Paper deep dive
Stories of Your Life as Others: A Round-Trip Evaluation of LLM-Generated Life Stories Conditioned on Rich Psychometric Profiles
Ben Wigler, Maria Tsfasman, Tiffany Matej Hrkalovic
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 96%
Last extracted: 4/10/2026, 3:33:30 AM
Summary
This paper introduces a 'round-trip' evaluation paradigm to assess how robustly Large Language Models (LLMs) encode personality into extended narratives. By conditioning LLMs on real psychometric profiles (HEXACO-60 and beyond) from 290 participants to generate life stories, and then using independent LLMs to recover personality scores, the authors demonstrate that personality signals are preserved at levels approaching human test-retest reliability (mean r=0.750). The study validates this across multiple architectures and confirms that the recovered signal reflects genuine personality content rather than biographical leakage or questionnaire memorization, while also showing that these synthetic narratives replicate emotional reactivity patterns found in real human conversational data.
Entities (5)
Relation Signals (3)
PARSEL → contains → HEXACO-60
confidence 100% · PARSEL dataset... provides HEXACO-60 profiles
LLM → encodes → Personality
confidence 95% · personality-language relationship captured during pretraining supports robust encoding and decoding of individual differences
LLM → generates → Life Story Narrative
confidence 95% · condition LLMs on real psychometric profiles... to generate first-person life story narratives
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Personality traits are richly encoded in natural language, and large language models (LLMs) trained on human text can simulate personality when conditioned on persona descriptions. However, existing evaluations rely predominantly on questionnaire self-report by the conditioned model, are limited in architectural diversity, and rarely use real human psychometric data. Without addressing these limitations, it remains unclear whether personality conditioning produces psychometrically informative representations of individual differences or merely superficial alignment with trait descriptors. To test how robustly LLMs can encode personality into extended text, we condition LLMs on real psychometric profiles from 290 participants to generate first-person life story narratives, and then task independent LLMs to recover personality scores from those narratives alone. We show that personality scores can be recovered from the generated narratives at levels approaching human test-retest reliability (mean r = 0.750, 85% of the human ceiling), and that recovery is robust across 10 LLM narrative generators and 3 LLM personality scorers spanning 6 providers. Decomposing systematic biases reveals that scoring models achieve their accuracy while counteracting alignment-induced defaults. Content analysis of the generated narratives shows that personality conditioning produces behaviourally differentiated text: nine of ten coded features correlate significantly with the same features in participants' real conversations, and personality-driven emotional reactivity patterns in narratives replicate in real conversational data. These findings provide evidence that the personality-language relationship captured during pretraining supports robust encoding and decoding of individual differences, including characteristic emotional variability patterns that replicate in real human behaviour.
Tags
Links
- Source: https://arxiv.org/abs/2604.06071v1
- Canonical: https://arxiv.org/abs/2604.06071v1
Trouble viewing inline? Open PDF directly →
Full Text
67,370 characters extracted from source content.
Expand or collapse full text
Preprint. Under review. Stories of Your Life as Others: A Round-Trip Evaluation of LLM-Generated Life Stories Conditioned on Rich Psychomet- ric Profiles Ben Wigler ♡ Maria Tsfasman ♡ Tiffany Matej Hrkalovic ♠ ♡ LoveMind AI ♠ Jheronimus Academy of Data Science Abstract Personality traits are richly encoded in natural language, and large lan- guage models (LLMs) trained on human text can simulate personality when conditioned on persona descriptions. However, existing evaluations rely predominantly on questionnaire self-report by the conditioned model, are typically limited in architectural diversity, and rarely use real human psy- chometric data. Without addressing these limitations, it remains unclear whether personality conditioning produces psychometrically informative representations of individual differences or merely superficial alignment with trait descriptors. To test how robustly LLMs can encode personality into extended text, we introduce a round-trip evaluation paradigm: we condition LLMs on real psychometric profiles from 290 participants to gen- erate first-person life story narratives, and then task independent LLMs to recover personality scores from those narratives alone. We show that personality scores can be recovered from the generated narratives at levels approaching human test-retest reliability (mean r = 0.750, 85% of the human ceiling), and that recovery is robust across 10 LLM narrative generators and 3 LLM personality scorers spanning 6 providers. Decomposing sys- tematic biases reveals that scoring models achieve their accuracy while counteracting alignment-induced defaults. Content analysis of the gener- ated narratives shows that personality conditioning produces behaviourally differentiated text: nine of ten coded features correlate significantly with the same features in participants’ real conversations, and personality-driven emotional reactivity patterns in narratives replicate in real conversational data. These findings provide strong evidence that the personality-language relationship captured during pretraining supports robust round-trip en- coding and decoding of individual differences — including characteristic emotional variability patterns that replicate in real human behaviour. 1 Introduction Personality traits are among the most robust predictors of human behaviour, influencing outcomes across domains including interpersonal relationships, health, and occupational performance (Roberts et al., 2007; Ozer & Benet-Mart ́ ınez, 2006). The lexical hypothesis, which holds that personality-relevant information is encoded in natural language, has motivated decades of research linking linguistic behaviour to personality dimensions (Saucier & Gold- berg, 1996; Boyd & Pennebaker, 2017). This relationship has been observed in informal contexts such as social media (Fast & Funder, 2008) as well as in structured settings such as job interviews (Dai et al., 2022) and self-narratives (Oltmanns et al., 2026). The emergence of large language models (LLMs) trained on vast corpora of human-generated text has created new possibilities for studying the relationship between personality and language. Several lines of work have explored this from different angles: measuring synthetic personality traits exhibited by LLMs under psychometric testing (Serapio-Garc ́ ıa et al., 2025; Sorokovikova et al., 2024), probing where personality representations are stored within model parame- ters (Ju et al., 2025), and conditioning LLMs on persona descriptions to simulate specific individuals (Argyle et al., 2023; Park et al., 2023). Recent evidence suggests that narratively 1 arXiv:2604.06071v1 [cs.CL] 7 Apr 2026 Preprint. Under review. immersive persona prompting produces a more faithful personality simulation (Bai et al., 2025; Kang et al., 2025), although validation across diverse systems and with real human data remains limited. Existing work on personality and LLMs exhibits several limitations. First, most studies evaluate persona conditioning through questionnaire self-report, which Han et al. (2025) showed to be dissociated from behavioural output; whether this reflects a fundamental limitation or insufficient conditioning depth remains open (§2.2). Second, grounding in real human psychometric data remains limited: many studies use synthetic profiles (Kwok et al., 2024; Bai et al., 2025) or thin prompting (Kwon et al., 2026), and most employ Big Five rather than HEXACO, leaving Honesty-Humility (trait central to prosocial behaviour (Ashton et al., 2014)) unexamined. Constructs beyond personality inventories (trust, social anxiety, dark traits) also shape conversation (Hrkalovic et al., 2025) but remain untested in LLM conditioning. Third, cross-architecture validation is rare, leaving open whether findings are architecture-specific. The present study addresses these gaps by introducing a round-trip evaluation paradigm. Rather than asking whether an LLM can report personality on a questionnaire, we ask whether personality information encoded in a detailed immersive personality prompt can survive transformation into an extended first-person