PCGEF: A Framework for Diagnosing Subjective Alignment in Human-Centered Persona-Conditioned Generation

Evaluating LLMs in subjective and expressive domains is challenging, as standard accuracy metrics overlook affect, style, and coherence. We present the Persona-Conditioned Generation Evaluation Framework (PCGEF), which disentangles the effects of persona and continuity controls (via summarization-based memory) across five axes: Affective Alignment, Preference Alignment, Stylistic Expressiveness, Semantic Grounding, and Contextual Coherence. These two levers influence alignment through different mechanisms, and the five axes capture key forms of subjective drift. Unlike prior persona-aware approaches, PCGEF compares model generations under controlled conditions rather than relying on gold standards or LLM judges. We instantiate PCGEF in a red-wine description task with a 2×2 factorial design involving 34 participants and four mid-scale, open-weight LLMs. Results show that persona control improves affective and preference alignment and tends to enhance style, continuity control stabilizes coherence, and semantic grounding remains weak. PCGEF offers a reusable, interpretable framework transferable to sensory/creative domains and interactive dialogue.