False-Belief Tasks

False-belief tasks separate what is true from what an agent believes. In a Sally-Anne style vignette, an object moves while Sally is away, so Sally's search location differs from reality. For language models, the story logic is only part of the challenge. Variants must prevent a shallow heuristic from reaching the same answer.

1. The shortcut is part of the task

A recency heuristic answers with the last-mentioned location. A copy heuristic answers by copying the queried object's current value. A visibility heuristic ignores belief unless the story explicitly says who saw the move. A good stimulus removes these shortcuts one by one, so success has to track the agent's information state.

2. Unexpected contents and change of location

Unexpected-contents tasks ask what someone thinks is inside a container after the contents are revealed to the reader. Change-of-location tasks ask where an agent will look after an object moves unseen. Both require a belief state that can diverge from reality, and both need minimal pairs where the reality answer and belief answer trade places.

3. Where the belief lives

The task keeps two answers about one object ready at once. Inside the model the two values sit in a shared value slot; a router at the query position selects which one the answer reads. Belief-versus-reality lives in the routing, not in the value: the mechanism worked out for belief–reality separation in language models.