TSPORec: Token Selection via Preference Optimization for LLM-Based Sequential Recommendation
2026-08-10 • Information Retrieval
Information RetrievalArtificial Intelligence
AI summaryⓘ
The authors study how large language models (LLMs) can help recommendation systems by using detailed text descriptions and user behavior. They point out that using full text is expensive to process, so current methods often use only the beginning of the text and lose useful information. To fix this, the authors propose TSPORec, a method that smartly selects the most important words throughout the full text to improve recommendations while saving computing power. Their experiments show that TSPORec improves accuracy by up to 31% and reduces computation time by over 60% compared to other methods.
Large Language ModelsSequential RecommendationToken SelectionUser PreferencesInference CostTextual InformationComputational EfficiencyProxy RewardRecommendation PerformanceDeep Learning
Authors
Wenqiao Zhu, Chao Xu, Haipang Wu, Ji Liu
Abstract
Large Language Models (LLMs) have emerged as powerful tools for improving recommendation systems. The effectiveness of LLMs arises from their ability to harness rich textual information and their capacity to model heterogeneous user preferences based on users' interaction history. However, due to the large-scale and deep architectures, LLM-based sequential recommendation approaches generally incur high inference costs, resulting in a low return on investment. To mitigate this cost, many existing approaches resort to using only the first few tokens of item descriptions, which inadvertently discards valuable information contained in the full text, thereby leading to suboptimal recommendation performance. To address this limitation, we propose a novel Token Selection approach for Preference Optimization in LLM-based sequential Recommendation, i.e., TSPORec, which accurately pinpoints informative tokens throughout the entire textual content to improve recommendation performance. Specifically, we design a three-stage pipeline to select informative tokens and introduce a novel proxy reward to facilitate the implementation. TSPORec not only enhances recommendation performance but also improves computational efficiency. Extensive experiments across two models and datasets demonstrate the superb performance (up to 31.25%) and efficiency (up to 63.4%) of our approach compared with six baseline approaches. Code is available at https://github.com/WNQzhu/TSPORec.git.