Towards Realistic Personalization: Evaluating Long-Horizon Preference Following in Personalized User-LLM Interactions

This paper introduces RealPref, a comprehensive benchmark designed to evaluate how well Large Language Models follow diverse user preferences in realistic, long-horizon interactions, revealing significant performance drops as context lengthens and preference expression becomes more implicit.

The Gist

Imagine you have a digital assistant, like a super-smart butler named "Alex." You've been chatting with Alex for months. You've told him you hate cilantro, you love jazz music, and you get anxious in crowded elevators. You've mentioned these things in passing, sometimes directly, sometimes by complaining about a salad, or by sighing when a jazz band plays too loudly.

Now, imagine you ask Alex: "What should I do this weekend?"

A truly personalized Alex wouldn't just suggest a generic jazz club or a crowded concert. He would remember your specific history, your subtle hints, and your long-term habits to suggest a quiet jazz listening party in a small, empty room.

The Problem:
Current AI assistants are great at following simple, one-time instructions ("Write a poem about a cat"). But when it comes to being a personal assistant over a long period, they often forget things, miss subtle hints, or get confused when the conversation gets very long. They act more like a robot that just read a manual than a friend who actually knows you.

The Solution (RealPref):
The authors of this paper built a giant test called RealPref to see how good these AI "butlers" really are at remembering and respecting your personal quirks over time.

Here is how they did it, using some fun analogies:

1. The "Memory Palace" Dataset

Instead of just asking the AI simple questions, they created 100 fake people with detailed lives.

• The Backstory: Each fake person has a biography, a job, and a history of life events (like graduating or starting a business).
• The Preferences: These people have 1300 specific likes and dislikes. Some are obvious ("I love pizza"), but many are hidden.
• The Conversation: They generated thousands of chat sessions. Sometimes the person says, "I hate spicy food" (Direct). Other times, they say, "Ugh, this curry is too hot, I'm sweating!" (Implicit). Sometimes, they only reveal a preference after talking about it for three different weeks (Long-Horizon).

2. The "Detective" Test

To test the AI, they didn't just ask, "Do you remember I hate spicy food?" That's too easy. Instead, they set up a mystery.

• The Scenario: The user asks, "I'm going to a new restaurant; what should I order?"
• The Trap: The AI has to look back through hundreds of pages of previous chat logs to find the clue that the user hates spicy food.
• The Challenge: The clues might be hidden in a conversation from three weeks ago, buried under 50 pages of talk about the weather, or hidden inside a metaphor ("I prefer my life as calm as a lake, not a boiling pot").

3. The Three Types of "Secrets"

The paper tested how well AI handles different ways people hide their preferences:

• The Shout: "I hate spicy food!" (Easy for AI).
• The Whisper: "I usually skip the spicy options on the menu." (Medium difficulty).
• The Riddle: "I prefer things that don't make my eyes water." (Hard for AI).
• The Slow Reveal: Over three months, the user complains about hot sauce, then mentions a bad experience with chili, then finally says they only eat mild food. The AI has to connect the dots over time.

4. What They Found (The Results)

The results were a bit of a wake-up call for the AI industry:

• The "Needle in a Haystack" Problem: As the conversation got longer (more "haystack"), the AI got worse at finding the "needle" (the preference). Even the smartest models started forgetting things when the chat history got too long.
• The "Subtlety" Gap: When people were direct, AI did okay. But when people were subtle or implied their preferences, the AI often failed completely. It's like a waiter who only understands "I want water" but gets confused if you say, "I'm feeling a bit parched."
• The "Stranger" Problem: If the AI learned you liked jazz, could it guess you'd also like a jazz-themed movie? Often, no. It struggled to apply what it learned about one thing to a totally new situation.
• The "Magic Trick" Fix: They tried helping the AI by giving it a "cheat sheet" (a reminder) or by letting it search its own memory (Retrieval-Augmented Generation). This helped, but it wasn't a perfect fix.

Why This Matters

Think of this paper as a driver's license test for AI personalization.

Right now, most AI assistants are like learner drivers who can drive in an empty parking lot (short, simple chats) but panic when you put them in rush-hour traffic with complex rules (long, nuanced, real-life conversations).

RealPref is the test that proves they aren't ready for the highway yet. It shows us that to build an AI that truly feels like a helpful, personal friend, we need to teach it to listen better, remember longer, and understand the "unsaid" things we say.

In short: We want AI that doesn't just answer questions, but knows us. This paper is the first big step in measuring how far we are from that dream.

Technical Summary

1. Problem Statement

While Large Language Models (LLMs) excel at general tasks, their ability to function as personal assistants in realistic, long-term scenarios remains underexplored. Existing benchmarks suffer from three critical limitations:

• Simplified Context: Most evaluations rely on short-term, single-turn dialogues rather than complex, multi-session interaction histories.
• Oversimplified Preferences: Real-world user preferences are often implicit, scattered, and expressed through emotional cues or lifestyle choices. Current benchmarks assume explicit, immediate instructions.
• Inadequate Evaluation: Existing methods often use classification tasks that allow models to "guess" answers by comparing options, or generation tasks lacking rigorous scoring criteria for open-ended responses.

