Language Models as Agent Models

Demonstrates that LLMs don't just generate text — they infer and represent properties of the agents who would have produced that text. They model the psychology behind the characters. A 98% author identity experiment shows the model tracking who would have written what. The key insight: 'beliefs exist only for individual agents' — the model reasons in terms of characters with beliefs, not abstract knowledge. The paper that turns the philosophical claim about LLMs-as-simulators into an empirical finding with testable predictions.