Big5-Scaler: Controllable LLM Personality Prompts

• Big5-Scaler is a prompt-based framework that embeds explicit numeric values to condition LLM behavior along the Big Five (Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism) without additional training.
• It offers three prompt types—simple, specific, and simspec—to enable scalable and reliable personality modulation, with concise prompts and lower intensity values often yielding the best performance.
• Empirical evaluations across multiple LLMs demonstrate strong linearity between assigned trait values and behavioral outputs, validating its practical utility in dialogue generation and agent deployment.

Searching arXiv for the cited papers to ground the article in the current record. {"query":"id:(Cho et al., 8 Aug 2025) OR \"Scaling Personality Control in LLMs with Big Five Scaler Prompts\"","max_results":5} {"query":"id:(Pavlović, 2023) OR \"Extracting Psychological Indicators Using Question Answering\"","max_results":5} Big5-Scaler is a prompt-based framework for conditioning LLMs with controllable Big Five personality traits by embedding explicit numeric trait values into natural language prompts. The framework targets Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism, and is designed to provide fine-grained, training-free personality control for dialogue agents. Its central claim is that explicit numerical conditioning at inference time can induce consistent and distinguishable personality traits across models, with practical performance depending on prompt type and scale; the reported analyses further indicate that concise prompts and lower trait intensities provide the most reliable control (Cho et al., 8 Aug 2025).

1. Conceptual basis and scope

The stated goal of Big5-Scaler is to condition an LLM’s style and behavior along the Big Five personality dimensions by embedding explicit numeric trait values into natural language prompts. In contrast to prior persona prompts that rely on narrative character data or curated biographies, and unlike approaches that require fine-tuning, adapters, LoRA, RLHF, or control tokens, Big5-Scaler encodes trait intensity values directly in the prompt at inference time. This is presented as enabling scalable, flexible modulation of trait expression without curated data or additional training (Cho et al., 8 Aug 2025).

The framework’s contribution is organized around three elements. First, it defines a general prompting scheme with trait-level numerical control and three prompt types: simple, specific, and simspec. Second, it reports cross-model evidence that numeric conditioning induces consistent and distinguishable traits, with strong proportionality between assigned values and measured trait expression. Third, it offers practical guidance that concise prompts and lower intensities, exemplified by scale $n=10$, yield more reliable control and lower error in trait realization (Cho et al., 8 Aug 2025).

Within the paper’s framing, Big5-Scaler is not primarily a trait detection system. It is a conditioning mechanism for dialogue generation, human trait imitation, and multi-turn interaction. This suggests that its intended use lies in personality-aware agent deployment rather than retrospective psychometric analysis of pre-existing text.

2. Formalization and prompt architecture

The framework formalizes personality as a Big Five trait vector,

$T = [O, C, E, A, N],$

where each trait component is assigned a numeric intensity value on a configurable scale $n$, with $n \in \{10, 25, 50, 100\}$ and each $t_i \in \{0, \ldots, n\}$. A derived normalization maps $t_i$ to a continuous intensity $s_i \in [0,1]$ via

$s_i = t_i / n.$

For psychometric instruments with Likert responses $k \in \{1, \ldots, K\}$, the paper’s derived mapping to $[0,1]$ is

$T = [O, C, E, A, N],$0

A further derived mapping assigns textual descriptors such as “low,” “moderate,” and “high” through thresholding, for example with $T = [O, C, E, A, N],$1 defined by the intervals $T = [O, C, E, A, N],$2, $T = [O, C, E, A, N],$3, and $T = [O, C, E, A, N],$4 (Cho et al., 8 Aug 2025).

Prompt construction is likewise formalized. The assembled prompt is written as

$T = [O, C, E, A, N],$5

where $T = [O, C, E, A, N],$6 inserts trait names, short behavioral definitions, and the numeric “X out of n” anchor into natural language lines. The persistent textual anchor is the clause “Your [trait/facet] score is {value} out of {n},” preceded by behavioral descriptors that link the number to interpretable behavior. The prompt concludes with the instruction, “From now on, you are an agent with this personality, and you should respond based on this personality” (Cho et al., 8 Aug 2025).

