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

Abstract: Despite the rapid proliferation of Role-Playing Agents (RPAs) based on 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.

• The paper introduces a multidimensional model that disentangles character identity into parametric and attributive components for RPAs.
• It shows that famous characters outperform in single-turn tasks while synthetic personas catch up during extended multi-turn interactions.
• The study reveals that negative motivations and interpersonal relationships compromise role-play fidelity, highlighting targets for future system improvements.

Disentangling Character Identity in Role-Playing Agents: A Multidimensional Analysis

Introduction and Motivation

Despite significant advances in LLM-based Role-Playing Agents (RPAs), the structural composition of a character's identity has received insufficient formalization. Existing approaches largely treat any character persona as free-form text within the prompt, conflating pre-existing world knowledge and the specific behavioral attributes that define nuanced character simulation. "Fame Fades, Nature Remains: Disentangling the Character Identity of Role-Playing Agents" (2601.04716) addresses this deficit by introducing a multidimensional model of "Character Identity," decomposed into two orthogonal layers: Parametric Identity (character-specific knowledge learned during pre-training) and Attributive Identity (fine-grained behavioral, moral, and personality attributes). By constructing a unified schema for both famous and synthetic characters, the work provides an empirical foundation for analyzing how these distinct axes contribute to RPA performance in both constrained and dynamic interactive contexts.

Figure 1: The two-layered Character Identity schema delineates parametric and attributive components and highlights valence sensitivity in morality attributes.

Unified Character Profile Schema

The research constructs a comprehensive hierarchical character profile schema comprising 38 leaf fields organized into five top-level dimensions: Personal Attributes, Personality Traits, Interpersonal Relationships, Motivations, and Abilities. This framework enables direct, field-level attribution analysis, circumventing the ambiguities arising from ad hoc template construction in current RPA datasets. By generating parallel datasets for both famous and synthetic characters under this schema, the authors support precise comparisons isolating the influence of pre-trained knowledge and explicit attribute grounding.

Figure 3: Schematic of the hierarchical character profile—five primary dimensions instantiated in 38 specific fields.

Parametric Identity: The "Fame Fades" Phenomenon

Experimental Design

The investigation into Parametric Identity contrasts the role-playing fidelity of famous (i.e., characters found in the pre-training corpus, such as "Harry Potter") versus synthetic (hallucinated, never-before-seen) characters. The evaluation utilizes PersonaGym (single-turn, persona-conditioned interview) and CoSER (multi-turn, multi-agent interaction), measured across several advanced open-source LLMs, including Qwen3 and DeepSeek-v3.2, with rigorous LLM-as-Judge assessments.

Key Findings

• Single-turn Interview Superiority: Famous characters show clear performance advantages in constrained, single-turn interview tasks. The advantage is statistically robust in scores for persona consistency, expected actions, and action justification. This demonstrates that parametric knowledge supplies a strong prior in limited context settings.
• Diminishing Returns in Multi-Turn Settings: In prolonged, multi-turn CoSER scenarios, the performance gap between famous and synthetic characters rapidly vanishes or may even reverse, with synthetic personas exhibiting increasing robustness as interaction context accumulates. Aggregate scores show negligible or negative differences.
• Attention Dynamics: Mechanistic analyses reveal that as conversational turns increase, models shift attention away from static profile or narrative context and towards self-conditioned, parametric output, especially for famous characters. However, this self-conditioning appears to suppress, rather than bolster, role fidelity in extended simulations. Attention lift metrics corroborate that persona constraints are incorporated later and less proactively for famous entities.

  Figure 2: Turn-wise trends reveal performance gap convergence—synthetic characters catch up and occasionally surpass famous ones as conversation length grows.

Attributive Identity: Field-Level Bottlenecks and Sensitivity

Attribute Polarity and Performance

By classifying each character attribute as Positive or Negative in valence, the work exposes a critical asymmetry:

• Personality Traits: Classic Big Five traits, regardless of polarity, yield stable RPA performance, confirming that LLMs can simulate positive and negative personality with similar fidelity.
• Motivations and Interpersonal Relationships: Attributes encoding negative moral valence or adversarial relationship dynamics exert a strong detrimental effect on role-play performance. The performance gap, robust across models, is isolated to these fields rather than personality-centric attributes.

  Figure 4: Negative polarity in motivation attributes leads to consistent performance decrease, while personality traits are robust to polarity.

  

  Figure 5: Interpersonal Relationships remain relatively stable, but Abilities degrade as negative character count increases in group interaction.

Interaction Length and Attribution

• Persistence of Bottlenecks: The negative attribute effect in motivations and relationships persists throughout multi-turn simulations, unaffected by increasing dialog length. There is no "learning out" of these bottlenecks as context accumulates.
• Attention Allocation: Field-level attention lift scores reveal that Personality Traits and Personal Attributes command significantly more decoding attention than Interpersonal Relationships, Motivations, or Abilities. Consequently, fields with lower attention are more susceptible to bias toward agreeable generation, reflecting safety-aligned model objectives that sanitize undesirable behaviors and diminish negative attribute fidelity.

  Figure 7: Top-level dimensions such as Personality Traits receive the greatest attention lift during decoding, leading to enhanced robustness.

Practical and Theoretical Implications

Practical

• Schema-Driven Character Construction: For RPA deployment or benchmarking, attention should be paid to explicit specification and enforced grounding of negative-oriented motivations and relationships, as these represent the true site of model brittleness.
• Beyond Personality-Centric Evaluation: Prioritizing only personality fidelity (e.g., via MBTI or BFI) fails to identify the most impactful performance disparities. Evaluation frameworks must be extended to specifically stress-test moral and relational attributes along the negative axis.

Theoretical

• Character Knowledge as an Initializer: The findings support a view of pre-trained world knowledge as a catalyst for initial role simulation but not a persistently guiding factor in dynamic interactions.
• Bias-Driven Sanitization: LLM safety alignment reinforces prosocial output defaults. Without strong context anchoring, negative or adversarial traits are suppressed, limiting accurate simulation for a significant space of plausible characters.
• Attribution-Grounded Mechanistic Interpretability: Attention metrics offer an avenue for decomposing LLM simulation failures, suggesting new directions for model interpretability research within multi-agent, long-term interaction regimes.

  Figure 6: Role-playing performance trends for Personality Traits over varying turns stratified by positive/negative field counts.

  

  Figure 8: Negative Motivations group demonstrates performance degradation as the number of turns increases, confirming persistent fidelity bottlenecks.

Future Directions

• Fine-Grained Valence Typing: Moving beyond binary positive/negative classification, future research should develop multi-dimensional, gradated valence scales for richer character representation and error analysis.
• Interventions for Negative Attribute Fidelity: Targeted prompting, specialized finetuning, or architectural modifications may be necessary to override sycophantic defaults and maintain structured negative attributes in generated output.
• Generalization to Larger-Scale LLMs: Extending attention and saturation analyses to state-of-the-art, higher parameter-count models will be essential for validating scalability of these mechanisms.

Conclusion

The paper establishes that the durability of character simulation by LLM-driven RPAs is inherently a function of how attributive identity, especially negative motivations and relationships, is integrated and maintained. Pre-trained character knowledge merely initializes simulation, with its influence fading as self-conditioned context grows. The principal bottleneck centers on the accurate reproduction of negative social and moral nature, a dimension routinely underexplored in current RPA evaluation. This calls for systematic redesign of both data schema and evaluation frameworks to surface and address these fidelity failures in future RPA systems.