Can LLMs predict character choices from narrative context?

Can LLMs predict how fictional characters will act at pivotal moments? The Character is Destiny paper (2024) constructs LIFECHOICE — a benchmark of 1,462 decision points from 388 novels, leveraging expert character analyses from literary scholars. The task: given the preceding narrative, predict which choice the character makes.

The architecture decomposes into two components. First, a character profile combining a static description (personality, experiences, values) with retrieved memories — specific passages from the preceding text. Second, a reasoning step using the profile to answer the decision question.

Three methods for constructing descriptions reveal a hierarchy: expert-written descriptions (from Supersummary) outperform both hierarchical merging (summarize chunks, merge summaries iteratively) and incremental updating (summarize sequentially, refine). This suggests that literary expertise captures something about character motivation — the relationship between personality, values, and action — that automated summarization misses.

The CHARMAP method adds persona-based memory retrieval: selecting narrative passages relevant to the character's psychological profile rather than just the decision context. This yields a 5.03% accuracy gain, indicating that who the character is determines which memories matter for predicting what they will do.

Character-driven motivations decompose into: personality and traits, emotions and psychological state, social relationships, values and beliefs, and desires and goals. This taxonomy suggests that persona simulation for decision-making requires richer internal models than the demographic + preference approaches used in most LLM persona work.

The connection to Why do LLMs fail to act on their stated beliefs? is instructive: when beliefs are extracted from rich narrative context rather than assigned through brief prompts, behavioral prediction improves.