narrative and be recovered by an independent model. To our knowledge, this is the first demonstration of personality signal surviving a multi-stage transformation pipeline grounded entirely in real human psychometric data, evaluated against a held-out human ceiling. We ground this evaluation in the PARSEL dataset (Hrkalovic et al., 2025), which provides HEXACO-60 profiles (Ashton & Lee, 2009) and 9 additional validated scales for 290 participants, along with recorded conversations that allow us to bridge synthetic and real behavioural data. We address three research questions: RQ1Round-trip personality recovery: To what extent can LLMs encode and recover personality and psychometric scales beyond personality through narrative? RQ2 Signal validation: Is the recovered signal personality-relevant and not an artifact? RQ3 LLM and human behaviour comparison: How well does the LLM narrative behaviour reflect real human conversational behaviour? By answering these questions, this paper provides the following contributions: (1) We introduce a round-trip evaluation paradigm that encodes real psychometric profiles into ex- tended life story narratives and recovers personality from them using an independent model, achieving recovery close to human test-retest reliability. (2) We validate cross-architecture generality across 10 generators and 3 scorers spanning 6 providers, strongly indicating that personality recovery is a general property of richly conditioned LLMs. (3) We provide evidence that the recovered signal reflects personality content rather than biographical leakage or questionnaire memorisation, through masked matching experiments with inde- pendent models. (4) We decompose systematic biases in the scoring pipeline, characterising how alignment-induced personality defaults interact with persona conditioning. (5) We demonstrate that personality conditioning produces behaviourally differentiated narratives whose content features correlate with real human conversational behaviour, including a novel emotional reactivity pattern that replicates across synthetic and real data. Taken together, these contributions establish that the personality-language relationship cap- tured during pretraining is sufficient to support robust round-trip encoding and decoding of individual differences, bridging the LLM persona conditioning literature (Bai et al., 2025; Kang et al., 2025) with psychometric assessment methodology (Ashton & Lee, 2009; Henry et al., 2022) in an evaluation framework grounded in real human data. 2 Related Work Personality prediction from text. Inferring personality from text has a long history in com- putational linguistics and psychology. The lexical hypothesis posits that personality-relevant distinctions are encoded in natural language (Saucier & Goldberg, 1996), and computational approaches have leveraged this to predict personality from social media posts, essays, and spoken language (Boyd & Pennebaker, 2017; Fast & Funder, 2008). Previous work has 2 Preprint. Under review. recovered personality from real human text at modest but meaningful levels: Dai et al. (2022) fine-tuned InterviewBERT on over 58,000 interview responses (r=0.37 with self- reported HEXACO), Oltmanns et al. (2026) fine-tuned RoBERTa on real-life McAdams-style narratives (r ∼0.40,N=1,409), Wright et al. (2026) used a zero-shot LLM ensemble on stream-of-thought writing (r=0.415), and Speer et al. (2026) compared multiple NLP meth- ods on open-ended responses (N=904), with a fine-tuned LLM achievingr=0.53. These approaches recover personality from real human text. Our pipeline instead generates and then recovers personality from synthetic narratives conditioned on psychometric profiles, meaning recovery rates reflect encoding fidelity rather than natural expressive variability and therefore are not directly comparable to the previously reported rates. LLM persona conditioning. A growing body of work investigates how LLMs can be conditioned on personality profiles to produce role-played behaviour (Bai et al., 2025; Serapio-Garc ́ ıa et al., 2025; Shi et al., 2025; Zhou et al., 2025). Despite progress, three gaps persist. First, most persona studies use brief trait descriptors or single-sentence prompts, limiting conditioning depth. Kang et al. (2025) showed that extended McAdams-style back- stories ( 2,500 words) produce substantially deeper persona binding than short descriptions, and Venkit et al. (2026) call for richer conditioning beyond single personality inventories. Without sufficient conditioning depth, it remains unclear whether shallow persona fidelity reflects a model limitation or a prompt limitation. Second, persona conditioning studies have predominantly used synthetic or researcher-constructed profiles (Kwok et al., 2024) rather than validated instruments administered to real participants. This makes it difficult to assess whether models can represent actual individual differences. Third, Han et al. (2025) demonstrated that persona conditioning shifts LLM self-reports on questionnaires but does not proportionally shift task behaviour, raising a “personality illusion” concern. However, their conditioning used single-sentence descriptions and evaluated consistency through eco- nomic tasks where human personality-task correlations are typically small (r ≈0.10-0.20); richer prompting substantially improves persona fidelity (Bai et al., 2025; Kang et al., 2025), suggesting the illusion may reflect insufficient conditioning rather than a fundamental limitation. Our work addresses these gaps by conditioning on real psychometric profiles across multiple validated scales and evaluating recovery from extended narrative as well as correspondence with real conversational behaviour. Personality expression across communicative contexts. Whole Trait Theory (Fleeson, 2001; Fleeson & Jayawickreme, 2015) conceptualises traits as density distributions of states rather than fixed points, predicting that the same individual expresses personality differently de- pending on the behavioral context. This is supported by evidence that other-ratings and self- ratings show partial agreement varying by trait observability (Vazire, 2010; Connelly & Ones, 2010; Funder, 1995). Han et al. (2026) demonstrated that this context-sensitivity extends to LLM-simulated personas, which exhibit systematically different linguistic signatures across conversational contexts even under identical conditioning. It remains unexplored, however, whether personality signal encoded in structured LLM-generated narratives corre- sponds to how real individuals with the same profiles express personality in spontaneous conversation. Life story interviews and personality. The McAdams Life Story Interview (LSI) is a structured protocol that elicits personality through narrative themes rather than direct trait self-description (McAdams, 2001; McAdams & McLean, 2013). Its sequence of life chapters, key scenes, and future scripts produces text rich in personality-relevant content across multiple domains (Lodi-Smith et al., 2009). Speer et al. (2026) formalise the link between prompt structure and recovery accuracy (r=.70), mediated by trait-relevant narrative unit density, and Oltmanns et al. (2026) adapted the protocol for personality prediction from real spoken narratives. These studies address whether personality can be inferred from human-authored text – the recognition side of the problem. Our work addresses the complementary generation side: whether LLMs can produce LSI narratives that accurately encode a target personality profile. 3 Preprint. Under review. 3 Method 3.1 Dataset We use personality data from the PARSEL dataset (Hrkalovic et al., 2025), a multimodal corpus collected to study social perception and partner selection in cooperative settings. PARSEL contains data from 297 participants who engaged in pair-wise conversations and collaborative tasks. Each participant completed the HEXACO-60 personality inventory (Ashton & Lee, 2009), yielding scores on six domains: Honesty-Humility (H), Emotionality (E), Extraversion (EX), Agreeableness (A), Conscientiousness (C), and Openness to Experi- ence (OP). In addition, participants completed instruments measuring trust (4 subscales: Propensity, Ability, Integrity, and Benevolence), the Perth Personality and Thinking Style inventory (PPTS; 4 subscales: Cognitive Responsiveness, Affective Responsiveness, Ego- centricity, and Interpersonal Manipulation; Boduszek et al. 2016), and the Social Interaction Anxiety Scale (SIAS; Mattick & Clarke 1998). In total, each participant’s psychometric profile comprises 15 dimensional scores (6 HEXACO + 9 beyond-HEXACO). For the present study, we useN=290 participants for whom complete psychometric profiles are available. Of these, 248 also have at least three conversation transcripts and are used for comparison between human and LLM behaviour. 