Kuaishou serves over 400 million daily active users, processing hundreds of millions of search queries daily against a repository of tens of billions of short videos. As the final decision layer, the reranking stage determines user experience by optimizing whole-page utility. While traditional score-and-sort methods fail to capture combinatorial dependencies, Generative Reranking offers a superior paradigm by directly modeling the permutation probability. However, deploying Generative Reranking in such a high-stakes environment faces a fundamental dual dilemma: 1) the structural trade-off where Autoregressive (AR) models offer superior Sequential modeling but suffer from prohibitive latency, versus Non-Autoregressive (NAR) models that enable efficiency but lack dependency capturing; 2) the optimization gap where Supervised Learning faces challenges in directly optimizing whole-page utility, while Reinforcement Learning (RL) struggles with instability in high-throughput data streams. To resolve this, we propose Dual-Rerank, a unified framework designed for industrial reranking that bridges the structural gap via Sequential Knowledge Distillation and addresses the optimization gap using List-wise Decoupled Reranking Optimization (LDRO) for stable online RL. Extensive A/B testing on production traffic demonstrates that Dual-Rerank achieves State-of-the-Art performance, significantly improving User satisfaction and Watch Time while drastically reducing inference latency compared to AR baselines.
Figures & tables
Figure 1. Comparison of Distribution Characteristics between NLP (Qwen2.5) and Industrial Reranking Tasks. (a) Distribution Shape Analysis : The curves plot the probability of the top- k tokens at each decoding step, averaged over 100 samples from the Alpaca dataset (NLP) and one day of production logs (Reranking). Unlike the high-entropy, long-tail distribution observed in NLP (Red), reranking exhibits a sharp Unimodal Concentration (Blue). (b) Quantitative metrics reveal a significant “Entropy-Consistency Gap,” validating the quasi-deterministic nature of ranking. Detailed explanation of the metrics are showed in Appendix C .
Figure 2. Overview of Dual-Rerank : joint online updates of an autoregressive Teacher and a non-autoregressive Student via (i) sequential distillation and (ii) online RL optimization (LDRO) under strict latency.
Figure 3. Teacher–Student PTAR during training. Distillation improves agreement and stability over a pure NAR baseline.
Figure 4. Stability under streaming drift. (a) & (b): Real-world logs reveal significant drifts in different time. (c): Comparison of Total Reward during training.
Figure 5. Overview of the Online Serving Phase. The framework leverages One-Step NAR decoding and Vectorized Gumbel-Max sampling to generate parallel candidates, enabling a high-throughput Best-of-N selection strategy.
Method
Avito (Public)
Kuaishou (Industrial)
AUC
NDCG
AUC
NDCG
Discriminative Methods
DNN ( Covington et al., 2016 )
0.6614
0.6920
0.6866
0.7122
DCN ( Wang et al., 2017 )
0.6623
0.7004
0.6879
0.7116
PRM ( Pei et al., 2019 )
0.6881
0.7380
0.7119
0.7396
Edge-rerank ( Gong et al., 2022 )
0.6953
0.7203
0.7143
0.7401
Table 1. Overall performance comparison on Avito and Kuaishou datasets. The best results are highlighted in bold , and the second-best are underlined . Improvement is calculated relative to the best baseline (GReF).
Figure 6. Structural Fidelity Analysis (Ranking Flip Rate). unlike PTAR which measures pointwise memorization, RFR evaluates the pairwise ranking logic . The divergence of the “Pure NAR” baseline reveals a Ranking Collapse phenomenon (loss of global order), while Dual-Rerank maintains structural consistency.
Model
Training
Inference
AUC
Pure NAR
MLE
Parallel
0.7254
Autoregressive Teachers
AR (Hybrid)
AR
Top-2 AR + Greedy
0.7325
AR (Full)
AR
Full AR Decoding
0.7387
Distilled NAR
Sequential Dist.
Parallel (NAR)
0.7391 ∗
vs. Pure NAR
(+1.89%)
Table 2. Impact of Sequential Modeling Strategies (Phase 1). * indicates parity with the AR Teacher.
Variant
AUC
Drop
Dual-Rerank (Full)
0.7448
-
w/ GRPO (Replace LDRO)
0.7395
-0.71%
w/o Distillation (Phase 1)
0.7254
-2.60%
w/o Double-Decoupling (Phase 2)
0.7402
-0.62%
w/o Rank-Weighting
0.7429
-0.26%
Table 3. Ablation study of Dual-Rerank components.
