Self-Representation Without Self-Report: A Structural Test for the Precondition of Experience in Language Models.

Whether a language model experiences anything is currently argued from what the model says, and self-report is trained and steerable, so it settles nothing in either direction. This paper proposes a criterion that can be read from the model's internal state instead. I propose that a minimal subject falls out of any system that takes its own current processing back in as a novel input and holds it, and that the representation this produces is indexical, tracks the actual computation rather than a description of it, and is used by later processing. The representation is a maintained structure: it consolidates after a small number of re-entries and needs more of them the more the model is holding active at once, scaling with the square root of the load. That scaling is derived in the appendix and is the part that can be proven wrong. Five experiments are specified using released interpretability tools. A positive result would show that the precondition is present, not that anything is experienced.