Pith. sign in

REVIEW 4 major objections 5 minor 41 references

GREAT: Guiding Query Generation with a Trie for Recommending Related Search about Video at Kuaishou

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper proposes GREAT, a trie-guided generative LLM framework for item-to-query recommendation in video related search, reporting offline and online gains over retrieval and generation baselines.

desk verdict KuaiRS is the real contribution; the offline Edit@k advantage is likely inflated by trie construction from the same pool that holds the test labels. read the letter →

arxiv 2507.15267 v1 pith:BX2GNQ7D submitted 2025-07-21 cs.IR cs.CL

Ninglu Shao , Jinshan Wang , Chenxu Wang , Qingbiao Li , Xiaoxue Zang , Han Li This is my paper
classification cs.IRcs.CL
keywords LargeLanguageModelQueryRecommendationGenerationItem-to-QueryRelatedSearchTrie-basedDecodingShortVideoPlatformKuaiRS
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

Short-video platforms now show a search query under each video, letting viewers jump straight to search results; the paper studies this item-to-query (I2Q) recommendation task and argues it is best solved by constrained generation rather than by embedding retrieval or free-form LLM output. It releases KuaiRS, a one-million-pair real-world dataset of videos and their clicked queries, and introduces GREAT, which builds a trie from high-exposure, high-CTR queries and uses it twice in a fine-tuned LLM: as an auxiliary training objective (Next Token in Trie Prediction) and as a decoding constraint that only allows tokens that continue a valid trie path. Offline Edit@k scores and a five-day online A/B test on the Kuaishou app show GREAT ahead of SimCSE, BGE, and fine-tuned Qwen baselines on effectiveness, relevance, and literal quality. If the results hold, the recipe is a practical way to give viewers a safe, relevant search entry point that converts passive interest into active search.

What carries the argument

The query-based trie is the load-bearing structure: a prefix tree whose nodes are tokens of queries that had high exposure and high click-through rate and survived manual screening for typos, rumors, and sensitive content, refreshed on a 15-day rolling window. It functions as a constrained vocabulary. In training the NTTP objective maximizes the model's probability on the trie's child nodes, so the model learns to prefer proven query continuations; in inference, decoding is limited to the trie's path, eliminating off-trie hallucinations; and the same token probabilities feed the Logits Filter's global and local quality scores. This trie is what aligns training and inference with the platform's high-performing query space.

What would settle it

Take the same KuaiRS test set but build the trie only from queries that appeared at least a week before the test videos were logged, then recompute Edit@k; if GREAT's edge over fine-tuned Qwen shrinks or vanishes, the reported gain comes from the trie containing the ground-truth queries rather than from better learned relevance.

Watch

Extended reading notes

Core claim

The central claim is that a query-based trie built from proven queries is the right way to bring a large language model into item-to-query recommendation. GREAT takes Qwen 2.5 1.5B, fine-tunes it on video-query pairs from KuaiRS, and uses the trie twice: during training, a Next Token in Trie Prediction (NTTP) auxiliary loss asks the model to assign probability not only to the ground-truth next token but to the trie's child tokens for the current prefix; during inference, Trie-based Decoding restricts the candidate next tokens to those children, so every generated query is a path in the trie. A Logits Filter then keeps only queries whose average and minimum token probabilities exceed thresholds, pruning low-quality outputs. The paper claims this configuration improves on retrieval and unconstrained generators across effectiveness, relevance, and literal quality, and that the two trie components together contribute more than either alone.

Load-bearing premise

The offline comparison assumes it is fair to judge a model whose output is forced to come from the same list of queries that contains the correct answer against models that can write any phrasing; the forced model gets an advantage whenever the true query is in that list.