Generative models offer a promising paradigm for the final stage reranking in multi-stage recommender systems, with the ability to capture inter-item dependencies within reranked lists. However, their practical deployment still faces two key challenges: (1) an inherent conflict between achieving high generation quality and ensuring low-latency inference, making it difficult to balance the two, and (2) insufficient interaction between user and item features in existing methods. To address these challenges, we propose a novel Personalized Semi-Autoregressive with online knowledge Distillation (PSAD) framework for reranking. In this framework, the teacher model adopts a semi-autoregressive generator to balance generation quality and efficiency, while its ranking knowledge is distilled online into a lightweight scoring network during joint training, enabling real-time and efficient inference. Furthermore, we propose a User Profile Network (UPN) that injects user intent and models interest dynamics, enabling deeper interactions between users and items. Extensive experiments conducted on three large-scale public datasets demonstrate that PSAD significantly outperforms state-of-the-art baselines in both ranking performance and inference efficiency.
Kai Cheng, Hao Wang, Wei Guo +4
University of Science and Technology of China, Hefei, China · Huawei, Shanghai, China · Shanghai Jiao Tong University, Shanghai, China +1
In multi-stage recommender systems, reranking optimizes overall utility by capturing intra-list contextual dependencies, yet its central challenge lies in exploring optimal sequences within an exponentially large permutation space. Recent studies have shifted towards end-to-end generative frameworks, which typically leverage list-wise rewards or preference alignment to guide generator training. However, these methods still face two critical issues. First is the heuristic label bias. Existing methods often construct training targets based on simple rules, such as promoting clicked items to the top, while ignoring causal dependencies within the list context. Second is the credit assignment problem. Sparse list-level posterior rewards fail to directly guide intermediate steps in sequence generation, leading to ambiguous optimization directions. To address these issues, we propose DeGRe (Dense-supervised Generative Reranking), a generative reranking framework that bridges the gap between offline exploration and online efficiency through dense supervision. The core of DeGRe lies in its offline-online decoupled design. During the offline phase, we introduce a Lookahead Evaluator based on cumulative regression, which leverages beam search to actively mine high-value lookahead sequences in the unexposed space. During training, we transform the step-wise value estimations from the evaluator into dense supervision signals and distill them into a lightweight Online Generator. This mechanism enables the generator to internalize lookahead planning capabilities, requiring only a single efficient greedy decoding pass during online inference to approximate the global optimum. Experiments demonstrate that DeGRe outperforms baseline models on public benchmarks and industrial datasets. We have successfully deployed DeGRe on Taobao Flash Shopping, significantly improving online recommendations.
Chaotian Song, Jingyao Zhang, Chenghao Chen +6
College of Software, Zhejiang University Hangzhou, China · Rajax Network Technology, Taobao Shangou of Alibaba Hangzhou, China · Rajax Network Technology, Taobao Shangou of Alibaba Beijing, China +1
Recent studies increasingly explore Large Language Models (LLMs) as a new paradigm for recommendation systems due to their scalability and world knowledge. However, existing work has three key limitations: (1) most efforts focus on retrieval and ranking, while the reranking phase, critical for refining final recommendations, is largely overlooked; (2) LLMs are typically used in zero-shot or supervised fine-tuning settings, leaving their reasoning abilities, especially those enhanced through reinforcement learning (RL) and high-quality reasoning data, underexploited; (3) items are commonly represented by non-semantic IDs, creating major scalability challenges in industrial systems with billions of identifiers. To address these gaps, we propose the Generative Reasoning Reranker (GR2), an end-to-end framework with a three-stage training pipeline tailored for reranking. First, a pretrained LLM is mid-trained on semantic IDs encoded from non-semantic IDs via a tokenizer achieving ≥99% uniqueness. Next, a stronger larger-scale LLM generates high-quality reasoning traces through carefully designed prompting and rejection sampling, which are used for supervised fine-tuning to impart foundational reasoning skills. Finally, we apply Decoupled Clip and Dynamic sAmpling Policy Optimization (DAPO), enabling scalable RL supervision with verifiable rewards designed specifically for reranking. Experiments on two real-world datasets demonstrate GR2's effectiveness: it surpasses the state-of-the-art OneRec-Think by 2.4% in Recall@5 and 1.3% in NDCG@5. Ablations confirm that advanced reasoning traces yield substantial gains across metrics. We further find that RL reward design is crucial in reranking: LLMs tend to exploit reward hacking by preserving item order, motivating conditional verifiable rewards to mitigate this behavior and optimize reranking performance.