This preprint from arXiv asks a deceptively simple question: do language models know what is going to happen in a story before they write it, and if so, how far ahead? The authors approach the problem at the level of narrative events—salient, state-changing moments—rather than individual tokens or whole-response attributes.

By framing lookahead in terms of events, the work aims to capture something closer to plot structure than to surface-level text statistics. The abstract indicates the authors extract event-level representations to study how much future information is present in the model's internal state.

The paper's title points to three related investigations: whether such narrative lookahead can be decoded from model representations, whether it provides a predictive advantage, and whether it can be used to steer generation. The abstract itself is cut off before presenting results, so this summary is limited to the stated research agenda.