Language models are increasingly used for tasks that require distinguishing real from imaginary scenarios, such as answering questions or writing fiction. A new paper on arXiv investigates whether these models internally represent event plausibility in a way that matches human judgments.
The study, titled "Is This Just Fantasy?", explores the modal category of sentences—whether an event is possible, impossible, or somewhere in between. The authors argue that for LMs to perform reliably, they must discern these distinctions.
According to the abstract, the research focuses on how LM representations reflect human judgments of event plausibility. While the full details are not yet available, the title and abstract suggest a positive alignment between model internals and human intuition.