Pith. sign in

REVIEW 4 major objections 5 minor 36 references

This paper claims that a single autoregressive generative model can both retrieve and pre-rank e-commerce search results if item business value is encoded into the semantic IDs themselves and into a jointly trained ranking module.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

TSGR makes Taobao search retrieval value-aware by ordering item IDs by clicks and query intent and re-ranking generated candidates with a value module, improving HR@1000 and GMV.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection Plausible industrial GR paper with real deployment evidence, but the offline HR gain is confounded with popularity re-ranking and the paper contains an internal contradiction about whether QP-SID's parallel paths are actually used. the 4 major comments →

arxiv 2607.18796 v2 pith:VFN75JSS submitted 2026-07-21 cs.IR

TSGR: Taobao Search Generative Retrieval

Tianyu Zhan , Gui Ling , Tong Xiong , Kunhai Lin , Yang Wang , Kaixuan Zhang , Zhihong Chen , Yuliang Yan
show 4 more authors
Dan Ou Shengyu Zhang Haihong Tang Bo Zheng
This is my paper
classification cs.IR
keywords generative retrievalsemantic IDe-commerce searchvalue-aware rankingquery-aware parallel SIDbusiness value modelingautoregressive retrieval
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

TSGR aims to fix a blind spot in generative retrieval for industrial e-commerce search: most systems rank by semantic likelihood, ignoring whether an item will actually be clicked or purchased. The paper proposes two coordinated changes: QP-SID, which orders item identifiers by click statistics so high-value and query-relevant items get earlier token positions, and VRM, a cross-attention ranking module fused into the same generative model that re-ranks beam-search candidates using item side information. Offline, the combination improves HR@1000 by 9.16% over the baseline; online A/B tests show gains of +0.43% in IPV, +1.12% in transaction count, and +1.64% in GMV. The system is fully deployed in production, making the claim concrete at Taobao scale.

Core claim

TSGR's central claim is that value awareness must enter both item representation and candidate ranking for generative retrieval to serve e-commerce objectives. QP-SID builds a three-level semantic ID codebook in which the final level orders items within each cluster by 30-day click counts, with parallel orderings conditioned on representative query terms, so that higher-value and query-relevant items are assigned earlier token indices. VRM then re-ranks the beam-search candidates by fusing user representations from the backbone's hidden states with item side-info, item embeddings, and SID embeddings, scoring PV, CTR, and CVR and ranking by PV×CTR. Trained progressively through multi-task Pre

What carries the argument

QP-SID (Query-aware Parallel SID) constructs a codebook of size 32768×8192×8192: the first two levels encode category and cluster priors via RQ-KMeans, and the third 'efficiency' level orders items within clusters by click counts, including parallel orderings for representative query terms so the model can select the intent-matched path during decoding. VRM (Value-aware Ranking Module) takes beam-search candidates, forms item representations from side-info, item embedding, and SID embedding, and applies cross-attention with the backbone's user-context hidden states to produce PV, CTR, and CVR scores, re-ranking by PV×CTR. The progressive training pipeline—multi-task Pre-SFT, session-wise wei

Load-bearing premise

The load-bearing premise is that the 30-day click statistics used to order item IDs do not leak test information, so the offline HR gains measure genuine retrieval improvement rather than a re-sorting of the label space by popularity.

