UniGRec: Unified Generative Recommendation with Soft Identifiers for End-to-End Optimization

TL;DR

UniGRec tackles the misalignment between tokenizer and recommender in generative recommendation by introducing differentiable soft identifiers and unifying optimization under the target recommendation objective. It proposes three mechanisms—Annealed Inference Alignment, Codeword Uniformity Regularization, and Dual Collaborative Distillation—to address training–inference discrepancy, codebook collapse, and deficient collaborative signals, respectively. Empirical results across Beauty, Pet, and Upwork show consistent improvements over state-of-the-art baselines, with ablations confirming the contribution of each component. The approach enables end-to-end, differentiable training and scalable generation while preserving coarse- and fine-grained signals essential for high-quality recommendations, offering practical impact for large-scale item corpora.

Abstract

Generative recommendation has recently emerged as a transformative paradigm that directly generates target items, surpassing traditional cascaded approaches. It typically involves two components: a tokenizer that learns item identifiers and a recommender trained on them. Existing methods often decouple tokenization from recommendation or rely on asynchronous alternating optimization, limiting full end-to-end alignment. To address this, we unify the tokenizer and recommender under the ultimate recommendation objective via differentiable soft item identifiers, enabling joint end-to-end training. However, this introduces three challenges: training-inference discrepancy due to soft-to-hard mismatch, item identifier collapse from codeword usage imbalance, and collaborative signal deficiency due to an overemphasis on fine-grained token-level semantics. To tackle these challenges, we propose UniGRec, a unified generative recommendation framework that addresses them from three perspectives. UniGRec employs Annealed Inference Alignment during tokenization to smoothly bridge soft training and hard inference, a Codeword Uniformity Regularization to prevent identifier collapse and encourage codebook diversity, and a Dual Collaborative Distillation mechanism that distills collaborative priors from a lightweight teacher model to jointly guide both the tokenizer and the recommender. Extensive experiments on real-world datasets demonstrate that UniGRec consistently outperforms state-of-the-art baseline methods. Our codes are available at https://github.com/Jialei-03/UniGRec.

Figures (6)

• Figure 1: Comparison of generative recommendation training paradigms. Staged training separates optimization into two phases with a frozen tokenizer during recommendation training. Alternating training updates the tokenizer and recommender asynchronously without a unified objective. End-to-end joint training enables a unified gradient flow, jointly optimizing both components under a single objective.
• Figure 2: Overview of the proposed UniGRec framework. By introducing soft identifiers, the tokenizer and recommender are seamlessly integrated, enabling end-to-end joint optimization under a unified recommendation objective. Temperature annealing and uniformity regularization are crucial for practical effectiveness, while a pre-trained lightweight teacher model offers additional collaborative supervision.
• Figure 3: Results of the collision rate curves during training under different temperature ($\tau$) schedules.
• Figure 4: Results of the collision rate and recommendation performance of UniGRec across different values of the uniformity regularization weight ($\lambda_\text{CU}$).
• Figure 5: Visualization analysis of codebook embeddings and codeword usage of each quantization layer, where “Entropy” measures the distribution of codeword usage, indicating how evenly the codewords are utilized.