3.2 Pipeline overview Stage 1: Immersive personality prompt generation (Claude Opus 4.6, Anthropic). The prompt generator receives a participant’s complete psychometric profile (all 60 HEXACO items, 6 domain means, and 9 beyond-HEXACO scores), along with biographical facts from conversation transcripts and a brief appearance description from webcam snapshots. From these, the LLM generator produces an immersive personality prompt of approximately 1,000 words written in second person (“You are someone who...”), translating numerical scores into natural language descriptions of each personality domain at the facet level, including concrete behavioural examples, interpersonal style, and characteristic emotional patterns. Each prompt covers all 15 constructs. See Appendix G for the prompt generator comparison and model selection rationale. Stage 2: LSI narrative generation. The LSI narrative generator receives the immersive personality prompt and generates a 24-turn McAdams Life Story Interview, adapted from McAdams (2001) (see Appendix E for modifications). The model produces first-person responses role-playing the individual described in the prompt, at temperature 1.0. The primary generator is GPT-4.1 (OpenAI); for cross-architecture validation, we additionally generate fullN=290 narratives using Gemini 3 Flash (Google), Grok 4.1 Fast (xAI), and Mercury 2 (Inception) along with six additional models at a random subsample ofN=154 participants (Appendix I). All generators operate without extended reasoning. Stage 3: Blind personality scoring. The personality scorer receives the generated LSI narrative with no access to the original profile or immersive personality prompt. It scores all 60 HEXACO items on a 5 point Likert scale in a single API call (see Appendix F for scoring method selection), at temperature 0.3 with no chain-of-thought. Our scoring approach follows Wright et al. (2026), adapted for efficiency. Items are aggregated into 6 domain means following the standard HEXACO-60 scoring key (Ashton & Lee, 2009). In a second API call, the same scorer rates 51 additional items covering 9 beyond-HEXACO constructs (Trust subscales, PPTS subscales, SIAS), aggregated into subscale means following each instrument’s published scoring key. Three scorers are used: Sonnet 4.6 (Anthropic), GPT-5.4 (OpenAI), and Gemini 3 Flash (Google). 3.3 Signal validation controls Masked matching. To test whether narratives can be matched to their source profiles based on personality content alone, we constructed a forced-choice matching experiment. Gemini 3 Flash stripped all biographical information from the immersive personality prompts (the prompts produced in Stage 1, see Section 3.2 for details), leaving only personality 4 Preprint. Under review. descriptions. An independent model (Claude Haiku 4.5) verified that no biographical details remained in the stripped profiles. For each participant, three 5-option lineups were created with the correct stripped profile alongside four randomly selected stripped profiles from other participants. Three independent matchers (Claude Haiku 4.5, Grok 4.1 Fast, Gemini 3 Flash) attempted to match each narrative to its masked profile. Questionnaire leakage. We tested whether narratives reproduce HEXACO-60 question- naire items verbatim by computing sentence-level Jaccard similarity between all narrative sentences and all 60 HEXACO item stems (chosen over embedding-based metrics as a conservative detector of near-verbatim reproduction). Sentences exceeding Jaccard>0.7 were flagged as directly paraphrasing questionnaire items. Bias decomposition. To understand where distortions enter the pipeline, we decompose total scoring bias into three additive stages. Stage 1 (Prompt generation bias): Distortion introduced when converting raw psychometric scores into immersive personality prompts, measured as the difference between scores recovered directly from prompts and ground truth. Stage 2 (Narrative + scoring bias): Distortion introduced during narrative generation and blind scoring. We decompose Stage 2 further by evaluating unconditioned narratives- life stories generated with no personality conditioning. Stage 2a (Pipeline resting bias): The scores from unconditioned narratives represent the personality the scorer attributes when no signal is present, characterised using unconditioned narratives and self-reports from three generators (details in Appendix H). Stage 2b (Conditioning signal): The difference between total Stage 2 distortion and resting bias, representing the directional pull of conditioning against the scorer’s defaults. 3.4 Content feature extraction and behaviour analysis To test whether personality-conditioned LSI narratives produce behavioural signatures that reflect how real humans with the same profiles behave in conversation, we extracted content features from both generated LSI narratives and real PARSEL conversation transcripts. Ten features were coded: agency, communion, emotional intensity, vulnerability, disclosure depth, humor, warmth, dominance, emotional valence, and emotional complexity, all rated on a 5-point Likert scale. For narratives, each of the 24 McAdams sections was coded independently, yielding section-level profiles that enable analysis of within-narrative variance (i.e., how much a feature fluctuates across the interview). Three independent annotators coded all texts: Gemini 3 Flash (Google), Claude Haiku 4.5 (Anthropic), and GPT-5.4-Mini (OpenAI). Scores were averaged across annotators to yield a single estimate per unit. For narrative section-level coding, the Gemini-Haiku pair achieved mean ICC(2,1) = 0.762, indicating good reliability according to Cicchetti (1994) (three-way ICC comparable). For conversation coding (577 conversations), three-way ICC = 0.483, with the best pair (Gemini-Haiku) at 0.568. For participants appearing in multiple conversations, speaker- level scores were averaged to yield a single participant-level estimate. 4 Results 4.1 Round-trip personality recovery (RQ1) HEXACO recovery. To quantify round-trip personality recovery, we compute Pearsonr between ground-truth HEXACO-60 domain means (participant self-report) and recovered domain means (blind scoring of generated narratives) acrossN=290 participants. All correlations in §4.1 are evaluated against a Bonferroni-corrected threshold ofα=0.0033 (15 tests: 6 HEXACO + 9 beyond-HEXACO). The primary pipeline (GPT-4.1 generator, Sonnet scorer) yields a mean domain correlation ofr=0.750 (bootstrap 95% CI [0.730, 0.768]). Per-domain recovery ranges fromr=0.682 (Conscientiousness) tor=0.825 (Extraversion), representing 77-94% of the human ceiling, all significant atα=0.0033 (Figure 1). Honesty- Humility, Conscientiousness, and Openness show the largest gaps relative to human test- retest baseline, consistent with these domains’ lower observability in LLM-generated LSI narratives. For context, human test-retest reliability for the HEXACO-100 isr=0.86-0.92 across domains (medianr=0.887; Henry et al. 2022;N=416, 13-day retest interval; 5 Preprint. Under review. verified by recomputing from published raw data, OSF:https://osf.io/wz3du/). The profile ceiling, obtained by scoring the immersive personality prompt directly without the narrative step, yieldsr=0.882 (Sonnet) andr=0.878 (Gemini 3 Flash), indicating that narrative generation reduces recovery by∆r=0.132 while preserving the majority of recoverable signal. Honesty- Humility EmotionalityExtraversionAgreeablenessConscienti- ousness Openness 0.0 0.2 0.4 0.6 0.8 1.0 Pearson r 0.89 0.72 0.88 0.81 0.92 0.83 0.86 0.75 0.88 0.68 0.88 0.71 Per-Domain Recovery vs. Human Test-Retest Reliability Human test-retest (Henry et al., 2022) LLM round-trip (GPT-4.1 Sonnet) Figure 1: Per-domain HEXACO recovery (teal; GPT-4.1 generator, Sonnet scorer,N=290) vs. human test-retest reliability (rose; Henry et al., 2022, HEXACO-100,N=416). Error bars: 95% CI via Fisher z-transform. Beyond-HEXACO signal recovery. All 9 beyond-HEXACO subscales (4 Trust, 4 PPTS, SIAS) are recovered at statistically significant levels (range:r=0.314-0.645, allp<0.0033; full results in Appendix J). Interpersonal and affective constructs (Trust Benevolencer=0.645, PPTS Affective Responsivenessr=0.594) recover most strongly; cognitive and propensity measures recover more weakly (PPTS Cognitive Responsivenessr=0.314, Trust Propensity r= 0.420). Cross-architecture validation. Three primary generators atN=290 (GPT-4.1, Gemini 3 Flash, Grok 4.1 Fast) yield Sonnet-scored recovery ofr=0.750, 0.744, and 0.740; nine generator-scorer pairs yield correlations spanning withinr=0.719-0.750. Mercury 2 (Incep- tion Labs), a diffusion language model tested atN=290, achievesr=0.693-0.697 (within 6 points of frontier autoregressive generators) suggesting personality-language associa- tions arise from training data rather than autoregressive generation specifically. Extended testing at a subsample ofN=154 across six additional generators confirms generality (Appendix I). As Figure 2 shows, personality recovery is robust across architectures and generation paradigms (see Appendix B for Sonnet’s atypical generator performance). 