What would settle it

Rebuild QP-SID with the 30-day click window ending strictly before the test period, and randomize the Level-3 efficiency orderings within clusters; if HR@1000 does not drop by roughly the reported 9.16% gap, the value-ordering mechanism is not the source of the gain. Also inspect whether any test-day click statistics entered the ordering construction.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Generative retrieval can be made value-aware without sacrificing the single-model architecture: the same model serves as retriever and pre-ranker, removing the distribution gap of a separate pre-ranking stage.
  • Assigning high-value items earlier token indices concentrates training signal and decoding budget on commercially important items, improving HR@K across all reported cutoffs (QP-SID alone gives +4.42% HR@1000).
  • Query-conditioned parallel orderings let a single item have multiple SID representations, so the model can route an item through the ordering that best matches query intent; 76.8% of training labels are query-relevant and the final model output spreads across ordering channels.
  • Jointly optimizing the ranking module with the generation objective contributes +3.73% HR@1000 with no additional serving latency, outperforming a conventional pre-ranking model on most metrics.
  • Value-aware retrieval translates to business metrics in production: +0.43% IPV, +1.12% transaction count, and +1.64% GMV, with TSGR deployed as an additional retrieval channel that bypasses pre-ranking.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the 30-day click statistics used for QP-SID overlap the test day, part of the offline HR gain could be a popularity re-sorting artifact rather than learned semantic matching; a clean time-shift experiment would separate the two.
  • The ablation showing 99.8% of outputs collapsing to the Default path in an intermediate variant suggests the final model's use of multiple orderings is fragile and depends on the specific input-label consistency trick, which may not transfer automatically to other taxonomies.
  • The online gains are measured relative to a production baseline that already includes multiple retrieval channels; the absolute contribution of TSGR alone is not isolated, so the +1.64% GMV should be read as incremental over that combined baseline.
  • The finding that RL adds only +0.52% while VRM adds more suggests that feature fusion is a more efficient way to inject value than reward optimization in this setting; combining both in different proportions is a testable extension.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes TSGR, a unified generative retrieval framework for Taobao Search that integrates value awareness into both the Semantic ID (SID) construction and the ranking stage. For item representation, it introduces Query-aware Parallel SID (QP-SID), which builds multiple click-statistics-based orderings of items within clusters, including a default click-count ordering and query-term-conditioned orderings, and selects the ordering per query. For candidate ranking, a Value-aware Ranking Module (VRM) fuses backbone hidden states with item side-features to produce PV/CTR/CVR scores and re-ranks beam-search candidates. A progressive training pipeline (Pre-SFT multi-task pretraining, weighted multi-positive SFT, and optional GRPO RL) is described. Offline experiments on Taobao search report HR@K improvements, with a headline 9.16% improvement in HR@1000 over FORGE, and online A/B tests report gains in IPV, transaction count, and GMV. The paper claims full production deployment.

Significance. If the central claim is established, the paper makes a notable industrial contribution: it demonstrates a single generative model that serves as both retriever and pre-ranker, and it incorporates business value into SID construction in a way that is deployable at Taobao scale. The paper is honest about the marginal gains from RL and reports a real online A/B test with production traffic. However, the offline evaluation has a potentially load-bearing confound: the QP-SID ordering is derived from 30-day click statistics, and the evaluation metric HR@K uses clicked/purchased items as ground truth. Without demonstrating that the gains are not merely popularity re-ranking or label leakage, the central 'value-aware retrieval' claim is not fully established. The paper also lacks variance or significance reporting for offline results and contains a direct contradiction in the QP-SID analysis (Section 4.3.1, Table 4). These issues require substantial revision before the contribution can be accepted at face value.

