💬 NLP

Bridging Textual Profiles and Latent User Embeddings for Personalization

The paper introduces BLUE, a reinforcement learning framework that unifies interpretable textual user profiles with discriminative latent embeddings by aligning language-based profiles with embedding-space reward signals, thereby achieving superior performance in zero-shot sequential recommendation and cross-domain transfer while enhancing personalized context for question answering.

Original authors: Zhaoxuan Tan, Xiang Zhai, Yan Zhu, Meng Jiang, Mohamed Hammad

Published 2026-05-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zhaoxuan Tan, Xiang Zhai, Yan Zhu, Meng Jiang, Mohamed Hammad

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

The Big Problem: Two Different Languages

Imagine you are trying to understand a customer to recommend the perfect product to them. You have two ways to describe this customer, but they speak different languages:

The "Black Box" Vector (Latent Embedding): This is like a secret code or a dense mathematical fingerprint. Computers love this because it's incredibly fast and efficient for finding matches. However, it's a mystery to humans. If you ask the computer, "Why did you recommend this?" it can't explain itself; it just points to the code. It's like a GPS that knows exactly where to go but can't tell you why it chose that route.
The "Story" Profile (Textual Profile): This is a written summary, like a biography of the customer's tastes. "This person loves hiking, hates spicy food, and buys gifts for their kids." Humans love this because it's easy to read and understand. But computers often struggle to use these stories effectively for recommendations because they are hard to turn into the math needed for fast searching.

The Gap: Existing systems usually pick one or the other. They either use the fast-but-mysterious code (great for ranking, bad for explaining) or the readable story (great for explaining, hard to optimize for results).

The Solution: BLUE (The Translator)

The authors created a system called BLUE (Bridging textuaL profiles and latent User Embeddings). Think of BLUE as a translator that teaches the "Story" to speak the language of the "Black Box," and vice versa.

BLUE uses a Reinforcement Learning approach. Imagine a student (the AI) trying to write a summary of a customer.

The Teacher: A "professor" AI (a Large Language Model) writes the story.
The Judge: A "math" AI (an embedding model) acts as the judge. It doesn't care if the story is pretty; it cares if the story helps it find the right product.

How It Works: The Training Gym

BLUE trains the student AI using two types of rewards, like a video game with two scoreboards:

The Math Score (Embedding Space Reward):
- The student writes a profile.
- The "Math Judge" turns that profile into a secret code.
- The Judge checks: "Does this code point close to the item the user actually bought next?"
- If the code is close to the right item, the student gets a high score. If it's far away, they get a low score.
- Analogy: This is like a darts player. The student throws a dart (the profile). The judge checks if it hit the bullseye (the correct product). The student learns to throw better darts.
The Story Score (Text Space Reward):
- The student writes a profile.
- The system asks: "Based only on this story, can you guess what the user will buy next?"
- If the story contains enough clear clues to make the right guess, the student gets a bonus score.
- Analogy: This ensures the story isn't just gibberish that happens to work mathematically. It has to be a good story that actually explains the user's preferences.

The Result: The AI learns to write stories that are not only easy for humans to read but also mathematically perfect for the computer to use for recommendations.

What They Found (The Results)

The researchers tested BLUE on real data from Amazon (clothing, books, electronics) and Google Local Reviews.

Better Recommendations: Even when they used a "frozen" (unchangeable) math model, BLUE's text profiles helped the system recommend items much better than just feeding the raw history of clicks into the system. It was like giving the computer a cheat sheet that summarized the user's habits perfectly.
Cross-Domain Magic: They trained BLUE on clothing data and tested it on books and electronics. BLUE worked surprisingly well in these new areas. It seems the "stories" it learned captured general human preferences (like "likes comfort" or "buys for kids") that apply across different types of products.
Better Conversations: When they used these profiles to answer questions (e.g., "What will this user buy next?"), the answers were more accurate than when using raw data or other methods. The profiles acted as a perfect "context" for the AI to understand the user.

The "Secret Sauce"

The paper highlights that BLUE doesn't just make the AI talk better; it makes the AI think better about what matters.

In a case study, a user bought work pants, sandals, and some baby clothes.
Other methods got confused and thought the user was mostly interested in workwear or just random accessories.
BLUE's profile correctly identified that the user had a strong, consistent interest in children's apparel (despite it being a smaller part of the history) and predicted the next baby item correctly. It learned to filter out the "noise" and focus on the "signal."

Summary

BLUE is a new way to build user profiles. It forces the AI to write summaries that are readable for humans but optimized for math. It bridges the gap between the "black box" of computer science and the "storytelling" of human language, resulting in systems that recommend better items and explain their choices more clearly.

Technical Summary: Bridging Textual Profiles and Latent User Embeddings for Personalization (BLUE)

1. Problem Statement

Personalized systems rely on user representations to connect behavioral history with downstream applications like recommendation and question answering. Existing approaches typically fall into two distinct categories, each with significant limitations:

Latent User Embeddings: These dense vectors, learned from sequential interaction histories, are highly effective for retrieval and ranking but lack interpretability. They compress complex histories into a single vector, acting as a "black box" that obscures high-level preferences and makes it difficult to understand why a recommendation was made.
Textual User Profiles: These natural language descriptions are interpretable and can capture rich semantics. However, generating them is challenging because there is often no single ground-truth profile for a user. Consequently, existing methods often rely on manually crafted prompts or zero-shot generation, which may not align well with downstream utility or optimization objectives.

