Pith. sign in

REVIEW 4 major objections 5 minor 31 references

SeqLLM: Augmenting LLMs with Behavioral-Sequence Modeling for High-Stakes Decisions at WeChat Pay

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

Pith's one-line read SeqLLM claims that prefix-guided supervised fine-tuning injects behavioral-sequence modeling into a pretrained LLM as effectively as continual pretraining while preserving language ability.

desk verdict A strong applied paper with a real empirical result, but the mechanism claim about prefix-guided forgetting avoidance rests on an under-specified control. read the letter →

arxiv 2608.03063 v1 pith:57N4M6AT submitted 2026-08-04 cs.CL

classification cs.CL
keywords behavioral-sequencemodelinglargelanguagemodelscatastrophicforgettingprefix-guidedcapabilityinjectionmerchantriskcontrolgenerativerecommendationtokengroundingWeChatPay
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that a pretrained large language model can gain native behavioral-sequence modeling—reading long, timestamped event streams—without losing its language and reasoning skills. The proposed framework, SeqLLM, does this with three pieces: a compact field-level behavior vocabulary, a lightweight projector that grounds behavior tokens in the model's semantic space, and prefix-guided capability injection, which replaces continual pre-training with task-prefixed supervised fine-tuning. The central claim is that this injection method matches continual pre-training on next-event prediction while retaining language ability, and that the resulting joint text–behavior model improves real decisions. In deployment at WeChat Pay, screening precision rises from 92.0% to 97.5%, and behavior-token embeddings improve a fraud detector's Precision@Top-0.01% by 26.8 percentage points. A sympathetic reader would care because it suggests a scalable way to add non-text capabilities to LLMs without the usual forgetting trade-off.

What carries the argument

The load-bearing mechanism is prefix-guided capability injection: next-event prediction is reformulated as generation conditioned on a task instruction, with loss computed only on the response suffix, in contrast to continual pre-training's unconditional next-token loss at every position. The framework also includes a field-level behavior vocabulary, which encodes each event with roughly nine dedicated tokens disjoint from the word vocabulary, and a lightweight residual projector $g_\psi(\mathbf{e})=\mathbf{e}+\mathrm{MLP}_\psi(\mathbf{e})$, trained by a translation-then-reasoning curriculum to ground behavior tokens in the LLM's semantic space. The projector's shared residual constraint and the task-prefixed loss are what the paper credits for compositional transfer and forgetting mitigation, respectively.

What would settle it

Train the same Qwen3-8B backbone on the same roughly 20 million merchant sequences with the same response-only loss, but replace the task prefix with a random token or remove it entirely; if C-Eval drops as much as under continual pre-training (from 0.789 toward 0.29), then the task prefix itself is not what preserves language.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that behavioral-sequence capability can be injected into a pretrained LLM through instruction-conditioned supervised fine-tuning rather than continual pre-training. SeqLLM recasts next-event prediction as conditional generation under a task prefix, applying loss only to the response suffix; on roughly 20 million merchant behavior sequences this matches sample-aligned continual pre-training on next-event prediction (HR@10 0.806 vs 0.804) while retaining C-Eval 0.789 versus continual pre-training's 0.293. The same recipe carries into production: a joint text–behavior screening model raises risk precision from 92.0% to 97.5% over a text-only DeepSeek baseline, and pretrained behavior-token embeddings improve an online fraud detector by 26.8 percentage points in Precision@Top-0.01%. On public benchmarks, SeqLLM surpasses User-LLM on MovieLens and Amazon Recall@5 and improves RecIF Pass@32 over the full OneRec-8B pipeline with 4.8 times fewer GPU-days, while retaining markedly stronger general language ability.

Load-bearing premise

The load-bearing premise is that applying the training loss only to the answer part of a task-prefixed example confines the behavior learning to a narrow part of the model, so language ability stays intact; if that confinement fails, the approach would need a recovery stage like continual pre-training.