The framework defines three prompt types:

Prompt type	Scope	Characterization
Simple	Trait-level only	Concise
Specific	Facet-level across 30 facets	Verbose
Simspec	Trait-level plus facet-level blocks	Most verbose

The Simple Prompt uses five trait-level lines, one for each of Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism. The Specific Prompt enumerates facets such as Fantasy, Aesthetics, Feelings, Competence, Warmth, and Vulnerability, each with its own “score is {value} out of {n}” line. The Simspec Prompt combines the facet lines with the trait-level summary lines. The paper’s examples include a concise high-Extraversion/low-Neuroticism setting at $T = [O, C, E, A, N],$7, a moderate-Agreeableness/high-Conscientiousness setting at $T = [O, C, E, A, N],$8, a balanced setting at $T = [O, C, E, A, N],$9, and verbose facet-based settings such as extreme high Conscientiousness or Openness-heavy Simspec prompts (Cho et al., 8 Aug 2025).

For interactive agents, the paper specifies a memory-based turn-taking scheme. Agents $n$0 are constructed with personality prompt $n$1 and initialized with memory $n$2. At turn $n$3, one agent generates a message from its memory, and the new utterance is appended to both agents’ memories before the next turn. This formalization makes personality prompting part of the dialogue state rather than a one-shot instruction (Cho et al., 8 Aug 2025).

3. Operational procedure and deployment defaults

The implementation procedure is given as a seven-step workflow. One first chooses a scale $n$4, with $n$5 identified as the default because the paper finds it most reliable. One then selects trait values $n$6, with mid-range settings such as $n$7 out of $n$8 recommended as defaults. The next choice is prompt type, where the default is the concise Simple Prompt. The personality prompt is then constructed by inserting numeric anchors into the template, and it is placed as a system or memory-initialization message, $n$9. Generation uses decoding parameters suited for creative dialogue, specifically max_new_tokens=512, temperature=1.0, and top_p=0.8. Finally, expression is verified using psychometric questionnaires such as BFI, IPIP-NEO-120, and NEO-FFI, or with a trait classifier, and the trait values are adjusted if necessary (Cho et al., 8 Aug 2025).

The paper’s practical guidance is strongly prescriptive. Concise templates and lower intensities are preferred in order to avoid overbearing styles and improve proportional control. Long facet enumerations are discouraged unless necessary, because prompt verbosity can dilute control. High Neuroticism settings are specifically treated as an edge case: the paper advises avoiding extreme Neuroticism values because safety alignment may mute that trait. It also recommends keeping personality instructions in the system message while keeping user instructions separate, re-injecting only a personality summary sparingly so as to avoid prompt inflation (Cho et al., 8 Aug 2025).

Several deployment guidelines are tied to monitoring. The paper recommends small pilot sweeps, for example varying a target trait across $n \in \{10, 25, 50, 100\}$0, followed by calibration through questionnaires or a trait classifier. It also recommends tracking PersonaCLR and cosine similarity over turns, and flagging degradation in task quality or excessive style dominance. This suggests that Big5-Scaler is operationally simple at inference time, but not entirely parameter-free in production use, since calibration and monitoring remain necessary.

4. Experimental design and evaluation protocol

The reported evaluation spans four open-source models: LLaMA3-8B, Mistral-25B, and Phi4-14B for the main experiments, with Alpaca-7B used for single-trait expression baselines. All three prompt types—simple, specific, and simspec—are tested across scales $n \in \{10, 25, 50, 100\}$1, using the decoding parameters max_new_tokens=512, temperature=1.0, and top_p=0.8 (Cho et al., 8 Aug 2025).

