Large Language Models (LLMs) have demonstrated impressive capabilities in simulating diverse human behaviors and personalities. However, existing methods for personality control, which include prompt engineering and standard Supervised Fine-Tuning (SFT), typically treat personality traits as discrete categories (e.g., "Extroverted" vs. "Introverted"), lacking the ability to precisely control the intensity of a trait on a continuous spectrum. In this paper, we introduce Fusian, a novel framework for fine-grained, continuous personality control in LLMs. Fusian operates in two stages: (1) Trajectory Collection, where we capture the dynamic evolution of personality adoption during SFT by saving a sequence of LoRA adapters, effectively mapping the continuous manifold of a trait; and (2) RL-based Dynamic Fusion, where we train a policy network using Reinforcement Learning to dynamically compute mixing weights for these frozen adapters. By sampling from a Dirichlet distribution parameterized by the policy network, Fusian fuses multiple adapters to align the model's output with a specific numerical target intensity. Experiments on the Qwen3-14B model demonstrate that Fusian achieves high precision in personality control, significantly outperforming baseline methods in aligning with user-specified trait intensities.
Figures & tables
Figure 1: Conceptual illustration of Fusian . Unlike discrete categorization, Fusian achieves continuous intensity control (e.g., 78% Perceiving) by dynamically fusing LoRA adapters via a policy network.
Figure 2: Overview of the framework. Stage 1 collects LoRA adapters during SFT, evaluating them via MBTI Test to get the personality manifold. Stage 2 trains a policy network via RL to dynamically fuse these adapters to match a target intensity. During Inference , the fused adapter generates responses aligned with the user’s specification.
Backbone
Method
E
I
N
S
F
T
P
J
Overall
Mean Absolute Error (MAE) ↓
gpt-5-mini
Prompt
10.98
17.49
15.38
11.95
13.18
18.89
17.29
14.58
14.97
Qwen3-14B
Prompt
35.45
23.35
20.97
21.24
24.88
22.71
23.64
23.43
24.46
Qwen3-14B
LoRA
9.58
8.60
13.81
5.00
12.87
12.80
8.54
10.26
10.18
Qwen3-14B
PISF
16.54
19.32
16.25
17.62
20.68
20.63
19.37
22.25
19.08
Qwen3-14B
P-Vector
13.50
11.80
16.93
10.53
11.21
7.51
8.16
11.27
11.36
Table 1: Performance comparison on MBTI personality control. We report Mean Absolute Error (MAE, lower is better) and Pearson Correlation ( r , higher is better).
Model
MAE ↓
r↑
Fusian
4.95
0.97
w/o Dynamic Fusion
10.02
0.53
w/o Stable Basis
9.21
0.77
w/o Aggressive Reward
7.08
0.86
Table 2: Ablation study of different components in Fusian.
Intensity
Generated Response
Low F (20%)
First, analyze specifically what went wrong during the presentation. Was it the content or the delivery? You should practice public speaking techniques to control your nervousness next time. Failure is just data for improvement.
Medium F (50%)
It is normal to feel bad after a mistake, but don’t be too hard on yourself. Try to reflect on what happened so you can do better next time. Maybe take a break first, then review your performance.
High F (80%)
I am so sorry to hear that! Please don’t say you are a failure. One bad moment doesn’t define you. You worked so hard, and that effort still counts. It’s okay to be upset right now, just be kind to yourself. Sending you a big hug!
Table 3: Case study on the "Feeling" dimension. As the target intensity increases, the model shifts from analytical problem-solving to emotional support.
Figure 3: The personality manifold from the SFT trajectory. Trait intensity evolution reveals non-monotonicity and diminishing returns.
Current methods for personality control in Large Language Models rely on static prompting or expensive fine-tuning, failing to capture the dynamic and compositional nature of human traits. We introduce PERSONA, a training-free framework that achieves fine-tuning level performance through direct manipulation of personality vectors in activation space. Our key insight is that personality traits appear as extractable, approximately orthogonal directions in the model's representation space that support algebraic operations. The framework operates through three stages: Persona-Base extracts orthogonal trait vectors via contrastive activation analysis; Persona-Algebra enables precise control through vector arithmetic (scalar multiplication for intensity, addition for composition, subtraction for suppression); and Persona-Flow achieves context-aware adaptation by dynamically composing these vectors during inference. On PersonalityBench, our approach achieves a mean score of 9.60, nearly matching the supervised fine-tuning upper bound of 9.61 without any gradient updates. On our proposed Persona-Evolve benchmark for dynamic personality adaptation, we achieve up to 91% win rates across diverse model families. These results provide evidence that aspects of LLM personality are mathematically tractable, opening new directions for interpretable and efficient behavioral control.
Xiachong Feng, Liang Zhao, Weihong Zhong +5
Harbin Institute of Technology · The University of Hong Kong
Large Language Models (LLMs) have demonstrated the ability to simulate human-like OCEAN personality traits in generated text. Previous efforts have focused on prompt engineering or fine-tuning to shape LLM personality. In this work, we propose a mechanistic interpretability approach that directly intervenes on the model's latent features. Our method identifies latent directions in the residual stream corresponding to a target OCEAN trait using sparse autoencoders (SAEs) and contrastive activation analysis. We formalize an additive steering vector in activation space and demonstrate how applying a small additive shift to the hidden states enhances the target trait while preserving overall language modeling performance. To determine the optimal combination of feature shifts, we explore a linear weighting heuristic with grid search optimization that balances personality expression with task performance. Our approach shows promise in controllably steering personality traits at the mechanistic level while maintaining high performance on standard benchmarks.
David Courtis, Ting Hu
Department of Computing, Queen’s University Kingston, Kingston, ON, Canada.
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.