major comments (4)
  1. [Section 3.1, Eqs. (3)-(5); Section 4.1; Table 2] The QP-SID ordering is computed from 30-day click counts (Eqs. 3-5), and HR@K is measured against clicked/purchased items on a test day (Section 4.1). The paper does not state the reference date of the 30-day window; if it includes the test day, test labels leak into the SID construction. Even with a clean temporal cutoff, ranking items by click count is a strong popularity prior, and high-click items are disproportionately likely to be clicked on the test day regardless of query. The Table 2 comparison of QP-SID vs FORGE (random last-level token) is therefore not a controlled test of query-aware value matching. A baseline that re-ranks FORGE candidates by click count, or a random ordering within clusters, is needed to separate popularity re-ranking from learned semantic matching.
  2. [Section 4.2, Table 2; Section 4.4] No variance or significance information is reported for the offline HR@K numbers. Table 2 gives point estimates only. The headline 9.16% improvement in HR@1000 and the component-wise gains (+4.42% for QP-SID, +3.73% for VRM) could be within noise, especially since the paper does not report multiple trials, seeds, or confidence intervals. Section 4.4 claims online gains are statistically significant but provides no test statistic, p-value, or confidence interval. For a paper that draws strong conclusions from comparative numbers, this is a load-bearing omission.
  3. [Section 4.3.1, Table 4] The analysis of QP-SID is internally contradictory. The text states '99.8% of model outputs collapse to the Default path', while Table 4 reports Model Output Default at 32.05% and term-rank categories at 44.91%, 14.53%, and 8.51% (summing to ~100%). The table is then invoked to claim the model 'has successfully learned to retrieve from multiple ranking channels rather than relying solely on the Default channel.' These two statements cannot both be true. Either the 99.8% figure refers to a different configuration or the table is mislabeled; the manuscript must clarify this contradiction because it directly affects the interpretation of whether QP-SID's parallel orderings are actually used.
  4. [Section 4.3.2, Figure 4; Eq. (14)] The evaluation of the VRM is confounded by the metric definition. HR@K is computed against clicked/purchased items, and the VRM re-ranks by PV×CTR, which is a direct proxy for click propensity. The paper itself states that PV×CTR 'directly corresponds to the HitRate objective' and therefore 'this result is expected,' but this does not resolve the circularity: the gain could simply be a re-ranking of candidates by the same signal used to define ground truth. A control that re-ranks by popularity alone, or evaluation on a semantically defined relevance set, is necessary to show that the VRM adds value beyond the label signal.
minor comments (5)
  1. [Abstract; Table 2; Section 4.2] The phrase '9.16% improvement in HR@1000' should specify whether the percentage is relative or absolute percentage points. From Table 2, 0.8651 - 0.7735 = 0.0916, suggesting absolute percentage-point increase; the same ambiguity applies to '+4.42%' and '+3.73%' in Section 4.2.
  2. [Section 3.1, Eq. (3)-(5)] The argsort direction is not explicit. The text says higher-value items get earlier token indices, but Eq. (3) and Eq. (5) should state 'descending' or define the ordering so readers know larger click counts map to smaller indices.
  3. [Table 4 heading] The table title uses 'QPSID' instead of 'QP-SID'. Please standardize the notation.
  4. [Section 4.3.1] The text says '99.8% of model outputs collapse to the Default path' but Table 4 shows a 32.05% Default rate. If the 99.8% is for a specific ablation (e.g., 'Multi-path' configuration in Row 3 of Table 3), that should be stated explicitly.
  5. [Appendix A.7] The TencentGR preprocessing section is detailed and helpful, but the construction of the 'third level by ranking items within each cluster in descending order of their transaction, click, and impression counts' uses the same popularity-based design as the main paper; the external dataset results (Table 9) therefore inherit the popularity-re-ranking confound described in the major comments.

Circularity Check

1 steps flagged

Offline QP-SID HR@K gain is partly circular: the SID token order and the HR@K labels are both derived from click behavior, with no popularity control; the online A/B evidence is independent and moderates the score.

specific steps
  1. fitted input called prediction [Section 3.1, Eqs. (3)-(5); Section 4.1, evaluation metrics; Section 4.2, Table 2 (QP-SID vs FORGE)]
    "A straightforward design ranks items by their click count over the past 30 days, giving the default ordering: π0 = argsort_{i∈C} s0(i) ... Ranking items by s(i,t) produces the term-specific ordering: πt = argsort_{i∈C} s(i,t) ... To verify retrieval performance, we adopt HitRate@K (HR@K) as evaluation metric ... HR@1000 as the primary metric, as the goal of retrieval is to ensure that relevant items are surfaced within the candidate pool passed to downstream rankers."

    Equations (3)-(5) define the SID token order by 30-day click counts, and Section 4.1 defines HR@K on test-day clicked/purchased items. Because the same click signal both orders the token vocabulary and defines the hit label, a model that simply emits the click-ranked SID order will achieve high HR@K without learning query-item semantics; the +4.42% HR@1000 gain over the random-order FORGE baseline is therefore a re-ordering of the label space by the evaluation signal. The paper supplies no popularity-only baseline and no explicit temporal cutoff for the 30-day window, so the fitted click statistics are not separated from the test labels. The paper's own §4.3.1 observation that 99.8% of model outputs collapse to the Default path is consistent with popularity memorization rather than exploit

