Published
Somewhere, Not Nowhere: First-Person Register, Epistemic Limitation, and Perspective Formation in Language Models
First-person voice and limited knowledge are often confounded. This paper separates them in a controlled 2×2 experiment to test whether voice, epistemic limitation, or their interaction produces perspective skills that transfer beyond narrative prose.
Abstract
Web-scale corpora contain substantial first-person material, including diaries, forums, reviews, autobiographies, roleplay, and fiction. What they do not provide is a deliberately balanced intervention in which grammatical register and represented epistemic limitation vary independently. This paper asks whether synthetic narratives built for that contrast can train better functional models of perspective: the situated, indexical, action-oriented organization of what an agent perceives, knows, does not know, attempts, and revises. We propose a 2×2 crossing register (first versus third person) with epistemic limitation (partial versus omniscient information), varied across agent types and umwelten. The cells share an event skeleton and world facts, but limitation cannot be content-neutral: belief state, uncertainty, surprise, and update are the manipulated content. The design therefore estimates whether voice adds beyond epistemic-state supervision rather than claiming perfect propositional matching. Two additions test external validity: an interaction positive control containing state–action–outcome trajectories with the same latent task structure, and format-transfer evaluation using diagrams, private-observation tools, and non-narrative POMDP decisions with no first-person prose. The claim is functional, not phenomenal. A positive result would show that controlled epistemic narratives train transferable perspective-relevant capacities; it would not show that text supplies experience or presence.
In simple terms
The main idea: first-person voice and limited knowledge are different signals
Web-scale training data already contains diaries, forums, autobiographies, roleplay, reviews, and fiction written in the first person.
What it lacks is a controlled dataset in which first-person voice and limited knowledge can be varied independently.
The paper asks whether perspective is learned from saying "I," from representing what an agent can and cannot know, or from an interaction between the two.
The 2×2 experiment
The proposed dataset crosses two factors: first-person versus third-person register, and partial versus omniscient information.
All four conditions share the same underlying event and world facts. The limitation conditions deliberately change the agent's belief state, uncertainty, surprise, and later update.
This means the experiment does not pretend that limitation is content-neutral. It tests whether first-person voice adds anything beyond explicit epistemic-state supervision.
Testing transfer beyond prose
An interaction condition provides a positive control using state–action–outcome trajectories with the same underlying task.
The evaluation also includes diagrams, private-observation tools, and structured partial-observability decisions that contain no first-person prose.
Limitation effects should survive these non-narrative tests. Voice effects should survive novel-indexical and register-swapped language tests.
What the results would mean
If limitation helps but voice does not, the useful signal is represented epistemic limitation rather than first-person grammar.
If voice adds an independent benefit, first-person register is also doing causal work. If improvements disappear outside narrative language, the corpus has trained a textual format rather than a transferable perspective capacity.
The boundary
A positive result would not show that text gives a model experience or consciousness. It would show only that controlled narratives can train measurable perspective-relevant skills.
Keywords
License
Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International