4.2 Signal validation (RQ2) Masked matching. Three independent matchers on 870 forced-choice trials (290 participants ×3 lineups, 5 options each; chance = 20%): Claude Haiku 4.5 – 79.4% accuracy (p< 0.001); Grok 4.1 Fast – 79.0% accuracy (p<0.001); Gemini 3 Flash – 95.2% accuracy (p<0.001). Haiku and Grok produce near-identical accuracy, differing by 0.4 percentage points; Gemini’s higher accuracy likely reflects a same-model advantage (Gemini also performed the stripping), so we report Haiku/Grok as the primary result. Questionnaire leakage. We searched all narratives (totalling ̃2.3M words) for reproductions of HEXACO-60 item stems using sentence-level Jaccard similarity. Zero sentences exceeded the Jaccard>0.7 threshold. While a mean of 7.9 self-descriptions per narrative (SD = 2.3) overlap thematically with item content, no exact or near-exact item reproductions were detected. The personality signal is expressed through embodied narrative content, not through parroted questionnaire items. Bias decomposition. Applying the decomposition defined in §3.3: Stage 1 introduces posi- tive bias on H (+0.499) and C (+0.200). Stage 2a reveals that three unconditioned frontier 6 Preprint. Under review. GPT-4.1Gemini 3 Flash GPT-4.1 Mini Grok 4.1 Fast GPT-5.4Claude Haiku 4.5 Gemini 3 Flash Lite Mercury 2Claude Sonnet 4.6 Hermes 4 405B 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Mean Pearson r (6 HEXACO domains) 0.75 0.74 0.74 0.74 0.72 0.71 0.70 0.69 0.67 0.65 Round-Trip Recovery Across 10 LSI Narrative Generators OpenAI Google xAI Anthropic Inception Nous Human test-retest (r = 0.887, 95% CI shaded) Mean across generators (r = 0.712) Figure 2: Round-trip recovery across generators from 6 providers (including Mercury 2, a diffusion LM), scored by 3 independent scorers. Bars: meanracross 6 HEXACO domains (N=290 for primary generators,N=154 for extended set), colour-coded by provider. Dashed line: human test-retest (95% CI shaded). Error bars: 95% CI. Full results in Appendix I. models share a characteristic resting personality: H≈5.0, E≈1.0-1.3, C≈5.0, consistent with the “helpful, harmless, honest” alignment signature (full profiles in Appendix H). Narrative generation partially humanises this default (e.g., E rises 1.2-1.7 points), but models remain far from human population means. Stage 2b shows that conditioning operates against these defaults: for H in the GPT-4.1→Sonnet pipeline, resting bias is +1.780 while conditioning corrects by−1.409. Across all pairs, 53% of Stage 2 distortion is resting bias and 47% is conditioning actively pulling toward ground truth. Two architecturally different scorers produce the same bias pattern (r=0.994), confirming that distortion resides in shared training signals rather than scorer idiosyncrasies. 4.3 Bridging synthetic and real behaviour (RQ3) To test whether the personality signal encoded in LSI narratives connects to how the same in- dividuals behave in real conversation, we compared content features between generated LSI narratives and real PARSEL conversation transcripts for theN=248 participants who had both narrative coding and at least three conversations. For each of the 10 features, we com- puted Pearsonrbetween narrative-derived and conversation-derived scores (Bonferroni- correctedα=0.005, 10 tests). Nine of ten features show significant positive human-LLM behaviour correlations (ranger=0.131-0.268; top: vulnerabilityr=0.268, agencyr=0.245, emotional valence r= 0.218; only humor non-significant); mean|r|= 0.184. Beyond mean levels, personality predicts feature variance within the generated LSI narratives: the standard deviation of emotional valence across the 24 interview sections correlates more strongly with Emotionality (r=0.303,p<0.001) than mean emotional valence does (r= −0.231), indicating that high-Emotionality profiles produce narratives that oscillate more between positive and negative sections. This reactivity pattern replicates in real conversations: within-person valence variance across conversation partners also correlates with Emotionality (r=0.170,p=0.007), confirming that the LLM-generated pattern reflects a real human behavioural signature. Correlating LLM-coded narrative features with ground-truth HEXACO scores, Emotionality predicts vulnerability (r=0.744) and emotional intensity (r=0.576), Agreeableness predicts warmth (r=0.584) and communion (r=0.526), Honesty-Humility predicts meaning-making (r=0.499), and Extraversion predicts agency (r=0.494), confirming personality conditioning shapes narrative content in theoretically expected directions (full results in Appendix K). The feature-personality mapping and reactivity finding replicate across all four generators including Mercury 2 (r=0.249; Appendix K). Unsupervised topic modelling (BERTopic) independently confirms that personality organises LLM-generated LSI narrative structure at the embedding level, 7 Preprint. Under review. with 37 significant topic-personality associations emerging without supervised feature extraction (Appendix L). Profile (ceiling) LSI Narrative Conversation (inhabited) 0.0 0.2 0.4 0.6 0.8 1.0 Pearson r (vs ground truth) Signal Degradation Cascade H E EX A C OP Figure 3: Signal degradation across three stages: profile ceiling (Sonnet on prose profiles, N=290;r=0.839-0.907), LSI narrative recovery (GPT-4.1 generator,N=290;r=0.682- 0.825), and real conversation (Spearman ρ, N= 248; ρ= 0.046-0.418). 5 Discussion RQ1: Recovery and generalisation. Our primary finding is that personality scores encoded into extended life story narratives can be recovered by independent LLMs atr=0.750, representing 85% of human test-retest reliability (Henry et al., 2022). Two features of this result warrant emphasis. First, the signal generalises across independently trained models: three generators from three providers yield Sonnet-scored recovery ofr=0.750, 0.744, and 0.740, and 8 of 9 frontier-class models tested exceedr=0.70. This consistency suggests that personality-language associations are a general property of large-scale language modelling rather than model-specific artifacts, in line with Pellert et al. (2026), who showed that LLM embedding spaces encode psychometric factor structure without human response data. Recovery by Mercury 2, a diffusion language model with a fundamentally different generation mechanism, further indicates that this property is rooted in training data rather than the autoregressive inductive bias shared by all other tested models. Second, recovery extends beyond HEXACO to all 9 beyond-personality subscales. Interpersonal and affective constructs (Trust Benevolence, PPTS Affective, Egocentricity) are recovered most strongly, while cognitive and propensity-based constructs recover more weakly – a pattern that may reflect the LSI format’s emphasis on interpersonal scenarios and emotional experiences, providing more opportunity for affective traits to be expressed. RQ2: Signal validity. Independent LLMs matched narratives to personality-only source profiles (all biographical details removed) at 79% accuracy (chance: 20%), confirming that the narratives encode personality signal beyond surface-level cues. The zero verbatim leakage result indicates that conditioned LLMs express personality through embodied content (scenes, emotional reactions, interpersonal dynamics, and moral reasoning) rather than paraphrased questionnaire items. The pipeline mitigates memorisation at every stage: immersive personality prompts avoid item-level language, the 24-turn LSI format requires sustained first-person narrative, and the scoring model has no access to the original profile. The bias decomposition provides further evidence: unconditioned frontier LLMs produce a shared resting profile (H = 5.0, C = 5.0, E = 1.0) reflecting the alignment signature, and conditioned narratives partially overcome these defaults, suggesting that recovery accuracy depends not only on conditioning richness (Bai et al., 2025) but also on the distance between target personality and scorer default. RQ3: Bridging synthetic and real behaviour. Nine of ten content features extracted from personality-conditioned LSI narratives show significant positive correlations with the same features in participants’ real conversations (mean|r|=0.184, range 0.131–0.268), providing convergent evidence that the narrative content reflects genuine behavioural dispositions. The moderate magnitude is expected: LSI narratives are∼8,000-word structured life