Evaluating Theory of Mind (ToM) in Large Language Models (LLMs) is an important area of research for understanding the social intelligence of AI. Recent ToM benchmarks have made significant strides in enhancing the complexity, comprehensiveness, and practicality of evaluation. However, while the focus has been on constructing “more difficult” or “more comprehensive” tasks, there has been insufficient systematic analysis of the structural factors that inherently determine the difficulty of ToM reasoning—that is, “what” makes reasoning difficult. To address this challenge, we propose a new dataset generation framework for ToM evaluation named AnaToM. To realize an “Anatomy of Difficulty” in ToM reasoning, AnaToM strictly controls structural parameters such as the number of entities and the timeline in a story. This parameter control enables the isolation and identification of factors affecting the ToM of LLMs, allowing for a more precise examination of their reasoning mechanisms. The proposed framework provides a systematic methodology for diagnosing the limits of LLM reasoning abilities and offers new guidelines for future benchmark design.

Large Language Models (LLMs) are considered to have potentially extensive knowledge, but because their internal processing is black-boxed, it has been difficult to directly edit the knowledge held by the LLMs themselves. To address this issue, a method called local modification-based knowledge editing has been developed. This method identifies the knowledge neurons that encode the target knowledge and adjusts the parameters associated with these neurons to update the knowledge. Knowledge neurons are identified by masking the o part from sentences representing relational triplets (s, r, o), having the LLM predict the masked part, and observing the LLM's activation during the prediction. When the architecture is decoder-based, the predicted o needs to be located at the end of the sentence. Previous local modification-based knowledge editing methods for decoder-based models have assumed SVO languages and faced challenges when applied to SOV languages such as Japanese. In this study, we propose a knowledge editing method that eliminates the need for word order constraints by converting the input for identifying knowledge neurons into a question where o is the answer. We conducted validation experiments on 500 examples and confirmed that the proposed method is effective for Japanese, a non-SVO language. We also applied this method to English, an SVO language, and demonstrated that it outperforms conventional methods.