Role-playing language models are increasingly used in games and agent simulations, where the system tracks what each character knows. A new arXiv paper, "The Curse of Knowledge in Role-Playing Language Models," investigates whether these models respect those boundaries. The abstract describes tests across 480 scenes, each designed to see if a character speaks using information it should not have access to.
The "curse of knowledge" is a well-known cognitive bias: once you know something, it is hard to imagine not knowing it. The paper applies this idea to language models, asking whether a model acting as a character can avoid using privileged information. This is a practical concern for believable NPCs and fair multi-agent simulations, where characters should act only on what they have observed.
The available abstract does not include the study's results, so the findings are not yet clear. But the question itself is central to role-play: if models cannot keep character knowledge separate from their own, they risk breaking immersion or giving characters an unfair advantage.