Editorial extensions

If this is right

  • On roughly 20 million unlabeled merchant sequences, prefix-guided SFT matches continual pre-training on next-event prediction (HR@10 0.806 vs 0.804) while retaining C-Eval 0.789, so sequence capability can be added at scale without a separate language-recovery stage.
  • A joint text–behavior model raises merchant screening precision from 92.0% to 97.5% in a three-month shadow evaluation, and the post-launch appeal rate drops from 12% to about 2% with zero exonerations.
  • Initializing a production fraud detector's behavior embeddings with SeqLLM's pretrained embeddings yields a 26.8 percentage-point gain in Precision@Top-0.01% and a 33.1 percentage-point gain in Recall@Top-1%.
  • On public recommendation benchmarks, SeqLLM outperforms User-LLM by up to 32% relative Recall@5 and improves RecIF Pass@32 by 14.2% over the full OneRec-8B pipeline while using 4.8 times fewer GPU-days.
  • Because injection is SFT rather than continual pre-training, the approach needs no distillation stage and can be applied to any pretrained LLM backbone, demonstrated here on Qwen3-8B and Qwen3-0.6B.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Inference: If the prefix-conditioning mechanism is what preserves language, the same recipe should transfer to other non-text skills such as graph reasoning or time-series forecasting by framing those tasks as instruction-conditioned generation.
  • Inference: The parameter diagnostics, including smaller per-layer weight changes, last-layer CKA 0.98, and a task-vector cosine of 0.14, suggest behavior learning occupies a low-dimensional subspace; a direct test would train with prefix-guided SFT, subtract the resulting task vector from the backbone, and check whether the base language model is recovered.
  • Inference: The projector's shared-residual constraint, shown to matter for multi-event composition, implies that grounding new token types through a shared interface matters more for compositional transfer than for per-token semantics; one could test this by adding a novel token type without a translation stage and predicting which capabilities fail.
  • Inference: The paper's matched-control design does not include a continual-pre-training variant with response-only loss but no task prefix; running that variant would isolate whether the loss mask or the task condition carries the forgetting-mitigation effect.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. SeqLLM proposes to add behavioral-sequence modeling to a pretrained LLM while preserving language ability. Events are encoded as compact field-level tokens; a residual projector with text-grounded initialization is aligned through translation and reasoning stages; then sequence capability is injected by task-prefixed SFT with response-only loss instead of continual pretraining. The paper claims that this prefix-guided injection matches CPT on next-event prediction (HR@10 0.806 vs. 0.804) while retaining C-Eval 0.789 instead of 0.293, that joint text-behavior modeling improves merchant risk screening in a three-month shadow evaluation (precision 97.5% vs. 92.0%) and fraud-detection embeddings in an A/B test, and that the approach transfers to public benchmarks where it outperforms User-LLM and OneRec-8B.

Significance. If the central claim holds, this is a significant applied contribution: it offers a scalable alternative to adapt-then-repair CPT for injecting a new modality into an LLM, with production evidence from two deployed systems. The paper includes a deliberately controlled RQ1 comparison (sample-aligned and gradient-aligned CPT with strictly more language replay, plus a no-alignment-stack condition), public-benchmark comparisons against released checkpoints, and a code release for the public experiments. The main gaps are that the control isolating the task prefix from the response-only mask is under-specified, and the production comparison confounds the framework with the base model; both are fixable with additional experiments or careful re-scoping of claims.

