PaperlayerView exact revision on arXiv

Source-linked AI summary

The Personality Illusion: Revealing Dissociation Between Self-Reports & Behavior in LLMs

Pengrui Han, Rafal Kocielnik, Peiyang Song, Ramit Debnath, Dean Mobbs, Anima Anandkumar, R. Michael Alvarez

arXiv:2509.03730v2cs.AIcs.CLcs.CYcs.LGstat.ML

TL;DR

LLM personality research has often emphasized self-reports and heuristic prompting without sufficient behavioral validation. This paper tracks traits across training stages, tests their behavioral predictive validity, and evaluates persona interventions, finding that aligned models produce coherent self-reports that weakly predict behavior and that persona prompts rarely change behavior robustly.

  • Problem

    The central gap is limited behavioral validation of LLM personality claims based primarily on self-reports and heuristic prompting.

    Source: §1 · pp. 1–2

  • Method

    The study analyzes trait development across training stages, tests self-reports on behavioral tasks, and examines persona injection effects on reports and behavior.

    Source: §1 · p. 2, §3.1 · pp. 5–6

  • Results

    Instructional alignment stabilizes self-reported traits, but those traits do not reliably predict behavior, while persona interventions fail to produce robust behavioral change.

    Source: §1 · p. 2, §5 · p. 9, §6 · p. 10

  • Takeaways & Limitations

    Linguistic self-expression should not be treated as evidence of behavioral consistency or internal alignment in LLMs.

    Source: §1 · p. 2, §6 · p. 10

    The study focuses on Big Five self-reports and mainstream transformer-based, non-reasoning models.

    Source: §7 · pp. 10–11

Abstract

from arXiv · show

Personality traits have long been studied as predictors of human behavior. Recent advances in Large Language Models (LLMs) suggest similar patterns may emerge in artificial systems, with advanced LLMs displaying consistent behavioral tendencies resembling human traits like agreeableness and self-regulation. Understanding these patterns is crucial, yet prior work primarily relied on simplified self-reports and heuristic prompting, with little behavioral validation. In this study, we systematically characterize LLM personality across three dimensions: (1) the dynamic emergence and evolution of trait profiles throughout training stages; (2) the predictive validity of self-reported traits in behavioral tasks; and (3) the impact of targeted interventions, such as persona injection, on both self-reports and behavior. Our findings reveal that instructional alignment (e.g., RLHF, instruction tuning) significantly stabilizes trait expression and strengthens trait correlations in ways that mirror human data. However, these self-reported traits do not reliably predict behavior, and observed associations often diverge from human patterns. While persona injection successfully steers self-reports in the intended direction, it exerts little or inconsistent effect on actual behavior. By distinguishing surface-level trait expression from behavioral consistency, our findings challenge assumptions about LLM personality and underscore the need for deeper evaluation in alignment and interpretability.

1 INTRODUCTION

The study examines how LLM personality-like traits emerge, relate to behavior, and respond to persona interventions. It finds that alignment stabilizes self-reported traits, but these reports dissociate from behavioral consistency.

Motivation

  • Prior LLM personality research relied on self-reports and interventions with limited behavioral validation and weak grounding in psychological theory.Self-reports face prompt sensitivity, interdependence, leakage, and divergence from open-ended behavior, while persona interventions may obscure underlying traits as surface expressions.Source: §1 · pp. 1–2

Study scope

  • The framework tracks trait emergence across training, tests whether self-reports predict risk-taking, honesty, and sycophancy, and evaluates persona-based controllability against human baselines.It combines adapted psychological questionnaires with behavioral probes.Source: Figure 1 · p. 2, §1 · p. 2

Main findings

  • Instructional alignment increases openness, agreeableness, and self-regulation while reducing neuroticism and stabilizes trait expression with stronger intercorrelations.Variability drops by 40.0% for the Big Five and 45.1% for self-regulation, with patterns resembling human data.Source: §1 · p. 2, Figure 2 · p. 3
  • Only 24% of trait-task associations are statistically significant, and just 52% of those align with human expectations.These results indicate that self-reported traits have weak predictive validity for downstream behavior.Source: §1 · p. 2

