Unleashing the Native Recommendation Potential: LLM-Based Generative Recommendation via Structured Term Identifiers

TL;DR

The paper tackles the challenge of grounding generative recommendations in open-world item catalogs by introducing Term IDs (TIDs), a semantically rich, native-vocabulary item representation. GRLM combines Context-aware Term Generation (CTG) with neighborhood guidance, Integrative Instruction Fine-tuning (IIFT) that jointly optimizes Generative Term Internalization and user sequence prediction, and Elastic Identifier Grounding (EIG) to map generated TIDs to real items. Empirical results show GRLM achieves state-of-the-art performance in both in-domain and cross-domain settings, while maintaining near-perfect grounding and reducing hallucinations, with gains amplified by larger LLM backbones and the potential for semantic compression. The approach demonstrates strong cross-domain transferability and scalability, suggesting TID-based generation as a generalizable path for high-performance, open-world recommender systems, grounded in the LLM’s native semantic space.

Abstract

Leveraging the vast open-world knowledge and understanding capabilities of Large Language Models (LLMs) to develop general-purpose, semantically-aware recommender systems has emerged as a pivotal research direction in generative recommendation. However, existing methods face bottlenecks in constructing item identifiers. Text-based methods introduce LLMs' vast output space, leading to hallucination, while methods based on Semantic IDs (SIDs) encounter a semantic gap between SIDs and LLMs' native vocabulary, requiring costly vocabulary expansion and alignment training. To address this, this paper introduces Term IDs (TIDs), defined as a set of semantically rich and standardized textual keywords, to serve as robust item identifiers. We propose GRLM, a novel framework centered on TIDs, employs Context-aware Term Generation to convert item's metadata into standardized TIDs and utilizes Integrative Instruction Fine-tuning to collaboratively optimize term internalization and sequential recommendation. Additionally, Elastic Identifier Grounding is designed for robust item mapping. Extensive experiments on real-world datasets demonstrate that GRLM significantly outperforms baselines across multiple scenarios, pointing a promising direction for generalizable and high-performance generative recommendation systems.

Figures (6)

• Figure 1: Comparison of Item Identifiers: Term IDs (TIDs) leverage standardized and structured text tokens to ensure precise semantic extraction and low hallucination, while maintaining native compatibility with LLMs vocabularies without the complex indexing pipelines or architectural modifications required by Semantic IDs (SIDs).
• Figure 2: Overall framework of GRLM.
• Figure 3: Context-aware Term Generation effectively ensures that Term IDs across items are globally consistent and locally discriminative.
• Figure 4: Performance scaling of GRLM with respect to different backbone model sizes (from 0.6B to 14B) on three in-domain datasets.
• Figure 5: Prompt for Context-aware Term Generation.