major comments (4)
  1. [§4.2, Eq. (5), Appendix C.2] The control intended to isolate the task prefix from the response-only loss is under-specified. 'Gradient-aligned CPT' is described only as 'matching the cumulative number of loss-bearing behavior tokens' (Section 4.2), and Appendix C.2 does not state which token positions receive loss or whether an instruction condition is present. Because LPrefix in Eq. (5) applies loss exclusively to the suffix after the event cutoff, a token-count-matched CPT could supervise a different set of positions (e.g., early events or packed-stream positions), so the C-Eval gap in Table 1 (e.g., 0.456 vs. 0.789 in the aligned rows) could be caused by the choice of supervised positions rather than by instruction conditioning. Please specify the exact loss-position schedule of the gradient-aligned control, and add controls that vary the task-prefix condition and the response-only mask independently, such as response-only loss with no task prefix and task prefix with full-sequence loss.
  2. [§4.4, Table 3; Table 13] The production screening comparison confounds the framework with the base model and adaptation recipe: SeqLLM uses Qwen3-8B while the production baseline is a DeepSeek-based LLM, so the +5.5 pp risk-precision gain in Table 3 cannot be attributed to the SeqLLM injection method. The offline deployment-candidate comparison in Table 13 has the same confound, as SeqLLM and the DeepSeek variants differ in backbone and in whether sequence capability was injected at all. Please provide a same-backbone controlled comparison of prefix-guided SFT against standard SFT or CPT on the same base model, or explicitly present the production result as an end-to-end system-level gain without component-level attribution.
  3. [§4.4 label definition] The shadow-evaluation label definition creates a potential circularity risk. A candidate is labeled positive if confirmed by 'the existing production risk system—an ensemble of expert strategies operated independently of all three evaluated scorers, including the DeepSeek baseline' (Section 4.4). If that production risk system includes the DeepSeek baseline as a component, then the label is partly derived from the baseline's outputs, making the precision comparison partially circular. Please state unambiguously whether the baseline contributes to the label source; if it does, re-run the evaluation with labels from an independent source or report the comparison under labels that exclude the baseline.
  4. [§3.4, Figure 2, Table 11] The mechanistic claim that the task prefix 'confines' behavioral learning to a specific parameter pathway is not directly tested. The evidence in Figure 2 and Table 11 consists of aggregate correlations (weight-change norms, CKA, task-vector cosine) between training objective and parameter displacement; these do not establish that the behavior objective updates the model only under the prefix. Please add an intervention, such as measuring next-event prediction and C-Eval with and without the task prefix after training, or comparing per-example gradient overlap with language examples, to test the pathway claim.
minor comments (5)
  1. [Appendix C.2] Please report the actual token counts of the raw-text and chat-format replay corpora, so the 'strictly more language supervision' claim can be verified against the ~1.1M text examples in Dinj.
  2. [Abstract vs. Appendix A] The abstract states the code is available at the GitHub URL, while Appendix A says the code 'will be open-sourced'; please reconcile these statements.
  3. [Table 1] The HR@10 differences between SeqLLM (0.806) and sample-aligned CPT (0.804) are within noise, but no confidence intervals or seed-level variance are reported; please provide them.
  4. [§3.4] The cutoff ratio k≈70% is a free parameter; no sensitivity analysis is provided, so the reader cannot tell whether the retention results depend strongly on this choice.
  5. [Figure 2] The figure lacks explicit axis labels and a definition of 'relative weight change' in the main text; please add the normalization formula or a pointer to Appendix C.2.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SeqLLM's headline comparisons rest on external benchmarks and an explicitly independent production-label protocol.

full rationale

Score 0. The paper's central derivation—prefix-guided SFT versus CPT—is an empirical comparison against released external baselines (User-LLM, OneRec-8B) and public benchmarks (MovieLens, Amazon, RecIF, C-Eval/MMLU/AGIEval). No equation in Section 3.4 defines the outcome in terms of the input: LPrefix and LCPT are distinct objectives, and the reported HR@10 and C-Eval differences are measured rather than constructed. The only self-citation (PANTHER, ref [10]) is used as a sequence-only baseline and as the source of the HR@10 metric; it is not load-bearing for any claimed result. The merchant-screening label protocol explicitly states that 'shadow outputs did not affect review, enforcement, or labels' and that the labeling ensemble is 'operated independently of all three evaluated scorers,' so the risk-precision comparison is not self-confirming. The fraud-detector A/B test changes only embedding initialization and leaves all other serving components unchanged, providing an independent external evaluation. The 'gradient-aligned CPT' control is under-specified in that loss-position matching is not guaranteed, but an underspecified control is an experimental-validity concern rather than circularity: it does not make the claimed result equivalent to its inputs by construction. No circular step can be exhibited with a specific equation or definition, so the appropriate finding is no significant circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 3 invented entities

