Every Preference Has Its Strength: Injecting Ordinal Semantics into LLM-Based Recommenders

2026-05-11

AI Summary

Existing large language model (LLM)-based recommender systems often simplify explicit ratings into implicit or purely positive feedback when incorporating collaborative filtering signals, thereby neglecting the ordinal preference structure inherent in rating data and struggling to capture fine-grained preference semantics. To address this limitation, this work proposes the Ordinal Semantic Anchoring (OSA) framework, which explicitly models the ordinal semantics of preference strength within LLM-based recommendation for the first time. OSA encodes the ordinal levels of user–item interactions as numerical text tokens and leverages their embeddings as semantic anchors to align user and item representations in the LLM’s latent space, enabling preference intensity-aware fusion of collaborative signals. Experiments on multiple real-world datasets demonstrate that OSA significantly outperforms state-of-the-art methods, particularly excelling in pairwise preference tasks, thereby validating its effectiveness in modeling fine-grained user preferences.

Abstract

Recent work has shown that large language models (LLMs) can enhance recommender systems by integrating collaborative filtering (CF) signals through hybrid prompting. However, most existing CF-LLM frameworks collapse explicit ratings into implicit or positive-only feedback, discarding the ordinal structure that conveys fine-grained preference strength. As a result, these models struggle to exploit graded semantics and nuanced preference distinctions. We propose Ordinal Semantic Anchoring (OSA), a hybrid CF-LLM framework that explicitly incorporates preference strength by modeling interaction-level user feedback. OSA represents ordinal preference levels as numeric textual tokens and uses their token embeddings as semantic anchors to align user-item interaction representations in the LLM latent space. Through strength-aware alignment across ordinal levels, OSA preserves preference semantics when integrating collaborative signals with LLMs. Experiments on multiple real-world datasets demonstrate that OSA consistently outperforms existing baselines, particularly in pairwise preference evaluation, highlighting its effectiveness in modeling fine-grained user preferences over prior CF-LLM methods.

Problem

• ordinal preference
• preference strength
• large language models
• recommender systems
• collaborative filtering

Innovation

• Ordinal Semantic Anchoring
• Preference Strength
• LLM-based Recommender
• Collaborative Filtering
• Semantic Alignment