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

Researchers introduce Avalon-ToM-Bench, a new benchmark for evaluating fine-grained Theory of Mind (understanding others' thoughts and motivations) in AI models. The benchmark uses the asymmetric-information mechanics of a popular game called 'The Resistance: Avalon'. The study reveals three key findings about large language models: 1) They have strong comprehension of game rules but struggle with social reasoning, 2) Models often represent correct mental-state inferences in their hidden states but fail to express them during generation, and 3) Improvements in Theory of Mind abilities come from learned reasoning policies rather than increased inference-time deliberation.

What the Document Says

• Theory of Mind (ToM) evaluation: oversimplified mental-state reasoning or limited diagnostic insight
• Avalon-ToM-Bench: operationalizes ToM through asymmetric game mechanics, decomposes ToM into a 2x2 taxonomy
• 28 LLMs analyzed: revealed three insights
• 1. Reasoning, not knowledge: models have strong rule comprehension but weak ToM abilities
• 2. Expression, not representation: models often represent correct mental-state inferences but fail to express them
• 3. Policy, not deliberation: dedicated reasoning training improves ToM significantly, test-time chain-of-thought provides minor gains

Key Concepts Mentioned

• Theory of Mind (ToM)
• Avalon-ToM-Bench
• large language models (LLMs)
• asymmetric game mechanics
• epistemic reasoning
• motivational reasoning
• inference
• action

Claims Made by the Document

• Existing ToM evaluations oversimplify mental-state reasoning or provide limited diagnostic insight
• Avalon-ToM-Bench decomposes ToM into a 2x2 taxonomy
• Models show strong game rule comprehension but weaker ToM abilities
• Models frequently represent correct mental-state inferences in their hidden states but fail to express them during generation
• Dedicated reasoning training yields substantial improvements in ToM abilities
• Test-time chain-of-thought provides only marginal gains in improving ToM abilities

Evidence Used

• oversimplified mental-state reasoning or limited diagnostic insight is a problem with existing evaluations
• Avalon-ToM-Bench uses asymmetric game mechanics and decomposes ToM into 2x2 taxonomy
• 28 LLMs analyzed and reveal three insights
• models show strong rule comprehension but weaker ToM abilities is one of the insights
• models often represent correct mental-state inferences but fail to express them is another insight
• dedicated reasoning training improves ToM significantly and test-time chain-of-thought provides only marginal gains is a third insight

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

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.