Theory of Mind and Persuasion Beyond Conversation: Assessing the Capacity of LLMs to Induce Belief States via Planning and Action

The paper's sharpest contribution is not the benchmark itself but the conceptual split it forces: a model can score well on conversational ToM tasks while being entirely untested on whether it can engineer belief changes by manipulating the environment around a person, without ever speaking to them. That distinction has been invisible in prior evals.

This connects directly to the MECoBench work covered the same day, which exposed how multiagent embodied systems create coordination and communication dynamics that single-model benchmarks miss entirely. NCP-ExploreToM is probing a parallel blind spot: once models act in environments rather than just respond in dialogues, the safety-relevant capabilities shift in ways current evals cannot see. The surrogate fidelity study from the same period adds a compounding concern, since if open models diverge internally from closed ones even when predictions match, then auditing agentic persuasion capacity in proprietary systems through proxy models becomes even less reliable than it already looks.

Watch whether any major safety team (Anthropic, DeepMind, or OpenAI) incorporates a non-conversational planning ToM task into their next published eval suite. Adoption there within six months would signal the field accepts agentic persuasion as a first-class safety property rather than an academic edge case.