The paper's central pipeline rests on three hand-set elements: the field-discretization buckets, the 70% prefix/suffix cutoff, and the scaled text-pooling initialization. None of these is derived; each is a design choice with plausible alternatives. The projector and the prefix-conditioned loss are the two genuinely new pieces, and the paper provides ablations supporting each. The production label definition is the most consequential domain assumption because it shapes the headline precision numbers.

free parameters (3)
  • Event cutoff ratio k at about 70% of sequence length = 70%
    The split between observed prefix and predicted suffix is chosen at about 70% of the sequence. The paper gives no ablation over this ratio, so it is a hand-set hyperparameter rather than a derived quantity.
  • Behavior vocabulary discretization buckets = 1,533 tokens across 28 fields
    The bucketing of continuous fields such as amount into log-scale buckets is a modeling choice that determines vocabulary size and information loss. The paper does not report a sweep over bucket counts.
  • Global batch configuration = e.g., 2*2*8*12 for PGCI
    Batch sizes, learning rates, and step counts are tuned choices for each stage and dataset. They affect results but are standard training hyperparameters.
assumptions (4)
  • domain assumption The Qwen3-8B pretrained backbone has a semantic embedding space into which text-pooled behavior tokens can be rescaled without disrupting the model.
    Section 3.3 assumes that mean-pooled tokenizer embeddings of readable field text, rescaled to backbone embedding statistics, land in a region usable by the LLM. This is a standard alignment assumption but not proven.
  • domain assumption The production ground-truth label, confirmed by an existing production risk system or user-reported harm within 30 days, is an unbiased measure of merchant risk.
    Section 4.4 and Appendix C.3 define positive labels this way. This assumes the production risk system detects most true risks and that user reports are reliable, which is plausible but not independently verifiable from the paper.
  • domain assumption GPU-day estimates can be computed from formula (7) with MFU in [0.4, 0.5] and shared hardware assumptions.
    Appendix C.3 uses an estimated MFU and measured token counts to compute the 4.4x/4.8x efficiency claims. The formula is reasonable, but the MFU range and the assumption that both models run at it are not verified.
  • standard math The standard math of autoregressive next-token prediction and supervised fine-tuning is sufficient to describe the method.
    Equations (4)-(6) are standard cross-entropy losses for translation, SFT, and next-event prediction.
invented entities (3)
  • Behavior projector g_psi(e) = e + MLP_psi(e) independent evidence
    purpose: Maps all behavior tokens into the LLM semantic space with a shared residual correction.
    The projector is verified by an ablation (Table 15) where removing it degrades multi-event reasoning, and by public RecProbe results. This is internal evidence, but the design is falsifiable by reproduction on public data.
  • Prefix-guided capability injection as a task condition independent evidence
    purpose: A training objective that confines behavior-sequence learning to a prefix-activated pathway.
    The mechanism makes a falsifiable prediction: matched CPT loses language ability while prefix-guided SFT retains it. The public benchmarks and the reported C-Eval numbers are testable.
  • RecProbe four-task evaluation suite independent evidence
    purpose: Measures behavior-token semantic understanding beyond ranking accuracy.
    The suite is constructed from public RecIF captions and SIDs, and the construction scripts are promised in the open-source release.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SeqLLM: Augmenting LLMs with Behavioral-Sequence Modeling for High-Stakes Decisions at WeChat Pay." pith.science (2026). https://pith.science/paper/57N4M6AT

@misc{pith2026260803063,
  author       = {Pith},
  title        = {Pith review of: SeqLLM: Augmenting LLMs with Behavioral-Sequence Modeling for High-Stakes Decisions at WeChat Pay},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/57N4M6AT}},
  note         = {Machine review of arXiv:2608.03063}
}
read the original abstract