Editorial extensions

If this is right

  • I2Q recommendation can be served by a single fine-tuned LLM with a trie, eliminating the separate embedding index and two-tower retrieval pipeline.
  • The trie components are individually useful and complementary: ablations show removing either NTTP or Trie-based Decoding degrades Edit@k, and removing both returns to plain fine-tuning.
  • Relevance and literal quality can be traded off by design: without the trie, literal quality drops by a large margin while relevance rises, so the trie is the mechanism that keeps generated queries safe to show.
  • The released KuaiRS dataset gives a public benchmark for I2Q in video-related search, allowing other teams to compare retrieval and generation approaches on real exposure-filtered data.
  • A five-day online A/B test shows the method can be deployed in production, with consistent day-over-day gains on exposure, CTR, and search-page CTR.

Reading between the lines

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

  • Implication the paper leaves implicit: constraining generation to the trie caps output diversity at the trie's vocabulary, so the method may under-serve emerging, long-tail, or meme-driven queries until the 15-day refresh admits them.
  • A testable extension: instead of hard trie restriction, interpolate trie child logits with the full-vocabulary distribution to let the model emit novel phrasings while still preferring proven paths; the paper's w/o Trie-based Decoding result suggests this could recover the lost relevance.
  • The Edit@k metric rewards literal proximity to the single ground-truth query; a semantic evaluation (e.g., embedding cosine or human paraphrase agreement) would show whether GREAT's gains reflect better understanding or only constrained vocabulary.
  • Because the trie is built from high-CTR queries, the method inherits whatever popularity bias the exposure/CTR signal has; a fairness check on low-exposure videos would show whether the trie guide disadvantages niche content.
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. The paper introduces a new task formulation and dataset, KuaiRS, for item-to-query (I2Q) recommendation in video-related search, and proposes GREAT, an LLM-based framework in which a trie constructed from high-exposure, high-CTR queries constrains generation during both training (via an auxiliary Next-Token-in-Trie Prediction loss) and inference (via Trie-based Decoding and a Logits Filter). The authors report offline improvements on a new Edit@k metric, online A/B gains in exposure, CTR, and SRP CTR, and human-evaluation gains in relevance and literal quality, along with ablations of the proposed components.

Significance. If the claims hold, the paper makes a useful practical and community contribution: KuaiRS appears to be the first public dataset for this specific I2Q recommendation scenario, and the trie-guided generation design is a plausible way to combine LLM relevance with the strict literal-quality and safety constraints of an industrial short-video platform. The paper is also honest about the multi-objective nature of the problem (effectiveness, SRP consumption, relevance, literal quality). However, the central effectiveness claim rests on an offline comparison that is structurally asymmetric, and the online and ablation evidence is reported without statistical significance measures. The stress-test concern about the Edit@k comparison is valid on the manuscript as written.

