Large language models may have a layered functional self based on role relations
What Is the "Self" of a Large Language Model?
Computation and Language
Summary
It is unclear what makes a large language model have a 'self' since it is not just one word or layer inside it. The authors suggest that the model's self comes from how it organizes different roles, like the user, system, and assistant, in its information processing. They studied a model with 27 billion parameters and found that parts of the model keep roles in distinct relationships that shift as you go deeper through layers. This means the model's 'self' might come from these evolving but connected role representations rather than a fixed part inside it.
What this means in practice
- •For llm developers: Design LLM architectures that separate roles like user, system, and assistant for clearer internal representation and improved model control.
- •For ai safety teams: Use role-conditioned geometry insights to monitor and interpret how LLMs process different inputs and maintain consistency over dialogues.
Authors
Yuchen Li, Zongqi Fan, Xintong Li, Ken-tye Yong, Nguyen H. Tran
Abstract
What could the "self" of a large language model be, if it is neither a particular word nor a single privileged layer? We propose a relational account: a functional self is an actor-centered organization of information that distinguishes the model's recorded actions from externally supplied material while content and reasoning operations change. Its representational candidate is a family of layer-specific source-binding components, rather than one unchanged vector throughout the network. Preliminary observations from a 27B BF16 model provide a geometric starting point. Across 150 matched message bodies, 450 role conditions, and a 75/75 discovery-held-out split, the combined Assistant-versus-external displacement transfers strongly across bodies and positions. Selected early components place User and System closer to one another than to Assistant, whereas this full-space distance relation does not persist throughout depth. These findings motivate a layered account: an early source-encoding candidate can be combined with role-specific information use and continuation preparation at later depths, without requiring monotonic overwriting of the common displacement. We connect this account to dynamic reasoning, historical continuity, and a computational model separating content operations, source binding, authority, and output. The reported evidence establishes a reproducible role-conditioned geometry; the proposed interpretation is that functional selfhood may reside in persistent relations among changing representations.