💬 NLP

Profiling What Matters: Context-Aware Item Profiles from Large-Scale Metadata for LLM Recommenders

This paper proposes CAIRO, a user context-aware item profiling framework that structures heterogeneous metadata and dynamically selects relevant information for each user-item pair to enhance LLM-based reranking performance.

Original authors: Dojun Hwang, Seunghan Lee, Cheonyoung Park, Sara Yu, SeongKu Kang

Published 2026-08-24

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Dojun Hwang, Seunghan Lee, Cheonyoung Park, Sara Yu, SeongKu Kang

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

In the digital age, recommendation systems act as the quiet curators of our online lives, suggesting the next video to watch, the book to read, or the product to buy. For years, these systems relied on simple tracking of what users clicked or purchased, treating items as anonymous codes without much depth. Recently, a new generation of artificial intelligence, known as large language models, has begun to change this landscape. These models possess a vast internal library of human knowledge and the ability to reason, allowing them to understand the nuance of a user's history and the details of an item in ways traditional systems could not. However, a significant hurdle remains: while these models are brilliant at understanding people, they struggle to make sense of the overwhelming, messy, and often unstructured information that describes the items themselves. Real-world products come with hundreds of attributes, from technical specifications to vague descriptions, and much of this data is buried in long paragraphs or scattered across different formats. If an artificial intelligence is fed too much of this raw, unorganized data all at once, it becomes confused, much like a person trying to read a novel while simultaneously listening to a radio broadcast; the signal gets lost in the noise.

Researchers at Korea University and KT Corporation have developed a new method called CAIRO to solve this specific problem. Their work focuses on the "item side" of the recommendation equation, aiming to teach artificial intelligence how to read a product's description and extract only the details that matter for a specific person at a specific moment. Instead of dumping the entire catalog of an item's features into the system, CAIRO acts as a skilled editor. It first organizes the chaotic raw data into two distinct categories: objective facts, such as battery size or material, and subjective traits, which are the feelings or experiences users describe in their reviews, like "seamless integration" or "vivid colors." The system then uses a lightweight, efficient process to decide which of these facts and feelings are relevant to the individual user currently browsing. For one person, the most important detail about a pair of headphones might be the noise-canceling technology; for another, it might be the brand's reputation for durability. CAIRO identifies these differences and constructs a concise, custom profile for the item that highlights only the evidence the user cares about.

The researchers tested this approach on massive datasets containing millions of interactions across video games, sports equipment, and electronics. They found that when the artificial intelligence received these carefully curated, user-specific profiles, its ability to rank items correctly improved significantly. In fact, the system consistently outperformed other methods that either ignored item details entirely or tried to feed the model all available information at once. The study explicitly ruled out the idea that simply adding more data helps; in their experiments, methods that injected raw, unstructured metadata actually performed worse than those using only basic titles. The success of CAIRO lies in its selectivity. It does not just summarize an item once and use that same summary for everyone; instead, it adapts the item's story to fit the listener. By combining structured facts with insights drawn from user reviews, and then filtering them through the lens of the current user's preferences, the system provides the artificial intelligence with a clear, focused narrative rather than a cluttered data dump.

To ensure this process was not just a theoretical exercise, the team compared their method against several existing approaches, including those that used complex, multi-step retrieval systems to find information. They discovered that while those complex systems could find relevant details, they were too slow for real-world use, taking hundreds of seconds to prepare a single list of suggestions. In contrast, CAIRO performed its selection in a fraction of a second, making it practical for live applications. The researchers also tested whether their method worked with different types of artificial intelligence models, ranging from smaller, less powerful versions to larger, more advanced ones. The results showed that the custom profiles improved performance across the board, suggesting that the value comes from the quality of the information presented, not just the size of the model reading it. Even for the smallest models tested, the structured profiles helped them make better decisions, proving that a well-organized presentation of facts can compensate for limited computing power.

One of the most compelling aspects of this work is how it handles the human element of shopping. The system recognizes that the same product can mean different things to different people. In a specific example involving a pair of wireless earbuds, the system generated two completely different profiles for the same item based on the user's history. For a user who prioritized technical performance and often criticized poor functionality, the system highlighted the earbuds' active noise cancellation and driver specifications. For a user who valued brand loyalty and ecosystem compatibility, the system instead focused on the product's seamless integration with other devices and the brand's reputation. When the artificial intelligence was given these tailored profiles, it ranked the correct item much higher for each user. Without this adaptation, the system struggled to distinguish which details were important, often burying the relevant signals under a mountain of irrelevant information.

The researchers also introduced a refinement step to ensure the system did not invent facts or misinterpret user intent. Since the subjective traits are inferred from reviews rather than taken directly from a spec sheet, there is a small risk the system could misread the tone of a review. To counter this, the system runs a diagnostic check where it simulates a recommendation scenario and identifies cases where its initial profile led to a wrong prediction. It then revises the profile, grounding the changes in the hard facts of the product's specifications to ensure accuracy. This process acts as a safety net, ensuring that the final profile remains faithful to the actual item while still capturing the nuances of user preference. The study concludes that the future of recommendation lies not in feeding artificial intelligence more data, but in teaching it how to find the right data. By structuring vast amounts of metadata into clear, context-aware profiles, CAIRO demonstrates that the key to better recommendations is not just knowing everything about an item, but knowing what matters to the person looking at it.