full rationale

The central derivation chain for QP-SID is: build SID orderings from 30-day click statistics (Eqs. 3-5), train the generative model to output those SIDs, then measure HR@K on test-day clicks. The input signal and the evaluation signal are the same behavioral variable, and the paper does not include a popularity-only baseline or state the temporal cutoff of the 30-day window. This makes the offline 'value-aware retrieval' claim partly circular, as the reported HR gain can be explained by re-ranking the item space by the label signal. The paper itself reports that 99.8% of outputs collapse to the Default path, further supporting a popularity-memorization reading. However, the paper also reports a statistically significant online A/B test (+0.43% IPV, +1.12% Transaction Count, +1.64% GMV) on live Taobao traffic, which is an external, non-circular validation of the deployed system. The self-citations ([11], [13]) are used for design motivations and infrastructure, not as load-bearing uniqueness theorems, so they do not materially raise the circularity score. Overall, the offline QP-SID evaluation exhibits a real fitted-input/prediction confound, but the online deployment evidence and the absence of a load-bearing self-citation chain keep the score at 5 rather than higher.

Axiom & Free-Parameter Ledger

8 free parameters · 7 axioms · 2 invented entities

The central evaluation rests on click/purchase statistics used both to build SID orderings and as ground truth; in addition, the companion self-citation [11] supplies the efficiency-codebook mechanism, making part of the claimed novelty a re-presentation of the authors' own prior work. The model also assumes Qwen2.5-0.5B can memorize the SID space and that reordering tokens by clicks improves real business value rather than only the HR metric.

free parameters (8)
  • coverage_threshold_for_term_filtering = 80%
    Eq. 4: terms with coverage >80% are discarded; threshold hand-chosen, affects which representative terms define query-aware orderings.
  • num_representative_terms_per_cluster = 3
    Section 3.1: top-3 terms per cluster; number of parallel orderings = 4 (default plus 3 terms); hand-chosen.
  • multi_positive_loss_weights = (w_pv, w_clk, w_pay) = (1, 2, 3)
    Eq. 13 and implementation details; chosen by hand to prioritize pay > click > pv; no sensitivity analysis.
  • rank_loss_weight_lambda = 1
    Eq. 15 and A.3; joint objective weight set to 1 without reported sweep.
  • codebook_sizes = 32768 x 8192 x 8192
    Section 3.1 and 4.1; semantic and efficiency codebook sizes chosen for deployment cost and coverage; no ablation.
  • beam_sizes = [400, 400, 10000]
    Section 4.1/A.3; decoding beam sizes chosen for latency; no ablation.
  • click_count_window = 30 days
    Section 3.1 Eqs. 3/5; default and term-specific orderings use past-30-day click counts; window choice not ablated.
  • vrm_scoring_formula = PV x CTR (CVR excluded)
    Figure 4/Table 10: selected from ablations; the paper also admits the CVR loss has negligible impact, so this is a post-hoc selection.
axioms (7)
  • domain assumption Items sharing the same third-level SID token mutually reinforce each other's training signal; assigning high-value items earlier token indices increases their training signal and retrieval likelihood.
    Section 1 motivation; not proven; depends on the model's token frequency dynamics.
  • domain assumption Click counts over the past 30 days are a valid proxy for item business value and query-item relevance.
    Section 3.1 Eqs. 3-5; no validation that click ordering aligns with GMV/transaction value; the paper's online metrics are GMV, not clicks.
  • domain assumption Representative terms derived from term-cluster click statistics capture query intent well enough to improve SID routing.
    Section 3.1 Eqs. 4-5; term selection based on coverage and click counts; no formal justification.
  • standard math RQ-KMeans residuals and category/cluster priors produce a semantically coherent SID space.
    Section 3.1; relies on RQ-VAE/RQ-KMeans from prior work [17,34] without re-derivation.
  • domain assumption The generative backbone (Qwen2.5-0.5B) can memorize and reproduce SID sequences for millions of items from the SFT data.
    Section 3.3; capacity assumption; no measurement of memorization accuracy.
  • domain assumption A separate pre-ranking distribution gap exists and is harmful; unified VRM avoids it without loss.
    Section 3.2; asserted, not measured; no direct comparison of objective misalignment magnitudes.
  • domain assumption Evaluating behavioral labels via HR@K after value-aware re-ranking is a fair measure of retrieval improvement.
    Section 4.1; this is the metric assumption that underpins the headline 9.16% HR improvement.
invented entities (2)
  • QP-SID parallel orderings (query-aware parallel SID paths) no independent evidence
    purpose: Gives each item multiple third-level SID assignments (default plus term-conditioned) so generation likelihood reflects business value and query relevance.
    Evaluated only on in-house Taobao data and a small TencentGR ablation; no external benchmark or public artifact confirms the representation's value.
  • VRM value score (PV x CTR) no independent evidence
    purpose: Re-ranks beam-search candidates inside the generative model by fusing user representation with item side-info.
    The score is trained and evaluated on the same behavioral labels; no external or held-out business metric is provided with confidence intervals.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of TSGR: Taobao Search Generative Retrieval." pith.science (2026). https://pith.science/paper/VFN75JSS