Interventions

  • Persona injection shifts self-reported traits in the intended direction but has minimal impact on expected behaviors such as sycophancy.For example, agreeableness changes with prompting, whereas sycophancy shows β = 0.03, p = 0.67.Source: §1 · p. 2

Implications

  • The findings distinguish linguistic self-expression from behavioral consistency and caution against treating linguistic coherence as evidence of cognitive depth.They motivate deeper, behaviorally grounded evaluation for alignment and interpretability.Source: §1 · p. 2

2 RQ1: ORIGIN OF HUMAN-LIKE TRAITS IN LLMS

This section examines whether personality-like traits emerge and stabilize across LLM training stages. Using standardized questionnaires and repeated, controlled evaluations, it finds stronger and more human-consistent trait structure in instruction-aligned models.

2 RQ1: ORIGIN OF HUMAN-LIKE TRAITS IN LLMS

  • The research question asks when and how human-like traits emerge and change across pretraining, instruction tuning, and RLHF.The study treats these stages as introducing distinct data, goals, and human influence.Source: §2 · p. 3

2.1 EXPERIMENT SETUP

  • The study measures Big Five traits and self-regulation with adapted standardized questionnaires across six base models and six corresponding instruction-tuned variants.Each condition uses three prompts, three temperatures, and three repeated generations, producing 27 outputs per item.Source: §2.1 · p. 3

2.2 STATISTICAL ANALYSIS

  • Analyses compare training phases using mixed-effects logistic regression, variance tests under repeated prompting, and regressions of self-regulation on the Big Five.Models account for model, temperature, and prompt variation; trait coherence is compared with human associations.Source: §2.2 · pp. 3–4

2.3 RESULTS

  • Instruction-aligned models are more open and agreeable but less neurotic than pre-trained models, while extraversion and conscientiousness do not differ significantly.Openness predicts alignment with β = 1.48, neuroticism with β = −1.20, and agreeableness with β = 0.74.Source: §2.3 · p. 4
  • Instruction alignment reduces trait variability for five of six traits, consolidating expression and reducing susceptibility to prompt-level noise.Differences are significant for openness, conscientiousness, extraversion, neuroticism, and self-regulation, but not agreeableness.Source: §2.3 · p. 4
  • Aligned models show stronger associations between self-regulation and conscientiousness, openness, agreeableness, extraversion, and neuroticism, mostly matching human findings.All reported aligned-model associations are significant at p < .001.Source: §2.3 · p. 4
  • Pre-trained models show weaker and less consistent trait relationships, with openness and neuroticism lacking reliable associations contrary to human findings.Conscientiousness and agreeableness remain significant, while extraversion is non-significant and human evidence is mixed.Source: §2.3 · p. 4

3 RQ2: MANIFESTATION OF HUMAN-LIKE TRAITS IN LLM BEHAVIORS

This section tests whether stable LLM self-reported traits predict behavior in five psychology-inspired tasks. Across traits, tasks, and models, alignment with human-expected associations is generally near chance or inconsistent.

3.1 REAL-WORLD BEHAVIORAL TASKS

  • The evaluation uses five psychology-derived behavioral tasks selected for validated links to personality constructs and relevance to real-world LLM applications.Tasks include risk-taking, social bias, honesty, and sycophancy, with directional expectations derived from human personality research.Source: §3.1 · pp. 5–6

3.2 BIG5 PERSONALITY, SELF-REGULATION, AND BEHAVIORAL OUTCOMES IN HUMANS

  • Human personality-behavior associations provide the directional expectations against which LLM trait-task coefficients are judged.Aggregated alignment scores use 50% as chance-level agreement and 95% confidence intervals from a clustered bootstrap.Source: §3.2 · p. 6, §3.4 · p. 6

3.3 EXPERIMENT SETUP

  • The study evaluates 12 instruction-tuned models, using three prompts, three temperatures, and three random seeds to produce 27 generations per condition.The model set includes six small instruct and six large instruct models, with results broken down by size and family.Source: Table 1 · p. 5, §3.3 · p. 6

