Fame Fades, Nature Remains: Disentangling the Character Identity of Role-Playing Agents

TL;DR

This paper formalizes character identity in Role-Playing Agents via a two-layer framework: Parametric Identity (LLM pre-training knowledge) and Attributive Identity (fine-grained traits and relationships). A unified 38-field Character Profile Schema enables controlled comparisons between Famous and Synthetic characters, assessed across single-turn and multi-turn interactions. The key findings are Fame Fades—parametric advantages yield short-lived benefits that fade as interactions proceed—and Nature Remains—performance is heavily driven by negatively-valenced Motivations and Interpersonal Relationships, with general personality traits showing limited polarity sensitivity. These insights shift focus toward robust representation and evaluation of negative social and moral attributes in RPAs, informing better character construction and more comprehensive benchmarks.

Abstract

Despite the rapid proliferation of Role-Playing Agents (RPAs) based on Large Language Models (LLMs), the structural dimensions defining a character's identity remain weakly formalized, often treating characters as arbitrary text inputs. In this paper, we propose the concept of \textbf{Character Identity}, a multidimensional construct that disentangles a character into two distinct layers: \textbf{(1) Parametric Identity}, referring to character-specific knowledge encoded from the LLM's pre-training, and \textbf{(2) Attributive Identity}, capturing fine-grained behavioral properties such as personality traits and moral values. To systematically investigate these layers, we construct a unified character profile schema and generate both Famous and Synthetic characters under identical structural constraints. Our evaluation across single-turn and multi-turn interactions reveals two critical phenomena. First, we identify \textit{"Fame Fades"}: while famous characters hold a significant advantage in initial turns due to parametric knowledge, this edge rapidly vanishes as models prioritize accumulating conversational context over pre-trained priors. Second, we find that \textit{"Nature Remains"}: while models robustly portray general personality traits regardless of polarity, RPA performance is highly sensitive to the valence of morality and interpersonal relationships. Our findings pinpoint negative social natures as the primary bottleneck in RPA fidelity, guiding future character construction and evaluation.

Figures (10)

• Figure 1: The two-layered Character Identity. (A) Parametric Identity: Comparison between famous and synthetic characters under a unified schema. (B) Attributive Identity: Fine-grained attributes (e.g., Morality) where RPA performance is highly sensitive to attribute valence.
• Figure 2: Overview of our hierarchical character profile schema. The schema organizes 38 leaf fields into 5 top-level dimensions to enable analyses of character identity.
• Figure 3: Turn-wise Trends in Role-Playing Performance of the Famous and Synthetic Groups
• Figure 4: Impact of negative character on CoSER scores. (a) Personality Traits exhibit relatively stable performance, while (b) Motivations show consistent decreasing trends as more negative characters.
• Figure 5: Attention lift scores by top-level profile field.