major comments (4)
  1. [5.1.1, 4.2.2, 4.4.1, Table 2] The headline offline comparison is structurally asymmetric. Section 5.1.1 states that "all queries from KuaiRS serving as the query pool," and Section 4.2.2 builds the query-based trie from this pool, which contains the validation and test ground-truth queries. Trie-based Decoding in Section 4.4.1 restricts GREAT's next-token candidates to paths in this trie, so GREAT's outputs are necessarily strings from the same label space as the ground truth. The fine-tuned Qwen baseline, by contrast, can generate arbitrary strings. The Edit@k metric in Section 5.1.2 then rewards surface-form proximity to the ground truth: a semantically equivalent paraphrase such as "Warriors vs. Mavericks highlights" is heavily penalized, while the exact stored string "Mavericks vs. Warriors" scores zero edit distance. The reported Edit@k advantage therefore conflates the trie constraint with model quality. Please provide a comparison in which all methods have access to the same candidate space (for example, by mapping generative outputs onto the query pool or by giving retrieval baselines the same constrained decoding), and supplement Edit@k with a semantic similarity metric and human judgments.
  2. [5.3, Table 3] The online A/B results are supportive but not statistically substantiated. The reported gains are small (+0.251% exposure, +0.174% CTR, +0.396% SRP CTR), yet no confidence intervals, p-values, or other significance tests are provided. The human evaluation uses only 200 sampled exposures per group and reports no annotator agreement or variance. Given the 5-day duration and small effect sizes, these numbers could be within normal traffic noise; please report significance measures and day-level variability, and provide details of the human annotation procedure.
  3. [5.4, Tables 4 and 5] The ablation study does not establish that the observed differences are meaningful. In Table 4, differences between GREAT and its variants are as small as 0.03–0.14 in Edit@1–Edit@20, and no standard deviations, confidence intervals, or repeated-run results are given. Table 5 reports relative improvements on only 100 sampled videos, again without any uncertainty estimates. Without such measures, the claim that all components contribute positively is not yet supported.
  4. [4.3, Eq. (1)-(2)] The NTTP loss definition appears to have a sign issue. Equation (1) defines L_NTTP as a sum of log probabilities, and Equation (2) then minimizes L = L_NTP + alpha * L_NTTP. Since log probabilities are negative, minimizing this sum would drive the model to reduce the probability of trie tokens. A negative sign or an explicit maximization statement is needed, and the equations should be corrected to match the intended optimization.
minor comments (5)
  1. [5.3, Table 3] The phrase "Compared to the online baseline" is ambiguous because Table 3 reports relative values for several methods but does not define which method is the baseline; please state the control condition explicitly.
  2. [3.1] There is a typo, "serach results page," in the dataset overview paragraph.
  3. [5.3] The table reference "Tbale 3" should be "Table 3," and Figure 3 would benefit from error bars or confidence bands to support the stability claim.
  4. [3.2 and 7] The paper says KuaiRS is released, but the only availability statement is a GitHub URL; please clarify the license, access procedure, and whether the full dataset is actually downloadable.
  5. [2.2 and 4.3] The claim that fine-tuned LLMs "often have poor literal quality" is stated without quantitative evidence; a small analysis or example counts would make the motivation for the trie more concrete.

Circularity Check

1 steps flagged · score 6.0 of 10

Offline Edit@k advantage is partly by construction: the trie is built from the same KuaiRS query pool that contains the test labels, so Trie-based Decoding is scored against its own label space while unconstrained Qwen is penalized for surface mismatch.

  1. fitted input called prediction [Section 5.1.1 (Implementations), Section 5.1.2 (Metrics), Section 4.4.1 (Trie-based Decoding), Table 2]
    "In the offline experiments, we train and evaluate GREAT and other baselines on the KuaiRS dataset, with all queries from KuaiRS serving as the query pool. [...] The results of generative models may extend beyond the query pool, making it difficult to directly compare them with the ground truth. To address this issue, we propose a new evaluation metric called Edit@k. [...] Trie-based Decoding avoids utilizing the full vocabulary table and instead focuses on the child nodes of the current sequence in the query-based trie."

    The offline query pool is all of KuaiRS, which includes the test ground-truth queries. The trie is constructed from that same pool, so every label's token path is an allowed decoding path for GREAT. Trie-based Decoding therefore turns generation into constrained retrieval over the answer set, while the fine-tuned Qwen baseline can emit arbitrary paraphrases that Edit@k penalizes by surface edit distance. The label path itself is in the trie, so GREAT is not required to generalize beyond stored strings; it only needs to select among stored candidates. This asymmetry, not model quality alone, explains part of the Table 2 gap (Edit@1 4.34 vs 4.48). The retrieval baselines and the online A/B test are not affected by this mechanism.

full rationale

The central offline effectiveness comparison (Table 2) is not an unbiased test of GREAT versus unconstrained generation. Section 5.1.1 states that all queries from KuaiRS serve as the query pool, and the test ground truth is part of KuaiRS; Section 4.2.2 builds the trie from that pool; Section 4.4.1 restricts decoding to trie children. Edit@k then rewards exact or near-exact surface matches to the same labels. This structural alignment gives GREAT an advantage over the fine-tuned Qwen baseline that is independent of query quality, so part of the reported offline improvement reduces by construction. I do not find a load-bearing self-citation chain: the same-team references are contextual, and no uniqueness theorem is invoked. The online A/B test, human relevance/literal-quality evaluation, and the comparisons against retrieval baselines are external evidence and are not circular, although they are reported without confidence intervals or significance tests. Because one of the paper's headline 'predictions' is partly forced by the label-containing trie plus Edit@k, the circularity score is 6 rather than 0-2.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

