💬 NLP

Can LLMs Discern the Traits Influencing Your Preferences? Evaluating Personality-Driven Preference Alignment in LLMs

This paper proposes a framework that leverages stable personality traits as latent signals to filter and align user preferences, demonstrating that conditioning LLMs on personality-consistent preferences significantly improves personalized question-answering accuracy and introducing PACIFIC, a new dataset of personality-labeled preference statements.

Original authors: Tianyu Zhao, Siqi Li, Yasser Shoukry, Salma Elmalaki

Published 2026-03-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Tianyu Zhao, Siqi Li, Yasser Shoukry, Salma Elmalaki

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to hire a personal assistant to help you make decisions, like choosing a vacation spot or picking a new coffee table.

The Problem: The "Forgetful" Assistant
Currently, Large Language Models (LLMs) are like assistants with a very short attention span. If you tell them, "I hate crowds," they might remember that for the first minute. But if you chat for an hour about other things, or if you mention a preference vaguely, they often forget or get confused. They try to remember every single specific rule you've ever said, which is like trying to carry a library of 10,000 books in your head. It's too much, and they drop the important ones.

The Solution: The "Personality Compass"
This paper proposes a smarter way. Instead of asking the assistant to remember every single rule you've ever made, why not just understand who you are?

Think of your personality (like being adventurous, organized, or anxious) as a Compass. Even if you forget to tell the assistant your specific preference for "avoiding crowded beaches" today, the assistant can look at your Compass (your personality) and say, "Ah, this person is generally anxious and dislikes crowds, so they probably wouldn't like a crowded beach."

The researchers call this approach Personality-Driven Preference Alignment.

The Three Main Ingredients

1. The New Map: PACIFIC Dataset

The team created a new map called PACIFIC. Imagine a giant library of 1,200 different scenarios (like buying a car, planning a trip, or choosing a movie).

The Old Way: Just writing down "I like action movies."
The PACIFIC Way: They labeled every preference with a "Big Five" personality tag (like Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism).
The Twist: They realized that for some traits, like Neuroticism (anxiety), people don't act like the trait suggests. An anxious person doesn't want to feel anxious; they want safety. So, they created a special rule: "High Anxiety" people actually prefer "Low Risk" options. It's like understanding that a person who is afraid of heights doesn't want to be told to "jump," but rather to "stay on the ground."

2. The Experiment: Guessing the Right Choice

They tested this on a game of "Multiple Choice."

Scenario: A user asks, "What coffee table should I buy?"
The Options:
1. A cheap, mass-produced metal table.
2. A sleek, glass table from a big store.
3. A flat-pack table from a box store.
4. A unique, handmade table from a local artist.
The Clue: The user previously said, "I hate mass-produced stuff; I like unique things."
The Test:
- Without Personality: The AI guessed randomly or picked the "safest" answer. It got it right only 29% of the time.
- With Personality: The AI looked at the user's "Compass" (e.g., "This user is High in Openness"). It realized, "High Openness people love unique, artistic things." It picked the handmade table. It got it right 76% of the time!

3. The Toolkit: How to Talk to the AI

The paper tested different ways to give the AI this "Compass" information:

The "Reminder" Method: Just telling the AI, "Hey, remember this user likes unique things," helped a little.
The "Label" Method: Giving the AI a cheat sheet that says, "User = High Openness, Low Conscientiousness," worked much better. It was like giving the assistant a clear instruction manual instead of a vague hint.
The "Search" Method (RAG): Sometimes the AI doesn't know the user's personality tags. So, they built a tool that searches the user's past chats, finds the most relevant "personality clues," and hands them to the AI. It's like a librarian finding the right book for you before you even ask for it.

The Big Takeaway

The paper found that understanding the "Why" (personality) is better than memorizing the "What" (specific preferences).

The "Social Desirability" Trap: The researchers also found a funny flaw. AI models are trained to be "nice" and "polite." Sometimes, they refuse to recognize that a user might be "low in Conscientiousness" (a bit messy) or "high in Neuroticism" (anxious) because those sound like "bad" traits. The AI tries to be too positive and ignores the user's actual messy or anxious nature. This is like a friend who refuses to admit you're tired and keeps telling you to "cheer up," even when you need a nap.

In Summary

This paper teaches us that to make AI truly personal, we shouldn't just feed it a list of your likes and dislikes. Instead, we should teach it to understand your personality compass. Once the AI knows who you are, it can predict what you want, even when you haven't said it out loud. It turns the AI from a forgetful note-taker into a thoughtful friend who truly "gets" you.

1. Problem Statement

While Large Language Models (LLMs) are increasingly used for personalized recommendations, current systems struggle to reliably leverage user preference signals for answer generation. Existing approaches face three critical challenges:

Noise and Incompleteness: User preferences are often stated sporadically, are incomplete, or are contradictory.
Contextual Fragility: Models fail to track preferences over long conversations; accuracy drops significantly (often below 10%) after just a few turns.
Lack of Latent Structure: Current benchmarks focus on explicit constraint following or memory retrieval but fail to evaluate a model's ability to infer latent personality traits that drive those preferences.