Technical Summary: CAIRO – Context-Aware Item Profiles for LLM Recommenders

Problem Definition
While Large Language Models (LLMs) have advanced the reranking stage of recommendation systems by leveraging world knowledge and reasoning capabilities, effectively utilizing item-side information remains a significant bottleneck. Real-world items are described by vast, heterogeneous, and unstructured metadata (ranging from tens to hundreds of attributes) and user reviews. The paper identifies three specific challenges in leveraging this data for LLM-based reranking:

Context Window Overload: The sheer volume and unstructured nature of item metadata (e.g., bullet lists, free-form text) overwhelm LLM context windows, leading to suboptimal processing if all data is injected.
Item-Dependent Feature Salience: The importance of specific features varies significantly across items within the same category (e.g., refresh rate is critical for gaming monitors but generic for standard ones). Static attribute sets obscure discriminative features.
User-Dependent Feature Salience: Even for a single item, different users prioritize different features based on their preferences and usage contexts (e.g., brand loyalty vs. functional performance). Existing methods often generate static item profiles that fail to capture this user-dependent variation.

Current approaches typically rely on item titles, fixed attribute subsets, or static summaries, which fail to scale to vast metadata or adapt to specific user contexts.

Methodology: The CAIRO Framework
The authors propose CAIRO (Context-Aware Item pROfiling), a framework designed to transform raw, large-scale metadata into concise, user-specific item profiles for LLM reranking. The framework operates in four main stages:

Structuring Item Information:
- Domain-Specific Key Exploration: Instead of using fixed keys, CAIRO employs a data-driven approach. It clusters item metadata embeddings and uses an LLM to extract representative keys for each cluster, which are then merged into a unified set of domain-specific keys ( $F$ ).
- Objective Feature Extraction: Using the unified keys, an LLM extracts factual, static attributes (e.g., weight, price) from raw metadata into a structured dictionary ( $O_i$ ). The process includes a verification step to prevent hallucinations, assigning null if evidence is missing.
- Subjective Trait Extraction: The framework extracts multi-faceted subjective traits ( $S_i$ ) from user reviews. Unlike single summaries, these traits capture diverse perspectives (e.g., "Character-driven Engagement" vs. "Long-term Loyalty"). Each trait is grounded in supporting objective features to ensure faithfulness.
User Context-Aware Item Profiler:
To avoid high latency from online LLM calls, CAIRO uses a lightweight, offline-trained profiler to select relevant information for each user-item pair.
- Objective Feature Selector: A controller network ( $g_w$ ) learns to quantify the importance of objective features based on collaborative signals (user-item interaction patterns). It assigns weights to features and selects the top- $k$ most predictive features for a specific user-item pair.
- Subjective Trait Selector: This submodule selects the single subjective trait most semantically aligned with the user's profile (constructed from their review history) using cosine similarity.
- Online Prompt Construction: The selected objective features and the aligned subjective trait (along with its supporting features) are assembled into a compact, user-specific item profile ( $P_{i|u}$ ) for the reranking prompt.
Feature-Grounded Trait Refinement (Optional):
To further improve quality, an offline refinement stage diagnoses weaknesses in subjective traits. It uses a proxy task (next-item prediction) to identify cases where a trait fails to support accurate predictions. The LLM then revises the trait based on these error diagnoses, grounded in the item's objective features to prevent hallucination.

Key Contributions

Problem Identification: The paper highlights the underexplored challenge of effectively leveraging vast, heterogeneous item metadata while capturing both item-specific and user-dependent feature salience in LLM reranking.
Framework Proposal: CAIRO introduces a novel pipeline that structures raw metadata into objective features and subjective traits, then employs lightweight, collaborative-signal-driven selectors to construct user-adaptive profiles.
Empirical Validation: Extensive experiments demonstrate that CAIRO consistently improves LLM-based reranking performance across diverse domains (Video Games, Sports, Electronics) compared to state-of-the-art baselines.

Experimental Results
Experiments were conducted on Amazon datasets using GPT-4o-mini for reranking.

Performance: CAIRO consistently outperformed baselines, including traditional models (BPR, SASRec), feature-based models (xDeepFM, AdaFS), and existing LLM reranking methods (LLMRank, EXP3RT, M-LLM3Rec). In the Video Games dataset, CAIRO improved nDCG@5 by 9.86% over LLMRank.
Feature Injection vs. Selection: Simply injecting raw features into existing LLM methods (denoted as "+Feat") often degraded performance, confirming that unstructured data overwhelms the model. CAIRO's selective approach was crucial for success.
Efficiency: Compared to RAG-based profiling methods (e.g., REAP), CAIRO achieved competitive or superior performance with significantly lower latency (0.25s–0.32s per user vs. hundreds of seconds for RAG) by performing selection via lightweight matrix operations rather than iterative LLM calls.
Ablation Studies: Removing the collaborative feature selection or using a single static trait per item resulted in performance drops, validating the necessity of user-adaptive selection and multi-faceted trait representation.
Transferability: The generated profiles improved performance across various LLM families (Qwen, Llama) and parameter scales (3B to 32B), indicating the profiles are not overfitted to a specific model architecture.

Significance and Claims
The paper claims that the organization and user-adaptive selection of item information are as critical as the information itself. By moving beyond static summaries and fixed attributes, CAIRO enables LLMs to access "what matters" for a specific user in a specific context without incurring prohibitive computational costs. The authors position this work as laying the groundwork for item-side profiling in recommendation, shifting focus from the prevailing user-side profiling to a more balanced, context-aware approach that fully exploits the potential of vast item metadata.

Technical Summary: CAIRO – Context-Aware Item Profiles for LLM Recommenders

More like this