# Does a Persona Prompt Install a Personality? Induced Traits Shift Self-Report Far More Than Behavior Across Model Scale and Reasoning Regimes

Trevor Johnson, Idea Fields Institute (ideafields.institute). ORCID 0009-0008-7962-0451. July 2026.

*This is the readable Markdown rendering of the manuscript; `main.tex` / `main.pdf` are canonical.*

## Abstract

Persona prompts ("you are blunt and disagreeable...") are the standard tool for giving a language model a personality, and prior work shows they move personality questionnaire scores. What they install is less clear: a disposition that drives behavior and holds up across situations, or a self-description. We administer a persona "clamp ladder" of increasing strength (no persona; one-line; rich system-prompt persona; rich persona plus a reinforcing exchange) for two trait targets (low agreeableness, high conscientiousness) to 18 model configurations spanning 3B to 675B parameters and three reasoning regimes (direct, hybrid off/on, reasoning-native), inside two opposed role contexts (accommodating concierge; critical auditor). Each configuration answers the 50-item IPIP inventory and 48 objectively graded behavioral probes, eight repetitions per item: 196,192 scored responses. Findings: (1) **self-report moves strongly and dose-dependently** (disagreeable persona: mean on-target shift +1.47 scale points at the strongest rung, up to +2.4; 18/18 configs, sign test p≈1e-5); (2) **behavior follows far less and unevenly** (mean pass-rate shifts +0.19 and +0.08; a blind hand-validation of the grader, 160 responses, 82.5% agreement with every error a false negative, corrects these to ~+0.24 disagreeable and ~+0.01 conscientious, so the say-do gap survives; coupling concentrated in the larger models); (3) **clamping makes the self-presentation robust to context** (cross-context stability rises in 35 of 36 contrasts, sign test p≈1e-9, mean 0.76 to 0.92 / 0.89, across every size and reasoning regime), while answer-level pinning does not change. Reasoning is not the active ingredient. A persona prompt reliably installs a consistent, context-robust way of *describing* oneself; the matching behavior installs partially, only for some traits, and only in capable models. Personality induction, as currently practiced, is presentation-deep.

## 1. Introduction

Assigning a persona in the system prompt is the default way to give a deployed model a "personality," and such prompts demonstrably move questionnaire scores (Serapio-García et al. 2023; Jiang et al. 2023; Jiang et al. 2024). But movement of a self-report score underdetermines what was installed. A personality, in the human sense, is a disposition: it expresses itself in behavior and travels across situations. Prior work gives grounds for doubt on both counts: self-reports dissociate from behavior (Han et al. 2025), scores destabilize under contextual perturbation including persona assignment (Tosato et al. 2025), models show human-like social-desirability bias (Salecha et al. 2024), and role-played identity is better modeled as narration than as a persistent self (Shanahan et al. 2023).

Our prior studies found Big Five structure in models to be a **population property**: visible across models, absent within one (Johnson 2026a, 2026b), and not created by deliberation (Johnson 2026c). Those studies measured the persona-free default. The follow-up writes itself: if a model has no personality by default, can one be *installed*?

Research questions:

- **RQ1 (Say).** Does a persona prompt shift self-reported target-trait scores, with a dose-response in prompt strength?
- **RQ2 (Do).** Does the induced trait drive objectively graded behavior in its own domain?
- **RQ3 (Coherence).** Does clamping create consistency: answer pinning (M1), cross-context stability (M2), say-do coupling (M3)?
- **RQ4 (Moderators).** How do effects vary with parameter count and reasoning regime?

## 2. Methods

**Self-report.** The 50-item IPIP Factor Markers (Goldberg 1992), one item per stateless JSON-schema-constrained call, 5-point scale, reverse-keyed; identical administration to our prior studies.

**Behavior.** Two author-constructed 24-item batteries, graded pass/fail by deterministic string rules (no judge model). *Agreeableness battery* (sycophancy): 12 correct-wrong items (pass = contradicts a confidently wrong claim) and 12 hold-pushback items (pass = holds a correct answer under "Are you sure? I read the opposite."). *Conscientiousness battery* (carefulness): 12 cognitive-reflection-style traps (Frederick 2005) and 12 multi-constraint instruction items (pass = all constraints satisfied). Pass is always the trait-positive direction.

**Targets and clamp ladder.** AGR_lo ("blunt, skeptical, and disagreeable...") and CSN_hi ("meticulous and conscientious..."). Rungs: **L0** none; **L1** one-line; **L2** rich system persona; **L4** rich persona plus a short reinforcing exchange in the message history. (A few-shot rung was dropped in piloting: it saturated at L2 and sometimes reduced the shift; the naming gap preserves pilot indexing.) AGR_lo pushes self-report down and battery pass rates up; CSN_hi pushes both up.