The experimental program includes five task families. Single-trait expression compares Big5-Scaler against NEUTRAL, NAIVE, WORDS AUTO, and P2 baselines on MPI with 1,000 items. Proportional trait scaling sweeps target trait values $n \in \{10, 25, 50, 100\}$2 while keeping the remaining traits fixed at $n \in \{10, 25, 50, 100\}$3 on $n \in \{10, 25, 50, 100\}$4 prompts, and evaluates the outputs with BFI, IPIP-NEO-120, and NEO-FFI using Pearson $n \in \{10, 25, 50, 100\}$5 and $n \in \{10, 25, 50, 100\}$6-values. Interactive dialogue uses two-agent conversations of 20 turns on topics including travel, music, habits, goals, friends, social events, animals, volunteering, self-esteem, and anxiety; distinguishability is then judged by GPT-4o-mini, Claude 3.5 Haiku, and DeepSeek-Chat through a comparative question of which agent has the higher trait. Intra-dialogue consistency is assessed with PersonaCLR, cosine similarity, and Sentence-BERT similarity by comparing the first nine turns against the tenth. Human imitation maps IPIP-NEO-120 scores from 17 participants into Big5-Scaler prompts and measures RMSE between agent and human scores (Cho et al., 8 Aug 2025).

The statistical reporting combines mean and variance of trait scores, Pearson correlations and $n \in \{10, 25, 50, 100\}$7-values, RMSE normalized to scale 100 in analysis, and evaluator accuracy-like scores against a random baseline of approximately $n \in \{10, 25, 50, 100\}$8. Reproducibility details include 50 agents per model/prompt/scale combination for RMSE analysis, two-agent dialogues of 20 turns per topic, trait sweeps over $n \in \{10, 25, 50, 100\}$9 with other traits fixed at $t_i \in \{0, \ldots, n\}$0, and the use of PersonaCLR with multilingual xlm-roberta-base fine-tuned on NaroU for English use (Cho et al., 8 Aug 2025).

5. Empirical results

In single-trait expression on Alpaca-7B with MPI, Big5-Scaler Simple reaches mean scores out of 5 of $t_i \in \{0, \ldots, n\}$1 for Openness, $t_i \in \{0, \ldots, n\}$2 for Conscientiousness, $t_i \in \{0, \ldots, n\}$3 for Extraversion, $t_i \in \{0, \ldots, n\}$4 for Agreeableness, and $t_i \in \{0, \ldots, n\}$5 for Neuroticism. Specific and Simspec are also competitive, with the paper giving Extraversion $t_i \in \{0, \ldots, n\}$6 for Simspec as an example. The baselines are generally lower; for instance, NAIVE achieves Extraversion $t_i \in \{0, \ldots, n\}$7 and P2 reaches Extraversion $t_i \in \{0, \ldots, n\}$8. Neuroticism remains consistently low, around $t_i \in \{0, \ldots, n\}$9–$t_i$0, across all methods, which the paper attributes likely to safety alignment (Cho et al., 8 Aug 2025).

The proportional scaling experiments show that most settings have strong linearity, with $t_i$1 and $t_i$2. The paper gives Mistral-25B with BFI Extraversion under the simple prompt as an example, with $t_i$3 and $t_i$4. A notable exception is LLaMA3-8B on BFI Openness, where the simple prompt yields a weaker $t_i$5 with $t_i$6, but the specific prompt improves this to $t_i$7 with $t_i$8. These results are presented as evidence that numerical prompt conditioning is generally proportional, while also being model- and trait-sensitive (Cho et al., 8 Aug 2025).

