Semantic Convergence: Harmonizing Recommender Systems via Two-Stage Alignment and Behavioral Semantic Tokenization

TL;DR

The paper tackles the gap between sparse collaborative signals in traditional recommender systems and dense language representations in large language models by introducing a two-stage alignment framework. Alignment Tokenization converts item IDs into a compact, semantically aligned token space via cascaded CodeBooks, while Alignment Task fine-tunes the LLM with sequential, textual, and query-oriented signals plus negative sampling. A dedicated inference strategy caches top-K item codes to reduce latency, enabling scalable, end-to-end LLM-assisted recommendations. Empirical results on three Amazon datasets demonstrate improved recall and NDCG, with ablations confirming the benefits of each component and a scaling trend suggesting larger LLMs further enhance performance. Overall, the approach offers a practical path to integrate LLM reasoning into recommender systems with improved efficiency and scalability.

Abstract

Large language models (LLMs), endowed with exceptional reasoning capabilities, are adept at discerning profound user interests from historical behaviors, thereby presenting a promising avenue for the advancement of recommendation systems. However, a notable discrepancy persists between the sparse collaborative semantics typically found in recommendation systems and the dense token representations within LLMs. In our study, we propose a novel framework that harmoniously merges traditional recommendation models with the prowess of LLMs. We initiate this integration by transforming ItemIDs into sequences that align semantically with the LLMs space, through the proposed Alignment Tokenization module. Additionally, we design a series of specialized supervised learning tasks aimed at aligning collaborative signals with the subtleties of natural language semantics. To ensure practical applicability, we optimize online inference by pre-caching the top-K results for each user, reducing latency and improving effciency. Extensive experimental evidence indicates that our model markedly improves recall metrics and displays remarkable scalability of recommendation systems.

Figures (3)

• Figure 1: Illustration of LLM-Based Recommendation Methods. (a) and (b) illustrates existing methods, where either textual information is input into the LLM to provide features for existing recommendation models ren2024representation in (a), or large-scale item IDs are directly fed into the LLM through instruction fine-tuning, enabling the LLM to directly output recommended item IDs li2023e4srec in (b). (c) represents our method, which can compress large-scale item IDs into a small-scale token representation that is easier for the LLM to process. Furthermore, our method allows the LLM to simultaneously handle both item IDs and textual information, directly outputting the recommended item IDs.
• Figure 2: The pipeline of ours two-stage alignment for recommendation. During the Alignment Tokenization phase, our objective is to obtain four token indices—$u$, $v$, $x$, and $y$ for each item. Subsequently, in the Alignment Task process, we utilize these item indices to introduce negative sampling tasks (highlighted by black boxes in the Alignment Task) alongside various positive prediction tasks. These tasks collectively fine-tune the LLM, integrating both behavioral and semantic signals into the model.
• Figure 3: Inference Stage. We cache valid codes in predictions for each user during the online inference phase.