Merchant risk control at large payment platforms screens tens of millions of merchants daily, where false positives harm legitimate merchants and false negatives leave harmful activity undetected. The hardest cases require jointly understanding a merchant's textual profile and long behavioral sequence. Large language models (LLMs) excel at text but cannot natively model such sequences, while adapting them often causes catastrophic forgetting. We present SeqLLM, a framework that adds behavioral-sequence modeling to a pretrained LLM while preserving its language ability. SeqLLM combines three components: a compact discrete vocabulary that represents behavioral events as native tokens; a lightweight projector, trained with a two-stage alignment curriculum, that grounds these tokens in the LLM's semantic space; and prefix-guided capability injection, which acquires sequence-modeling ability through task-prefixed supervised fine-tuning rather than continual pre-training. SeqLLM is deployed at WeChat Pay, screening millions of merchants daily. Against the production DeepSeek-based LLM baseline, it raises screening precision from 92.0% to 97.5%. Its pretrained behavior-token embeddings also improve Precision@Top-0.01% by 26.8 percentage points in a production fraud detector serving billion-scale transaction traffic. Beyond payments, SeqLLM achieves state-of-the-art results on public recommendation benchmarks. On MovieLens and Amazon, it surpasses the strong User-LLM baseline by up to 32% relative Recall@5 while retaining markedly stronger language ability. On RecIF, it improves Pass@32 by 14.2% over the full OneRec-8B pipeline using only one-fifth of its GPU-days.

Figures

Figures reproduced from arXiv: 2608.03063 by the authors.

Figure 1
Figure 1. Overview of SeqLLM. (a) At inference, projected behavior tokens are interleaved with the task prefix and textual context for joint text–behavior reasoning. (b) Phase 1 grounds text-initialized behavior tokens through token translation and multi-event reasoning. Phase 2 jointly mixes task-prefixed sequence completion with general-instruction SFT, applying loss only to response tokens. This injects sequence modeling w… view at source ↗
Figure 2
Figure 2. Per-layer relative weight change after sample-aligned CPT and prefix-guided SFT training from the same Qwen3-8B backbone. 4.3 Multimodal Merchant Risk Modeling (RQ2) To answer RQ2, we compare controlled text-only, behavior-only, and joint variants that use the same SeqLLM architecture, training data, and supervision. This isolates whether text and behavioral sequences provide complementary evidence for merchant risk… view at source ↗
Figure 3
Figure 3. Online A/B gains from SeqLLM-initialized behavior em￾beddings. Blue bars show absolute percentage-point gains over the baseline; absolute values are withheld for business confidentiality. 4.5 Generalization and Comparison with Strong Public Baselines (RQ4) We assess generalization along three dimensions: MovieLens/Amazon test transfer from payment behavior to standard recommendation; RecIF provides a stage-matched c… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 14 canonical work pages

  1. [1]

    Bruce Croft

    Qingyao Ai, Yongfeng Zhang, Keping Bi, Xu Chen, and W. Bruce Croft. 2017. Learning a Hierarchical Embedding Model for Personalized Product Search. In Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval, Shinjuku, Tokyo, Japan, August 7-11, 2017, Noriko Kando, Tetsuya Sakai, Hideo Joho, Hang Li, A...

  2. [2]

    Jesús Bobadilla, Fernando Ortega, Antonio Hernando, and Abraham Gutiérrez

  3. [3]

    Davide Caffagni, Federico Cocchi, Luca Barsellotti, Nicholas Moratelli, Sara Sarto, Lorenzo Baraldi, Marcella Cornia, and Rita Cucchiara. 2024. The revolution of multimodal large language models: A survey.Findings of the association for computational linguistics: ACL 2024(2024), 13590–13618

  4. [4]

    Jiaxin Deng, Shiyao Wang, Kuo Cai, Lejian Ren, Qigen Hu, Weifeng Ding, Qiang Luo, and Guorui Zhou. 2025. Onerec: Unifying retrieve and rank with generative recommender and iterative preference alignment.arXiv preprint arXiv:2502.18965 (2025)

  5. [5]

    Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. 2022. Recommendation as Language Processing (RLP): A Unified Pretrain, Personalized Prompt & Predict Paradigm (P5). InRecSys ’22: Sixteenth ACM Conference on Recommender Systems, Seattle, WA, USA, September 18 - 23, 2022, Jennifer Golbeck, F. Maxwell Harper, Vanessa Murdock, Michael D...

