Avalon-ToM-Bench: Evaluating Fine-Grained Theory of Mind via Asymmetric Game Mechanics
2026-08-10 • Artificial Intelligence
Artificial IntelligenceComputation and LanguageComputers and SocietyComputer Science and Game Theory
AI summaryⓘ
The authors created Avalon-ToM-Bench, a new test to better measure Theory of Mind (ToM) by using a social game called The Resistance: Avalon. They break down ToM into four parts involving knowledge and motivation, inference and action, with special questions to test each part. Testing 28 large language models, they found these models understand game rules well but struggle with social reasoning itself. Also, models often have the right mental understanding internally but fail to show it when producing answers. Finally, training specifically for reasoning helps a lot more than just thinking longer during testing.
Theory of MindThe Resistance: Avalonasymmetric informationlarge language modelsepistemic reasoningmotivational reasoninginferencechain-of-thoughtlinear probingactivation steering
Authors
Yen-Shan Chen, Yu Chian Duan, Chih-En Kuo, Jian-Bin Wu, Yun-Nung Chen
Abstract
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.