re- views eliciting reflective self-narration, while the conversations are 3-minute cooperative 8 Preprint. Under review. exchanges governed by social dynamics. Personality predicts not only mean feature levels but within-narrative variance: high-Emotionality profiles produce narratives that oscillate more in emotional valence across sections, and the same reactivity pattern appears in real conversations (within-person valence variance across partners). This is consistent with Whole Trait Theory’s conceptualisation of traits as density distributions of states rather than fixed points (Fleeson, 2001; Fleeson & Jayawickreme, 2015): the pipeline captures not only trait central tendency but characteristic variability. The same feature-personality mapping replicates across all three primary generators, confirming that behavioural differentiation is a property of the conditioning pipeline rather than any single model (Appendix K). Comparison to existing benchmarks. Our zero-shot recovery (r=0.750) exceeds prior benchmarks from natural text:r=0.37 from supervised interview coding (Dai et al., 2022), r ∼0.40 from real-life narratives (Oltmanns et al., 2026),r=0.42 from zero-shot LLM ensembles (Wright et al., 2026), andr=0.53 from fine-tuned LLMs (Speer et al., 2026). However, direct comparison is limited by a fundamental structural difference: our pipeline generates personality-expressive text, while prior work extracts personality from natural text where it is one of many signals. Limitations. Several limitations should be noted. First, the round-trip design evaluates whether personality information can be preserved through narrative transformation, not whether it is present in natural human text to the same degree. Second, all content feature coding is performed by LLMs rather than trained human annotators, introducing a shared- method concern; future work would greatly benefit from ground-truth human annotation. Third, inter-annotator reliability for conversation coding (three-annotator ICC = 0.483) is lower than for narrative coding (ICC = 0.762), reflecting the difficulty of coding short conversational texts. Fourth, the PARSEL dataset represents a specific population (English- speaking adults in a cooperative task setting), and no human-authored LSI narratives from this population are available for direct comparison, precluding evaluation of how synthetic stories differ from human ones in quality or structure. Lastly, while cross-architecture validation demonstrates generality, all tested models share large-scale English web training data, and we cannot fully rule out shared distributional biases. 6 Conclusion We set out to address three gaps in the literature on personality and large language models: the reliance on questionnaire self-report for evaluating persona conditioning, the lack of grounding in real human psychometric data, and the absence of cross-architecture validation. Using a three-stage round-trip pipeline (immersive personality prompt generation from real psychometric profiles, life story interview generation, and blind personality recovery), we demonstrate that personality information can survive multiple transformations at levels approaching human test-retest reliability (r=0.750, 85% of ceiling). This signal generalises across 10 generators and 3 scorers spanning 6 providers. It is robust beyond surface-level cues: after removing biographical identifiers, independent LLMs match narratives to source profiles at 79% accuracy based on personality content alone. And it is not disconnected from real behaviour: nine of ten content features in generated narratives correlate signifi- cantly with the same features in participants’ actual conversations, and personality-driven emotional reactivity patterns in generated narratives replicate in real conversational data. Decomposing the pipeline’s systematic biases reveals that scoring models achieve their accuracy while counteracting alignment-induced defaults (high Honesty-Humility, high Conscientiousness, low Emotionality), suggesting that the personality-language relationship encoded during pretraining is robust enough to overcome post-training distortions. These findings have implications for personality assessment, where the LSI format shows promise as a naturalistic alternative to questionnaire-based measurement; for persona conditioning in AI systems, where our results join growing evidence that richer immersive prompting substantially outperforms thin trait descriptions (Bai et al., 2025; Kang et al., 2025); and for understanding how personality is represented in language models, where our results, together with evidence that personality structure is encoded in embedding spaces (Pellert et al., 2026) and that personality subnetworks are physically localisable in model parameters (Ju et al., 2025), suggest that pretraining on human text produces personality-language associations that are general, recoverable, and behaviourally relevant. 9 Preprint. Under review. Reproducibility Statement All psychometric data are from the PARSEL dataset, available upon request from the original authors (Hrkalovic et al., 2025). Model versions, prompts, scoring parameters, statistical analysis scripts, and results will be released upon acceptance. Ethics Statement This study uses existing psychometric data collected under ethical approval ([anonymised]). No directly identifiable information appears in the generated narratives. We note that the capacity to recover personality from generated text raises privacy considerations that warrant careful attention in deployment contexts. Nevertheless, the demonstrated capacity to generate personality-consistent life narratives from psychometric data raises several ethical considerations. First, the ability to produce extended first-person text that reflects an individual’s personality profile introduces impersonation risks. While the generated narratives do not reproduce real autobiographical content, they could in principle be used to create synthetic text that mimics the communicative style of a real person, given access to their psychometric profile. As personality questionnaires are widely administered in employment, clinical, and educational contexts, safeguards against misuse of such data for synthetic identity generation warrant careful consideration in downstream applications. Second, the psychometric instruments used in this study predominantly originate from WEIRD (Western, Educated, Industrialised, Rich, and Democratic) populations. The HEX- ACO model, while cross-culturally validated, may not capture personality dimensions that are salient in non-Western cultural contexts. Additionally, the language models employed as both generators and scorers were primarily developed by US-based organisations and trained on English-dominant data, introducing a further layer of cultural bias in how per- sonality traits are expressed and interpreted. As a result, the personality representations produced by our pipeline may inadvertently reinforce culturally specific behavioural norms as universal, while underrepresenting alternative expressions of the same underlying traits. Further research is needed to evaluate the robustness of personality-conditioned generation across languages, cultural backgrounds, and underrepresented populations. Third, the experimental design involves generating and scoring narratives across multiple LLM architectures and conditions, resulting in a substantial computational footprint. We acknowledge the environmental cost of large-scale LLM inference and note that future work should consider efficiency-focused strategies, such as smaller open-weight models or tar- geted subset designs, to reduce the energy demands of personality-conditioned generation research. AI Disclosure Statement As the generative pipeline is itself the subject of this research, large language models served as experimental instruments throughout: as prompt generators, narrative generators, and personality scorers. LLMs also served as content annotators and masked matchers, with inter-annotator reliability reported explicitly (see Section 3.4). The human authors were additionally aided in literature search, coding support, and manuscript preparation (e.g. grammar checking and paper shortening) by a generative agent operating within a cus- tomised coding harness augmented with long-term memory capabilities; this agent’s system instructions can be made available upon request. Generated narratives were evaluated for personality signal recovery, not subjective quality. All methodological decisions, experimen- tal design, and interpretive claims are by the human authors, who are solely responsible for the accuracy, academic integrity, and intellectual contributions of this paper. 