LLMs Could Better Personalize Responses by Understanding Your Personality Traits

Personalizing language model responses based on what users prefer has emerged as a key goal, but relying on raw preference signals often backfires—user choices can be contradictory, incomplete, or simply wrong. Researchers have now proposed rethinking this problem by anchoring personalization in the stable psychological traits that actually shape how people make everyday decisions.

The Personality-Driven Approach#

The new framework, called PACIFIC, leverages the Big-Five personality model (openness, conscientiousness, extraversion, agreeableness, neuroticism—or OCEAN) as a principled foundation for understanding and organizing user preferences. Rather than treating each preference in isolation, the system uses these five psychological dimensions as a latent signal that connects disparate choices into a coherent picture of who the user is. The intuition is straightforward: if you know someone scores high on openness, you can better predict their aesthetic and intellectual choices across domains, even when you lack explicit preference data.

Testing the Framework#

To rigorously test this approach, the researchers built a dataset of 1,200 preference-query pairs spanning travel, movies, education, and other domains, carefully constructed to represent both high and low ends of each Big-Five trait. The results were striking: when language models were given clean, trait-aligned preference context, they achieved near-perfect accuracy—in some cases reaching 99%—demonstrating that reasoning ability is not the limiting factor.

The Real Bottleneck: Finding Relevant Preferences#

The challenge emerges in the real world, where preference histories are messy, mixed-trait, and unlabeled. The team found that the true obstacle is retrieval—identifying which past preferences actually matter for a given query. They developed a persona-aware contrastive retriever (PiRAG) that improved label-free accuracy from 30% to 43% compared to standard semantic search, all without requiring personality trait annotations at inference time.

Why It Matters#

This work suggests that understanding user personality is not merely a nice-to-have for personalization; it may be essential for reliably surfacing the most relevant past choices. By anchoring preference alignment in psychological principles rather than treating preferences as atomized signals, LLMs could deliver more coherent and trustworthy personalized answers.