Meta-Consistent Persona Modulation for Large Language Models without Fine-Tuning

Personalized large language models (LLMs) should maintain a stable, recognizable persona across contexts, yet existing approaches based on fine-tuning or prompt templates often drift in style and are costly to deploy at scale. We introduce \emph{Meta-Consistent Persona Modulation} (\textbf{PersonaReg}), a lightweight framework that keeps the LLM backbone frozen while a meta-network generates layer-wise feature-wise affine parameters $(\gamma,\beta)$ conditioned on user and context embeddings. The modulation is applied after LayerNorm at selected mid-to-high layers and parameterized in a low-rank subspace to minimize degrees of freedom. Training optimizes language modeling with (i) a cross-context consistency loss on hidden states and (ii) a semantic-preservation loss given by symmetric KL between base and modulated output distributions, plus a stability penalty on modulation magnitude. On \textsc{PersonaChat} with a frozen Qwen-2.5B backbone, PersonaReg improves persona consistency from $0.71$--$0.78$ to $0.84$ while keeping semantic fidelity (Sym-KL $\approx 0.039$) and perplexity ($20.6$) on par with baselines, using only $\sim 0.8$M additional parameters ($\approx 0.03\%$ of the backbone). Layer-wise analysis shows modulation intensity peaks at layers $24$--$30$, aligning with the largest gains in consistency. These results demonstrate that meta-conditioned feature-wise modulation of a frozen backbone is an effective and scalable route to stable persona control in LLMs.