Whole-page ranking jointly optimizes advertisements and organic content to balance business outcomes and user engagement under complex layout and business constraints. Existing non-autoregressive (NAR) approaches enable efficient page generation but typically score candidate items independently of previously selected items, limiting their ability to capture order-dependent interactions. Autoregressive (AR) methods address this limitation through sequential conditioning, but incur substantial inference overhead that challenges large-scale deployment. Bridging this quality-efficiency gap remains a challenge in industrial whole-page ranking. We propose Multi-Item Scoring (MIS), a prefix-aware NAR-style framework that incorporates sequential context into efficient multi-item scoring. MIS integrates global candidate interactions, localized prefix conditioning, and position-aware scoring within a shared representation architecture. By computing reusable representations once per request and performing lightweight prefix-conditioned scoring during constrained beam search, MIS captures important sequential dependencies without repeated full-model inference. Experiments on KuaiRand-1K and RL4RS demonstrate that MIS approaches the generation quality of an autoregressive CGA baseline while achieving 4.07x and 3.54x faster CPU decoding, respectively. In a randomized evaluation on a large-scale social-video platform, MIS improves the realization metric by 0.104% and watch time by 0.058%, while reducing ranking-service CPU consumption by 9.73%. These results demonstrate that prefix-aware multi-item scoring can narrow the quality gap between NAR and AR generation while retaining the efficiency required for industrial scalability.
Figures & tables
Figure 1. One-pass MIS and constrained generation. (a) Item and slot encoders supply global and position context to prefix-conditioned task heads; optional RQ-VAE descriptors preserve separate ordered lags. PA- H has H+2 input axes; the cubes depict the optional PA-1 domain. (b) Task values guide beam search over active prefixes, with feasibility masks and top- B selection. Only the B×C live score rows are materialized. Encoder repetition labels denote stack depth; numerical beam values are illustrative. Shared attention-based item and position encoders feed prefix-conditioned task heads. The decoder combines task values and performs constrained beam search using only active-prefix score rows.
KuaiRand-1K
RL4RS
Efficiency (K/R)
Model
AUC
NE
Lift
AUC
NE
Lift
FLOPs (M)
Latency (ms)
DNN †
0.5597
0.9858
–
0.6169
0.9324
–
–
–
DCN ( Wang et al., 2017 )
0.6042
0.9757
–
0.7696
0.8091
–
–
–
Seq2Slate † ( Bello et al., 2019 )
0.6198
0.9673
16.325%
0.7865
0.7979
24.327%
14.86/14.37
1.00/0.91
NAR ( Zheng et al., 2025 )
0.6038
0.9760
15.847%
0.7698
0.8081
23.697%
13.10/12.08
0.87/0.63
CGA–AR ( Zhu et al., 2024 )
0.6424
0.9553
18.279%
0.8115
0.7503
33.276%
650.50/476.79
27.54/20.28
Table 1. Five-seed prediction (12 epochs) and generation. AUC/NE use held-out labels; a disjoint scorer evaluates beam-32 lift relative to random order. Bold marks the best quality point estimate; underlining marks the runner-up. † Separately fitted tower; Seq2Slate also fits its own evaluator. Compute costs are full decoding by one seeded model at B=32 , L=C : profiled FLOPs (millions) and mean CPU latency (ms), as KuaiRand/RL4RS pairs. Dashes denote prediction-only baselines. Appendix C details architectures and aggregation. KuaiRand PA- N uses Identity history and is reproduced verbatim in Table 2 ; the PA- N Identity reference in Appendix Table 5 also uses these figures.
Figure 2. RQ-VAE versus Identity history on KuaiRand-1K across prefix lengths. The fixed length-sweep scorer uses exact historical vectors (Identity) or RQ-VAE reconstructions. Panels report AUC, NE, and beam-32 generation lift relative to random order. States coincide at H=1 ; N=9 denotes the full prefix. This sweep is separate from the controlled full-prefix study in Table 2 and the candidate-query results in Table 1 .