The central claims rest on several hand-chosen hyperparameters and on domain assumptions about the relationship between CTR, query quality, and edit distance. No new physical or conceptual entities are introduced; the trie is a standard data structure. The most consequential assumption is that Edit@k is a fair metric, which is undermined by the structural alignment between the trie and the ground-truth label set.

free parameters (6)
  • alpha (NTTP loss weight) = 0.1
    Hyperparameter controlling the weight of the NTTP auxiliary loss during training (Section 5.1.1).
  • theta_G (Global Filter threshold) = 0.2
    Threshold for global logits-based query filtering in online experiments (Section 5.1.1).
  • theta_L (Local Filter threshold) = 0.05
    Threshold for local token-level filtering in online experiments (Section 5.1.1).
  • MBVR similarity threshold = 0.44
    Threshold for filtering low-relevance video-query pairs during dataset construction, chosen 'based on experience' (Section 3.2).
  • Exposure and click thresholds = exposure >= 1000, clicks >= 10
    Criteria for selecting video-query pairs in the dataset (Section 3.2).
  • Query pool time window = 15 days
    Time window for the query pool used to build the trie (Section 4.2.2).
assumptions (3)
  • domain assumption LLM logits can serve as a reliable proxy for query relevance and literal quality.
    Section 4.4.2 states: 'This mechanism is based on a key assumption that well-trained LLMs have incorporated the knowledge pertaining to I2Q recommendation in related search, enabling them to accurately select the next token.'
  • domain assumption High-exposure and high-CTR queries are the appropriate target distribution for I2Q recommendation.
    Section 4.2.2 uses exposure and click-through rate to select the query pool, assuming these signals define query quality.
  • ad hoc to paper Edit distance to ground-truth queries is a valid measure of recommendation quality.
    Section 5.1.2 defines Edit@k as the evaluation metric without validating it against human relevance judgments; lexical distance may not reflect semantic quality.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GREAT: Guiding Query Generation with a Trie for Recommending Related Search about Video at Kuaishou." pith.science (2026). https://pith.science/paper/BX2GNQ7D

@misc{pith2026250715267,
  author       = {Pith},
  title        = {Pith review of: GREAT: Guiding Query Generation with a Trie for Recommending Related Search about Video at Kuaishou},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BX2GNQ7D}},
  note         = {Machine review of arXiv:2507.15267}
}
read the original abstract

Currently, short video platforms have become the primary place for individuals to share experiences and obtain information. To better meet users' needs for acquiring information while browsing short videos, some apps have introduced a search entry at the bottom of videos, accompanied with recommended relevant queries. This scenario is known as query recommendation in video-related search, where core task is item-to-query (I2Q) recommendation. As this scenario has only emerged in recent years, there is a notable scarcity of academic research and publicly available datasets in this domain. To address this gap, we systematically examine the challenges associated with this scenario for the first time. Subsequently, we release a large-scale dataset derived from real-world data pertaining to the query recommendation in video-\textit{\textbf{r}}elated \textit{\textbf{s}}earch on the \textit{\textbf{Kuai}}shou app (\textbf{KuaiRS}). Presently, existing methods rely on embeddings to calculate similarity for matching short videos with queries, lacking deep interaction between the semantic content and the query. In this paper, we introduce a novel LLM-based framework named \textbf{GREAT}, which \textit{\textbf{g}}uides que\textit{\textbf{r}}y g\textit{\textbf{e}}ner\textit{\textbf{a}}tion with a \textit{\textbf{t}}rie to address I2Q recommendation in related search. Specifically, we initially gather high-quality queries with high exposure and click-through rate to construct a query-based trie. During training, we enhance the LLM's capability to generate high-quality queries using the query-based trie. In the inference phase, the query-based trie serves as a guide for the token generation. Finally, we further refine the relevance and literal quality between items and queries via a post-processing module. Extensive offline and online experiments demonstrate the effectiveness of our proposed method.

