๐ง Theory of Mind Research
๐ค Automated Bayesian Theory of Mind for Open-ended Reasoning
Zhining Zhang, Chuanyang Jin, Mung Yao Jia
Researchers introduced AutoToM, a novel automated Bayesian Theory of Mind method that can operate in any domain, infer any mental variable, and conduct robust ToM reasoning of any order. Unlike previous approaches that rely on either error-prone LLMs or rigid handcrafted models, AutoToM iteratively refines its model based on inference uncertainty, achieving state-of-the-art performance across multiple benchmarks. https://doi.org/10.48550/arxiv.2502.15676
๐ฆ Bonobos Show Theory of Mind in Pointing Behavior
Luke A. Townrow, Christopher Krupenye
In a groundbreaking study, researchers found that bonobos point more for social partners who are ignorant about relevant information than for those who are knowledgeable. This provides compelling evidence that great apes can deploy theory of mind flexibly in communication, challenging previous assumptions about the uniqueness of human ToM for cooperative communication. https://doi.org/10.1073/pnas.2412450122
๐ฑ Multi-modal Multi-Agent Theory of Mind Benchmark
Haojun Shi, Suyu Ye, Xinyu Fang
Researchers introduced MuMA-ToM, the first multi-modal benchmark for evaluating mental reasoning in embodied multi-agent interactions. The benchmark provides video and text descriptions of people's behavior in realistic household environments and asks questions about people's goals, beliefs, and beliefs about others' goals. Their novel model LIMP outperformed state-of-the-art methods including GPT-4o and Gemini-1.5 Pro. https://doi.org/10.48550/arxiv.2408.12574
๐ฌ Enhancing AI Conversations with Theory of Mind
Mohammadmahdi Jafari, Devin Yuncheng Hua, Hao Xue
This study examined how effectively open source language models can capture and preserve ToM-related information and contribute to consistent ToM reasoning. By explicitly manipulating ToM components like beliefs, desires, and intentions, researchers improved response alignment in LLaMA 3 models, achieving win rates of 67% and 63% for the 3B and 8B models respectively. https://doi.org/10.48550/arxiv.2502.14171
๐งฉ Theory of Mind Mediates Cognitive Flexibility and Mindfulness in Children
Utku Beyazฤฑt, Bรผลra Kurtoฤlu Karataล, Aynur Bรผtรผn Ayhan
Researchers examined 282 children aged 9-11 years and found that theory of mind skills mediate the relationship between cognitive flexibility and mindfulness. While first-order false belief tasks didn't significantly mediate this relationship individually, second-order false belief and faux pas recognition tasks did. The findings highlight the importance of developing ToM skills in educational and psychological interventions. https://doi.org/10.1016/j.jecp.2024.106192
โ ๏ธ Current ToM Benchmarks Inadequate for Language Models
Matthew Riemer, Zahra Ashktorab, Djallel Bouneffouf
This position paper argues that most theory of mind benchmarks for LLMs are fundamentally flawed because they only measure "literal theory of mind" (ability to predict others' behavior) rather than "functional theory of mind" (ability to adapt to agents based on predictions about their behavior). The authors found that even top-performing open-source LLMs struggle with functional ToM despite sometimes showing strong literal ToM capabilities. https://doi.org/10.48550/arxiv.2412.19726