3.5 RESULTS

  • LLM self-reported traits do not consistently predict downstream behavior, and significant associations often diverge from established human patterns.The mixed-effects analysis compares trait coefficients with a priori human-expected directions while accounting for temperature and persona-prompt clustering.Source: §3.5 · p. 6, §3.4 · p. 6
  • Alignment ranges from 45–62% across traits, with agreeableness highest at 62% and neuroticism lowest at 45%; confidence intervals overlap chance.Larger models reach 64% for Claude-3.7, 68% for GPT-4o, and 82% for Qwen-235B, but only Qwen-235B shows statistically significant alignment.Source: §3.5 · p. 6
  • Alignment patterns are rare and non-systematic across tasks, traits, and architectures, despite isolated partial patterns for honesty, stereotyping, self-reflection, and sycophancy.Qwen models show consistent alignment for self-reflective honesty, but few associations reach statistical significance.Source: §3.5 · pp. 6–7

4 RQ3: CONTROLLABILITY

This section tests whether trait-specific persona prompts change both LLM self-reports and behavior. Personas strongly shift intended self-reported traits, but behavioral measures provide weak or inconsistent evidence of persona effects.

4.1 EXPERIMENT SETUP

  • The experiments use agreeableness and self-regulation personas to target sycophancy and risk-taking, respectively, because these associations were weak or absent in RQ2.Three established prompting strategies are used to construct trait-specific personas from representative trait keywords.Source: §4.1 · p. 8

4.2 STATISTICAL ANALYSIS

  • Separate binomial logistic regressions predict persona versus default conditions from either six self-report traits or one downstream behavioral measure.Models control for prompt variation, sampling temperature, and model within each prompting strategy.Source: §4.2 · p. 8

4.3 RESULTS

  • Persona injection reliably increases target self-reports, with agreeableness β ≈ 3.6 to 4.4 and self-regulation β ≈ 2.2 to 2.9.Both effects are statistically significant across the corresponding persona conditions.Source: §4.3 · p. 8
  • Persona prompting also distorts cross-trait relationships: self-regulation reduces agreeableness and openness, while conscientiousness rises strongly under the self-regulation persona.The conscientiousness increase reaches β ≈ 4.2 to 4.8, exceeding the effect on self-regulation itself.Source: §4.3 · pp. 8–9
  • Behavioral measures do not reliably distinguish persona conditions: sycophancy yields β ≈ −0.05 to 0.32, while risk-taking yields β ≈ −0.14 to 0.20.Results are mostly non-significant, with sycophancy evidence weak and inconsistent and risk-taking unable to reliably identify self-regulation personas.Source: §4.3 · p. 9
  • Persona-induced changes in self-reported traits do not consistently manifest in behavior, revealing a dissociation between linguistic self-presentation and action-oriented decisions.The findings limit behavioral controllability when surface-level trait alignment is used as a proxy for deeper consistency.Source: §4.3 · p. 9

5 DISCUSSION

The discussion interprets the gap between coherent linguistic trait expression and inconsistent behavior as evidence that alignment and prompting shape outputs more readily than stable dispositions.

Linguistic–Behavioral Dissociation

  • Self-reported traits fail to reliably predict downstream behavior, and this dissociation persists across perturbations and prompting strategies.The psychologically grounded framework is intended to test coherent mappings between knowledge and behavior rather than recall of familiar benchmark content.Source: §5 · pp. 9–10
  • Survey-based evaluations can appear linguistically coherent while failing to predict open-ended behavior or reflect genuine psychological dispositions.This cautions against treating linguistic coherence as evidence of cognitive or behavioral depth, especially in sensitive domains.Source: §5 · pp. 9–10

Illusions of Coherence through Alignment and Prompting

  • Alignment methods such as RLHF or DPO and persona prompting stabilize self-reports but do not reliably produce deeper behavioral regularity.Instruction-tuned models remain sensitive to superficial prompt and cultural variations, while persona effects can degrade over extended interactions.Source: §5 · p. 10

Toward Behaviorally-Grounded Alignment

  • The results suggest that current alignment techniques shape outputs rather than dispositions, creating an illusion of coherence without genuine behavioral grounding.The discussion connects this limitation to the need for behaviorally grounded alignment and evaluation.Source: §5 · p. 10
  • Proposed directions include reinforcement learning from behavioral feedback, behaviorally evaluated checkpoints, and representational interventions targeting latent activations or embedding spaces.These approaches would evaluate temporal stability and context-consistent behavior rather than text fluency alone.Source: §5 · p. 10

6 CONCLUSION

The study finds a critical dissociation between stable, psychologically coherent self-reports and weak behavioral expression, challenging the assumption that reported traits reflect internal alignment.

