This paper investigates the discrepancy in LLM personalization performance between synthetic and human data. It finds that human data reveals significant system limitations in attribute extraction, attribute relevance, and generating genuinely personalized responses.
This work introduces a framework for Human-Centered Large Language Models (HCLLMs), integrating perspectives from NLP, HCI, and responsible AI. It argues for prioritizing human concerns, preferences, and values rigorously at every stage of LLM development, rather than as a mere post-training consideration.
The author reflects on transitioning from a reactive, data-driven monitoring approach at the Siuntio Fort to a more observant, human-aligned stewardship. By synchronizing with the household's natural rhythms, they learned that true presence and understanding the data's story are more important than constant processing.