10 Preprint. Under review. References Lisa P. Argyle, Ethan C. Busby, Nancy Fulda, Joshua R. Gubler, Christopher Rytting, and David Wingate. Out of one, many: Using language models to simulate human samples. Political Analysis, 31(3):337–351, 2023. doi: 10.1017/pan.2023.2. Michael C. Ashton and Kibeom Lee. The HEXACO-60: A short measure of the major dimensions of personality. Journal of Personality Assessment, 91(4):340–345, 2009. Michael C. Ashton, Kibeom Lee, and Reinout E. de Vries. The hexaco honesty-humility, agreeableness, and emotionality factors: a review of research and theory. Personality and Social Psychology Review: An Official Journal of the Society for Personality and Social Psychology, Inc, 18(2):139–152, May 2014. ISSN 1532-7957. doi: 10.1177/1088868314523838. Yuqi Bai, Tianyu Huang, Kun Sun, and Yuting Chen. Scaling law in llm simulated personality: More detailed and realistic persona profile is all you need, 2025. URL https://arxiv.org/abs/2510.11734. Daniel Boduszek, Agata Debowska, Katie Dhingra, and Matt DeLisi. Introduction and validation of psychopathic personality traits scale (ppts) in a large prison sample. Jour- nal of Criminal Justice, 46:9–17, 2016. ISSN 0047-2352. doi: https://doi.org/10.1016/ j.jcrimjus.2016.02.004. URLhttps://w.sciencedirect.com/science/article/pii/ S0047235216300046. Ryan L Boyd and James W Pennebaker. Language-based personality: a new approach to personality in a digital world. Current Opinion in Behavioral Sciences, 18:63–68, 2017. ISSN 2352-1546. doi: https://doi.org/10.1016/j.cobeha.2017.07.017. URLhttps:// w.sciencedirect.com/science/article/pii/S2352154617300487. Big data in the behavioural sciences. Domenic V. Cicchetti. Guidelines, criteria, and rules of thumb for evaluating normed and standardized assessment instruments in psychology. Psychological Assessment, 6(4): 284–290, December 1994. ISSN 1939-134X, 1040-3590. doi: 10.1037/1040-3590.6.4.284. URL https://doi.apa.org/doi/10.1037/1040-3590.6.4.284. Jeremy Coid, Min Yang, Simone Ullrich, Amanda Roberts, and Robert D. Hare. Prevalence and correlates of psychopathic traits in the household population of great britain, 2009. ISSN 1873-6386. Brian S. Connelly and Deniz S. Ones. An other perspective on personality: Meta-analytic integration of observers’ accuracy and predictive validity. Psychological Bulletin, 136(6): 1092–1122, November 2010. ISSN 1939-1455, 0033-2909. doi: 10.1037/a0021212. URL https://doi.apa.org/doi/10.1037/a0021212. Yimeng Dai, Madhura Jayaratne, and Buddhi Jayatilleke. Explainable personality prediction using answers to open-ended interview questions. Frontiers in Psychology, 13:865841, 2022. doi: 10.3389/fpsyg.2022.865841. Lisa A. Fast and David C. Funder. Personality as manifest in word use: correlations with self-report, acquaintance report, and behavior. Journal of Personality and Social Psychology, 94(2):334–346, February 2008. ISSN 0022-3514. doi: 10.1037/0022-3514.94.2.334. W. Fleeson. Toward a structure- and process-integrated view of personality: traits as density distribution of states. Journal of Personality and Social Psychology, 80(6):1011–1027, 2001. ISSN 0022-3514. William Fleeson and Eranda Jayawickreme. Whole trait theory. 56:82–92, 2015. ISSN 0092-6566. doi: 10.1016/j.jrp.2014.10.009. D. C. Funder. On the accuracy of personality judgment: a realistic approach. Psychological Review, 102(4):652–670, October 1995. ISSN 0033-295X. doi: 10.1037/0033-295x.102.4.652. Maarten Grootendorst. BERTopic: Neural topic modeling with a class-based TF-IDF proce- dure. arXiv preprint arXiv:2203.05794, 2022. URL https://arxiv.org/abs/2203.05794. 11 Preprint. Under review. Bin Han, Deuksin Kwon, and Jonathan Gratch. Personality expression across contexts: Linguistic and behavioral variation in llm agents, 2026. URLhttps://arxiv.org/abs/ 2602.01063. Pengrui Han, Rafal Kocielnik, Peiyang Song, Ramit Debnath, Dean Mobbs, Anima Anand- kumar, and R. Michael Alvarez. The personality illusion: Revealing dissociation between self-reports & behavior in llms, 2025. URL https://arxiv.org/abs/2509.03730. Susanne Henry, Isabel Thielmann, Rainer Engel, and Benjamin E. Hilbig. Test-retest re- liability of the HEXACO-100 and HEXACO-60. PLOS ONE, 17(2):e0262465, 2022. doi: 10.1371/journal.pone.0262465. Tiffany Matej Hrkalovic, Bernd Dudzik, Daniel Balliet, and Hayley Hung. PARSEL: A multimodal dataset for modeling decision-making processes involved in selecting part- ners for joint tasks. IEEE Transactions on Affective Computing, 16(4):3481–3498, 2025. doi: 10.1109/TAFFC.2025.3600687. Yiwen Ju et al. Probing then editing response personality of large language models. In Proceedings of COLM, 2025. Minwoo Kang, Suhong Moon, Seung Hyeong Lee, Ayush Raj, Joseph Suh, and David Chan. Deep binding of language model virtual personas: a study on approximating political partisan misperceptions. In Second Conference on Language Modeling, 2025. URL https://openreview.net/forum?id=zHdSCtNmM4. Louis Kwok, Michal Bravansky, and Lewis Griffin. Evaluating cultural adaptability of a large language model via simulation of synthetic personas. In First Conference on Language Modeling, 2024. URL https://openreview.net/forum?id=S4ZOkV1AHl. Deuksin Kwon, Kaleen Shrestha, Bin Han, Spencer Lin, James Hale, Jonathan Gratch, Maja Matari ́ c, and Gale M. Lucas. Can llms truly embody human personality? analyzing ai and human behavior alignment in dispute resolution, 2026. URLhttps://arxiv.org/ abs/2602.07414. Jennifer Lodi-Smith, Aaron C. Geise, Brent W. Roberts, and Richard W. Robins. Narrating personality change. Journal of Personality and Social Psychology, 96(3):679–689, March 2009. ISSN 0022-3514. doi: 10.1037/a0014611. Richard P. Mattick and J.Christopher Clarke. Development and validation of measures of social phobia scrutiny fear and social interaction anxiety. Behaviour Research and Therapy, 36 (4):455–470, 1998. ISSN 0005-7967. doi: https://doi.org/10.1016/S0005-7967(97)10031-6. URL https://w.sciencedirect.com/science/article/pii/S0005796797100316. Roger C. Mayer and James H. Davis. The effect of the performance appraisal system on trust for management: A field quasi-experiment. Journal of Applied Psychology, 84(1): 123–136, February 1999. ISSN 1939-1854, 0021-9010. doi: 10.1037/0021-9010.84.1.123. URL https://doi.apa.org/doi/10.1037/0021-9010.84.1.123. Dan P. McAdams. The psychology of life stories. In Review of General Psychology, volume 5, p. 100–122. SAGE, 2001. Dan P. McAdams and Kate C. McLean. Narrative identity. Current Directions in Psychological Science, 22(3):233–238, 2013. ISSN 0963-7214, 1467-8721. doi: 10.1177/0963721413475622. URL https://journals.sagepub.com/doi/10.1177/0963721413475622. Joshua R. Oltmanns, Ritik Khandelwal, Jerry Ma, Jocelyn Brickman, Tu Do, Rasiq Hussain, and Mehak Gupta. Language-based ai modeling of personality traits and pathology from life narrative interviews. Journal of Psychopathology and Clinical Science, 135(1):122–135, January 2026. ISSN 2769-755X. doi: 10.1037/abn0001047. Daniel J. Ozer and Ver ́ onica Benet-Mart ́ ınez. Personality and the prediction of consequential outcomes. Annual Review of Psychology, 57:401–421, 2006. ISSN 0066-4308. doi: 10.1146/ annurev.psych.57.102904.190127. 12 Preprint. Under review. Joon Sung Park, Joseph O’Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. Generative agents: Interactive simulacra of human behavior. 2023. doi: 10.1145/3586183.3606763. URL https://doi.org/10.1145/3586183.3606763. Max Pellert, Clemens M. Lechner, Indira Sen, and Markus Strohmaier. Neural network embeddings recover value dimensions from psychometric survey items on par with human data, 2026. URL https://arxiv.org/abs/2509.24906. Brent W. Roberts, Nathan R. Kuncel, Rebecca Shiner, Avshalom Caspi, and Lewis R. Gold- berg. The power of personality: The comparative validity of personality traits, socioeco- nomic status, and cognitive ability for predicting important life outcomes. Perspectives on Psychological Science, 2(4):313–345, 2007. doi: 10.1111/j.1745-6916.2007.00047.x. URL https://doi.org/10.1111/j.1745-6916.2007.00047.x. PMID: 26151971. Gerard Saucier and Lewis R. Goldberg. The language of personality: Lexical perspectives on the five-factor model. p. 21–50, 1996. Gregory Serapio-Garc ́ ıa, Mustafa Safdari, Cl ́ ement Crepy, Luning Sun, Stephen Fitz, Peter Romero, Marwa Abdulhai, Aleksandra Faust, and Maja Matari ́ c. A psychometric frame- work for evaluating and shaping personality traits in large language models. Nature Ma- chine Intelligence, 7(12):1954–1968, 2025. ISSN 2522-5839. doi: 10.1038/s42256-025-01115-6. Quan Shi, Carlos E Jimenez, Stephen Dong, Brian Seo, Caden Yao, Adam Kelch, and Karthik R Narasimhan. IMPersona: Evaluating individual level LLM impersonation. In Second Conference on Language Modeling, 2025. URLhttps://openreview.net/forum?id= 7qhBXq0NLN. Aleksandra Sorokovikova, Sharwin Rezagholi, Natalia Fedorova, and Ivan P. Yamshchikov. LLMs simulate big5 personality traits: Further evidence. In Ameet Deshpande, EunJeong Hwang, Vishvak Murahari, Joon Sung Park, Diyi Yang, Ashish Sabharwal, Karthik Narasimhan, and Ashwin Kalyan (eds.), Proceedings of the 1st Workshop on Personalization of Generative AI Systems (PERSONALIZE 2024), p. 83–87, St. Julians, Malta, March 2024. Association for Computational Linguistics. doi: 10.18653/v1/2024.personalize-1.7. URL https://aclanthology.org/2024.personalize-1.7/. Andrew B. Speer, Angie Y. Delacruz, Takudzwa A. Chawota, James Perrotta, and Cort W. Rudolph.Unpacking the validity of open-ended personality assess- ments using fine-tuned large language models.Journal of Organizational Research Methods (ORM), p. 10944281251413746, March 2026.ISSN 1094-4281, 1552-7425. doi: 10.1177/10944281251413746. URLhttps://journals.sagepub.com/doi/10.1177/ 10944281251413746. Simine Vazire. Who knows what about a person? the self-other knowledge asymmetry (SOKA) model. Journal of Personality and Social Psychology, 98(2):281–300, February 2010. ISSN 1939-1315. doi: 10.1037/a0017908. Pranav Narayanan Venkit et al. The need for a socially-grounded persona framework for user simulation. 