On July 11, 2026, researchers Oliver Steele, Jiangtao Wen, and Yuxing Han published a paper exploring the separation of belief and reality in language models. Their findings reveal how these models can distinguish between what a character believes and what is true.
Understanding Belief-Reality Separation
The paper demonstrates that advanced language models can hold a character's beliefs apart from objective reality. For instance, if a character named Anna believes a cup is blue while it is red in reality, models can accurately respond with blue for Anna's belief and red for the actual situation. This capability hinges on two distinct mechanisms operating within the models.
According to the authors, the separation resides in two specific positions within the model's computation. A generic value slot binds the attributed value, while a router at the query position determines whether to access the character's belief or the actual state of the world.
Mechanisms Behind the Model's Functionality
The study identifies two routes that fill the value slot: an asserted belief, which is directly supplied by the text, and a derived belief that requires inference based on visibility. This means that the model can deduce what the character could see to arrive at the derived belief.
Notably, the research indicates that a subspace trained on one route can influence the other. Interestingly, only the derived route relies on described visibility, suggesting a complex interplay between the two mechanisms. The authors emphasize that the slot itself does not carry a belief-reality tag; altering it affects both belief and reality readouts equally.
Implications for Future Research
The findings have significant implications for understanding how language models operate. The research highlights a dissociated pair of routing subspaces that can switch a query between belief and reality without altering the underlying values. This innovative approach was validated across three different model architectures and various stimuli.
Additionally, the authors note that their work contributes to a broader understanding of the belief-reality axis, with a companion paper indicating that this slot-and-router format is applicable to various non-actual contexts, including counterfactual and fictional scenarios.
- Publication Date: July 11, 2026
- Authors: Oliver Steele, Jiangtao Wen, Yuxing Han
- Paper Length: 21 pages
- Figures and Tables: 6 figures, 6 tables
🤖 This article was rewritten by Feed and Figures' editorial AI from a report originally published by arXiv NLP. Facts and quotes are preserved from the original; the rewrite focuses on clarity and structure. For the unedited original, see the source link below.