  6. [6]

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2020. Measuring massive multitask language under- standing.arXiv preprint arXiv:2009.03300(2020)

  7. [7]

    Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk

  8. [8]

    Yuzhen Huang, Yuzhuo Bai, Zhihao Zhu, Junlei Zhang, Jinghan Zhang, Tangjun Su, Junteng Liu, Chuancheng Lv, Yikai Zhang, Yao Fu, et al . 2023. C-eval: A multi-level multi-discipline chinese evaluation suite for foundation models. Advances in neural information processing systems36 (2023), 62991–63010

Show all 31 references
  1. [9]

    Wang-Cheng Kang and Julian J. McAuley. 2018. Self-Attentive Sequential Recommendation. InIEEE International Conference on Data Mining, ICDM 2018, Singapore, November 17-20, 2018. IEEE Computer Society, 197–206. doi:10.1109/ICDM.2018.00035

  2. [10]

    Guilin Li, Yun Zhang, Xiuyuan Chen, Chengqi Li, Bo Wang, Linghe Kong, Wenjia Wang, Weiran Huang, and Matthias Hwai Yong Tan. 2025. PANTHER: Generative Pretraining Beyond Language for Sequential User Behavior Modeling.CoRR abs/2510.10102 (2025). arXiv:2510.10102 doi:10.48550/AR...

  3. [11]

    Jianghao Lin, Xinyi Dai, Yunjia Xi, Weiwen Liu, Bo Chen, Hao Zhang, Yong Liu, Chuhan Wu, Xiangyang Li, Chenxu Zhu, et al. 2025. How can recommender systems benefit from large language models: A survey.ACM Transactions on Information Systems43, 2 (2025), 1–47

  4. [12]

    Bin Liu, Chenxu Zhu, Guilin Li, Weinan Zhang, Jincai Lai, Ruiming Tang, Xi- uqiang He, Zhenguo Li, and Yong Yu. 2020. Autofis: Automatic feature interaction selection in factorization models for click-through rate prediction. Inproceedings of the 26th ACM SIGKDD international ...

  5. [13]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual instruction tuning.Advances in neural information processing systems36 (2023), 34892–34916

  6. [14]

    Zhanyu Liu, Shiyao Wang, Xingmei Wang, Rongzhou Zhang, Jiaxin Deng, Honghui Bao, Jinghao Zhang, Wuchao Li, Pengfei Zheng, Xiangyu Wu, Yifei Hu, Qigen Hu, Xinchen Luo, Lejian Ren, Zixing Zhang, Qianqian Wang, Kuo Cai, Yunfan Wu, Hongtao Cheng, Zexuan Cheng, Lu Ren, Huanjie Wang...

  7. [15]

    Eric WT Ngai, Yong Hu, Yiu Hing Wong, Yijun Chen, and Xin Sun. 2011. The application of data mining techniques in financial fraud detection: A classification framework and an academic review of literature.Decision support systems50, 3 (2011), 559–569

  8. [16]

    Lin Ning, Luyang Liu, Jiaxing Wu, Neo Wu, Devora Berlowitz, Sushant Prakash, Bradley Green, Shawn O’Banion, and Jun Xie. 2025. User-LLM: Efficient LLM Contextualization with User Embeddings. InCompanion Proceedings of the ACM on Web Conference 2025, WWW 2025, Sydney, NSW, Aust...