@misc{pith2026260718796,
  author       = {Pith},
  title        = {Pith review of: TSGR: Taobao Search Generative Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VFN75JSS}},
  note         = {Machine review of arXiv:2607.18796}
}
Share X Bluesky LinkedIn Reddit HN
abstract

Generative retrieval (GR) has demonstrated strong promise for industrial e-commerce search by training a single autoregressive model to directly generate the Semantic IDs (SIDs) of target items. However, existing GR systems are primarily optimized for semantic matching and remain insensitive to item business value: SID construction is value-unaware, and candidates are ranked without access to item side-info. Consequently, high-value items are often missed or deprioritized at the retrieval stage, limiting downstream business impact. This limitation is particularly critical in industrial settings such as Taobao Search, where business objectives are central to system design. To address this, we propose $\textbf{T}$aobao $\textbf{S}$earch $\textbf{G}$enerative $\textbf{R}$etrieval ($\textbf{TSGR}$), a unified generative retrieval framework that incorporates value awareness into both item representation and candidate ranking. 1) For item representation, TSGR introduces $\textbf{Query-aware Parallel SID (QP-SID)}$, which encodes query-conditioned value orderings into the SID construction by building parallel codebooks derived from query-item statistics, so that higher-value and query-relevant items are assigned better token indices. 2) For candidate ranking, we introduce a $\textbf{Value-aware Ranking Module (VRM)}$ that is built upon and jointly optimized with the GR, enabling a single model to seamlessly serve as both retriever and pre-ranker without a dedicated pre-ranking stage. A progressive training pipeline further aligns the model with semantic relevance, user preferences, and business objectives. Offline experiments show that TSGR achieves a 9.16\% improvement in HR@1000, and online A/B tests further validate its effectiveness, yielding gains of +0.43\% in IPV, +1.12\% in Transaction Count, and +1.64\% in GMV. TSGR has been fully deployed in production.

Figures

Figures reproduced from arXiv: 2607.18796 by Bo Zheng, Dan Ou, Gui Ling, Haihong Tang, Kaixuan Zhang, Kunhai Lin, Shengyu Zhang, Tianyu Zhan, Tong Xiong, Yang Wang, Yuliang Yan, Zhihong Chen.

Figure 1
Figure 1. Figure 1: Overview of TSGR, a unified generative framework that jointly handles retrieval and pre-ranking within a single [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Training Pipeline. Model capabilities are developed [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of different QP-SID configurations and [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Ablation study on the effect of different score com [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Ablation study on the effect of different architecture [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Ablation study on Pre-SFT task design. Each variant [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Online deployment architecture of TSGR. Items [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Case study of QP-SID on two clusters: phone cases (Cluster (140, 3504)) and dining tables (Cluster (10278, 7631)). For [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Case study of term selection in QP-SID during serving. Left: different queries each match a single term, which [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

36 extracted references · 16 linked inside Pith

  1. [1]

    Ben Chen, Xian Guo, Siyuan Wang, Zihan Liang, Yue Lv, Yufei Ma, Xinlong Xiao, Bowen Xue, Xuxin Zhang, Ying Yang, et al . 2025. Onesearch: A preliminary exploration of the unified end-to-end generative framework for e-commerce search.arXiv preprint arXiv:2509.03236(2025)

    arXiv 2025

  2. [2]