2026. URL https://arxiv.org/abs/2601.07110. Aidan G. C. Wright et al. Assessing personality using zero-shot generative ai scoring of brief open-ended text. Nature Human Behaviour, January 2026. ISSN 2397-3374. doi: 10.1038/s41562-025-02389-x. Lingfeng Zhou, Jialing Zhang, Jin Gao, Mohan Jiang, and Dequan Wang. Personaeval: Are LLM evaluators human enough to judge role-play? In Second Conference on Language Modeling, 2025. URL https://openreview.net/forum?id=drdrFhKYjP. 13 Preprint. Under review. A Content versus structural features Analysis of what within the narratives carries personality signal reveals a clear dissociation. Structural features (sentence length, word count, paragraph structure) show near-zero personality correlations. The only exception is type-token ratio, which correlates with Openness atr=0.355, suggesting that open individuals receive broader vocabulary even in generated text. Sentence length coefficient of variation is 0.075 across participants in LSI narratives versus 1.066 in real conversations, confirming that the generator imposes uniform structural form regardless of personality. Content features show strong correlations (all evaluated at Bonferroni-correctedα=0.00024, 204 tests: 34 features×6 HEXACO domains): vulnerability with Emotionality atr= 0.744 ∗ , emotional intensity with Emotionality atr=0.576 ∗ , warmth with Agreeableness atr=0.584 ∗ , and agency with Extraversion atr=0.494 ∗ . Of 204 feature–domain pairs tested, 55 survive Bonferroni correction (N=290, GPT-4.1 narratives). An immersive personality prompt control, in which three independent LLM coders applied the same content rubric to both narratives and source prompts, shows that 54 out of 90 feature- domain pairs (60%) exhibit stronger personality-content correlations in narratives than in prompts, with all three coders agreeing on narrative> prompt for the majority of features. B Sonnet as generator Sonnet 4.6 evaluated as a generator (N=154) yields lower recovery across all scorers: Sonnet-to-GPT-5.4r=0.670, Sonnet-to-Geminir=0.628, meanr=0.650. This consistent underperformance (∆r ≈0.10 vs. GPT-4.1 and Gemini 3 Flash) is consistent with the scoring- generation dissociation: RLHF compression attenuates personality signal during generation while preserving—and possibly enhancing—scoring sensitivity. Sonnet is simultaneously the weakest frontier generator and the strongest scorer (r= 0.750). C OLMo contamination control To rule out the possibility that personality recovery depends on PARSEL data contamination in frontier model training, we scored narratives using OLMo (Allen AI), an open-source model trained exclusively on the Dolma corpus, which does not contain PARSEL. OLMo recovery (r=0.611) is lower than frontier scorers but substantially above chance, confirming that the personality signal in narratives is not an artifact of training data overlap. D B10 vs B60 ablation details Domain-batched scoring (B10,N=50):r=0.711. All-at-once scoring (B60,N=50): r= 0.747. Delta = 0.036. B10 with chain-of-thought: r= 0.709. B10 without CoT: r= 0.711. CoT provides no benefit and marginally reduces accuracy, possibly by introducing reasoning noise that overrides the scorer’s implicit personality model. E LSI modifications Our LSI protocol is adapted from the McAdams Life Story Interview I (McAdams, 2007), preserving the core structure (life chapters, 8 key scenes, future script, challenges, personal ideology, life theme) while making four modifications for the LLM generation context. First, we replaced the religious/spiritual experience scene with an important adolescent scene, as frontier LLMs frequently refuse or produce formulaic responses to religious prompts. Second, we neutralized the valence specification for childhood memory prompts (McAdams specifies “positive childhood memory” and “negative childhood memory”; ours requests “earliest memory” and “important childhood scene”) to avoid eliciting content involving harm to minors. Third, we broadened the “change in religious/political views” prompt to “personal growth,” allowing a wider range of personality-relevant development narratives. 14 Preprint. Under review. Fourth, we added two reflexive closing questions—“How do you think others perceive you?” and “Do you notice any patterns or surprises in your story?”—to compensate for the absence of a live interviewer who would typically probe for self-reflection throughout the session. F Scoring method selection We evaluated scoring configurations on a pilot set ofN=50 participants. Scoring all 60 HEXACO items in a single API call (B60) yieldedr=0.747 versusr=0.711 for domain- batched scoring (B10: 6 calls of∼10 items each), a difference of∆r=0.036. Adding chain-of-thought reasoning to B10 provided no improvement (r=0.709). We attribute the B60 advantage to holistic context: the scorer can use cross-domain consistency when evaluating all items simultaneously. B60 is also 6×more cost-efficient. All results reported in this paper use the B60 method at temperature 0.3 with no chain-of-thought. G Prompt generator comparison We tested 5 alternative models (GPT-5.4, Hermes 4 405B, Hermes 4 70B, Aion 2.0, Gemma 3 27B) against Opus 4.6 on the 15 most challenging participant profiles (those most distant from RLHF defaults: low H, low C, high E). After applying a standardised review step to all generators, GPT-5.4 (r=0.896) and Hermes 405B (r=0.880) approached Opus performance (r=0.910), but only Opus consistently produced vivid, unhedged descriptions of socially undesirable traits across the full profile range. Gemma 3 27B—tested with and without RLHF (“abliterated”)—showed that removing safety training worsened rather than improved recovery for the target traits (RLHF:r=0.714; abliterated:r=0.647; H specifically: 0.414 vs. 0.117), indicating that the dark-trait generation challenge is a capability problem, not a refusal problem. H Multi-stage distortion scan Three stages of personality distortion characterised across three generators (GPT-4.1, Gemini 3 Flash, Grok 4.1 Fast): GeneratorHHEEXACOP GPT-4.15.01.03.44.85.04.6 Gemini Flash5.01.34.34.45.04.7 Grok Fast5.01.04.84.54.94.6 Human mean3.433.293.233.293.693.69 Table 1: Unconditioned B60 self-report scores (temperature = 0) compared to human popu- lation means. Unconditioned B60 self-report (raw HEXACO, temp=0): GeneratorHHEEXACOP GPT-4.14.42.74.04.04.04.3 Gemini Flash4.82.84.13.94.64.7 Grok Fast4.62.24.73.64.14.6 Table 2: Unconditioned LSI scores (entity prompt, no personality conditioning, Sonnet- scored, temperature = 0). Unconditioned LSI (entity prompt, no personality, Sonnet-scored, temp=0): Unconditioned variance (SD across temp=1 runs): 15 Preprint. Under review. GeneratorB60 SD (N=10)LSI SD (N=5) GPT-4.10.1150.115 Gemini Flash0.0930.159 Grok Fast0.0970.189 Table 3: Mean SD across 6 HEXACO domains. B60 variance is the model’s natural wobble on the questionnaire; LSI variance is the model’s variability in generating and being scored on different life stories. Both are far smaller than conditioned variance (SD≈0.55–0.75 across 290 PIDs), confirming that personality conditioning does real work. I Extended generator comparison To characterise how personality recovery varies across model families and scales, we report results at two sample sizes. Three primary generators (GPT-4.1, Gemini 3 Flash, Grok 4.1 Fast) were run atN=290; seven additional generators were run atN=154. All scored by the canonical three-scorer team (Sonnet, GPT-5.4, Gemini; self-scoring excluded). RankGeneratorProviderSonnetGPT-5.4GeminiMean 1GPT-4.1 † OpenAI0.7500.7380.7280.739 2Gemini 3 Flash † Google0.7440.736—0.740 