VaLiDRec: Variable-Length LLM-Aligned Semantic IDs for Generative Recommendation

2026-07-27

AI Summary

This work addresses critical limitations of traditional generative recommender systems, which rely on fixed-length semantic IDs and suffer from excessive semantic compression, incompatibility with large language model (LLM) vocabularies, and inefficient autoregressive decoding. To overcome these issues, the authors propose VaLiDRec, a novel framework that introduces variable-length semantic IDs natively aligned with the LLM’s vocabulary and models user preferences through graph-aware soft prompts and non-autoregressive token set prediction. The approach incorporates importance estimation, semantic quality pruning, and conflict-aware optimization to achieve both efficiency and accuracy. Extensive experiments demonstrate that VaLiDRec significantly outperforms state-of-the-art baselines across four real-world datasets, exhibits strong zero-shot cold-start performance, and achieves an 87.49× speedup in inference over LC-Rec.

Abstract

Generative recommendation commonly represents items using fixed-length semantic identifiers (SIDs) constructed through clustering and quantization. However, these artificial codes may overcompress item semantics, remain misaligned with pretrained LLM vocabularies, and require costly autoregressive decoding. In light of this, we propose VaLiDRec, a generative recommendation framework based on variable-length, LLM-aligned semantic identifiers. VaLiDRec constructs SIDs directly from informative native LLM vocabulary tokens via token importance estimation, semantic-quality-aware pruning, and collision-aware refinement, allowing identifier lengths to adapt to item semantic complexity. To model user preferences, VaLiDRec incorporates graph-aware soft prompts and reformulates recommendation as token-set prediction with token-level item scoring, eliminating autoregressive SID generation and beam search. Experiments on four real-world datasets show that VaLiDRec consistently outperforms strong sequential and generative recommendation baselines across all evaluation metrics. It further achieves superior zero-shot item cold-start performance and 87.49$\times$ faster inference than LC-Rec. These results demonstrate that LLM-native variable-length semantic identifiers provide a more expressive and efficient paradigm for generative recommendation.

Problem

• generative recommendation
• semantic identifiers
• LLM alignment
• variable-length representation
• autoregressive decoding

Innovation

• variable-length semantic identifiers
• LLM-aligned tokens
• generative recommendation
• non-autoregressive decoding
• token-set prediction