LLMs trained to predict what readers expect next in stories

A new approach to understanding reader expectations#

How do readers decide whether to keep reading a story? Much of that decision hinges on their predictions about what happens next. A new research framework harnesses large language models to estimate these forward-looking beliefs, offering a scalable way to model how audiences anticipate narrative developments.

How the framework works#

The method generates multiple imagined continuations of a story using a pre-trained LLM, then extracts meaningful features from those variations—including emotional tone and narrative trajectory. The researchers grounded these features in narrative theory to ensure they capture psychologically relevant dimensions of story expectation.

To validate the approach, they employed two complementary strategies:

• Survey-based validation: Comparing LLM-derived predictions to expectations that readers reported directly in controlled lab settings
• Rational-expectations validation: Testing predictions against actual story outcomes observed in practice

What the data showed#

Testing on both lab surveys and real-world reading behavior from an online platform, the team found that LLM-generated expectations correlated meaningfully with human beliefs across all features examined. Notably, forward-looking predictions explained reader engagement even after accounting for features already present in the consumed portion of the narrative.

Implications for storytelling and platforms#

The scalable nature of this framework opens possibilities for multiple applications: helping writers refine narrative structures, informing platform recommendations, and deepening academic understanding of how narrative media shapes audience behavior. The work demonstrates that generative models can approximate the cognitive process of anticipation that drives audience engagement with stories.