Recently, Large Language Model (LLM) agents have been widely deployed as conversational assistants, tutors, and productivity tools, where the personality of the agents emerges in generated sentences that reflect the extent of their extraversion or introversion. The style of the generated sentences can be adjusted to match a target extraversion level while preserving semantic fidelity. Traditional approaches , such as few-shot prompting and prompt tuning, have limited abilities to reflect the degree of extraversion of an agent. Therefore, it is required to extract and emphasize words that reflect their personalities. We propose a new framework, termed Keyword-centered Rescheduling, that inserts soft prompt embeddings before sentence-level keywords based on personality traits. This design enables fine-grained trait-specific modulation at the lexical level while maintaining semantic consistency and remains fully compatible with diverse LLM architectures without the need for additional fine-tuning or instruction alignment. On benchmark datasets, the method improved the normalized composite F1-Score by 60% and increased the human-judged Top-2 Semantic Fidelity by 26% (86% vs. 68%) over prompt tuning while maintaining comparable Behavioral Alignment. Beyond improving controllability, our approach provides a practical foundation for deploying robust, user-aligned LLM agents with on-demand, continuous personality control and stable semantics.
Ji et al. (Fri,) studied this question.