to underweight certain useful, but ambiguous, information. Our theory of mind ability measure predicts both individual- and team-level performance. Observing teams' first 25% of messages explains about 8% of the variation in final team performance, a 170% improvement compared to the current state of the art.

Motivation & Objective

• To develop a real-time, computationally grounded measure of theory of mind (ToM) ability in human teams.
• To model how humans and AI agents form mental representations of teammates' beliefs during collaborative decision-making.
• To test whether AI agents can improve team performance by triggering targeted interventions based on belief divergence.
• To evaluate the predictive power of ToM ability on both individual and team-level performance in cognitive tasks.
• To extend collective intelligence theory by embedding dynamic, Bayesian inference into multi-agent systems for real-time adaptation.

Proposed method

• A network of Bayesian agents models the mental states of human teammates using observed chat-based communication in a Hidden Profile task.
• Each agent maintains an Ego Model (self-beliefs) and multiple Alter Models (teammates' beliefs), updated via Bayesian inference on incoming messages.
• The agents compute belief divergence between Ego and Alter Models using Kullback-Leibler (KL) divergence to identify optimal intervention messages.
• Interventions are selected as the message that minimizes KL divergence between Ego and Alter posteriors over answer options.
• A counterfactual simulation evaluates performance under random vs. targeted AI interventions, using a generative model trained on real human communication.
• The framework enables real-time ToM measurement by analyzing early communication patterns and their predictive power for final team performance.

Experimental results

Research questions

• RQ1 Can a Bayesian ToM model accurately infer and represent the evolving mental states of teammates from naturalistic communication in human-AI teams?
• RQ2 To what extent does individual theory of mind ability, measured in real time, predict both individual and team-level performance in collaborative cognitive tasks?
• RQ3 Can AI agents use belief divergence to identify and send targeted messages that significantly improve team performance beyond human-only performance?
• RQ4 How much of the variation in team performance can be explained by early communication patterns using the proposed real-time ToM measure?
• RQ5 What is the performance gain of AI-driven, ToM-based interventions compared to random or non-targeted interventions in human-AI teams?

Key findings

• The proposed real-time theory of mind measure explains 8% of the variation in final team performance after observing only the first 25% of team messages, a 170% improvement over the current state of the art.
• Human teams underperform due to cognitive biases, particularly underweighting ambiguous but useful information, especially under high communication load.
• AI agents using the Bayesian ToM framework achieved 82.1% team performance in counterfactual simulations with targeted interventions, significantly outperforming random interventions (79.0%) with p < 0.0001.
• The model successfully captures human decision-making patterns, showing that humans use Bayesian inference and ToM, albeit imperfectly, in collaborative settings.
• The framework enables real-time prediction of collective intelligence by measuring belief alignment and surprise minimization across team members.
• The study demonstrates that AI agents can autonomously identify and send messages that reduce uncertainty in teammates’ beliefs, thereby enhancing team performance without pretraining.