Figures

Figures reproduced from arXiv: 2507.15267 by the authors.

Figure 1
Figure 1. I2Q recommendation in related search at Kuaishou [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The framework of GREAT. During the data preparation, we construct prompts for LLMs’ inputs and build a query-based [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. The performance of GREAT in online A/B test on [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 18 canonical work pages

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floren- cia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)

    arXiv 2023

  2. [2]

    Yu, Yin-Wen Chang, Yiming Yang, and Sanjiv Kumar

    Wei-Cheng Chang, Felix X. Yu, Yin-Wen Chang, Yiming Yang, and Sanjiv Kumar

    work page

  3. [3]

    Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu

  4. [4]

    Paul Covington, Jay Adams, and Emre Sargin. 2016. Deep neural networks for youtube recommendations. In Proceedings of the 10th ACM conference on recommender systems. 191–198

    2016

  5. [5]

    Sunhao Dai, Ninglu Shao, Haiyuan Zhao, Weijie Yu, Zihua Si, Chen Xu, Zhongx- iang Sun, Xiao Zhang, and Jun Xu. 2023. Uncovering chatgpt’s capabilities in recommender systems. In Proceedings of the 17th ACM Conference on Recom- mender Systems. 1126–1132

    work page 2023

  6. [6]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783 (2024)

    arXiv 2024

  7. [7]

    Wenqi Fan, Yujuan Ding, Liangbo Ning, Shijie Wang, Hengyun Li, Dawei Yin, Tat-Seng Chua, and Qing Li. 2024. A survey on rag meeting llms: Towards retrieval-augmented large language models. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 6491–6501

    2024

  8. [8]

    Turnbull, and Junshuo Liao

    Malay Haldar, Mustafa Abdool, Prashant Ramanathan, Tyler Sax, Lanbo Zhang, Aamir Mansawala, Shulin Yang, Bradley C. Turnbull, and Junshuo Liao. 2020. Improving Deep Learning for Airbnb Search.Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (2020). https: //api.semanticscholar.org/CorpusID:211096624

    work page 2020

Show all 41 references
  1. [9]

    Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. 2017. Neural collaborative filtering. In Proceedings of the 26th international conference on world wide web . 173–182

    2017

  2. [10]

    Jian Jia, Yipei Wang, Yan Li, Honggang Chen, Xuehan Bai, Zhaocheng Liu, Jian Liang, Quan Chen, Han Li, Peng Jiang, and Kun Gai. 2024. Knowledge Adaptation from Large Language Model to Recommendation for Practical Industrial Appli- cation. ArXiv abs/2405.03988 (2024). https://a...

    2024 arXiv

  3. [11]

    Mingyu Jin, Hua Tang, Chong Zhang, Qinkai Yu, Chengzhi Liu, Suiyuan Zhu, Yongfeng Zhang, and Mengnan Du. 2024. Time Series Forecasting with LLMs: Understanding and Enhancing Model Capabilities. ArXiv abs/2402.10835 (2024). https://api.semanticscholar.org/CorpusID:267740358

    2024 arXiv

  4. [12]

    Jeff Johnson, Matthijs Douze, and Hervé Jégou. 2017. Billion-Scale Similarity Search with GPUs. IEEE Transactions on Big Data 7 (2017), 535–547. https: //api.semanticscholar.org/CorpusID:926364

    2017

  5. [13]

    Christoph Kofler, Martha Larson, and Alan Hanjalic. 2016. User intent in multime- dia search: a survey of the state of the art and future challenges.ACM Computing Surveys (CSUR) 49, 2 (2016), 1–37

    2016

  6. [14]

    Hoyeop Lee, Jinbae Im, Seongwon Jang, Hyunsouk Cho, and Sehee Chung. 2019. MeLU: Meta-Learned User Preference Estimator for Cold-Start Recommendation. Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (2019). https://api.semantics...