  9. [17]

    Clifton Phua, Vincent Lee, Kate Smith, and Ross Gayler. 2010. A compre- hensive survey of data mining-based fraud detection research.arXiv preprint arXiv:1009.6119(2010)

  10. [18]

    Tran, Jonah Samost, Maciej Kula, Ed H

    Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Hulikal Kesha- van, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Q. Tran, Jonah Samost, Maciej Kula, Ed H. Chi, and Mahesh Sathiamoorthy. 2023. Recom- mender Systems with Generative Retrieval. InAdvances in Neural In...

  11. [19]

    Steffen Rendle. 2010. Factorization machines. In2010 IEEE International confer- ence on data mining. IEEE, 995–1000

  12. [20]

    Fatemeh Sarvi, Nikos V oskarides, Lois Mooiman, Sebastian Schelter, and Maarten de Rijke. 2020. A comparison of supervised learning to match methods for product search.arXiv preprint arXiv:2007.10296(2020)

  13. [21]

    Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang. 2019. BERT4Rec: Sequential Recommendation with Bidirectional Encoder Representations from Transformer. InProceedings of the 28th ACM International Conference on Information and Knowledge Management, C...

  14. [22]

    OneRec Team. 2025. OpenOneRec Technical Report.CoRRabs/2512.24762 (2025). arXiv:2512.24762 doi:10.48550/ARXIV .2512.24762

  15. [23]

    OneRec Team. 2026. OneReason Technical Report.CoRRabs/2606.06260 (2026). arXiv:2606.06260 doi:10.48550/ARXIV .2606.06260

  16. [24]

    Qwen Team. 2025. Qwen3 Technical Report.CoRRabs/2505.09388 (2025). arXiv:2505.09388 doi:10.48550/ARXIV .2505.09388

  17. [25]

    Christophe Van Gysel, Maarten de Rijke, and Evangelos Kanoulas. 2016. Learning latent vector spaces for product search. InProceedings of the 25th ACM interna- tional on conference on information and knowledge management. 165–174

  18. [26]

    Shoujin Wang, Liang Hu, Yan Wang, Longbing Cao, Quan Z Sheng, and Mehmet Orgun. 2019. Sequential recommender systems: challenges, progress and prospects.arXiv preprint arXiv:2001.04830(2019)

  19. [27]

    Jiaqi Zhai, Lucy Liao, Xing Liu, Yueming Wang, Rui Li, Xuan Cao, Leon Gao, Zhaojie Gong, Fangda Gu, Jiayuan He, Yinghai Lu, and Yu Shi. 2024. Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Genera- tive Recommendations. InForty-first Internationa...

  20. [28]

    Duzhen Zhang, Yahan Yu, Jiahua Dong, Chenxing Li, Dan Su, Chenhui Chu, and Dong Yu. 2024. Mm-llms: Recent advances in multimodal large language models.Findings of the Association for Computational Linguistics: ACL 2024 (2024), 12401–12430

  21. [29]

    pets / cute animals

    Wanjun Zhong, Ruixiang Cui, Yiduo Guo, Yaobo Liang, Shuai Lu, Yanlin Wang, Amin Saied, Weizhu Chen, and Nan Duan. 2024. Agieval: A human-centric benchmark for evaluating foundation models. InFindings of the association for computational linguistics: NAACL 2024. 2299–2314. KDD ...

  22. [2013]

    Recommender systems survey.Knowledge-based systems46 (2013), 109– 132

  23. [2016]

    In4th International Conference on Learning Representations, ICLR 2016, San Juan, Puerto Rico, May 2-4, 2016, Conference Track Proceedings, Yoshua Bengio and Yann LeCun (Eds.)

    Session-based Recommendations with Recurrent Neural Networks. In4th International Conference on Learning Representations, ICLR 2016, San Juan, Puerto Rico, May 2-4, 2016, Conference Track Proceedings, Yoshua Bengio and Yann LeCun (Eds.). http://arxiv.org/abs/1511.06939

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.