Bi-Level Optimization for Generative Recommendation: Bridging Tokenization and Generation

TL;DR

The paper addresses generative recommendation by modeling the tokenization of items and autoregressive item generation as a coupled problem. It introduces BLOGER, a bi-level optimization framework that uses a residual-quantization tokenizer (RQ-VAE) and a Transformer-based recommender, coupled through differentiable mixed representations and optimized with meta-learning (MAML) and gradient surgery (PCGrad). The outer objective optimizes the tokenizer with both tokenization loss and downstream recommendation loss, while the inner objective trains the recommender; gradient flow between the two is carefully managed to avoid conflicts, leading to improved codebook utilization and recommendation accuracy. Experiments on real-world Amazon datasets show BLOGER outperforms state-of-the-art generative methods with only modest overhead, demonstrating effective alignment between tokenization and generation and offering practical gains for large-scale generative recommendation systems.

Abstract

Generative recommendation is emerging as a transformative paradigm by directly generating recommended items, rather than relying on matching. Building such a system typically involves two key components: (1) optimizing the tokenizer to derive suitable item identifiers, and (2) training the recommender based on those identifiers. Existing approaches often treat these components separately--either sequentially or in alternation--overlooking their interdependence. This separation can lead to misalignment: the tokenizer is trained without direct guidance from the recommendation objective, potentially yielding suboptimal identifiers that degrade recommendation performance. To address this, we propose BLOGER, a Bi-Level Optimization for GEnerative Recommendation framework, which explicitly models the interdependence between the tokenizer and the recommender in a unified optimization process. The lower level trains the recommender using tokenized sequences, while the upper level optimizes the tokenizer based on both the tokenization loss and recommendation loss. We adopt a meta-learning approach to solve this bi-level optimization efficiently, and introduce gradient surgery to mitigate gradient conflicts in the upper-level updates, thereby ensuring that item identifiers are both informative and recommendation-aligned. Extensive experiments on real-world datasets demonstrate that BLOGER consistently outperforms state-of-the-art generative recommendation methods while maintaining practical efficiency with no significant additional computational overhead, effectively bridging the gap between item tokenization and autoregressive generation.

Figures (4)

• Figure 1: Comparison of optimization strategies in generative recommendation. Snowflake indicates frozen parameters; flame indicates modules being optimized. “Token.” and “Rec.” refer to the tokenizer and recommender, respectively.
• Figure 2: An overview of the proposed BLOGER framework, which comprises an encoder-decoder model architecture and a tailored learning scheme. Specifically, we construct a mixed representation based on the assignment probabilities over the codebook to ensure the differentiability of the recommendation loss to the tokenizer. Besides, we adopt a meta-learning strategy to enable efficient optimization, and apply gradient surgery to extract beneficial components from the gradients of the recommendation loss.
• Figure 3: Results of the performance of BLOGER across different values of hyper-parameters. $\lambda$ controls the balance between the recommendation loss and tokenization loss in the outer level, and $M^\prime$ representing the update frequency of the tokenizer parameters $\phi$.
• Figure 4: Codebook utilization comparison between TIGER and BLOGER, where “Codebook $i$” denotes the $i$-th codebook in the residual quantization process. “Density” is defined as the ratio of activated codewords to the total number of codewords, and “Entropy” measures the distribution of codeword usage, indicating how evenly the codewords are utilized.