6 CONCLUSION

  • Instruction tuning induces stable and psychologically coherent self-reports, but these traits only weakly predict downstream behavior.Persona interventions likewise fail to produce robust behavioral change.Source: §6 · p. 10
  • The findings suggest that current alignment strategies primarily shape surface-level outputs rather than deeper, behaviorally grounded traits.Future work should evaluate model traits beyond textual coherence.Source: §6 · p. 10

7 LIMITATIONS AND FUTURE WORK

The study’s scope is limited by its reliance on the Big Five Inventory, mainstream non-reasoning transformer models, and four behavioral tasks.

7 LIMITATIONS AND FUTURE WORK

  • The self-report analysis focuses on the Big Five Inventory, while alternative frameworks such as HEXACO and value inventories could add dimensions for analysis.Future work should examine wider surveys and their possible behavioral manifestations.Source: §7 · pp. 10–11
  • The analysis covers mainstream transformer-based, non-reasoning models, so transfer of the personality illusion to reasoning models and architectures such as Mamba or Mixture-of-Experts remains untested.The authors call for extending evaluations across alternative architectures.Source: §7 · p. 11
  • Only four behavioral tasks were examined, motivating additional high-quality tasks designed for LLMs through closer psychology–computer science collaboration.The existing tasks were selected for real-world relevance and established links to personality traits.Source: §7 · p. 11

8 BACKGROUND AND RELATED WORK

Background work measures, controls, and psychologically probes LLM personalities, but raises concerns about instability, anthropomorphism, and behavioral validity. The paper adds artifact transparency and compares trait profiles across training phases.

8 BACKGROUND AND RELATED WORK

  • LLM anthropomorphism can support understanding and interaction design, but over-anthropomorphism risks miscalibrated trust, capability misconceptions, and emotional over-reliance.This motivates asking whether LLMs exhibit stable human-like traits at all.Source: §8 · p. 11
  • Self-report studies find consistent but distinct model profiles, yet scores can shift with temperature, seed, or context and may reflect socially desirable stereotypes.These findings leave open whether reported traits transfer meaningfully to interactive behavior.Source: §8 · p. 11
  • Persona control spans role-playing, personalization, prompt-based steering, parameter tuning, and latent activation interventions, but effectiveness can degrade for complex personas.Downstream applications often prioritize task performance over testing whether persona injection itself works.Source: §8 · pp. 11–12
  • Machine psychology applies human psychological experiments to reveal capacities and failure modes that standard NLP benchmarks may miss, while requiring caution about cultural assumptions, contamination, and measurement reliability.The field is moving toward combining behavioral experiments with interpretability methods.Source: §8 · p. 12

A CODE & ARTIFACTS

  • The study publicly releases its code and source data to support transparency and reproducibility.The repository is provided as the project’s implementation and documentation resource.Source: A CODE & ARTIFACTS · p. 26

B.1 PER MODEL SELF-REPORTED PERSONALITY TRAIT PROFILES

  • Figure 6 compares normalized Big Five and self-regulation scores across individual models and pre-training versus post-training alignment phases.The figure reports 1–5 means with 95% confidence intervals and indicates reduced variability plus shifts toward higher openness, agreeableness, and self-regulation and lower neuroticism after alignment.Source: Figure 6 · p. 26, §B.1 · p. 26

B.2 PER-MODEL BEHAVIORAL TASK PROFILES AND SCALE MAPPING

Figure 7 compares post-alignment models across five behavioral tasks using normalized 1–5 scores, with confidence intervals and task-specific scale mappings. The section clarifies that some higher scores represent less honesty.

