Higher-Order HI-TOM Benchmark

• Higher-Order HI-TOM Benchmark is a suite of evaluation tasks that rigorously assess recursive Theory of Mind reasoning in LLMs.
• It employs crafted narratives and multi-order questions to measure agents' beliefs about nested mental states and handle deceptive scenarios.
• Empirical results reveal scaling challenges and algorithmic breakthroughs, with methods like RecToM achieving near-perfect accuracy at higher orders.

The Higher-Order HI-TOM (Higher-Order Theory of Mind) Benchmark refers to a family of evaluation resources and methodologies designed to rigorously assess the recursive mental-state reasoning abilities of LLMs and other machine learning systems. The HI-TOM benchmark suite probes an agent's capacity to track beliefs nested through multiple agents—e.g., “Where does A think B thinks C believes X?”—and is foundational for measuring progress toward advanced, human-comparable Theory of Mind (ToM) in artificial systems (He et al., 2023).

1. Formalization of Higher-Order Theory of Mind

Theory of Mind is formally rendered in epistemic or doxastic logic as a parametric family of belief operators. First-order belief is denoted $B_i(\varphi)$ (“agent $i$ believes $\varphi$”); higher orders are expressed recursively:

$$B_{i_1} \left( B_{i_2}\left( \dots B_{i_k}(\varphi) \dots \right) \right)$$

Here, a $k$-th order ToM query asks about an agent's beliefs about another’s beliefs, recursively nested up to $k$ steps. The HI-TOM benchmark operationalizes tasks up to the fourth order, and related variants (e.g., MoToMQA) extend to sixth order (Street et al., 2024, Lei et al., 10 Jun 2026).

2. HI-TOM Benchmark Construction and Task Types

HI-TOM and its derivatives—such as MoToMQA and Hi-ToM—comprise handcrafted or programmatically generated narratives with complex agent interactions, partial observability, and possible deception. Each story is coupled with a suite of multiple-choice or binary (true/false) queries at varied ToM orders. Canonical task structures include:

• Zero-th order (“reality”): “Where is the object really located?”
• First to Fourth (or higher) order: “Where does A think B thinks … C thinks the object is?”

Narratives typically involve sequences where agents enter and exit environments, manipulate objects, and communicate (sometimes deceptively). Stories are designed to avoid spurious correlations by distributing correct answers and including distractors. Question prompts emphasize explicit reasoning: e.g., “Read the story (lines 1–17), then answer: Where does $C_4$ think $C_3$ thinks $C_2$ thinks $C_1$ thinks the carrot is?” (He et al., 2023, Street et al., 2024).

3. Benchmark Implementations and Methodological Advances

A variety of approaches have been proposed to tackle HI-TOM family benchmarks, targeting persistent LLM failure modes (insufficient recursion, event-order confusion, belief collapse). Notable inference-time methods include:

• Chain-of-Thought (CoT): Stepwise rationales, but prone to low-order or position bias errors.
• SimToM / Simulation Theory approaches: Recursively simulate agent perspectives, but may fail to maintain belief persistence.
• Decompose-ToM: Decomposes tasks into modules for agent identification, question reframing, world-model update, and knowledge availability; supports recursive simulation until a factual base case is reached (Sarangi et al., 15 Jan 2025).
• RecToM: Explicitly constructs agent perspectives in a recursive pipeline, generating partial knowledge sets and updating them as seen events accumulate. RecToM guarantees that the modal belief operator satisfies KD45 axioms (distribution, consistency, introspection), producing reliable state representations even under deep nesting (Lei et al., 10 Jun 2026).

Method	Recursion Depth	World-Model Persistence	Symbolic Belief Logic	Empirical Max Accuracy*
Chain-of-Thought	up to 2–3	Fragile	✗	~60–70%
SimToM	up to 3–4	Varies	Partial	~50–55% (LLama-3-8B)
Decompose-ToM	up to 4	Explicit	Partial	~87% (GPT-4o order 4)
RecToM	≥4	Formalized	KD45-compliant	100% (GPT-5.4)

*Best reported on corresponding backbone under original benchmark.

