First-Order Theory of Mind

Agents with first-order ToM do not just predict actions based on environmental stimuli; they predict actions based on inferred internal goals. The reasoning chain is:

1. Observe Agent B's actions and context.
2. Infer Agent B's likely goal or desire.
3. Infer Agent B's belief about how to achieve that goal.
4. Predict Agent B's future actions based on that belief-desire pair.

This moves prediction from simple behavioral pattern matching to model-based intention recognition.

The system must maintain a clear ontological distinction between its own knowledge/beliefs and its model of the other agent's knowledge/beliefs. This requires:

• Egocentric Model: The agent's own, presumably accurate, view of the world.
• Allocentric Model: A separate, attributed view representing the other agent's perspective.

Failure of this separation leads to egocentric bias, where the agent incorrectly assumes others know what it knows, a common failure mode in early AI systems and young children.

First-order ToM is a prerequisite for pragmatic language understanding. To interpret an utterance, an agent must model the speaker's:

• Informational Goal: What they are trying to convey (communicative intent).
• Knowledge State: What they believe the listener already knows (modeled via ToM).

This enables understanding of non-literal speech, sarcasm, and indirect requests. It underpins the Gricean Maxims of cooperative conversation, as effective communication requires estimating the listener's mental state.

A dominant computational method for achieving first-order ToM is inverse planning (or Bayesian inverse reinforcement learning). The agent solves:

'Given that I observed these actions from Agent B, what goals and beliefs would make these actions rational?'

• Process: The observer assumes the other agent is a bounded rational planner.
• Mechanism: It uses a generative model of planning (a 'forward' model) and runs it in reverse via probabilistic inference.
• Output: A posterior distribution over the other agent's possible goals and world beliefs.

Second-order Theory of Mind is the recursive ability to attribute mental states about mental states. It involves reasoning like, "Alice believes that Bob believes X." This is critical for deception, bluffing, and complex coordination.

• Key Difference: First-order: "I think you think X." Second-order: "I think you think I think X."
• Strategic Necessity: Essential for agents operating in competitive multi-agent environments like trading, negotiation, or adversarial games where predicting an opponent's prediction of your move is vital.

Mental state attribution is the active computational process of ascribing specific internal cognitive or emotional states to another agent. First-order ToM is its primary manifestation.

• Process Flow: Observe behavior/communication → Infer underlying beliefs, knowledge, goals → Predict future actions.
• Technical Implementation: Often involves Bayesian inverse planning or fine-tuned language models that map observable cues to a distribution over possible hidden mental states.

The Belief-Desire-Intention (BDI) model is a foundational software architecture for intelligent agents that operationalizes Theory of Mind concepts for decision-making.

• Beliefs: The agent's knowledge about the world (including beliefs about others' beliefs).
• Desires: The agent's goals or motivational state.
• Intentions: Committed plans of action chosen to achieve desires.
• Architectural Link: A first-order ToM-capable agent uses its BDI architecture to model the BDI states of other agents, creating nested agent models for prediction.

Inverse planning is a primary Bayesian computational method for implementing mental state attribution. It works backwards from observed actions to infer the hidden goals and beliefs that likely caused them, assuming the observed agent is approximately rational.

• Mechanism: Given a sequence of actions and a model of how a rational planner would act, it calculates the posterior probability over possible goals and world beliefs.
• Application: Core to plan recognition and intent inference in collaborative and assistive AI systems, allowing an agent to discern a human user's objective without explicit instruction.