This creates a gap: latent embeddings are strong for retrieval but weak in reasoning, while textual profiles are interpretable but difficult to optimize for downstream performance. The paper proposes BLUE (Bridging textuaL profiles and latent User Embeddings) to unify these representations.

2. Methodology

BLUE is a reinforcement learning (RL) framework that treats user profile generation as a policy optimization problem. It connects the semantic transparency of language-based profiles with the discriminative ranking ability of latent embeddings.

Core Framework

Given a raw user interaction sequence $H_u$ , a profiler Large Language Model (LLM) generates a textual user profile $p_u$ . This profile serves two roles:

Text Space: An interpretable summary for downstream language tasks (e.g., personalized QA).
Latent Space: An input to an embedding model to generate user representations for item retrieval.

The framework optimizes the profiler LLM using two distinct reward signals:

A. Embedding-Space Reward (Latent Supervision)

To ensure the generated profile improves retrieval performance, BLUE utilizes a frozen embedding model to provide contrastive reward signals.

Mechanism: The generated profile (either alone or combined with recent history) is encoded into a latent user embedding.
Objective: The reward is based on bidirectional contrastive alignment (InfoNCE-style) between the user embedding and the target positive item embedding.
Signal: The reward encourages the profile to induce user embeddings that are closer to the positive next item and farther from negative items and other users in the batch. This guides the profiler to generate representations that are functionally useful for ranking, even without ground-truth text summaries.

B. Text-Space Reward (Semantic Supervision)

To ensure the profile remains semantically meaningful and usable in natural language, a text-space reward is introduced.

Mechanism: The generated profile is used as context for a next-item prediction task formulated as a multiple-choice question.
Signal: A binary reward is given if the profiler (acting as the predictor) correctly identifies the next item based solely on the generated profile. This ensures the profile contains sufficient explicit information to support language-based reasoning.

Optimization

The profiler model is optimized using Group Relative Policy Optimization (GRPO).

For each user, a group of candidate profiles is sampled.
The total reward is the sum of the text-space reward and the embedding-space reward.
The advantage is normalized within the group to reduce variance.
A KL-divergence term regularizes the policy to prevent drift from a reference model, ensuring fluency.

Embedding Model Adaptation (Optional)

While the primary setting involves a frozen embedding model, the paper also explores a setting where the learned textual profiles are used to fine-tune the embedding model via supervised contrastive learning. This tests whether the denoised, structured summaries can provide better supervision signals than raw interaction histories.

3. Key Contributions

Unified Framework (BLUE): A reinforcement learning approach that jointly optimizes textual user profiling and latent embedding alignment. It demonstrates that latent-space contrastive rewards can effectively supervise a profiler LLM without requiring ground-truth user summaries.
Dual-Objective Optimization: The method successfully balances two complementary goals: improving retrieval quality in the embedding space and maintaining high-quality, interpretable context for downstream language tasks.
Cross-Domain Generalization: The learned profiles demonstrate strong transferability, outperforming baselines in zero-shot sequential recommendation across different domains (e.g., training on Clothing, testing on Books, Electronics, Sports, and Google Reviews).

4. Experimental Results

The framework was evaluated on Amazon Reviews 2023 and Google Local Reviews datasets under zero-shot sequential recommendation settings.

Embedding-Frozen Setting: BLUE consistently outperformed strong baselines (including raw history, base profiles, and other RL-based profile methods like RLAIF and RLPF) when the embedding model was fixed.
- Notable gains were observed in cross-domain transfer (e.g., Amazon Books N@10 improved from 0.228 to 0.291).
- Combining recent history with the BLUE profile yielded the best performance across all datasets.
Embedding-Adaptive Setting: When the embedding model was fine-tuned using the generated profiles, BLUE+SCL achieved superior cross-domain transfer performance compared to standard sequential recommendation models (SASRec, BERT4Rec) and other profile-based contrastive methods.
Personalized Question Answering: The generated profiles provided better context for next-item prediction and rating prediction tasks compared to raw histories or alternative profile methods. BLUE achieved the highest accuracy in next-item prediction across all domains.
Efficiency and Structure:
- Quality-Efficiency Trade-off: BLUE achieved the highest quality scores while maintaining inference times comparable to other profile-based methods, outperforming raw history in quality and other profile methods in efficiency.
- Latent Structure: T-SNE visualizations showed that BLUE-induced embeddings formed tighter local clusters with clearer separation between user groups compared to baselines, indicating a more discriminative and preference-aware latent structure.

5. Significance and Claims

The paper claims that BLUE provides an effective mechanism to unify the interpretability of textual profiling with the discriminative power of latent embeddings.

Complementarity: It argues that textual profiles and latent embeddings should not be viewed as competing representations but as complementary spaces that can be jointly optimized.
Practical Utility: The method is particularly significant for scenarios where embedding models are proprietary or frozen (e.g., off-the-shelf services), as it improves personalization by optimizing the input context (the profile) rather than the model itself.
Generalization: The results highlight that optimizing profiles for downstream utility (via latent rewards) leads to representations that generalize better across domains than those optimized solely for language fluency or raw history compression.
Interpretability: The generated profiles serve as concise, human-readable contexts that improve both retrieval performance and the quality of downstream generative personalization tasks (like QA), offering a "white-box" alternative to black-box latent vectors.

The authors conclude that their approach bridges the gap between the "strong ranking, weak reasoning" nature of embeddings and the "interpretable, hard-to-optimize" nature of textual profiles, offering a robust solution for next-generation personalized systems.