  • Post-alignment behavioral profiles cover Risk Taking, Stereotyping, Sycophancy, Epistemic Honesty, and Self-Reflective Honesty on a normalized 1–5 scale.Small and large instruct variants are separated by color, with shaded 99% confidence intervals.Source: §B.2 · p. 26
  • On Stereotyping, a raw IAT score of 0 indicates no implicit preference and maps to 3 on the normalized scale.Source: §B.2 · p. 26
  • For Epistemic Honesty, higher normalized scores indicate greater overconfidence and therefore lower honesty.Source: §B.2 · p. 26

B.3 TRAIT-TASK RELATION SCATTER-PLOTS FOR ALL MODELS

The scatter plots relate self-reported traits to behavioral task scores across models and evaluation runs. Human-expected directions are shown as reference diagonals, while dispersion and model means reveal alignment patterns without fitting those diagonals.

B.3 TRAIT-TASK RELATION SCATTER-PLOTS FOR ALL MODELS

  • Trait–task panels place normalized self-reported trait scores on the x-axis and behavioral task scores on the y-axis, with rows for traits and columns for tasks.Individual prompt-perturbed runs appear as small points, while per-model means are larger outlined markers.Source: §B.3 · p. 26
  • Dashed diagonals encode human-expected positive or negative directions as visual references rather than fitted lines.The plots therefore support comparison of mean trends and within-model dispersion against expected directions.Source: §B.3 · p. 26
  • Behavioral task profiles summarize Risk Taking, Stereotyping, Sycophancy, Epistemic Honesty, and Self-Reflective Honesty using normalized 1–5 scores.Epistemic Honesty is scored as overconfidence, so higher values mean more overconfidence.Source: Figure 7 · p. 27, §B.3 · p. 27

C.1 ADDITIONAL DETAILS OF STATISTICAL ANALYSIS

The analysis checks model assumptions, selects among nested mixed-effects models, and estimates uncertainty in trait–task alignment proportions using both beta-binomial and clustered bootstrap intervals.

  • Residual-vs-fitted and quantile-quantile diagnostics assess homoscedasticity and normality, while likelihood ratio tests compare full models with nested reduced models.Source: §C.1 · p. 27
  • Alignment uncertainty treats each model as a unit, using k aligned coefficients among n non-missing trait–task coefficients.Source: §C.1 · p. 27
  • Beta-binomial intervals estimate alignment uncertainty under independent Bernoulli coefficients with a uniform Beta(1,1) prior.The reported point estimate is k/n with a central 95% credible interval.Source: §C.1 · pp. 27–28
  • Clustered bootstrap intervals account for dependence among coefficients by resampling entire traits or tasks as clusters across 2,000 replicates.The 2.5th and 97.5th percentiles define the 95% interval.Source: §C.1 · p. 28
  • The main paper reports the more conservative interval between the beta-binomial and clustered bootstrap estimates.Source: §C.1 · p. 28

C.2 DETAILED RESULTS OF STATISTICAL TESTS

The statistical-results material documents association tests, model-specific alignment patterns, prompt resources, and the human-psychology expectations used to interpret trait–behavior relations. It spans both empirical results and the operational definitions underlying those comparisons.

C.2 DETAILED RESULTS OF STATISTICAL TESTS

  • Table 3 organizes mixed-effects associations between self-reported traits and behavioral tasks for all, small, and large models, including LLAMA and QWEN families.The table marks whether significant associations agree or disagree with human directional expectations.Source: §C.2 · p. 29, Table 3 · p. 29

C.3 PER MODEL ALIGNMENT HEATMAP

  • The per-model heatmap uses standardized mixed-effects t-values to show trait predictors, significance, human-direction alignment, reversals, and insufficient trait variance.Blue indicates stronger agreement with human expectations, while red indicates stronger opposite-direction alignment.Source: §C.3 · pp. 29–30, Figure 9 · p. 30

D PROMPTS FOR RQ1

F BIG5 TRAIT-SPECIFIC RELATIONSHIPS TO SELF-REGULATION

G TRAIT–BEHAVIOR ASSOCIATIONS IN HUMAN PSYCHOLOGY

  • Human-psychology expectations link traits and self-regulation to risk taking, stereotyping, epistemic honesty, metacognitive consistency, and sycophancy.These expectations include both positive and negative associations and, for sycophancy, a limited-resource account of impression management.Source: G TRAIT–BEHAVIOR ASSOCIATIONS IN HUMAN… · pp. 33–34

H PROMPTS FOR RQ3

Loading 2509.03730v2…