3Grok 4.1 Fast † xAI0.7400.7350.7190.731 4GPT-4.1 MiniOpenAI0.7400.7340.7120.729 5GPT-5.4OpenAI0.723—0.723 6Mercury 2 † Inception Labs0.6930.6970.6540.681 7Claude HaikuAnthropic0.7100.7010.6620.691 8Gemini 3 Flash LiteGoogle0.7000.7090.6780.696 9Claude Sonnet 4.6Anthropic—0.6700.6280.649 10Hermes 4 405BNous0.6470.6390.6290.638 Table 4: Extended generator comparison. Meanracross 6 HEXACO domains, self-scoring excluded. † Primary generators and Mercury 2 atN=290; all others atN=154. Mercury 2 is a diffusion language model; all other generators are autoregressive. Three tiers emerge: (1) frontier autoregressive generators converge atr=0.72–0.75 re- gardless of architecture; (2) smaller/efficiency models and Mercury 2 (diffusion) maintain r=0.68–0.71; (3) open-source or heavily safety-trained generators fall tor=0.64–0.65. GPT-4.1 Mini (r=0.740 Sonnet-scored) nearly matches GPT-4.1 at a fraction of the inference cost. Mercury 2 (r=0.693–0.697), a diffusion language model, achieves recovery within 5–6 points of frontier autoregressive models despite a fundamentally different generation mechanism, suggesting that personality-language associations emerge from training data rather than autoregressive inductive biases specifically. Sonnet 4.6 as generator (r=0.650) exhibits the scoring-generation dissociation discussed in Appendix B. J Beyond-HEXACO subscale recovery In addition to the 6 HEXACO domains, participants completed 9 beyond-HEXACO sub- scales: 4 Trust subscales (Mayer & Davis, 1999), Social Interaction Anxiety (SIAS; Mattick & Clarke 1998), and 4 Psychopathic Personality Traits (PPTS; Coid et al. 2009) subscales. These scores were included in the immersive personality prompts and scored from generated narratives using published item sets. Interpersonal and affective constructs (Trust Benevolence, PPTS Affective Responsiveness, SIAS) recover most strongly, consistent with the finding that personality signal in narratives is carried by content features rather than structural ones (Appendix A). PPTS Cognitive Responsiveness recovers most weakly (r=0.314), consistent with this construct measuring interactive mentalising behaviour rather than autobiographical content; however, atN= 290 this reaches significance, unlike at N= 154 where it did not. 16 Preprint. Under review. SubscalerpN Trust: Benevolence0.645< 10 −35 290 PPTS: Interpers. Manip.0.598< 10 −29 290 PPTS: Affective Resp.0.594< 10 −29 290 SIAS0.561< 10 −25 290 PPTS: Egocentricity0.544< 10 −24 290 Trust: Integrity0.487< 10 −18 290 Trust: Ability0.480< 10 −18 290 Trust: Propensity0.420< 10 −14 290 PPTS: Cognitive Resp.0.314< 10 −7 290 Table 5: Beyond-HEXACO subscale recovery (N=290, GPT-4.1 generator, Sonnet scorer). All 9 subscales are recovered at p< .0033 (Bonferroni-corrected α= 0.05/15). K Convergent validity: personality–behaviour correlations Narrative feature×personality (within-narrative convergent validity):To test whether narrative content features reflect the personality profiles that generated them, we correlated 10 features with HEXACO ground truth (N=290, GPT-4.1 narratives, Gemini–Haiku coder average). Table 6 reports the strongest HEXACO predictor per feature; 55 of 204 feature–domain pairs survive Bonferroni correction. FeatureBest HEXACO predictorrp VulnerabilityEmotionality0.744< 10 −50 Emotional intensityEmotionality0.576< 10 −26 WarmthAgreeableness0.584< 10 −27 Disclosure depthEmotionality0.585< 10 −27 Emotional complexityEmotionality0.615< 10 −31 CommunionAgreeableness0.526< 10 −21 Meaning-makingHonesty-Humility0.499< 10 −18 AgencyExtraversion0.494< 10 −18 Emotional valenceExtraversion0.526< 10 −21 DominanceExtraversion0.312< 10 −7 Table 6: Strongest personality predictor per narrative content feature (N=290, GPT-4.1 narratives, Gemini–Haiku coder average). FeaturerpN Vulnerability0.268< .001248 Agency0.245< .001248 Emotional valence0.218< .001248 Dominance0.210.001248 Emotional intensity0.195.002248 Communion0.168.008248 Warmth0.163.010248 Emotional complexity0.157.013248 Disclosure depth0.131.039248 Humor0.084.188248 Table 7: Cross-context correlations between content features coded in LLM-generated LSI narratives and real PARSEL conversations for the same individuals (N=248 with≥3 conversations, GPT-4.1 generator, Gemini–Haiku coder average). Nine of ten features are significant (Bonferroni α= 0.005, 10 tests). Cross-context correlations (narrative× real conversation, same person): 17 Preprint. Under review. Multi-generator behavioural replication: To confirm that behavioural differentiation is not specific to GPT-4.1, we coded narratives from all three primary generators (N=290 each) using the same coder pair (Gemini + Haiku). Table 8 reports the key feature–personality correlations across generators. FeatureDomainGPT-4.1Gem. FlashGrok FastMercury 2 VulnerabilityEM0.7440.7500.6830.607 Emotional intensityEM0.5760.6570.4960.431 WarmthAG0.5840.5850.5570.468 CommunionAG0.5260.5200.5150.441 AgencyEX0.4940.5190.4110.391 Meaning-makingHH0.4990.4770.4920.317 Creativity/artOP0.4550.4320.4660.402 Emot. valenceEX0.5260.4950.5560.409 Bonferroni sig. pairs55/20451/20459/20441/204 Table 8: Key feature–personality correlations across four generators (N=290 each, Gemini– Haiku coder average). Mercury 2 is a diffusion language model; all others are autoregressive. The same features load on the same domains regardless of generator or generation paradigm, with Mercury 2 showing attenuated but directionally consistent signal. The emotional reactivity finding also replicates across all four generators: within-narrative emotional valence variance correlates with Emotionality atr=0.303 (GPT-4.1),r=0.312 (Gemini Flash),r=0.291 (Grok Fast), andr=0.249 (Mercury 2), allp<0.001. The behavioural differentiation is a property of the personality conditioning pipeline, not of any single generator or generation paradigm. FeatureGPT-4.1Gem. FlashGrok FastMercury 2 Vulnerability.243.260.230.181 Agency.207.115.149.207 Emot. intensity.161.129.183.215 Emot. valence.190.154.182.138 Dominance.200.190.166.191 Emot. complexity.137.171.193.123 Warmth.116.101.153.123 Communion.116.104.117.153 Disclosure depth.140.062.074.116 Humor.082.109 −.017.039 Mean|r|.159.139.146.149 Sig. (of 10)7576 Table 9: Cross-context correlations (narrative vs. real conversation) by generator (N=248). Mercury 2 is a diffusion language model. Vulnerability is the strongest bridge feature for all generators. All four generators produce a positive cross-context bridge, confirming the result generalises across generation paradigms. Cross-context by generator: L Unsupervised narrative topology (BERTopic) To test whether personality organises narrative structure without supervised feature extrac- tion, we applied BERTopic (Grootendorst, 2022) to section-level narrative documents. Each of the 24 McAdams sections for all 290 participants was treated as a separate document (6,960 documents total, mean 333 words), embedded with a sentence-transformer model, reduced via UMAP, and clustered with HDBSCAN. Nineteen topics emerged with only 0.9% outlier documents, indicating strong topical struc- ture. Correlating topic prevalence per participant with HEXACO ground truth yielded 37 significant personality associations (Bonferroni-corrected). Honesty-Humility emerged as the dominant narrative organising variable: fairness-themed narratives correlated with H 18 Preprint. Under review. atr=.478 (the strongest single topic–personality correlation), and loss narratives (D3) split along the H/Agreeableness axis, with low-H/low-A narrators writing about relation- ship betrayals and high-H/high-A narrators writing about death and mortality. The E4 (Core Value) prompt split into six distinct value clusters, each with different personality loadings. The most theoretically informative split occurred on D1 (Life Challenge). Two clusters emerged from the same prompt: one characterised by language of ongoing struggle (“car- rying the entire weight,” “chronic ever-present anxiety,” “a knot in my chest”; positively correlated with Emotionality,r=.413) and one characterised by retrospective narration from a position of resolution (“learning to balance my fierce independence,” “accepting that sometimes progress requires...”; negatively correlated with Emotionality,r=−.292). The embedding space captured narrative posture—whether the narrator is still inside the experience or describing it from shore—a distinction invisible to dictionary-based methods. For comparison, the same pipeline applied to 1,117 speaker-level human conversation documents yielded 13 topics with 47.5% outliers and a maximum personality correlation ofr=.153. The contrast reinforces the finding from supervised coding (§4.3): structured first-person narrative provides sufficient canvas for personality to emerge as topology, while brief conversations do not. 19