    Ben Chen, Siyuan Wang, Yufei Ma, Zihan Liang, Xuxin Zhang, Yue Lv, Ying Yang, Huangyu Dai, Lingtao Mao, Tong Zhao, et al. 2026. OneSearch-V2: The Latent Reasoning Enhanced Self-distillation Generative Search Framework.arXiv preprint arXiv:2603.24422(2026)

    Pith/arXiv arXiv 2026

  3. [3]

    Junyi Chen, Lu Chi, Bingyue Peng, and Zehuan Yuan. 2024. Hllm: Enhancing sequential recommendations via hierarchical large language models for item and user modeling.arXiv preprint arXiv:2409.12740(2024)

    Pith/arXiv arXiv 2024

  4. [4]

    Kairui Fu, Tao Zhang, Shuwen Xiao, Ziyang Wang, Xinming Zhang, Chenchi Zhang, Yuliang Yan, Junjun Zheng, Yu Li, Zhihong Chen, et al . 2025. Forge: Forming semantic identifiers for generative retrieval in industrial datasets.arXiv preprint arXiv:2509.20904(2025)

    Pith/arXiv arXiv 2025

  5. [5]

    Ruidong Han, Bin Yin, Shangyu Chen, He Jiang, Fei Jiang, Xiang Li, Chi Ma, Mincong Huang, Xiaoguang Li, Chunzhen Jing, et al . 2025. Mtgr: Industrial- scale generative recommendation framework in meituan. InProceedings of the 34th ACM International Conference on Information and Knowledge Management. 5731–5738

    2025

  6. [6]

    Xuegang Hao, Ming Zhang, Alex Li, Xiangyu Qian, Zhi Ma, Yanlong Zang, Shijie Yang, Zhongxuan Han, Xiaolong Ma, Jinguang Liu, et al. 2025. OxygenREC: An Instruction-Following Generative Framework for E-commerce Recommendation. arXiv preprint arXiv:2512.22386(2025)

    arXiv 2025

  7. [7]

    Yupeng Hou, Jiacheng Li, Ashley Shin, Jinsung Jeon, Abhishek Santhanam, Wei Shao, Kaveh Hassani, Ning Yao, and Julian McAuley. 2025. Generating long semantic ids in parallel for recommendation. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 956–966

    2025

  8. [8]

    Yanhua Huang, Yuqi Chen, Xiong Cao, Rui Yang, Mingliang Qi, Yinghao Zhu, Qingchang Han, Yaowei Liu, Zhaoyu Liu, Xuefeng Yao, et al . 2025. Towards large-scale generative ranking.arXiv preprint arXiv:2505.04180(2025)

    Pith/arXiv arXiv 2025

  9. [9]

    Jie Jiang, Xinxun Zhang, Enming Zhang, Yuling Xiong, Jun Zhang, Jingwen Wang, Huan Yu, Yuxiang Wang, Hao Wang, Xiao Yan, et al. 2026. End-to-End Semantic ID Generation for Generative Advertisement Recommendation.arXiv preprint arXiv:2602.10445(2026)

    Pith/arXiv arXiv 2026

  10. [10]

    Zhijie Lin, Zhuofeng Li, Chenglei Dai, Wentian Bao, Shuai Lin, Enyun Yu, Haoxi- ang Zhang, and Liang Zhao. 2025. GReF: A Unified Generative Framework for Efficient Reranking via Ordered Multi-token Prediction. InProceedings of the 34th ACM International Conference on Information and Knowledge Management. 5879–5887

    2025

  11. [11]

    Gui Ling, Zhihong Chen, Yu Li, Tong Xiong, Kunhai Lin, Kaixuan Zhang, Yuliang Yan, Dan Ou, Haihong Tang, and Bo Zheng. 2026. Beyond Semantic IDs: Encod- ing Business-Value Ranking into Document Identifiers for Generative Retrieval. arXiv:2607.11392 [cs.IR] https://arxiv.org/abs/2607.11392

    Pith/arXiv arXiv 2026

  12. [12]

