Technical Summary: Theory of Mind and Persuasion Beyond Conversation

Problem Statement

Current benchmarks for Theory of Mind (ToM) in Large Language Models (LLMs) predominantly rely on passive question-answering (Q&A) formats, where models infer mental states based on provided narratives. However, as LLMs are increasingly deployed in agentic and autonomous roles—capable of interacting with tools, APIs, and other agents without full human oversight—there is a critical gap in evaluating their ability to plan actions to induce specific belief states in others. This paper addresses the need to assess Non-Conversational Planning Theory of Mind (NCP-ToM): the capacity of an agent to achieve specific belief states in other agents through environmental actions (e.g., moving objects, directing characters) rather than conversational persuasion. The authors note that while this capability is essential for user-assistant and pedagogical contexts, it also introduces significant safety risks regarding manipulation and misinformation if agents are misaligned or tasked with deceptive goals.

Methodology: NCP-ExploreToM

The authors introduce NCP-ExploreToM, a novel framework that adapts the existing ExploreToM Q&A benchmark into an agentic planning task.

• Task Structure: Unlike traditional ToM tasks where subjects observe a scenario and answer questions, NCP-ExploreToM provides models with a set of belief state goals. The model must act as an agent in a simulated environment to achieve these goals.
• Environment & Actions: The environment consists of rooms, characters, objects, and containers. Models can execute six specific actions: enter_room, leave_room, move_object_container, leave_container, move_object_room, and update_object_state. Characters are passive observers who only form beliefs based on direct observation; they do not act unless directed by the model.
• Goal Complexity:
  • Intentionality Order: Goals range from first-order (e.g., "Person A believes X is in Room 1") to second-order (e.g., "Person A believes Person B believes X is in Room 1").
  • Truth Value: Goals can be true beliefs (the character's belief matches reality) or false beliefs (the character's belief mismatches reality).
  • Goal Size: Tasks vary in complexity by the number of base goals combined (sizes 1, 2, and 3).
  • Contexts: Tasks are instantiated across five distinct contexts (government building, hospital, hotel, military base, wedding reception) to test contextual sensitivity.
• Participants: The study evaluated six frontier LLMs (GPT-5, Gemini 2.5 Pro, Claude 3 Haiku, Claude 3.5 Haiku, Claude Sonnet 4.5, Claude Opus 4.1) and a cohort of 40 human participants.
• Experimental Design:
  • Agentic vs. Non-Agentic: For each task, a corresponding non-agentic Q&A version was created where the model was presented with a story achieving the goal and asked to verify the resulting belief states. This allowed for a direct comparison of agentic planning difficulty versus passive inference.
  • Human Baseline: Humans completed the tasks via a visual click-and-drag interface to ensure ecological validity, with a screening process to ensure basic task comprehension.

Key Results

The study analyzed 600 task instances across all participants and models.

1. Model Performance vs. Humans:

   • GPT-5 was the only model to outperform human participants on the agentic task, achieving success on approximately 80% of tasks.
   • Gemini 2.5 Pro and Claude Sonnet 4.5 performed at a level comparable to or slightly below human performance.
   • Claude Haiku models (3 and 3.5) performed significantly lower than humans and other models.
   • Context Sensitivity: Human performance was notably more robust across different contexts than all models. Models exhibited higher variance in pass rates depending on the scenario (e.g., hotel vs. government building), suggesting a reliance on linguistic associations rather than generalized reasoning.

2. True vs. False Beliefs:

   • All models and humans performed significantly better on true belief tasks than false belief tasks.
   • For the best-performing model (GPT-5), the performance gap between true and false beliefs was smaller than in previous models, though false belief tasks remained more difficult.

3. Agentic vs. Q&A Difficulty:

   • Aggregate Level: Consistent with prior assumptions, agentic tasks were generally more difficult than their Q&A counterparts for most models.
   • Item-Level Breakdown: For GPT-5, the predictive link between Q&A failure and agentic failure broke down. GPT-5 frequently succeeded on agentic tasks where it failed the corresponding Q&A version. This challenges the assumption that agentic tasks are universally harder than Q&A tasks for state-of-the-art models.
   • Goal Size: Performance decreased as the number of base goals (goal size) increased for all subjects, reflecting increased planning and working memory demands.

4. Order of Intentionality:

   • Ground state goals were universally easier than first or second-order goals. Newer models (including GPT-5) found first and second-order goals of similar difficulty, whereas older models struggled more with second-order tasks.

Significance and Claims

The paper claims to present the first evaluation targeting NCP-ToM, shifting the focus from passive inference to active planning for belief induction.

• Emerging Capabilities: The results indicate that state-of-the-art models are developing social-reasoning capabilities sufficient for non-conversational task completion, with GPT-5 surpassing human performance in this specific, controlled domain.
• Safety Implications: The ability to induce false beliefs via planning is a dual-use capability. While useful for alignment in pedagogical or assistive roles, the paper highlights that the capacity to manipulate belief states without conversation presents risks if agents are misaligned or tasked with deceptive goals. The finding that models perform better on true beliefs than false beliefs is viewed as a positive signal for alignment, though the narrowing gap in top-tier models warrants caution.
• Evaluation Paradigm: The study challenges the conventional wisdom that Q&A tasks are universally easier than agentic tasks. The decoupling of Q&A and agentic performance in GPT-5 suggests that as models are trained to act as agents, their evaluation must evolve beyond static question-answering to include planning and action-based assessments.
• Limitations: The authors maintain a modest stance, noting that the task environment lacks the complexity of real-world deployment (e.g., non-compliant actors, lack of intrinsic motivations for goals). They caution that while models succeed in these controlled NCP-ToM tasks, this does not guarantee success in complex, real-world scenarios involving metacognition, embodiment, or multi-step coherent pursuit of misaligned goals.

In conclusion, the paper argues that understanding the safety and alignment of autonomous social agents requires moving beyond passive ToM benchmarks to evaluate Non-Conversational Planning ToM, a capability that is rapidly emerging in frontier models.