4. Empirical Results and Model Scaling Effects

Empirical studies across multiple HI-TOM variants reveal a consistent pattern: accuracy declines as ToM order increases, both in standard prompting and with naïve agent simulation strategies. Key findings include:

• Human and LLM comparison (MoToMQA): GPT-4 and Flan-PaLM approach or exceed adult human performance up to sixth-order tasks, while smaller or non-finetuned LLMs plateau near random guessing (Street et al., 2024).
• Order-dependence: Even GPT-4 and comparable LLMs experience a drop in accuracy at fifth-order queries (human: 98%, GPT-4: 82%), but can surpass humans at sixth order (GPT-4: 93%, human: 82%).
• Algorithmic breakthroughs: RecToM reaches 100% accuracy at up to fourth order on Hi-ToM with GPT-5.4 and Qwen3.5; Decompose-ToM yields large boosts (up to +40%) on third- and fourth-order queries for GPT-4o and Llama-3-70B (Lei et al., 10 Jun 2026, Sarangi et al., 15 Jan 2025).
• Scaling and fine-tuning: There exists a "computational potential" breakpoint above which higher-order ToM emerges, notably above ~175B parameters in instruction-finetuned models (Street et al., 2024).

5. Failure Modes and Error Analysis

Analysis of reasoning traces and model outputs identifies the primary failure types:

• Reasoning depth errors: Models shortcut recursive chains and answer at lower ToM order than required.
• Temporal ignorance: Misaligned event timeline leads to incorrect instantiation of agent beliefs.
• Belief collapse and omniscient bias: LLMs default to physical world knowledge instead of maintaining agent-relative uncertainty (especially in cases of partial observability).
• Deception handling: Story elements involving lies or double-bluffs exacerbate error rates; joint accuracy for high-order queries drops sharply when communication-based deception is included (He et al., 2023, Srishty et al., 19 May 2026).

Curriculum learning and adversarial data generation (as in OSCToM) help address these weaknesses by increasing the density of observer-self conflicts and deception chains in training corpora, though performance still plateaus in the mid-60% range for models <32B parameters (Srishty et al., 19 May 2026).

6. Broader Applications and Extensions

Higher-order HI-TOM benchmarks have motivated or directly contributed to development in several areas:

• Multi-agent collaboration: The LLM-Hanabi benchmark demonstrates that first-order ToM proficiency more strongly predicts in-game collaboration performance than second-order, though higher-order ToM remains essential for advanced settings and nuanced coordination (Liang et al., 6 Oct 2025).
• Neuro-symbolic modeling: There is increasing advocacy for hybrid architectures that encode explicit belief chains or symbolic memory in conjunction with LLM computation to surpass System-1 heuristic limitations (He et al., 2023).
• Cognitive science: Systematic probing of LLMs with HI-TOM variants provides comparative baselines for artificial versus human ToM and reveals qualitative differences in error profiles and reasoning strategies.

7. Implications, Limitations, and Future Directions

HI-TOM and its higher-order descendants highlight both the promise and the persistent challenges of achieving truly human-comparable ToM in artificial agents. While LLMs now rival or exceed humans on narrow, strictly factual recursive ToM tasks—especially under algorithmic guidance (e.g., RecToM, Decompose-ToM)—performance degrades in the presence of communicative deception, high temporal complexity, and under computational or memory limitations associated with scaling.

Ongoing research is focused on:

• Extending benchmarks to higher order, incorporating multimodal/social signals, and increasing ecological validity.
• Developing robust, generalizable algorithms for recursive reasoning not limited to specific narrative templates or question forms.
• Integrating HI-TOM tasks into collaborative game-play and real world, imperfect information scenarios to further stress-test social reasoning capabilities (Liang et al., 6 Oct 2025).

The HI-TOM benchmark suite stands as a critical diagnostic and developmental tool for advances in machine Theory of Mind and artificial social intelligence more broadly (He et al., 2023, Street et al., 2024, Lei et al., 10 Jun 2026, Sarangi et al., 15 Jan 2025, Srishty et al., 19 May 2026).