Controllable and explainable personality sliders for LLMs at inference time

Abstract

Aligning Large Language Models (LLMs) with specific personas typically relies on expensive and monolithic Supervised Fine-Tuning (SFT) or RLHF. While effective, these methods require training distinct models for every target personality profile. Inference-time activation steering offers a parameter-efficient alternative, yet naive approaches fail to control multiple traits simultaneously due to destructive vector interference. In this work, we propose a modular framework for continuous, multi-dimensional personality control. Our key innovation is Sequential Adaptive Steering (SAS): a method that orthogonalizes steering vectors by training subsequent probes on the residual stream shifted by prior interventions. This approach transforms steering vectors into reusable primitives, allowing users to instantly synthesize complex, high-fidelity personality profiles by simply adjusting coefficients alpha. We validate our framework on the Big Five personality traits, demonstrating that it outperforms naive baselines in both goal adherence and coherence, enabling precise, holistic personality modulation without updating model parameters.

Figures (18)

• Figure 1: System Overview: Users can dynamically adjust model personality via five-factor traits (OCEAN): Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism. Our Sequential Adaptive Steering method enables the composition of these traits to synthesize novel personality profiles and steering behaviors. \ref{['sec:procedure']} explains the procedure for measuring the personality profile.
• Figure 2:
• Figure 3: Single-Trait Controllability. The impact of steering intensity ($\alpha$) on the resulting personality score (range 1--5, where 1 = low trait expression and 5 = high). Only points with $<50\%$ perplexity degradation are shown. The monotonic increase confirms that probes provide fine-grained control over individual dimensions.
• Figure 4: Radar plot comparing the Baseline, DPO, and our adaptive steering approach against a target multi-dimensional personality profile. Our method successfully achieves the target configuration with minimal cross-trait interference. The naive chaining approach is omitted as it causes rapid degeneration of model coherence before achieving meaningful multi-dimensional shifts.
• Figure 5: The Pareto Frontier of Personality Score vs. Perplexity. SAS achieves a superior trade-off, maintaining coherence (low perplexity) even at high steering intensities. In contrast, naive activation addition fails to reach significant alignment scores, causing perplexity to explode before meaningful personality shifts can be realized.