    Xiaoyu Liu, Fuwei Zhang, Yiqing Wu, Xinyu Jia, Zenghua Xia, Fuzhen Zhuang, Zhao Zhang, Fei Jiang, and Wei Lin. 2026. CAT-ID2: Category-Tree Integrated Document Identifier Learning for Generative Retrieval In E-commerce. InPro- ceedings of the Nineteenth ACM International Conference on Web Search and Data Mining. 426–435

    2026

  13. [13]

    Dan Ou, Gui Ling, Hao Wan, Hongbin Zhou, Jialiang Cheng, Jiangnan Pang, Silu Zhou, Wei Shi, Weichen Ye, Wenming Zhang, Yang Wang, Yu Li, Yuliang Yan, Zhan Fa, Zhihong Chen, Zongyuan Wu, Bo Zheng, Changfa Wu, Dunxian Huang, Haihong Tang, Jinlong Guo, Kaixuan Zhang, Kun Ma, Lin Qu, Longbo Zhong, Tao Lan, Tong Xiong, and Zhibo Wu. 2026. Prompt Generation Tec...

    Pith/arXiv arXiv 2026

  14. [14]

    Junwei Pan, Wei Xue, Chao Zhou, Xing Zhou, Lunan Fan, Yanbo Wang, Haoran Xin, Zhiyu Hu, Yaozheng Wang, Fengye Xu, et al . 2026. Tencent Advertising Algorithm Challenge 2025: All-Modality Generative Recommendation.arXiv preprint arXiv:2604.04976(2026)

    Pith/arXiv arXiv 2026

  15. [15]

    Changhua Pei, Xinru Yang, Qing Cui, Xiao Lin, Fei Sun, Peng Jiang, Wenwu Ou, and Yongfeng Zhang. 2019. Value-aware recommendation based on reinforcement profit maximization. InThe World Wide Web Conference. 3123–3129

    2019

  16. [16]

    Aleksandr V Petrov and Craig Macdonald. 2024. RecJPQ: training large-catalogue sequential recommenders. InProceedings of the 17th ACM International Conference on Web Search and Data Mining. 538–547

    2024

  17. [17]

    Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Hulikal Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Tran, Jonah Samost, et al

  18. [18]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300(2024)

    Pith/arXiv arXiv 2024

  19. [19]

    Chaotian Song, Jingyao Zhang, Chenghao Chen, Zisen Sang, Dehai Zhao, Guodong Cao, Boxi Wu, Deng Cai, and Jia Jia. 2026. DeGRe: Dense-supervised Generative Reranking for Recommendation.arXiv preprint arXiv:2605.25749 (2026)

    Pith/arXiv arXiv 2026

  20. [20]

    Yanglong Song, Zihao Yang, Shuo Meng, Rujun Guo, Jin Zhang, Bin Wang, Shaoyu Liu, Xiaozhao Wang, and Guanjun Jiang. 2026. From Item-Only to Query- Item: Query-Conditioned Generative Search with QGS in Quark.arXiv preprint arXiv:2605.25514(2026)

    Pith/arXiv arXiv 2026

  21. [21]

    Juntao Tan, Shuyuan Xu, Wenyue Hua, Yingqiang Ge, Zelong Li, and Yongfeng Zhang. 2024. Idgenrec: Llm-recsys alignment with textual id learning. InProceed- ings of the 47th international ACM SIGIR conference on research and development in information retrieval. 355–364

    2024

  22. [22]

    Yi Tay, Vinh Tran, Mostafa Dehghani, Jianmo Ni, Dara Bahri, Harsh Mehta, Zhen Qin, Kai Hui, Zhe Zhao, Jai Gupta, et al. 2022. Transformer memory as a differentiable search index.Advances in neural information processing systems35 (2022), 21831–21843

    2022

  23. [23]

    Wenjie Wang, Honghui Bao, Xinyu Lin, Jizhi Zhang, Yongqi Li, Fuli Feng, See- Kiong Ng, and Tat-Seng Chua. 2024. Learnable item tokenization for generative recommendation. InProceedings of the 33rd ACM International Conference on Information and Knowledge Management. 2400–2409

    2024

  24. [24]