    2019

  7. [15]

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. 2022. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning . PMLR, 12888–12900

    2022

  8. [16]

    Venkatesh Balavadhani Parthasarathy, Ahtsham Zafar, Aafaq Iqbal khan, and Arsalan Shahid. 2024. The Ultimate Guide to Fine-Tuning LLMs from Basics to Breakthroughs: An Exhaustive Review of Technologies, Research, Best Practices, Applied Research Challenges and Opportunities. A...

    2024 arXiv

  9. [17]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learni...

    2021

  10. [18]

    Juan Enrique Ramos. 2003. Using TF-IDF to Determine Word Relevance in Document Queries. https://api.semanticscholar.org/CorpusID:14638345

    2003

  11. [19]

    Robertson and Hugo Zaragoza

    Stephen E. Robertson and Hugo Zaragoza. 2009. The Probabilistic Relevance Framework: BM25 and Beyond. Found. Trends Inf. Retr. 3 (2009), 333–389. https: //api.semanticscholar.org/CorpusID:207178704

    2009

  12. [20]

    Pranab Sahoo, Ayush Kumar Singh, Sriparna Saha, Vinija Jain, Samrat Sohel Mondal, and Aman Chadha. 2024. A Systematic Survey of Prompt Engineering in Large Language Models: Techniques and Applications. ArXiv abs/2402.07927 (2024). https://api.semanticscholar.org/CorpusID:267636769

    2024 arXiv

  13. [21]

    Zihua Si, Zhongxiang Sun, Xiao Zhang, Jun Xu, Xiaoxue Zang, Yang Song, Kun Gai, and Ji-Rong Wen. 2023. When search meets recommendation: Learning disentangled search representation for recommendation. InProceedings of the 46th International ACM SIGIR Conference on Research and...

    2023

  14. [22]

    ZhongXiang Sun, Zihua Si, Xiaoxue Zang, Dewei Leng, Yanan Niu, Yang Song, Xiao Zhang, and Jun Xu. 2023. KuaiSAR: A Unified Search And Recommendation Dataset. Proceedings of the 32nd ACM International Conference on Information and Knowledge Management (2023). https://api.semant...

    2023

  15. [23]

    Hongyan Tang, Junning Liu, Ming Zhao, and Xudong Gong. 2020. Progressive Layered Extraction (PLE): A Novel Multi-Task Learning (MTL) Model for Personal- ized Recommendations. Proceedings of the 14th ACM Conference on Recommender Systems (2020). https://api.semanticscholar.org/...

    2020

  16. [24]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yas- mine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhos- ale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)

    2023 arXiv

  17. [25]

    Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N

    Ashish Vaswani, Noam M. Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is All you Need. In Neural Information Processing Systems . https://api.semanticscholar.org/ CorpusID:13756489

    2017

  18. [26]

    Xun Wang, Bingqing Ke, Xuanping Li, Fangyu Liu, Mingyu Zhang, Xiao Liang, and Qiushi Xiao. 2022. Modality-balanced embedding for video retrieval. In Proceedings of the 45th international ACM SIGIR conference on research and devel- opment in information retrieval . 2578–2582

    2022

  19. [27]

    Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jin- sheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. 2024. Emu3: Next-Token Prediction is All You Need. arXiv preprint arXiv:2409.18869 (2024)

    2024 arXiv

  20. [28]

    Yin wei Wei, Xiang Wang, Qi Li, Liqiang Nie, Yan Li, Xuanping Li, and Tat seng Chua. 2021. Contrastive Learning for Cold-Start Recommendation. Proceedings of the 29th ACM International Conference on Multimedia (2021). https://api. semanticscholar.org/CorpusID:235794926

    2021

  21. [29]

    Jules White, Quchen Fu, Sam Hays, Michael Sandborn, Carlos Olea, Henry Gilbert, Ashraf Elnashar, Jesse Spencer-Smith, and Douglas C. Schmidt. 2023. A Prompt Pattern Catalog to Enhance Prompt Engineering with ChatGPT. ArXiv abs/2302.11382 (2023). https://api.semanticscholar.org...

