Does reasoning training improve theory of mind or just stability?

The paper's central claim is a reinterpretation of why reasoning models score better on Theory of Mind tests. Using "novel adaptations of machine psychological experiments together with results from established benchmarks," the authors report that reasoning models "consistently exhibit increased robustness to prompt variations and task perturbations." Their analysis suggests the gains come "at least partly" from models being "more robust at reaching the correct answer under prompt and task variation," and they read this "as evidence for a robustness-based account rather than for a new ToM-specific ability." The discussion also says models "have improved substantially in ToM tasks since 2023," and that part of this is "likely just that models have become more capable overall." Robustness is offered as a second factor, not the only one.

The mechanism is stated as a hypothesis rather than a demonstrated result. The authors borrow a finding from Yue et al., that the reasoning paths of reasoning models stay bounded by their base models and that reasoning training "does not appear to add fundamentally new capabilities." Read alongside their own results, this yields the claim 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." The clearest support in the excerpt is the thinking-on version of Claude, which was more robust to prompt and task variation than its thinking-off counterpart. On this view, thinking makes a latent ToM skill more reliable, not larger.

This sits in an unusual place among the neighbors. Two existing notes report that reasoning training does not help social cognition: Why do reasoning models fail at theory of mind tasks? finds regression on Decrypto, and Why do reasoning models struggle with theory of mind tasks? finds no link between reasoning effort and accuracy. This paper agrees that reasoning training adds no ToM-specific mechanism, but it does not agree that nothing improves: it reports better performance since 2023 and a real robustness effect. One way to hold both is that reasoning may steady an existing ability without extending it, so a gain shows up under perturbation more clearly than on a fixed benchmark. That reconciliation is my reading, and the excerpt does not make it. The robustness account also bears on Can language models solve ToM benchmarks without real reasoning? and on Do large language models genuinely simulate mental states?. All three ask what a ToM score measures, and this paper's answer is that it can partly measure how stable a model is under rewording.

The excerpt leaves a lot unestablished. It names no benchmarks, model list, sample sizes or effect sizes beyond the Claude thinking-on versus thinking-off contrast, and the "at least partly" wording leaves open how much of the gain robustness explains. The authors state that the setup "compares thinking and non-thinking models rather than RLVR-trained and non-RLVR-trained ones," so the effect cannot be attributed to RLVR specifically. The bounded-by-base-model premise comes from another paper and is used here as supporting context, not as something this study tests. What follows at this strength is a caution about how to describe ToM results: a reasoning model passing a perturbed ToM task is evidence of reliability, and the excerpt gives no ground for calling it a new capacity to model minds.