Toward User Preference Alignment in LLM Recommendation via Explicit Context Feedback

Technical Summary: Towards User Preference Alignment in LLM Recommendation via Explicit Context Feedback

Problem Statement
Traditional recommender systems (RecSys) predominantly rely on implicit signals (clicks, views, purchases) to infer user preferences. This approach often neglects rich explicit contextual feedback provided by users through verbal text, such as reviews, comments, and critiques. Consequently, conventional systems fail to capture the nuanced "semantic context" and underlying reasons behind user decisions. This oversight leads to misaligned preferences, reinforces "filter bubbles" by over-personalizing based on narrow behavioral patterns, and limits the explainability of recommendations. While recent Large Language Models (LLMs) offer the capacity to understand natural language, current LLM-based recommenders largely underutilize user-generated text, focusing instead on item metadata or interaction logs.

Methodology and Framework
The paper advocates for a paradigm shift toward prioritizing explicit context feedback in the next generation of LLM-based RecSys. It proposes a comprehensive framework (illustrated in Figure 1) that integrates explicit user signals through four key stages:

1. Taxonomy of Explicit Feedback: The authors categorize explicit feedback into four distinct types to guide modeling:

   • Positive Preferences: Direct expressions of favor (e.g., "loved the cleanliness") which serve as primary supervisory signals for aspect-level preferences.
   • Negative Constraints: Explicit dislikes or boundaries (e.g., "too sweet," "no jump scares") that define what to avoid, acting as protective alignment layers.
   • User-Specific Attributes: Self-declared identity or situational context (e.g., "vegetarian," "professional photographer") that grounds recommendations in stable user traits.
   • Item-Level Comments: Holistic assessments combining pros and cons, offering structured signals about trade-offs and collective sentiment.

2. LLM-Driven Framework: The proposed architecture utilizes LLMs as comprehension engines throughout the recommendation pipeline:

   • Profile Construction: LLMs act as agentic memory managers to extract rationales from text and update user profiles dynamically. This includes translating unstructured text into structured heterogeneous knowledge graphs where edges represent evolving user-aspect relations.
   • Context-Aware Retrieval: Moving beyond simple semantic similarity, retrieval leverages explicit profiles (likes, dislikes, real-time signals) to perform rationale reasoning. LLMs encode reasoning-infused embeddings to align candidate items with user intent.
   • Reranking and Rule Adaptation: The framework integrates explicit feedback into scoring via personalized contextual rules. Hard constraints (e.g., allergies) deterministically filter items, while soft preferences adjust scores. LLMs synthesize user rationales and item features to generate final, explainable scores.
   • Asynchronous Taxonomy Alignment: To address latency constraints, the paper proposes an asynchronous backend mechanism where LLMs map user feedback to an evolving taxonomy of semantic tags (e.g., #LongBatteryLife). These tags, associated with affinity scores, inform real-time retrieval and ranking without compromising system speed.

Key Contributions

• Paradigm Advocacy: The paper positions explicit context feedback as a critical, underutilized resource for achieving fine-grained user preference alignment, arguing that it is essential for moving beyond behavioral patterns to understanding user rationale.
• New Evaluation Directions: The authors propose rethinking benchmarks and metrics. They suggest moving beyond traditional accuracy metrics (RMSE, NDCG) to include:
  • Preference Alignment Metrics: Such as "dislike avoidance rate" and "preference fulfillment score" to measure how well recommendations adhere to stated likes/dislikes.
  • Dynamic Adaptability: Evaluating diversity, novelty, and the system's ability to adapt to evolving user interests revealed through text.
• Systematic Framework: The paper provides a structured roadmap for integrating explicit signals into scalable LLM-driven RecSys, detailing specific mechanisms for profile updating, retrieval, and ranking.

Results and Claims
As a visionary and review paper, this work does not present new experimental results or empirical benchmarks. Instead, it synthesizes existing literature and identifies critical gaps in current industrial and academic practices. The paper claims that:

• Current LLM-based recommenders largely fail to integrate the "wealth of user-authored context," relying instead on implicit logs.
• Explicit feedback enables systems to distinguish between general category preferences and specific attribute preferences (e.g., disliking "jump scares" within the horror genre), thereby mitigating filter bubbles.
• Integrating these signals transforms recommendations from opaque, behavior-based predictions into transparent, explainable, and value-aligned interactions.

Significance
The paper argues that aligning recommendations with the "textual voice" of the user is the key to the next generation of RecSys. By leveraging LLMs to interpret the "why" behind user actions, systems can achieve higher fidelity in preference modeling. This shift promises to foster more personalized, transparent, and trustworthy online platforms, transforming the recommendation process from a passive inference task into an active collaboration between human intent and AI systems. The authors call for the community to prioritize this domain, develop new benchmarks, and build frameworks that treat explicit feedback as a primary data modality.