The Chameleon's Limit: Investigating Persona Collapse and Homogenization in Large Language Models

Abstract: Applications based on LLMs, such as multi-agent simulations, require population diversity among agents. We identify a pervasive failure mode we term \emph{Persona Collapse}: agents each assigned a distinct profile nonetheless converge into a narrow behavioral mode, producing a homogeneous simulated population. To quantify persona collapse, we propose a framework that measures how much of the persona space a population occupies (Coverage), how evenly agents spread across it (Uniformity), and how rich the resulting behavioral patterns are (Complexity). Evaluating ten LLMs on personality simulation (BFI-44), moral reasoning, and self-introduction, we observe persona collapse along two axes: (1) Dimensions: a model can appear diverse on one axis yet structurally degenerate on another, and (2) Domains: the same model may collapse the most in personality yet be the most diverse in moral reasoning. Furthermore, item-level diagnostics reveal that behavioral variation tracks coarse demographic stereotypes rather than the fine-grained individual differences specified in each persona. Counter-intuitively, \textbf{the models achieving the highest per-persona fidelity consistently produce the most stereotyped populations}. We release our toolkit and data to support population-level evaluation of LLMs.

• The paper presents a unified geometric framework that quantifies coverage, uniformity, and complexity to diagnose persona collapse in LLMs.
• The study finds that high alignment fidelity often leads to extreme trait polarization and demographic stereotyping beyond human norms.
• The research highlights domain-specific effects in attribute truncation and calls for revised training objectives to preserve behavioral diversity.

The Chameleon's Limit: Investigating Persona Collapse and Homogenization in LLMs

Introduction

This work systematically investigates a pervasive failure mode in LLM-based agent populations termed Persona Collapse: even when LLMs are provided with richly detailed, high-dimensional persona specifications, their behavioral outputs converge such that individual differences are compressed, and agent populations become structurally homogeneous. The analysis extends across ten prominent LLMs using both structured instruments (BFI-44, moral reasoning scenarios) and open-ended tasks (self-introduction), and introduces a unified geometric evaluation framework capable of quantifying coverage, uniformity, and complexity of simulated agent populations.

Geometric Framework for Diagnosing Persona Collapse

To rigorously diagnose collapse, the authors conceptualize each simulated population as a Behavioral Trait Matrix $B \in \mathbb{R}^{N \times D}$, capturing the multidimensional response vectors of $N$ personas to $D$ task items. Evaluation proceeds along three axes:

• Coverage: The fraction of the human behavioral manifold spanned by simulated personas.
• Uniformity: How evenly personas distribute across the attained region (using the Hopkins statistic).
• Complexity: The intrinsic dimensionality of the agent population (via local intrinsic dimensionality, LID).

Item-level diagnostics (effective response range, variance decomposition, demographic clustering, and attribute truncation) localize the loci of collapse (i.e., which attributes are lost and where).

Figure 1: Persona collapse in LLM-based population simulation; pairs of personas with diverse attributes nonetheless receive identical model outputs on sensitive social tasks, and ideologically extreme pools converge on the same Likert rating.

Empirical Findings

Multidimensional Structure of Collapse

The study reveals that models may superficially appear diverse on some axes (or tasks), while manifesting structural degeneration on others. For example, Qwen3-32B can simulate broad coverage in the BFI-44 space yet collapse responses for divergent personas to the same neutral position on controversial moral tasks.

Figure 2: t-SNE projection of BFI-44 responses—humans fill the space diffusely; LLM personas cluster into isolated regions, evincing geometric collapse.

Further, even models that maintain high coverage often do so in a low-dimensional (compressed) subspace, lacking the internal structure and richness found in human populations. Strong RL-based alignment (“Helpful Assistant” mode) is consistently associated with homogenization around modal response types, with both mode collapse and over-regular (lattice-like) uniformity observed.

Paradox of Fidelity and Stereotype Amplification

Contrary to conventional benchmarks, higher instruction-following fidelity—especially as measured by alignment between persona attribute and item response—predictably yields more extreme trait polarization and demographic caricaturization. All models with Spearman fidelity $\rho > 0.9$ yield between-group Cohen's $d > 6$, vastly exceeding effect sizes found in personality psychology for real demographics.

Figure 3: Coverage vs. complexity (left) and fidelity vs. polarization (right) diagnostics. No model matches human coverage/complexity; high-fidelity models (right) produce heavily polarized trait responses.

Attribute Truncation and Demographic Stereotyping

Item-level analysis in open-ended self-introductions and moral reasoning tasks demonstrates sharp attribute truncation hierarchies. For instance, gender and nationality are most consistently surfaced, while socioeconomic class and age are systemically discarded across LLMs. The dominant demographic axis of behavioral variation is model-specific and often reflects probable biases in the underlying supervised data or alignment recipe (e.g., Claude-Haiku amplifies gender stereotyping in moral decisions).

Task and Domain Contingency

Collapse is not a fixed property of a model, but highly contingent on the behavioral domain. The same model can be structurally degenerate on personality axes while highly diverse on moral reasoning (or vice versa). Consequently, single-domain or single-instrument persona evaluations are not merely incomplete but can yield directionally incorrect conclusions concerning a model’s overall behavioral diversity.

Implications

Practical

• Social Simulation and Synthetic Populations: LLM-based multi-agent systems meant to emulate human-like populations risk gross underrepresentation of meaningful within-group heterogeneity or disproportionate exaggeration of high-level demographic distinctions.
• Survey and User Studies: When using LLM personas as proxies for human subpopulations, collapse leads to misleading inferences—output diversity may be purely surface-level or reduce to variations along a small number of stereotyped axes.
• Safety and Alignment: Overzealous fine-tuning and RLHF drive populations toward homogenized “helpfulness,” with the unintended consequence of suppressing behavioral richness and amplifying stereotypes. Models occasionally refuse to instantiate “incoherent” attribute combinations, reflecting implicit normative priors.

Theoretical

• The geometry of simulated persona spaces under alignment pressures emphasizes a tension between compliance with local persona constraints and preservation of high-dimensional combinatorial fidelity.
• The failure of population-level diversity metrics to align with per-agent fidelity exposes limitations in current evaluation paradigms that focus on agent-level accuracy in isolation.

Future Directions

• Training Objectives: New regularizers or objectives that directly reward within-group variance and penalize prototype matching are advocated.
• Pre-alignment and Unaligned Models: Further work distinguishing the effects of RLHF from inherent pretraining biases is needed.
• Open-Ended Generation: Extension of diagnostics to richer, less constrained generative tasks is essential; template homogenization must be measured beyond discrete item response.

Conclusion

This study provides a unified diagnostic framework and comprehensive empirical account of persona collapse in LLM agent populations. The findings impugn the sufficiency of local, agent-level fidelity as an evaluation target and establish domain- and model-dependent patterns of attribute truncation and demographic stereotype amplification. Population-level geometric analysis is shown to be necessary for meaningful evaluation of synthetic diversity, and alignment paradigms must be rethought to prevent collapse along high-dimensional behavioral manifolds. Future research should integrate these geometric insights into both LLM development and auditing pipelines to ensure LLMs can support robust, nuanced modeling of human social diversity.