On the Intelligence and Knowledgeability of Virtual Agents

Paper Title

On the Intelligence and Knowledgeability of Virtual Agents

Publication Info

• Topic area: Human-agent interaction in immersive virtual reality environments.
• Keywords: Virtual agents, intelligence, knowledgeability, human-agent interaction, virtual reality, co-presence, trust, uncanny valley, collaborative tasks.

Background and Problem

• Problem / challenge: Intelligence and knowledgeability in virtual agents are often treated interchangeably, yet their distinct roles and combined effects on human perceptions and interactions remain underexplored.
• Significance: Understanding these traits is crucial for designing virtual agents that foster meaningful and efficient human-agent collaboration, especially in immersive environments.
• Motivation and related work: Previous research has focused on virtual agent intelligence and knowledgeability separately, exploring their impacts on user experiences, trust, and collaboration. However, the interaction between these traits in situated collaborative settings, such as VR environments, has not been thoroughly examined.

Solution

• Proposed approach: A VR jigsaw puzzle application featuring conversational virtual agents with manipulated intelligence (puzzle-solving ability) and knowledgeability (depth of art-related knowledge).
• Novelty:
  1. Disentangling intelligence and knowledgeability as distinct yet interrelated traits in virtual agents.
  2. Investigating their combined effects on human perceptions, social/emotional experiences, and interaction dynamics.
  3. Designing a 2 × 2 within-group study to systematically evaluate these traits in a collaborative VR setting.
• Procedure and key techniques:
  • Intelligence manipulation: Agents solve puzzles with varying accuracy (25% for low intelligence, 100% for high intelligence).
  • Knowledgeability manipulation: Agents provide surface-level or in-depth art-related knowledge using LLM-based prompt engineering.
  • Experimental design: Four conditions (low/high intelligence × low/high knowledgeability) with balanced Latin square order.
  • Data collection: Self-reported ratings, application-logged data (e.g., gaze distribution, puzzle completion metrics), and conversational transcripts.

Results

• Concrete findings:
  • High intelligence agents were perceived as more intelligent, trustworthy, and elicited higher co-presence and rapport.
  • High knowledgeability agents were perceived as more knowledgeable and trustworthy but could reduce rapport due to robotic delivery.
  • Interaction effects showed interdependence between intelligence and knowledgeability in shaping perceptions.
  • Low intelligence agents led participants to complete more puzzle pieces and spend more time on the task.
• Advantage over baselines:
  • Demonstrated distinct yet interrelated impacts of intelligence and knowledgeability on human-agent interaction.
  • Highlighted design considerations for balancing task competence and conversational delivery in virtual agents.
• Experiments / evaluation:
  • Participants: 24 university students (balanced gender, age M = 26.79).
  • Metrics: Perceived intelligence, knowledge, trust, co-presence, uncanny valley, rapport; puzzle completion time, gaze distribution; conversational response types.
  • Statistical analysis: RM-ANOVA, Pearson correlation, thematic analysis of open-ended feedback.
• Limitations and future work:
  • Limited scope of intelligence and knowledgeability definitions (e.g., only puzzle-solving and art knowledge).
  • Gender bias in agent design (female agent only).
  • Interaction design confounds (e.g., agent holding puzzle pieces while speaking).
  • Short interaction durations and small sample size.
  • Lack of mutual gaze measurements and extended task complexity exploration.

Summary

This study investigates the interplay between intelligence and knowledgeability in virtual agents within a VR collaborative puzzle-solving task. Results show that intelligence enhances perceptions of competence, trust, and co-presence, while knowledgeability boosts perceived expertise but may reduce rapport due to robotic delivery. Interaction effects highlight their interdependence, offering actionable insights for designing virtual agents with balanced cognitive traits. Future work should explore broader cognitive dimensions, multitasking complexity, and adaptive conversational strategies to further enhance human-agent interaction in immersive environments.