FATE-5: Faithful Alignment of Trait Expression for Continuous Big Five Personality Control

Gunhee Cho; Hongyeob Kim; Yun-Gyung Cheong

cs.CLOct 26, 2026

FATE-5: Faithful Alignment of Trait Expression for Continuous Big Five Personality Control

Authors: Gunhee Cho, Hongyeob Kim, Yun-Gyung Cheong

Organizations: Department of Artificial Intelligence, Sungkyunkwan University, Suwon, South Korea

Abstract

Controlling personality in large language models is important for building conversational agents that can be customized to different users and interaction settings. However, existing personality control methods often face a tradeoff between strong trait expression and natural dialogue generation. We introduce FATE-5, Faithful Alignment of Trait Expression for the Big Five, a two-stage framework that separates personality encoding from naturalness alignment. FATE-5 first learns continuous soft prompt representations of Big Five profiles using dialogue supervision and an auxiliary survey objective. It then improves dialogue naturalness through preference alignment based on both dialogue quality and personality agreement. Across three language model backbones, FATE-5 maintains strong trait expression while achieving substantially higher dialogue naturalness than other trait-strong methods. On held-out questionnaires, it achieves monotonicity correlations of 0.97-0.99 and recovery RMSE of 8.9-12.4, substantially improving over the evaluated baselines. These results demonstrate effective continuous personality control while maintaining natural conversational behavior.

Figures & tables

Figure 1: Overview of FATE-5. Stage 1 learns a continuous personality representation from a target Big Five profile using a lightweight encoder while keeping the backbone LLM frozen. The encoder is trained jointly with dialogue loss and survey loss to strengthen trait expression. Stage 2 freezes the encoder and backbone, attaches a trainable LoRA adapter, and performs offline DPO using preference pairs scored by an external LLM judge. This stage improves dialogue naturalness while preserving the personality representation learned in Stage 1.

Configuration	Trait Acc.	Nat.
FATE-5 (Encoder+DPO, $\lambda{=}1$ )	75.3	43.0
$-$ DPO (Encoder only)	71.3	37.5
$-$ Encoder (numeric + DPO)	60.0	48.8
$-$ Enc $-$ DPO (numeric)	69.0	46.2
$-$ Survey loss	51.3	34.2
DPO $\rightarrow$ SFT	69.7	41.3

Table 3: Ablation of FATE-5 on Llama-3-8B (RQ4), scored using the same LLM judges as Table 1 , which are distinct from the judge used to build the DPO preference pairs. Trait Acc. measures target-trait accuracy, and Nat. measures dialogue naturalness on a

0

–

100

scale. The full model row is taken from the Llama-3-8B column of Table 1 . Each other row removes or substitutes one component.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

$\lambda$	Usable pairs	Mean reward gap	Mean utterance length	Nat.
$0$	$37$	$33.8$	$12.3$	$19.4$
$1$	$198$	$12.8$	$36.3$	$43.0$
$2$	$200$	$25.7$	$34.4$	$32.5$

Appendix

Table 4: Preference pair statistics for different values of

\lambda

on Llama-3-8B. With

\lambda{=}0

, the reward is limited to five discrete levels. This produces many ties between candidates and removes most profiles from preference training, resulting in a much smaller preference-training set. Mean reward gap is reported only as a within-setting diagnostic of pair separation and is not directly comparable across different values of

\lambda

, since the reward definition changes with

\lambda

. Mean utterance length is measured in tokens.

Personality Modeling in Language Models

Personality modeling in language models studies how traits are measured, represented, simulated, and controlled in model behavior.

See all 167 papers

Scaling Personality Control in LLMs with Big Five Scaler Prompts
Aug 8, 2025Gunhee Cho, Yun-Gyung CheongPersona-Conditioned Generation Controllable Text Generation
Sungkyunkwan University, South Korea
Fusian: Multi-LoRA Fusion for Fine-Grained Continuous MBTI Personality Control in Large Language Models
Mar 16, 2026Zehao Chen, Rong PanPersonality Modeling in Language Models Persona-Conditioned Generation
School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China
Personality, Role, and Expressive Style in Large Language Models: An Interactionist Analysis
May 27, 2026Moe Nagao, Koichiro Terao, Mikio Nakano +1Large Language Model-Based Role-Play Simulation LLM Prompting
Okayama Prefectural University · AI & Humans Lab · C4A Research Institute

Aug 8, 2025J/K move · Enter open · S save

Scaling Personality Control in LLMs with Big Five Scaler Prompts

Gunhee Cho, Yun-Gyung Cheong

Sungkyunkwan University, South Korea

Persona-Conditioned Generation Controllable Text Generation Personality Modeling in Language Models LLM Prompting

Open paper AskPDF arXiv

We present Big5-Scaler, a prompt-based framework for conditioning large language models (LLMs) with controllable Big Five personality traits. By embedding numeric trait values into natural language prompts, our method enables fine-grained personality control without additional training. We evaluate Big5-Scaler across trait expression, dialogue generation, and human trait imitation tasks. Results show that it induces consistent and distinguishable personality traits across models, with performance varying by prompt type and scale. Our analysis highlights the effectiveness of concise prompts and lower trait intensities, providing a efficient approach for building personality-aware dialogue agents.

Abstract

Personality Modeling in Language Models

Explore similar work

Scaling Personality Control in LLMs with Big Five Scaler Prompts

Fusian: Multi-LoRA Fusion for Fine-Grained Continuous MBTI Personality Control in Large Language Models

Personality, Role, and Expressive Style in Large Language Models: An Interactionist Analysis