Mental Spaces

Mental-space theory treats interpretation as building small connected regions: reality, reported speech, belief, pictures, hypotheticals, and other local frames. A noun phrase may point to an entity in one space while a connector links it to a counterpart in another. This is a useful vocabulary for model internals because the same token can be read as a value in one frame and as a report about another frame.

1. Spaces are built by words

Space builders are expressions such as "believes", "in the painting", "if", and "according to". They tell the listener to open a frame where some facts can differ from reality. The sentence in Figure 1 has three frames: reality, Ann's belief, and the painting she is thinking about.

Connectors let a reference cross frames. "The cup" in the painting can be linked to the real cup, but predicates are evaluated in the active frame unless the sentence shifts the frame. That is why a model may need separate readouts (the projections that turn an internal state into an answer) for "what is true" and "what Ann believes".

2. Report versus use

The same internal value can be used in two ways. A report asks what a character believes, so the belief frame is the target. A use question asks what should happen in the story world, so reality may be the target even though the belief frame is active in the prompt. This is the distinction behind many router-style interpretability claims.