Avalon-ToM-Bench: Evaluating Fine-Grained Theory of Mind via Asymmetric Game Mechanics

Source

• Source name: arXiv
• Source kind: paper
• Authors / organization: Yen-Shan Chen, Yu Chian Duan, Chih-En Kuo, Jian-Bin Wu, Yun-Nung Chen
• Venue: arXiv: cs.AI
• Published: 2026-08-10
• Collected: 2026-08-11
• URL: http://arxiv.org/abs/2608.09638v1
• PDF: https://arxiv.org/pdf/2608.09638v1
• DOI: UNKNOWN
• Registry group: academic_apis_and_preprints
• Evidence role: canonical_academic_index
• Language: en
• Country / region: UNKNOWN

Plain-Language Brief

A new benchmark called Avalon-ToM-Bench is introduced to evaluate fine-grained Theory of Mind (ToM) abilities in agents. The benchmark operates through the asymmetric-information mechanics of a game named The Resistance: Avalon, and decomposes ToM into four categories: epistemic vs motivational reasoning, and inference vs action. The study reveals three insights about large language models (LLMs): 1) Models show strong understanding of game rules but weak abilities in social reasoning. 2) Models often represent correct mental-state inferences in their hidden states but fail to express them during generation. 3) Dedicated reasoning training leads to substantial improvements, while test-time chain-of-thought provides only minor gains.

What the Document Says

• Theory of Mind (ToM) is evaluated using a new benchmark called Avalon-ToM-Bench.
• The benchmark uses the asymmetric-information mechanics of The Resistance: Avalon game.
• It decomposes ToM into four categories: epistemic vs motivational reasoning, and inference vs action.
• Study reveals three insights about large language models (LLMs).
• Insight 1: Models show strong understanding of game rules but weak abilities in social reasoning.
• Insight 2: Models often represent correct mental-state inferences in their hidden states but fail to express them during generation.
• Insight 3: Dedicated reasoning training leads to substantial improvements, while test-time chain-of-thought provides only minor gains.

Key Concepts Mentioned

• Theory of Mind
• Avalon-ToM-Bench
• The Resistance: Avalon
• large language models (LLMs)
• epistemic reasoning
• motivational reasoning
• inference
• action

Claims Made by the Document

• Existing evaluations oversimplify mental-state reasoning or provide limited diagnostic insight.
• Avalon-ToM-Bench operationalizes ToM using asymmetric game mechanics.
• The study reveals three insights about LLMs.

Evidence Used

• The text describes the purpose and design of Avalon-ToM-Bench.
• The text states that the study reveals three insights about LLMs.

Methods / Approach

• Human-crafted, perspective-constrained queries are used to evaluate ToM.
• Linear probing and activation steering are used for mechanistic analyses.

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

The study only evaluates large language models and doesn't provide any information about the performance of other types of agents.

Open Questions

• How does Avalon-ToM-Bench compare to other evaluation methods?
• What are the potential implications of these findings for AI development?

Notes for Future Q&A

May include interpretation. Labeled where it does.

The text doesn't provide direct comparisons with other evaluation methods, but it suggests that Avalon-ToM-Bench offers a more fine-grained approach to evaluating Theory of Mind.

Raw Abstract / Excerpt

Theory of Mind (ToM) is essential for agent interactions, yet existing evaluations either rely on static scenarios that oversimplify mental-state reasoning or interactive settings that provide limited diagnostic insight. We present Avalon-ToM-Bench, a fine-grained benchmark that operationalizes ToM through the asymmetric-information mechanics of The Resistance: Avalon. Rather than evaluating end-to-end gameplay, it decomposes ToM into a 2$\times$2 taxonomy -- epistemic versus motivational reasoning crossed with inference versus action -- using human-crafted, perspective-constrained queries. Benchmarking 28 LLMs reveals three insights: 1) Reasoning, not knowledge. Models show strong game-rule comprehension but markedly weaker ToM abilities, isolating failures to social reasoning rather than missing domain knowledge. 2) Expression, not representation. Mechanistic analyses via linear probing and activation steering show that models frequently represent correct mental-state inferences in their hidden states but fail to express them during generation -- linear probes recover 77-82% accuracy versus 62-70% from the models' own chain-of-thought. 3) Policy, not deliberation. Dedicated reasoning training yields substantial improvements whereas test-time chain-of-thought provides only marginal gains (+11.0 versus +1.1 points on average), suggesting that robust ToM depends on a learned reasoning policy rather than increased inference-time deliberation.