Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note2 min read

The Dual Nature of LLM Persona: Aggregated Tendencies and Frame-Dependent Geometry

Source

arXiv

arxiv.org/abs/2607.02368v1

Read the source

Source

Plain-Language Brief

In this study, the authors analyze the structure of Large Language Model (LLM) personas by examining how they respond to psychometric questionnaires. They find that the geometry of these responses is not inherent but rather depends on the frame or context. The authors argue that this reveals a dual nature of LLM personas, consisting of both frame-dependent geometric features and more stable, aggregated features.

What the Document Says

  • The study evaluates LLM personas using psychometric questionnaires and finds that the within-instance correlation structure is typically discarded in aggregate scoring.
  • The authors construct correlation matrices from IPIP-50 responses and analyze geometry on SPD manifolds under manipulated question orderings for GPT-4 simulating American and Chinese-American personas.
  • They find that persona expression comprises two components: aggregated features (Big Five scores) that degrade under randomization but are frame-robust, and geometric features (SPD manifold) that collapse under frame misalignment but recover significantly under shared frames, performing better than aggregated features.
  • The authors conclude that the collapse-recovery pattern suggests that persona geometry is not inherent but a frame-dependent coordination pattern encoding information invisible to aggregation.

Key Concepts Mentioned

  • LLM personas
  • psychometric questionnaires
  • IPIP-50 responses
  • SPD manifolds
  • GPT-4
  • Big Five scores

Claims Made by the Document

  • Aggregated scores discard within-instance correlation structure in LLM persona evaluations.
  • Persona expression has two dissociable components: aggregated features and geometric features.
  • Aggregated features degrade under randomization but are frame-robust.
  • Geometric features collapse under frame misalignment but recover significantly under shared frames, performing better than aggregated features.

Evidence Used

  • Construction of correlation matrices from IPIP-50 responses and analysis on SPD manifolds.
  • Analysis of the behavior of aggregated features and geometric features under different conditions.

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

Evaluations of LLM personas via psychometric questionnaires typically rely on aggregate scores, discarding within-instance correlation structure. We test whether this geometric structure is intrinsic or frame-dependent. Constructing within-instance correlation matrices from IPIP-50 responses, we analyze geometry on SPD manifolds under manipulated question orderings in GPT-4o simulating American and Chinese-American personas. We find that persona expression comprises two dissociable components: aggregated features (Big Five scores) degrade under randomization (21% drop) but are frame-robust; geometric features (SPD manifold) collapse under frame misalignment (42% drop) but recover substantially (to 84%) under shared frames, surpassing aggregated features (76%). This collapse-recovery pattern reveals that persona geometry is not intrinsic but a frame-dependent coordination pattern encoding information invisible to aggregation. Our findings establish a dual-nature framework for LLM personas, frame-dependent geometry versus frame-robust aggregates, necessitating frame-aware evaluation and challenging static trait conceptions.