Tracing the Persona Circuit: How Large Language Models Encode and Express Character Traits

Large Language Models (LLMs) demonstrate remarkable potential in role-playing tasks but frequently suffer from personality decay—termed "Out-of-Character" (OOC) behavior—during prolonged interactions. While heuristic strategies exist to align model behaviors, the internal computational dynamics driving personality expression remain opaque. A fundamental barrier to decoding these mechanisms is a *metric gap*: while standard causal attribution paradigms target atomic, single-token outcomes, personality manifests as a holistic, multi-token behavioral tendency. We bridge this gap via the *Latent Persona Vector*, a differentiable proxy enabling the first fine-grained causal tracing of personality circuits. This metric reveals a structured "Preparation-Establishment-Expression" dynamic and identifies a mechanistic contributor to OOC behavior: competition between persona-specific signals and an assistant-like default direction during the critical "Establishment" phase. Guided by this diagnosis, we propose surgically recalibrating the signal magnitude in fewer than $5\\%$ of attention heads. This targeted intervention effectively strengthens the persona signal, significantly restoring character consistency while preserving general reasoning capabilities.

Large language models can act as fictional characters or follow a requested personality, but they often drift back to a generic "helpful assistant" style during longer conversations. This paper studies why that happens by looking inside the model while it is generating role-playing responses. We find that the model appears to handle personality in three broad steps: first gathering clues about the requested character, then combining them into a stable character intention, and finally turning that intention into words. Character drift often happens when the requested personality is present but too weak to overcome the model's default assistant-like behavior. Based on this finding, we propose a small, targeted adjustment that strengthens the character signal during generation, without retraining the model. Experiments show that this improves character consistency in long conversations while largely preserving the model’s general reasoning ability.