The core research question is: Can LLMs effectively capture, retain, and follow user preferences across extended interaction contexts, and can they generalize these preferences to unseen scenarios?

2. Methodology: The RealPref Benchmark

The authors propose RealPref, a comprehensive benchmark designed to simulate authentic, long-horizon user-LLM interactions.

A. Dataset Construction

RealPref is generated via a multi-stage pipeline using synthetic data to ensure privacy and diversity:

• User Profiles: 100 curated user personas with detailed demographics, biographies, and life events (e.g., graduation, career changes) to establish a timeline.
• Preferences: 1,300 total preference-query pairs per user.
  • Original Preferences: 10 per user, covering diverse topics (7 high-level fields, 37 subtopics).
  • Generalized Preferences: 3 per user, logically inferred from original preferences to test reasoning across unseen scenarios (e.g., inferring a preference for eco-friendly cleaning supplies from a preference for cruelty-free beauty products).
• Conversation Sessions:
  • Life-Event Sessions: Establish user background.
  • Random Sessions: User-irrelevant conversations inserted to control context length and dilute personal signals.
  • Preference-Expression Sessions: The core of the dataset, featuring four distinct expression types ranging from explicit to implicit:
    1. Explicit - Direct Statement: Direct declaration of preference.
    2. Explicit - Contextualized Mention: Preference stated within a broader daily topic.
    3. Implicit - Stylistic Expression: Preference implied via metaphors, rhetorical questions, or emotional tone (no explicit "like/dislike" words).
    4. Implicit - Experience Feedback: Preference revealed gradually over multiple sessions through shared experiences and feedback.
• Context Lengths: Configurations range from 2K tokens (Simple) to 247K tokens (Extreme), with random conversations inserted at intervals or appended to the tail to test retrieval capabilities.

B. Evaluation Framework

RealPref utilizes three task types and a granular rubric system:

1. Multiple-Choice Questions (MCQ): Often prone to "option comparison" artifacts where models guess the outlier.
2. True-or-False Questions: A novel classification task where the model evaluates a single option against the user's history, preventing option comparison.
3. Open-Ended Generation: The most challenging task, requiring the model to proactively generate a response aligned with preferences.
   • Scoring Rubrics (1-5 scale):
     • Preference Awareness: Does the model recognize and mention the preference?
     • Preference Alignment: Does the content strictly adhere to the preference (e.g., avoiding disliked items)?
     • Answer Quality: Is the response helpful and constructive?

3. Key Contributions

1. RealPref Dataset: The first user-centered, long-horizon dialogue dataset incorporating diverse preference expressions (explicit to implicit) and generalization scenarios.
2. Robust Evaluation Protocol: A multi-faceted framework combining classification and generation tasks with systematic, multi-dimensional rubrics for open-ended evaluation.
3. Comprehensive Empirical Analysis: Extensive experiments analyzing the impact of context length, expression complexity, and improvement techniques on state-of-the-art LLMs.

4. Experimental Results

The authors evaluated models including GPT-5, GPT-5 mini, Qwen3, Gemini 2.5 Flash-Lite, and Llama 3.3.

• Task Type Comparison:
  • True-or-False outperformed MCQs in reflecting true preference-following abilities, as MCQs allowed models to exploit "odd-one-out" logic without understanding the user.
  • Open-Ended Generation provided the best differentiation between models, as it requires proactive reasoning rather than passive selection.
• Effect of Expression Type:
  • Performance dropped significantly as preferences became more implicit. Models struggled most with Stylistic Expression and Experience Feedback, indicating a difficulty in inferring preferences from non-literal cues.
• Long-Context Performance:
  • Context Length Sensitivity: As context length increased (from 2K to 142K+ tokens), performance degraded sharply, particularly in Preference Awareness.
  • Position Matters: Placing random conversations at the tail (after the preference) caused a larger drop than inserting them in the middle, suggesting that the distance between the preference cue and the query is a critical factor.
• Improvement Methods:
  • Reminder: A simple prompt to recall preferences was effective and cost-efficient.
  • Few-Shot CoT: Provided moderate improvements.
  • Retrieval-Augmented Generation (RAG): Proved most effective for very long contexts (exceeding 142K tokens), helping models retrieve relevant history when the full context window was too large for effective attention.
• Preference Generalization:
  • Models performed worse on generalized preferences (unseen scenarios) compared to original ones.
  • While reminders helped, models still struggled to infer new preferences based on underlying reasons, highlighting a lack of proactive reasoning.

5. Significance and Future Directions

• Foundation for User-Aware AI: RealPref provides a necessary baseline for developing LLM assistants that can adapt to individual needs over time, moving beyond static instruction following.
• Identified Gaps: The study reveals that current SOTA models are not yet robust enough for real-world personalization, particularly regarding implicit cues and long-term memory retention.
• Future Work: The authors suggest expanding the dataset to include dynamic preferences (changing over time), multimodal interactions, and more complex feedback mechanisms (e.g., thumbs up/down). They also emphasize the need for responsible AI practices regarding privacy and over-personalization.

In conclusion, RealPref demonstrates that while LLMs show promise, significant challenges remain in maintaining consistency and reasoning over long, complex, and implicit user interaction histories. The benchmark serves as a critical tool for driving the next generation of personalized AI assistants.