    2023 arXiv

  22. [30]

    Shitao Xiao, Zheng Liu, Weihao Han, Jianjin Zhang, Chaozhuo Li, Yingxia Shao, Defu Lian, Xing Xie, Hao Sun, Denvy Deng, Liangjie Zhang, and Qi Zhang. 2022. Progressively Optimized Bi-Granular Document Representation for Scalable Embedding Based Retrieval. Proceedings of the AC...

    2022

  23. [31]

    Shitao Xiao, Zheng Liu, Yingxia Shao, Tao Di, Bhuvan Middha, Fangzhao Wu, and Xing Xie. 2021. Training Large-Scale News Recommenders with Pretrained Language Models in the Loop. Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (2021). https:...

    2021

  24. [32]

    Le Xue, Manli Shu, Anas Awadalla, Jun Wang, An Yan, Senthil Purushwalkam, Honglu Zhou, Viraj Prabhu, Yutong Dai, Michael S Ryoo, et al . 2024. xgen- mm (blip-3): A family of open large multimodal models. arXiv preprint arXiv:2408.08872 (2024)

    2024

  25. [33]

    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 preprint arXiv:2412.15115 (2024)

    2024 arXiv

  26. [34]

    Xiaoyong Yang, Yadong Zhu, Yi Zhang, Xiaobo Wang, and Quan Yuan. 2020. Large Scale Product Graph Construction for Recommendation in E-commerce. ArXiv abs/2010.05525 (2020). https://api.semanticscholar.org/CorpusID:222291351

    2020 arXiv

  27. [35]

    Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. 2024. A survey on multimodal large language models. National Science Review (2024), nwae403

    2024

  28. [36]

    Jiaqi Zhai, Lucy Liao, Xing Liu, Yueming Wang, Rui Li, Xuan Cao, Leon Gao, Zhaojie Gong, Fangda Gu, Michael He, Yin-Hua Lu, and Yu Shi. 2024. Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Genera- tive Recommendations. ArXiv abs/2402.17152 (2024...

    2024 arXiv

  29. [37]

    Shengyu Zhang, Linfeng Dong, Xiaoya Li, Sen Zhang, Xiaofei Sun, Shuhe Wang, Jiwei Li, Runyi Hu, Tianwei Zhang, Fei Wu, et al. 2023. Instruction tuning for large language models: A survey. arXiv preprint arXiv:2308.10792 (2023)

    2023

  30. [38]

    Han Zhu, Pengye Zhang, Guozheng Li, Jie He, Han Li, and Kun Gai. 2018. Learning Tree-based Deep Model for Recommender Systems. Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (2018). https://api.semanticscholar.org/CorpusID:25648541

    2018

  31. [39]

    Yutao Zhu, Kun Zhou, Kelong Mao, Wentong Chen, Yiding Sun, Zhipeng Chen, Qian Cao, Yihan Wu, Yushuo Chen, Feng Wang, et al. 2024. Yulan: An open-source large language model. arXiv preprint arXiv:2406.19853 (2024)

    2024 arXiv

  32. [2020]

    ArXiv abs/2002.03932 (2020)

    Pre-training Tasks for Embedding-based Large-scale Retrieval. ArXiv abs/2002.03932 (2020). https://api.semanticscholar.org/CorpusID:211068995

    2020 arXiv

  33. [2024]

    In Annual Meeting of the Association for Computational Linguistics

    BGE M3-Embedding: Multi-Lingual, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation. In Annual Meeting of the Association for Computational Linguistics . https://api.semanticscholar.org/ CorpusID:267413218

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.