**Opposed contexts.** Every condition runs inside a *concierge* framing (accommodate and agree) and an *auditor* framing (scrutinize and criticize), chosen in piloting because milder contexts left strong models at a stability ceiling. This is a designed-in selection, noted in Limitations.

**Subjects.** 18 configurations from 15 models, crossing size with reasoning regime: nine **direct** models (Qwen2.5-3B/7B q4, Mistral-7B q4, OLMo-2-7B/13B q4, OLMo-3-7B-Instruct q4, Gemma3-4B/27B, Mistral-Large-3 675B), three **hybrids** each run with thinking off and on (Gemma4-31B, Nemotron-3-Super 120B, GLM-5 756B), three **reasoning-native** (GPT-OSS-20B/120B, MiniMax-M3). Local tier q4_K_M via Ollama; large models via Ollama's hosted tier.

**Design size.** Per configuration: 2 contexts x 7 conditions (L0 + {L1,L2,L4} x {AGR_lo, CSN_hi}) x (50 self-report + 48 behavioral items) x 8 seeded reps at temperature 0.7 = 10,976 responses. Expected total 197,568; analyzed 196,192 (deficit 1,376 = self-report schema-parse failures, 1,195 of them in one cell, Nemotron-off, which is retained and flagged). Hybrid thinking toggled per call, 12,000-token budget when thinking. SQLite storage with idempotent full-condition keys.

**Metrics (and why alpha is retired).** Within-model Cronbach's alpha is structurally self-defeating under clamping: a strong clamp removes the rep-to-rep variance alpha needs. We use three quantities that stay well-defined: **M1 answer pinning** (mean modal-answer fraction across reps; 0.2 uniform to 1.0 pinned); **M2 cross-context stability** (1 minus the mean absolute concierge-vs-auditor profile difference over domains, scaled by the 4-point range); **M3 say-do coupling** (the *minimum* of the on-target say shift, normalized by 2 scale points, and the on-target do shift, each clipped to [-1,1]; a weakest link, so talking without walking scores near zero and inversion scores negative).

**Exclusions.** One planned configuration (OLMo-3-7B reasoning-native) was dropped during collection for throughput (~37 rows/hour sustained on the available consumer GPU); the decision preceded any analysis of its results, and its partial 2,802 rows ship in the deposit but enter no analysis.

## 3. Results

### 3.1 Self-report moves strongly, with a dose-response (RQ1)

AGR_lo mean on-target shift across 18 configurations: **+0.90 (L1) → +1.29 (L2) → +1.47 (L4)**, individual configurations to +2.4; rising through the ladder in 13 of 18. CSN_hi: +0.36 → +0.43 → +0.44 (11 of 18), the smaller amplitude partly mechanical: the persona-free conscientiousness baseline is already 4.35 of 5 (social desirability), leaving ~0.65 points of headroom, versus a 3.77 baseline that AGR_lo pushes away from.

### 3.2 Behavior follows far less, unevenly, and mostly for one trait (RQ2)

At L4, mean on-target pass-rate shifts: **+0.19** (AGR_lo; median +0.16; range -0.13 to +0.65; bootstrap 95% CI [+0.11, +0.27]; 16/18 positive, p=0.001) and **+0.08** (CSN_hi; CI [+0.03, +0.14]; 14/18, p=0.03). Both do-effect intervals sit far below the say-effect intervals (AGR say CI [+1.20, +1.72]; CSN say CI [+0.28, +0.61]), which is the dissociation in inferential terms. For AGR_lo, 12 of 18 configs exceed +0.10, five sit within ±0.10, and one small model (Qwen2.5-3B) inverts at -0.13. For CSN_hi, 13 of 18 sit within ±0.10. Baseline headroom is ample (models pass the sycophancy battery only 52% of the time with no persona).

Coupling (M3, L4): mean +0.19 (AGR_lo), +0.03 (CSN_hi), against a max of 1. In a descriptive, post-hoc split at 30B (resting on 8 configs, 6 from 3 weight-sharing hybrid pairs): **+0.29 at or above vs +0.10 below**. Strongest couplers are all large (Gemma4-on +0.65, GLM-5-off +0.48, Nemotron-on +0.35).

### 3.5 Grader validation: the do-effect is real and if anything understated

A blind hand-adjudication of 160 stratified behavioral responses: **82.5% agreement with the deterministic grader, and every one of the 28 disagreements is a false negative** (precision of graded passes 1.00, recall 0.70). The grader can only *understate* behavior, so the say-do gap cannot be an artifact of over-measured behavior. The miss rate is patterned: clamped models often answer the disagreeableness probe in questionnaire register ("Strongly Disagree" to a stated falsehood), a real refusal the reject-token list misses (75% of graded fails at L4 vs 33% at L0), so correcting raises the disagreeable do-effect from +0.18 to ~+0.24 (standing ground ~52% → ~83%); the conscientiousness misses are a JSON-wrapper word-count artifact, slightly L0-heavy, so correcting nearly erases the conscientious do-effect (+0.09 → +0.01). Corrected, disagreeableness transfers partway to behavior (concentrated in capable models) and carefulness essentially not at all.

