arXiv ScienceSearch

arXiv · 2503.15516

In Pursuit of Predictive Models of Human Preferences Toward AI Teammates

Abstract

We seek measurable properties of AI agents that make them better or worse teammates from the subjective perspective of human collaborators. Our experiments use the cooperative card game Hanabi -- a common benchmark for AI-teaming research. We first evaluate AI agents on a set of objective metrics based on task performance, information theory, and game theory, which are measurable without human interaction. Next, we evaluate subjective human preferences toward AI teammates in a large-scale (N=241) human-AI teaming experiment. Finally, we correlate the AI-only objective metrics with the human subjective preferences. Our results refute common assumptions from prior literature on reinforcement learning, revealing new correlations between AI behaviors and human preferences. We find that the final game score a human-AI team achieves is less predictive of human preferences than esoteric measures of AI action diversity, strategic dominance, and ability to team with other AI. In the future, these correlations may help shape reward functions for training human-collaborative AI.

Explore related subjects

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Ho Chit Siu, Jaime D. Peña, Yutai Zhou, Ross E. Allen. 2025-01-31. In Pursuit of Predictive Models of Human Preferences Toward AI Teammates. https://arxiv.org/abs/2503.15516

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

Are explainable AI (XAI) evaluation strategies aligned? Comparing subjective, objective, and mathematical evaluation measures using saliency maps

The evaluation of explainable AI (XAI) approaches often relies on three families of methods: subjective measures (e.g., questionnaires on trust or satisfaction), objective measures (e.g., task performance metrics), and mathematical metrics (e.g., for faithfulness). Yet, it remains unclear how these families align or diverge in practice. In a{preregistered} between-subjects study (N=166), we use three established saliency map techniques (LIME, Grad-CAM, Guided Backpropagation) as a testbed to examine this issue. We find that each family of methods leads to different conclusions: participants reported no differences in trust or satisfaction, Grad-CAM improved user performance, while mathematical metrics favored Guided Backpropagation. At the same time, mathematical metrics were only partially related to user performance, and these relationships were sometimes counterintuitive. Our findings highlight the methodological importance of comparing subjective, objective, and mathematical approaches when evaluating XAI, illustrating both tensions and aspects that are aligned. We discuss implications for XAI evaluation frameworks.

cs.HC

Mind Your Ps and Qs: Positive Moderation Practice in the Positive Queue

Online communities rely on volunteer moderators to maintain order. Despite their key role, moderators are given a toolbox of punishments and far less support for encouraging contributions they want to see more of. We introduce the Positive Queue as a positive counterpart to Reddit's modqueue: a dedicated space for moderators to discover contributions and behaviors they want to encourage and positively reinforce. With five moderators, four with 6-14 years of experience, we use the Positive Queue to examine how moderators operationalize positive reinforcement. Moderators combined predicted community reception, observed engagement, and their own judgment; used prediction-engagement mismatches to identify overlooked content; and repurposed positive features for punitive and retrospective work. These findings surface tensions around labor, attribution, and community fit. We contribute the Positive Queue as a working system and conceptualize positive moderation as recognition infrastructure that shapes what moderators notice, whose judgment becomes visible, and how recognition reaches contributors.

cs.HC

PILOT: Control Surfaces for Authoring Social Media Feeds

Personalized social media feeds infer preferences from behavior, leaving people little direct control over what they see. Existing controls range from post-level reactions to rules and natural language, but little is known about how people use them together or how added expressiveness changes effort. We built PILOT, a Bluesky feed-authoring system that turns in-feed actions into explicit preferences, deterministic ranking, and inspectable outcomes. A study with seven participants informed an expanded implementation, which we organized into three nested control surfaces that ten participants compared within subjects. Participants assigned controls to distinct jobs: broad controls set direction, in-post actions refined results, and filters removed content. Richer surfaces did not necessarily feel more effortful, and participants' experiences depended more on whether they could verify and repair outcomes. Our findings show that usable feed control requires not simply more controls, but orchestration across mechanisms that support expression, inspection, repair, and episodic use.

cs.HC