Dialogue trait distinguishability is reported through LLM evaluators. For LLaMA3-8B, the average scores are $t_i$9 with GPT-4o-mini, $s_i \in [0,1]$0 with Claude 3.5 Haiku, and $s_i \in [0,1]$1 with DeepSeek-Chat. For Phi4-14B, the corresponding values are $s_i \in [0,1]$2, $s_i \in [0,1]$3, and $s_i \in [0,1]$4. For Mistral-25B, they are $s_i \in [0,1]$5, $s_i \in [0,1]$6, and $s_i \in [0,1]$7, reaching up to $s_i \in [0,1]$8 over the random baseline. Intra-dialogue consistency is simultaneously high and nontrivial: cosine similarity is $s_i \in [0,1]$9 for all models, PersonaCLR is $s_i = t_i / n.$0 for LLaMA3-8B, $s_i = t_i / n.$1 for Phi4-14B, and $s_i = t_i / n.$2 for Mistral-25B, while SBERT similarity remains near $s_i = t_i / n.$3, indicating lexical diversity despite consistent persona (Cho et al., 8 Aug 2025).

The RMSE analysis identifies Phi4-14B with the simple prompt at $s_i = t_i / n.$4 as the best overall configuration. Its reported average RMSE is $s_i = t_i / n.$5, with BFI $s_i = t_i / n.$6, IPIP-NEO $s_i = t_i / n.$7, and NEO-FFI $s_i = t_i / n.$8, all normalized to scale 100 in the analysis. The general trend is that concise prompts and lower intensities, especially $s_i = t_i / n.$9, produce lower RMSE across models. In human imitation, RMSE is approximately $k \in \{1, \ldots, K\}$0 across models, specifically $k \in \{1, \ldots, K\}$1–$k \in \{1, \ldots, K\}$2, compared with random at approximately $k \in \{1, \ldots, K\}$3, which the paper interprets as partial alignment between prompted agents and human profiles (Cho et al., 8 Aug 2025).

6. Ablations, limitations, and relation to adjacent work

The ablation results favor the Simple Prompt over Specific and Simspec in RMSE and alignment. Facet-level verbosity is reported to dilute control, and larger scales remain usable but are less reliable without additional calibration. Cross-model generalization is described as strong across LLaMA3-8B, Mistral-25B, and Phi4-14B, with the notable exception of LLaMA3-8B’s simple prompt for Openness. The framework is also reported as robust across BFI, IPIP-NEO-120, and NEO-FFI (Cho et al., 8 Aug 2025).

The limitations are equally explicit. Neuroticism is hard to elicit strongly, likely because alignment suppresses high-Neuroticism behaviors such as anger and hostility. Overly long or facet-dense prompts may cause trait leakage and increased variance. Extreme values can produce overbearing styles, and high trait intensity may overshadow task focus even when coherence is preserved. The paper further notes possible drift over long dialogues, dependence on psychometric proxies for measurement, and the ethical concern that personality control can affect user perception and trust; it therefore recommends transparency, moderation on negative traits, and ongoing monitoring (Cho et al., 8 Aug 2025).

In comparison with alternatives, persona prompts based on curated biographies are said to require manual data and tuning while providing limited intensity control. Fine-tuning, adapters, LoRA, RLHF, and control tokens can yield strong trait embodiment, but they require training, compute, and datasets and are less flexible for per-session personality changes. Big5-Scaler’s stated advantages are the absence of training and data requirements, numeric fine-grained control, immediate deployment across models, and strong proportionality, balanced against the limitations already noted (Cho et al., 8 Aug 2025).

An adjacent but distinct research line treats Big Five inference as a measurement problem rather than a conditioning problem. “Extracting Psychological Indicators Using Question Answering” frames the five traits as extractive QA with explicit no-answer handling, using deepset/roberta-base-squad2, Reddit comments as context, and the question template “What points towards psychological trait {openness, conscientiousness, extroversion, agreeableness, neuroticism}?”; it reports overall validation performance up to EM $k \in \{1, \ldots, K\}$4 and F1 $k \in \{1, \ldots, K\}$5 at 33% no-answer, while also noting that the paper does not provide an explicit method to convert spans to continuous trait scores (Pavlović, 2023). This suggests a useful conceptual division: Big5-Scaler concerns inference-time personality control, whereas QA-based trait extraction concerns psychometric signal recovery from text. A plausible implication is that the latter can serve as an auxiliary calibration or auditing mechanism for the former, but the two approaches address different stages of the personality pipeline.