A new arXiv paper introduces ToMStories, a synthetic narrative benchmark designed to probe theory-of-mind in language models. The authors frame the task around a common human ability: understanding what people feel or want from how they behave, even when they never say so directly. The work asks whether language models can make the same kind of inference, and what mechanisms underlie it.

Because the narratives are synthetic, the setup allows for controlled testing of mental-state reasoning without the confounds of real-world text. The paper's contribution appears to be the dataset itself, along with an interpretive lens for examining model behavior. The abstract does not report specific results, so the significance lies in the new resource and the questions it enables.

The emphasis on interpreting how models infer mental states sets this work apart from purely accuracy-based benchmarks. By focusing on behavior-driven inference in synthetic stories, ToMStories offers a way to isolate and study a core component of social reasoning in language models. Further details on dataset size, model performance, and findings would require access to the full paper.