Bringing Reasoning to Generative Recommendation Through the Lens of Cascaded Ranking

TL;DR

Generative Recommendation (GR) often suffers from bias amplification where token-level probabilities over frequent tokens grow as generation proceeds, reducing diversity. The authors identify two core causes—homogeneous reliance on a fixed encoded history and fixed computation per token—and propose CARE, a cascaded reasoning framework with progressive history encoding and query-anchored reasoning to enable heterogeneous information use and richer computation. CARE is trained with a generation loss and a diversity loss, combined as $L = L_{gen} + \alpha L_{div}$, and instantiated on three GR backbones across four real-world datasets, delivering improved accuracy and diversity with only modest efficiency overhead. The results demonstrate that debiased GR through cascaded reasoning can achieve higher-quality, more diverse recommendations suitable for resource-constrained deployments.

Abstract

Generative Recommendation (GR) has become a promising end-to-end approach with high FLOPS utilization for resource-efficient recommendation. Despite the effectiveness, we show that current GR models suffer from a critical \textbf{bias amplification} issue, where token-level bias escalates as token generation progresses, ultimately limiting the recommendation diversity and hurting the user experience. By comparing against the key factor behind the success of traditional multi-stage pipelines, we reveal two limitations in GR that can amplify the bias: homogeneous reliance on the encoded history, and fixed computational budgets that prevent deeper user preference understanding. To combat the bias amplification issue, it is crucial for GR to 1) incorporate more heterogeneous information, and 2) allocate greater computational resources at each token generation step. To this end, we propose CARE, a simple yet effective cascaded reasoning framework for debiased GR. To incorporate heterogeneous information, we introduce a progressive history encoding mechanism, which progressively incorporates increasingly fine-grained history information as the generation process advances. To allocate more computations, we propose a query-anchored reasoning mechanism, which seeks to perform a deeper understanding of historical information through parallel reasoning steps. We instantiate CARE on three GR backbones. Empirical results on four datasets show the superiority of CARE in recommendation accuracy, diversity, efficiency, and promising scalability. The codes and datasets are available at https://github.com/Linxyhaha/CARE.

Figures (15)

• Figure 1: Illustration of (1) bias amplification on the popular token groups; and (2) increased bias amplification from initial to later token generation steps (i.e., 1st to 2nd token). G1-G8 denotes token groups sorted by popularity.
• Figure 2: Paradigm comparisons between traditional recommendation and generative recommendation, where both essentially perform cascaded item ranking.
• Figure 3: Cascaded reasoning framework with two key considerations, i.e., heterogeneous information incorporation and enriched computations allocation.
• Figure 4: Illustration of CARE, including progressive history encoding and query-anchored reasoning.
• Figure 5: Illustration of progressive attention mask, where the query vectors only interact with selected tokens in user history. The example shows one historical item and one query vector for each step.