History
AUC
NE
Lift (%)
MiB
Identity
0.6423
0.9554
18.284
187.81
RQ-VAE
0.6418
0.9557
18.120
2.207
Table 2. Matched KuaiRand PA- N compression. The Identity baseline (AUC 0.6423, NE 0.9554, lift 18.284%) exactly matches the KuaiRand MIS–PA- N row of Table 1 ; only historical representations change. Catalog payload assumes FP32 vectors and packed 4-bit assignments, excluding networks and working states. Storage scales as O(Ud) for Identity and O(Kprimd+UR) for RQ-VAE.
(a) Next-day prediction
Task
Pointwise
NAR
MIS-PA1
MIS-PA4
Impression
0.7641
0.7623 ( −0.23% )
0.7601 ( −0.52% )
0.7597 ( −0.58% )
Ad click
0.7655
0.7648 ( −0.09% )
0.7644 ( −0.14% )
0.7642 ( −0.17% )
Organic share
0.4542
0.4512 ( −0.66% )
0.4504 ( −0.84% )
0.4503 ( −0.86% )
Organic 15-second view
0.7452
0.7439 ( −0.17% )
0.7438 ( −0.19% )
0.7436 ( −0.21% )
Organic valuable view
0.9486
0.9485 ( −0.01% )
0.9484 ( −0.02% )
0.9483 ( −0.03% )
Table 3. Industrial A–O results. MIS-PA4 is the candidate used for page generation and the scalability experiment; MIS-PA1 is the first-order variant. (a) All five displayed next-day prediction tasks use normalized entropy (NE). Organic 15-second view and valuable view correspond to VPV and VPVD. Changes are relative to pointwise. (b) Beam-64 page-value lift (%) over logged order under identical constraints. Bold values mark the lowest prediction error or highest lift, including ties.
Metric
Result
Realization metric
+0.104 [ +0.054 , +0.153 ]
Overall video watch time
+0.058 [ +0.001 , +0.114 ]
Linked sessions
−0.002 [ −0.054 , +0.055 ]
Serving metrics: relative change (%)
Ranking-service CPU (mean)
−9.73 [ −10.06 , −9.46 ]
Slate-evaluation CPU
−9.73 [ −10.13 , −9.40 ]
Table 4. Industrial scale evaluation results show relative changes (%) versus an industrial ranker; brackets give 95% confidence intervals. Outcome and seven-day serving metrics use their specified populations (Appendix C ). Negative service deltas indicate lower CPU consumption. These are service, not isolated model-inference, metrics.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Model
Scoring architecture
Parameters
DNN
32→1→1 MLP
35/35
DCN
3 cross layers + MLP
8,097/8,065
NAR
Item–slot inner product
320/288
MIS–PA-1
96→32→1 residual
4,513/4,481
CGA
2 causal layers + scalar head
18,497/18,465
MIS–PA- N
2 causal layers + attention
29,857/29,825
Appendix
Table 6. Public scoring implementations. Counts exclude frozen catalog embeddings and tower fitting. Counts are KuaiRand/RL4RS, whose slot embedding sizes differ. All layers use width 32 unless stated otherwise.
A recommendation page is consumed as a whole: the items on it compete for attention and complement one another, so the best page is not simply the individually best items placed side by side. We study whole-page optimization, which selects and orders the items on a page jointly to maximize the score of the page. The number of possible pages grows combinatorially, so good pages must be generated rather than enumerated. Some generative re-ranking methods use the same page-level training signal for every choice in a sampled page. Others imitate target pages built before training, which remain largely fixed as the generator improves. We address both limitations with one local page-level objective, which repeatedly moves the current generator toward pages with higher score. This objective splits exactly into one training target per position, and each target compares candidates at the same partial page by the scores of the complete pages they lead to. Every position therefore learns to choose the item that is better for the whole page, rather than the item that looks best on its own. We call this procedure whole-page policy improvement. On two public datasets, KuaiRand-1K and RL4RS-SeqSlate, which cover multi-objective item selection and constrained ordering, and on an industrial dataset of logged pages, our method outperforms the evaluated generative re-ranking baselines. On a synthetic benchmark with a known optimal page, it also comes closest to the optimum, showing that its improvements come from better whole-page decisions.
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.
Chao Zhang, Shuai Lin, ChengLei Dai +5
Kuaishou Technology, Hangzhou, China · Kuaishou Technology, Beijing, China · Unaffiliated, Beijing, China
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