Disentangling Style and Semantics for Calendar-Driven Text Generation: A Knowledge Graph-Guided Activation Steering Approach

Personalizing LLMs for structured communications such as event invitations faces a fundamental tension: generated text must preserve an exact factual schema (dates, venues, hosts) while reflecting the user’s distinct stylistic register. Existing methods: fine-tuning, RAG, and vanilla activation steering, conflate style and content in representation space, a failure we term semantic leakage. We present a framework that structurally reduces leakage by pairing Symbolic Knowledge Graph (KG) guidance with Contrastive Activation Steering. Before any activation is extracted, the KG defactualizes email pairs by replacing every entity with a typed placeholder; style vectors derived from these clean pairs are injected during generation, after which the KG refactualizes the output with verified entity values. On LLaMA-2-7B, the approach achieves ROUGE-L 0.2261 and METEOR 0.2418 (+11.7% and +23.3% over pure steering). KG-guided refactualization reduces Entity Hallucination Rate (EHR) by 61% (0.90 \(\rightarrow \) 0.35), with BERTScore F1 \(=+\) 0.193. Non-LLM overhead is <0.05% of inference latency; per-user style storage scales to 10k users within 1% of model memory.