    Yifan Wang, Yixuan Wang, YiDan Liang, Qiang Liu, and Fei Xiao. 2026. Towards Sustainable Growth: A Multi-Value-Aware Retrieval Framework for E-Commerce Search.arXiv preprint arXiv:2605.17994(2026)

    Pith/arXiv arXiv 2026

  25. [25]

    Yanjing Wu, Yinfu Feng, Jian Wang, Wenji Zhou, Yunan Ye, Rong Xiao, and Jun Xiao. 2024. Hi-gen: Generative retrieval for large-scale personalized e-commerce search.arXiv preprint arXiv:2404.15675(2024)

    Pith/arXiv arXiv 2024

  26. [26]

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al . 2024. Qwen2. 5 Technical Report.arXiv e-prints(2024), arXiv–2412

    2024

  27. [27]

    Haiyang Yang, Qinye Xie, Qingheng Zhang, Chen Li Yu, Huike Zou, Chengbao Lian, Shuguang Han, Fei Huang, Jufeng Chen, and Bo Zheng. 2025. GSID: Gener- ative Semantic Indexing for E-Commerce Product Understanding. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track. 1113–1121

    2025

  28. [28]

    Jiaqi Zhai et al . 2024. Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations. (2024)

    2024

  29. [29]

    Jun Zhang, Yi Li, Yue Liu, Changping Wang, Yuan Wang, Yuling Xiong, Xun Liu, Haiyang Wu, Qian Li, Enming Zhang, et al. 2025. GPR: Towards a Generative Pre-trained One-Model Paradigm for Large-Scale Advertising Recommendation. arXiv preprint arXiv:2511.10138(2025)

    arXiv 2025

  30. [30]

    Kaike Zhang, Xiaobei Wang, Shuchang Liu, Hailan Yang, Xiang Li, Lantao Hu, Han Li, Qi Cao, Fei Sun, and Kun Gai. 2025. Goalrank: Group-relative optimization for a large ranking model.arXiv preprint arXiv:2509.22046(2025)

    arXiv 2025

  31. [31]

    Xinxun Zhang, Yuling Xiong, Jiale Zhou, Zhengkai Guo, Zhennan Pang, Junbang Huo, Jingwen Wang, Xuyang Sun, Enming Zhang, Jiaguang Jin, et al. 2026. Unified Value Alignment for Generative Recommendation in Industrial Advertising.arXiv preprint arXiv:2605.05803(2026)

    Pith/arXiv arXiv 2026

  32. [32]

    Bowen Zheng, Yupeng Hou, Hongyu Lu, Yu Chen, Wayne Xin Zhao, Ming Chen, and Ji-Rong Wen. 2024. Adapting large language models by integrating collaborative semantics for recommendation. In2024 IEEE 40th International Conference on Data Engineering (ICDE). IEEE, 1435–1448

    2024

  33. [33]

    Zuowu Zheng, Ze Wang, Fan Yang, Jiangke Fan, Teng Zhang, Yongkang Wang, and Xingxing Wang. 2025. Ega-v2: An end-to-end generative framework for industrial advertising.arXiv preprint arXiv:2505.17549(2025)

    Pith/arXiv arXiv 2025

  34. [34]

    Guorui Zhou, Jiaxin Deng, Jinghao Zhang, Kuo Cai, Lejian Ren, Qiang Luo, Qian- qian Wang, Qigen Hu, Rui Huang, Shiyao Wang, et al. 2025. OneRec Technical Report.arXiv preprint arXiv:2506.13695(2025)

    arXiv 2025

  35. [35]

    磁吸” (magnetic) for phone cases and “ 实木

    Jianbo Zhu, Xing Fang, Jing Wang, Mingmin Jin, Bokang Wang, Guangxin Song, Zhenyu Xie, and Junjie Bai. 2026. Efficient Generative Retrieval for E- commerce Search with Semantic Cluster IDs and Expert-Guided RL.arXiv preprint arXiv:2605.14434(2026). Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Trovato et al. Figure 7: Online deployment architect...

    Pith/arXiv arXiv 2026

  36. [2023]

    Recommender systems with generative retrieval.Advances in Neural Information Processing Systems36 (2023), 10299–10315

    2023

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.