Mentalizing Interface: Using LLM to Explore How Humans Attribute 'Mind' to Non-Humanoid Robots

The research team from the University of Pisa developed an experimental platform that uses large language models (LLMs) to generate self-explanations of different styles for non-humanoid robots, studying when and why humans attribute psychological properties like beliefs, desires, and intentions to machines. The innovation lies in abandoning the traditional path of inducing mind attribution through anthropomorphic appearance; instead, it uses linguistic frameworks (intentional, teleological, mechanistic) to trigger humans' intentional stance, providing a new experimental approach for exploring the mechanism of mind attribution in human-robot interaction.

Traditional Path Limitations

Traditional human-robot interaction research relies on the appearance of humanoid robots (e.g., iCub) to induce mind attribution, but it has problems such as many confounding variables (intertwined processes like anthropomorphism and empathy), low ecological validity (most real-world robots are non-humanoid), and ethical risks (over-reliance).

New Idea: Language as a Carrier of Mind

Based on Daniel Dennett's "Intentional Stance" theory, humans explain behavior through narratives of mental states. The study hypothesizes that if non-humanoid robots describe their behavior using belief-desire-intention (BDI) language, humans may adopt an intentional stance toward them, and the differences in mind attribution driven by language versus anthropomorphic appearance can be separated.

Three-Layer Architecture

1. Bottom Layer: ROS2 + Gazebo simulate non-humanoid robots (TurtleBot3), with Nav2 enabling autonomous movement;
2. Middle Layer: BDI state system (belief base includes confidence/source, desire set with priority, intention queue records action status);
3. Upper Layer: Local Llama3.2 generates three explanatory frameworks:
   • Intentional: Uses psychological vocabulary like "I believe/want/intend";
   • Teleological: Functional description of goals and behaviors;
   • Mechanistic: Only reports physical parameters (e.g., odometer, Twist messages).

Keep the robot's behavior consistent, only changing the language style to accurately measure the framework's impact.

Bookstore Scenario

• User instruction: "I'm a Tolkien fan, go to the relevant book section" → The robot associates "Tolkien" with "fantasy literature section" and expresses reasoning and plans using intentional language;
• User instruction: "Post content about health books online" → The robot distinguishes the core of the task (internet area) from background information (health books) to avoid misleading.

Apartment Scenario

• User says: "You look dirty, go clean at the sink" → The robot accepts the anthropomorphic description and responds;
• User requests: "Choose a random place" → The robot autonomously selects based on internal states (e.g., table).

Demonstrates semantic reasoning, indirect instruction understanding, and natural dialogue capabilities, verifying the triggering effect of linguistic frameworks on mind attribution.

Sociopsychological Interpretability

Proposes a new definition: The interpretability of social robots lies in their ability to express themselves using folk psychology terms, different from traditional technical transparency (showing internal mechanisms), emphasizing the use of human-familiar psychological narrative frameworks to describe behavior.

Stance Separation

For the first time, experiments separate Dennett's three explanatory stances (physical, design, intentional), avoiding the confusion between anthropomorphic appearance and intentional stance in traditional research, providing a method for precise study of mind attribution.

1. Trust Calibration: Adjust users' trust levels in robots through linguistic frameworks (avoiding over/under trust);
2. Ethical Experiments: No need to deceive participants (honestly explain LLM-generated explanations), resolving ethical disputes in traditional research;
3. Cross-Cultural Research: Explore differences in responses to robot psychological narratives across different cultures;
4. Special Population Research: Such as mind attribution patterns in autistic spectrum populations.

Technical Limitations

• Positioning uncertainty leads to dialogue contradictions;
• Simplified distinction between desires and intentions (directly promoting desires to intentions);
• Randomness in LLM generation.

Conceptual Challenges

• Confusion between linguistic stance and intentional stance (does fluent psychological vocabulary constitute an independent stance?);
• Definition of mind attribution (self-report vs. true belief);
• Unknown impact of long-term interaction on mind attribution patterns.