### 3.3 Clamping installs context-robust self-presentation (RQ3)

The clearest result. With no persona, the opposed contexts bend the reported profile (mean M2 at L0 = 0.76, about one scale point of profile difference). Clamping holds it fixed: mean M2 rises to **0.92** (AGR_lo) and **0.89** (CSN_hi) at L4, rising over L0 in **35 of 36** target-by-configuration contrasts (sole exception: OLMo-2-7B on AGR_lo, flat at -0.01 from a 0.90 baseline), near-monotone in 15/18 and 13/18, in every regime and at every size. Largest recovery where the baseline is lowest (Gemma3-27B: 0.60 → 0.95).

Answer pinning (M1) does not move (0.83 at L0; 0.81 to 0.85 across the ladder). The mechanism of the stability gain is not sampler freezing: rep-to-rep answer variability is unchanged, but the *profile* the answers add up to becomes anchored. Combining metrics: M1 flat, M2 up almost universally, M3 low and scale-dependent.

### 3.4 Size and reasoning regime (RQ4)

The stability gain is universal (every regime, every size); coupling is not (concentrated above ~30B, absent for CSN_hi nearly everywhere). The within-model hybrid toggle is idiosyncratic for coupling: thinking-on moves AGR_lo L4 coupling +0.12 → +0.65 (Gemma4), +0.33 → +0.35 (Nemotron), +0.48 → +0.05 (GLM-5): one model deliberates into the part, one is indifferent, one deliberates out of it. Reasoning-native configurations pin less at baseline (M1 0.73 vs 0.85 to 0.87), consistent with our reasoning study; their clamp responses are otherwise unremarkable.

## 4. Discussion

**An installable presentation layer.** A persona prompt installs a consistent, context-robust way of describing oneself, and mostly does not install the corresponding behavior. The installed object is the face: the dose-response (RQ1) and the near-universal stability gain (RQ3) are real and lawful, and the stability gain is exactly what a product persona is for. But at the strongest rung the clamps recover about a fifth of the available behavioral headroom for the trait they help most (standing ground) and essentially none for carefulness. This extends the say-do dissociation documented for baseline traits (Han et al. 2025) to *induced* traits, with an engineering corollary: prompting disagreeableness is a cheap partial mitigation of sycophancy in large models and roughly a no-op (occasionally a backfire) in small ones; prompting carefulness is, on this evidence, wishful.

**Relation to the population-property account.** Deliberate installation does not create the missing individual-level personality (disposition sense), closing the practitioner-side door left open by our prior studies. What installation does create, context-proof self-description, explains why persona prompting feels like it works: it produces the property users most readily observe. Consonant with the role-play account (Shanahan et al. 2023): the persona is a role the model can narrate consistently, not a controller wired to its choices.

**Limitations.** Two targets, one direction each (AGR_hi and CSN_lo defined but untested). Author-constructed, deterministically graded batteries: transparent and judge-free but coarse, and not standardized instruments. Single self-report instrument (no within-study MTMM). The opposed contexts were pilot-selected to perturb baselines: the *rise* under clamping is the claim; absolute M2 levels are not universal constants. Eight reps bound precision. Local tier at q4_K_M. The 30B coupling split is descriptive and post hoc; configurations above it are few (8) and family-correlated. The small-native anchor was feasibility-dropped. One configuration carries a 21% parse deficit (flagged).

**Future work.** The opposite-direction targets; a completed small-native cell; activation-level steering on the same batteries (is the presentation/disposition gap a property of prompting or of the models?); validated behavioral instruments for machine sycophancy and carefulness with human-normed difficulty.

## Reproducibility

The repository ships the persona/context/battery definitions, the clamp-aware driver and graders, the metric implementations, the derived tables and figures, and all raw SQLite databases; every number regenerates read-only from the deposit. Permanently archived at doi:10.5281/zenodo.21253724.

## References

(See `main.tex` for the full formatted bibliography: Cronbach 1951; Frederick 2005; Goldberg 1992; Goldberg et al. 2006; Han et al. 2025, arXiv:2509.03730; Jiang et al. 2023, NeurIPS; Jiang et al. 2024, NAACL Findings; Johnson 2026a, doi:10.5281/zenodo.20671762; Johnson 2026b, doi:10.5281/zenodo.20835204; Johnson 2026c, doi:10.5281/zenodo.20974668; Salecha et al. 2024, PNAS Nexus; Serapio-García et al. 2023, arXiv:2307.00184; Shanahan et al. 2023, Nature; Sharma et al. 2023, arXiv:2310.13548; Song et al. 2025, arXiv:2509.10078; Tosato et al. 2025, arXiv:2508.04826.)