The authors propose that personality traits (specifically the Big Five/OCEAN model) serve as a stable, high-level "latent signal" that can organize and interpret noisy preference data more robustly than storing raw preference statements verbatim.

2. Methodology

A. The PACIFIC Dataset

To address the lack of psychometrically grounded benchmarks, the authors introduce PACIFIC (Preference Alignment Choices Inference for Five-factor Identity Characterization).

Scale: 1,200 distinct human preference statements across 20 diverse domains (e.g., travel, finance, home maintenance).
Structure: Each instance consists of a triple $(p_i, q_i, a_i)$ $(p_{i}, q_{i}, a_{i})$ :
- $p_i$ : A user preference statement (explicit or implicit).
- $q_i$ : A multiple-choice question with 4 options.
- $a_i$ : The ground-truth answer aligned with the user's preference.
Annotation Strategy (Dual-Strategy):
- Mirroring Strategy (O, C, E, A): For Openness, Conscientiousness, Extraversion, and Agreeableness, preferences are scored based on Self-Congruity Theory. High scores indicate alignment with the trait (e.g., a highly Conscientious user prefers structured tools).
- Compensatory Strategy (N): For Neuroticism, the approach is inverted based on the "Safety vs. Efficiency" trade-off. High Neuroticism users seek safety/risk-avoidance (low risk), while Low Neuroticism users prioritize efficiency even at the cost of risk.
Quality Control: A strict consistency filtering protocol ( $\tau=0.7$ confidence threshold) ensures that both the preference and the correct answer choice explicitly exhibit the target trait polarity, filtering out common-sense answers.

B. Persona-Driven Preference-Following Framework

The authors propose a framework where LLMs infer a user's latent personality profile from historical interactions and use this profile to retrieve or select relevant preferences for new queries. They evaluate four prompting strategies:

Few-Shot (Preference Only): Providing raw preference statements aligned with the query's trait.
Few-Shot + Persona Hints: Augmenting preferences with explicit trait labels (e.g., "High Openness") and/or trait scores for the answer choices.
Reminder: Providing raw preferences with a natural language instruction to consider the user's implied personality.
Retrieval-Augmented Generation (RAG): Using a Dense Passage Retriever (DPR) to automatically fetch trait-aligned preferences from a database based on the query embedding. They compare a pretrained retriever against one fine-tuned on the PACIFIC dataset to prioritize persona consistency over semantic similarity.

3. Key Contributions

Empirical Validation of Personality as a Signal: The study demonstrates that conditioning on personality-aligned preferences significantly outperforms using random or mixed preferences, proving that personality traits act as a robust prior for preference prediction.
PACIFIC Dataset: The first benchmark explicitly designed to evaluate the link between Big Five personality traits and downstream user preferences, addressing the distributional skew and lack of psychometric validity in prior datasets (like PrefEval).
Novel Framework: A comprehensive evaluation of methods to inject persona information into LLMs, showing that explicit trait labeling and retrieval mechanisms can substantially improve personalization accuracy.

4. Experimental Results

Experiments were conducted using the Gemma-3-4B-IT model on 200 sampled questions per trait category.

Baseline Performance: Without preference context, accuracy was ~25.75%. With random/mixed preferences, it dropped to 29.25%.
Trait-Aligned Preferences: Using preferences aligned with the user's inferred trait increased accuracy to 63.00%.
Impact of Explicit Labels: Augmenting preferences with ground-truth trait labels (e.g., "High Openness") further boosted accuracy to 76.00%.
- Note: Providing trait labels for the answer choices actually reduced performance (to 57.25%), suggesting a "positivity bias" where models struggle to select "Low" trait options (e.g., Low Conscientiousness) when explicitly labeled.
RAG Performance: A pretrained retriever achieved ~30% accuracy (similar to random). However, a fine-tuned retriever improved this to 43.00%, demonstrating that models can learn to retrieve persona-consistent contexts without explicit labels.
Social Desirability Bias: The model exhibited significant difficulty in identifying "Low" poles of traits (e.g., Low Openness, Low Conscientiousness), often conflating them with negative behavior due to RLHF safety constraints. This led to misaligned recommendations for users with these traits.

5. Significance and Implications

Shift from Memory to Abstraction: The paper argues that effective user modeling in agentic AI should shift from "lossless storage" of raw history to semantic abstraction via personality traits. This reduces computational overhead and noise while improving long-term consistency.
Robust Personalization: By leveraging stable personality traits, LLMs can make reliable decisions even when specific user preferences are sparse, outdated, or noisy.
Future Directions: The work highlights the need for better handling of "Low" trait signals in RLHF-trained models to avoid social desirability bias and suggests that future retrieval systems must be explicitly trained on persona consistency, not just semantic similarity.

In conclusion, the paper establishes that personality-driven alignment is a superior paradigm for LLM personalization, offering a 46%+ improvement in accuracy over baseline methods and providing a foundational dataset and framework for future research in psychometrically grounded AI.