Do LLMs use inferred beliefs to adapt their game strategies?

The paper asks a question that theory-of-mind tests leave open: not whether LLMs can report what another agent believes, but whether they can use inferred beliefs and intentions "to guide adaptive behaviour." It tests individual LLM agents from four model families (DeepSeek, GPT-4.1, GPT-5 and Gemini 2.0 Flash, N = 2,099) in two economic games against opponents of varying sophistication, with human participants (N = 251) as a comparison. Across both games, the LLMs "showed clear behavioural and computational signatures of mentalizing that differed markedly by model provider and size." The discussion adds that mentalization strategies varied across the two games, with significant differences across providers.

The mechanism is a change of measurement. Earlier studies, the authors say, assessed theory of mind by "choice accuracy in response to story-based prompts, obscuring the latent mechanisms underlying observed behaviour." The inspection game and rock-paper-scissors were previously validated in humans, give a "normative assessment of machine intelligence," and allow behavior to be examined "at specific depths of mentalizing." Cognitive computational modeling then recovers the latent strategy behind the choices. Humans showed "both recursive and adaptive mentalization," which the authors take as replication and evidence that the tasks are robust. A prompting strategy built to elicit strategic reasoning, Social Chain-of-Thought (SCoT), produced "robust improvements in performance, reflecting more sophisticated reasoning," though the abstract says the benefit "differed across the two tasks."

Against the library, this sits close to Do large language models use one reasoning style or many?, which also finds that strategic behavior is not one capability. That note varies the game and reads reasoning chains; this paper varies provider and size and fits latent strategy models to behavior, so the two agree that a single score hides structure. It supplies a behavioral answer to the evaluation-format worry in Do large language models genuinely simulate mental states?: instead of making questions more open-ended, it moves to interactive play. It also shares the concern of Can models recognize how individuals reason differently? that matching outputs is not the same as sharing a process. A prompting gain from SCoT is worth setting beside Why do advanced reasoning models fail at understanding minds?. That note concerns models trained for reasoning, while this one concerns a prompt aimed at the social task itself.

The excerpt does not say which provider or model mentalized more deeply, how any LLM compared quantitatively with the humans, how many agents ran in each game, or which of the two tasks gained more from SCoT and why. It also does not show that the fitted strategies reflect mental-state representation rather than good policies for these two games. What it supports is narrower: mentalizing in LLMs looks provider-dependent and task-dependent in interactive settings, so a claim that "LLMs have theory of mind" needs the model family and the task attached to it.