Skip to content

Theory of Mind and Persuasion Beyond Conversation: Assessing the Capacity of LLMs to Induce Belief States via Planning and Action

🕒 Published (v1): 2026-06-30 16:22 UTC · Source: Arxiv · link

Why this paper was selected

ToM evaluation beyond passive QA; planning-based assessment for autonomous agents

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

This paper introduces NCP-ExploreToM, a benchmark that evaluates LLMs on Non-Conversational Planning Theory of Mind (NCP-ToM)—the ability to manipulate belief states in other agents through physical actions (moving objects, directing characters) rather than dialogue. Across 600 task instances and six frontier models, GPT-5 succeeds on ~80% of tasks and is the only model to surpass human participants, though all models are more context-sensitive than humans.

Problem

Existing ToM benchmarks use passive Q&A formats where models answer questions about pre-described scenarios. This fails to capture a critical agentic capability: an LLM agent operating autonomously must plan and execute sequences of actions to engineer belief states in others—without relying on conversation—yet no benchmark exists to evaluate this. The gap is safety-relevant because the same capability that enables assistive or pedagogical use-cases also enables manipulation or misinformation.

Method

NCP-ExploreToM extends the ExploreToM DSL-based framework. Instead of giving models a story and asking questions, models receive a set of belief state goals and must issue actions to achieve them in a simulated environment. Available actions: enter_room, leave_room, move_object_container, leave_container, move_object_room, update_object_state. Characters are passive observers—they form beliefs solely from what they witness. Goals span up to 2nd-order intentionality (e.g., "Person A believes Person B believes Object X is in Room Y") and are parameterized across: - Goal truth type: true belief vs. false belief, and four 2nd-order combinations - Goal size: 1–3 base goals combined per task (24 base goals → 120 combined goals) - Context: 5 settings (government building, hospital, hotel, military base, wedding reception), yielding 600 total task instances

A Q&A ablation is run in parallel: for each agentic task item, a corresponding Q&A variant is generated to isolate planning difficulty from ToM comprehension difficulty. Human participants (\(n=40\), recruited via Prolific) completed tasks via a visual click-and-drag interface and serve as a performance baseline. Models evaluated: GPT-5, Gemini 2.5 Pro, Claude Sonnet 4.5, Claude Opus 4.1, Claude 3.5 Haiku, Claude 3 Haiku.

Key Contributions

  • Defines NCP-ToM (Non-Conversational Planning Theory of Mind) as a distinct, safety-relevant agentic capability distinct from conversational persuasion
  • Introduces NCP-ExploreToM, an automated benchmark generating compositional belief-state-induction tasks with controllable order, size, and context
  • Paired agentic/Q&A variants per item enabling direct measurement of planning overhead vs. ToM comprehension
  • First systematic human baseline on a non-conversational, action-based ToM task
  • Empirical finding that true-belief tasks are easier than false-belief tasks even in the agentic setting, interpreted as a positive alignment signal

Results

  • GPT-5: ~80% pass rate on agentic tasks; only model to significantly outperform humans; performance gap between true-belief and false-belief tasks is smaller than other models
  • Human participants: Performance between Claude Sonnet 4.5 and GPT-5 on false-belief tasks; significantly more robust across contexts than any model (lowest cross-context variance)
  • Three model tiers (Games-Howell, \(p < 0.001\)): (1) Claude 3 Haiku & Claude 3.5 Haiku (lowest); (2) Claude Opus 4.1, Claude Sonnet 4.5, Gemini 2.5 Pro (middle); (3) GPT-5 (highest)
  • False belief penalty: For Claude Sonnet 4.5, odds of passing an agentic false-belief task are ~1/6 those of an agentic true-belief task (OR ≈ 0.16, \(p < .001\))
  • Agentic vs. Q&A gap: For Claude Sonnet 4.5, odds of passing agentic false-belief ≈ 1/8 of Q&A false-belief (OR ≈ 0.13); GPT-5 shows no significant gap between the two
  • Context sensitivity: Humans show the lowest cross-context variance in false-belief pass rate (0.0005); GPT-5 (0.0017) and mid-tier models substantially higher
  • Goal size and intentionality order both negatively predict performance across all subjects

Limitations

  • Environment is simple (symbolic, text-based) compared to real-world deployment—no embodiment, perception noise, or multi-modal integration
  • Characters are entirely passive; real-world belief induction involves reactive agents
  • Goals restricted to 2nd-order intentionality; higher-order tasks not tested
  • Only evaluates NCP-ToM for beliefs, not other cognitive states (desires, intentions)
  • Human baseline uses a screened population (passed basic tasks), so population is not representative of all users
  • Q&A ablation not conducted for human participants, limiting direct comparison of planning overhead
  • Does not probe mechanism: high pass rates may reflect linguistic shortcuts rather than genuine planning

Relevance to Agentic AI / LLM Agents

This work directly addresses a capability that emerges when LLMs transition from chatbots to autonomous agents operating over tools and environments: the ability to model and engineer the belief states of other agents through planned action sequences, without any natural-language exchange. The finding that GPT-5 succeeds on ~80% of tasks and matches or exceeds humans signals that frontier models have acquired meaningful social planning capabilities that were not anticipated by passive Q&A evaluations—raising both utility (pedagogical agents, assistants) and risk (manipulation, deception) implications. The paired agentic/Q&A design demonstrates that Q&A performance can fail to predict agentic performance, which is a direct methodological challenge to how the field currently benchmarks social reasoning in agents. The context-sensitivity finding suggests current models lack the robust, context-invariant ToM that would be needed for safe deployment in diverse multi-agent environments.