Evaluating Theory of Mind in Reasoning Models: Robustness over Reasoning

Abstract. Large language models (LLMs) have recently shown strong performance on Theory of Mind (ToM) tests, prompting debate about the nature and validity of the underlying capabilities. At the same time, reasoning-oriented LLMs trained via reinforcement learning with verifiable rewards have demonstrated notable improvements across a range of benchmarks. In this work, we examine the behavior of such reasoning models in ToM tasks using novel adaptations of machine psychological experiments together with results from established benchmarks. We observe that reasoning models consistently exhibit increased robustness to prompt variations and task perturbations. Our analysis suggests these gains come at least partly from models being more robust at reaching the correct answer under prompt and task variation. We read this as evidence for a robustness-based account rather than for a new ToM-specific ability.

Introduction. In recent years, large language models (LLMs) have become sufficiently capable to be part of our daily lives, and are increasingly used in social and agentic settings [16]. Users tend to attribute human-like intentionality and reasoning to those models, even though the extent to which the models are capable of complex reasoning remains, to a large extent, an open research problem [29,19,5,17]. Due to the complexity of these systems, even if all information about their architecture is known, it is still difficult to determine and predict their actual behavior. Considering this, researchers in a multitude of fields started investigating those systems and their intelligence by interacting with them and understanding their behavior, rather than by analyzing their architecture. [19] In this context, previous work has debated the extent to which LLMs demonstrate Theory of Mind (ToM) behavior, that is, the ability to reason about mental states, beliefs, intentions, and desires. Some early claims were optimistic of such a skill emerging to some extent [11].

Discussion / Conclusion. The results show that models have improved substantially in ToM tasks since 2023. Part of this is likely just that models have become more capable overall, but in what follows we focus on a second factor: a newfound robustness in their ToM skills. Yue et al. [28] show that the reasoning paths produced by reasoning models stay bounded by their base models, reaching asymptotic performance without lowering perplexity on the base model, reasoning training does not appear to add fundamentally new capabilities. Read alongside our own results, this points to a specific interpretation. In our experiments the thinking-enabled models, and most clearly the thinking-on version of Claude compared to its thinkingoff counterpart, were more robust to prompt and task variation. We therefore hypothesize that the main effect of this kind of reasoning training is improved stability in reaching a solution the model could already, in principle, reach, rather than an expansion of representational capacity. We stress that our setup compares thinking and non-thinking models rather than RLVR-trained and non- RLVR-trained ones, so we cannot attribute this effect to RLVR specifically (see Section 5.2).