Bringing Reasoning to Generative Recommendation Through the Lens of Cascaded Ranking

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.

💡 Research Summary

The paper investigates a critical shortcoming of modern generative recommendation (GR) systems: as the model autoregressively generates the token sequence that represents a recommended item, it progressively amplifies popularity bias. Empirical analysis shows that the probability of generating frequent tokens can increase by more than 200 % from the first to the second token, leading to reduced accuracy, limited diversity, and an echo‑chamber effect in the final recommendations.

To understand the root cause, the authors compare GR with traditional multi‑stage recommendation pipelines that employ cascaded ranking. In a multi‑stage pipeline, early stages use lightweight models and coarse‑grained features, while later stages allocate richer features and more computation to capture fine‑grained user preferences. By contrast, a GR model uses the same encoded user history and a fixed single‑forward pass for every token, which results in two key limitations: (1) homogeneous information – the model repeatedly relies on the same history representation, preventing deeper preference modeling, and (2) fixed computation – the model cannot devote additional processing power to later, finer‑grained tokens where bias is most harmful.

The authors propose CARE (Cascaded REasoning), a framework that directly addresses these limitations. CARE introduces two complementary mechanisms:

1. Progressive History Encoding – a stage‑wise attention mask that gradually expands the portion of the user history considered as token generation proceeds. Early tokens attend only to recent interactions, while later tokens incorporate the full interaction history, thereby providing increasingly heterogeneous information aligned with the granularity of the token being generated.

2. Query‑Anchored Reasoning – before generating each token, the model performs multiple parallel forward passes, each conditioned on a query vector derived from the current token prefix. This injects additional computation at every generation step without breaking real‑time constraints, because the passes are executed in parallel. A diversity loss is added to encourage the parallel reasoning paths to produce distinct probability distributions, which helps the model assign higher probability to less‑frequent tokens.

CARE is instantiated on three representative GR backbones (TIGER, LETTER, SETRec) and evaluated on four large‑scale datasets covering short‑form video (Kuaishou, Douyin), movies, and e‑commerce. Results demonstrate consistent improvements: Recall@5 and NDCG increase by 12 %–18 % on average, diversity metrics (ILD, coverage) improve by 15 %–22 %, and the additional reasoning steps add less than 5 % latency thanks to parallel execution. FLOPs‑per‑accuracy efficiency rises by more than 1.4×, and scaling experiments confirm that CARE remains effective as token length grows.

The paper’s contributions are threefold: (i) identification and quantitative analysis of bias amplification in GR, (ii) a novel cascaded reasoning architecture that supplies heterogeneous information and dynamic computation, and (iii) extensive empirical validation showing that CARE simultaneously boosts accuracy, diversity, and efficiency. Limitations include increased memory demand for parallel reasoning and the need for domain‑specific query design; future work will explore memory‑efficient implementations and automated query generation.

Overall, CARE offers a practical solution for deploying generative recommenders in real‑time systems while mitigating the long‑standing popularity‑bias problem, bridging the gap between end‑to‑end generation and the proven benefits of traditional cascaded ranking pipelines.