New study reveals how LLMs separate belief from reality via routing

A new arXiv paper (2607.11945) by Steele, Wen, and Han uncovers the computational mechanism behind language models' ability to separate a character's beliefs from objective reality. The researchers demonstrate that this ability relies on two separable components: a generic value slot that binds the attributed value (e.g., the color blue), and a router at the query position that selects whether to read out the belief frame or the reality frame. Two distinct routes fill the slot: an 'asserted belief' route, where the text directly supplies the value, and a 'derived belief' route, where the value is inferred via a visibility-gated lookback from what the character could observe. Crucially, the slot itself carries no belief-reality tag—intervening on it shifts a reality readout as strongly as a belief readout. Instead, separation is achieved through a dissociated pair of routing subspaces that flip the query between frames without injecting the donor's value.

The findings hold across three architectures (likely transformer-based) and emerge between 3B and 7B parameters in five model families, with stimuli carefully de-confounded against theory-of-mind benchmarks. Only the derived route depends on described visibility, revealing a modular structure. This work is accompanied by a companion paper showing the same slot-and-router format applies to other non-actual contexts like counterfactual, fictional, and temporal frames. For AI practitioners, this reveals a principled internal mechanism underlying theory-of-mind reasoning, offering a path to more controllable and interpretable belief handling in large language models.