pith. sign in

arxiv: 2605.16474 · v1 · pith:PDZIM2VJnew · submitted 2026-05-15 · 💻 cs.IR · cs.AI

LERA: LLM-Enhanced RAG for Ad Auction in Generative Chatbots

Haoran Sun , Xinrui Song , Xinyu Zhang , Zhaohua Chen , Xu Chu , Zhilin Zhang , Chuan Yu , Jian Xu
show 2 more authors
Bo Zheng Xiaotie Deng
This is my paper

Pith reviewed 2026-05-19 21:57 UTC · model grok-4.3

classification 💻 cs.IR cs.AI
keywords LLM chatbotsad auctionsretrieve-then-generaterelevance logitstruthful mechanismscritical-value paymentsRAG advertising
0
0 comments X

The pith

LERA combines embedding filters with LLM logits for relevance and applies a critical-value payment to run truthful ad auctions inside generative chatbots.

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

The paper proposes LERA to insert ads into LLM-based chatbots without relying only on text embeddings. Pure embedding methods risk commercial misinterpretation and repetitive insertions. LERA first uses embeddings for coarse candidate filtering, then prompts the LLM to output logits that serve as refined organic relevance scores. These scores are combined with bids under a payment rule that accounts for both filtering and ranking thresholds. The design supports multiple ad insertions in long responses and maintains truthfulness for utility-maximizing advertisers while showing gains in accuracy and diversity on synthetic benchmarks.

Core claim

LERA is a two-stage retrieve-then-generate auction framework for LLM chatbots. Embedding-based coarse filtering pre-selects a small set of candidate advertisers. The LLM is then queried with a carefully designed prompt to produce logits over candidates that serve as refined organic relevance scores. These scores are combined with bids, and a critical-value payment rule accounts for both the coarse-filtering and fine-ranking thresholds to ensure truthfulness for utility-maximizing advertisers. The framework extends to multiple ad insertions within dynamic dialogue flows and long responses.

What carries the argument

The critical-value payment rule that incorporates both coarse-filtering and fine-ranking thresholds when combining LLM logits with bids to enforce truthful bidding.

If this is right

  • The mechanism extends naturally to multiple ad insertions within ongoing dialogue flows and long responses.
  • Ad selection accuracy and insertion diversity improve substantially on synthetic advertiser-query benchmarks.
  • Latency overhead stays controllable despite the additional LLM query in the second stage.
  • Commercial misinterpretation and repetitive insertions decrease relative to pure text-embedding retrieval.

Where Pith is reading between the lines

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

  • If the logit scores remain stable across different LLMs and prompt variations, the same scoring step could be reused for non-advertising ranking tasks inside generative systems.
  • Real-world deployment would require measuring whether the improved selection translates into higher user engagement or satisfaction metrics.
  • Advertisers might develop new bidding heuristics once they know the payment explicitly depends on both the coarse and fine thresholds.

Load-bearing premise

Logits produced by a carefully designed prompt to the LLM constitute reliable and unbiased organic relevance scores that can be directly combined with bids without introducing new commercial misinterpretation or repetitive insertion problems.

What would settle it

A controlled test in which advertisers submit bids that deviate from their true values and the resulting payments fail to make truth-telling the dominant strategy, or where the LLM logits produce more repetitive insertions than an embedding-only baseline.

Figures

Figures reproduced from arXiv: 2605.16474 by Bo Zheng, Chuan Yu, Haoran Sun, Jian Xu, Xiaotie Deng, Xinrui Song, Xinyu Zhang, Xu Chu, Zhaohua Chen, Zhilin Zhang.

Figure 1
Figure 1. Figure 1: Comparison of ad selection strategies for LLM-based chatbots. Each panel shows a different [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Latency and throughput tradeoff. Comparison of LERA and its ablation variants under different concurrency levels using 30B and 235B Gen Models. Each point corresponds to a concurrency configuration and illustrates the efficiency cost of more expensive auction mechanisms. 0 5 10 15 20 K (Number of Retrieved Candidates) 50 55 60 65 70 75 80 Avg Satisfaction (%) Qwen3-4B 0 5 10 15 20 K (Number of Retrieved Ca… view at source ↗
read the original abstract

The integration of advertising auction mechanisms into large language model (LLM)-based chatbots presents a significant opportunity for commercialization, yet poses unique challenges in balancing relevance, efficiency, and user experience. Recently, Feizi et al.~\citep{feizi2023online} and Hajiaghayi et al.~\citep{hajiaghayi2024ad} outlined a retrieve-then-generate paradigm that decouples retrieval and generation, offering lightweight ad insertion and payment determination. However, current retrieval relies solely on text embedding similarity, which may lead to commercial misinterpretation and issues such as repetitive insertions. In this paper, we propose LERA, a two-stage retrieve-then-generate auction framework tailored for LLM chatbots. In the first stage, embedding-based coarse filtering pre-selects a small set of candidate advertisers. In the second stage, the LLM itself is queried with a carefully designed prompt to produce logits over candidates, which serve as refined organic relevance scores. These scores are combined with bids, and a critical-value payment rule accounts for both the coarse-filtering and fine-ranking thresholds, ensuring truthfulness for utility-maximizing advertisers. The framework naturally extends to multiple ad insertions within dynamic dialogue flows and long responses. Experiments on a synthetic advertiser-query benchmark show that LERA substantially improves ad selection accuracy and insertion diversity while incurring only controllable latency overhead.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes LERA, a two-stage retrieve-then-generate auction for ad insertion in LLM chatbots. Stage 1 uses embedding similarity for coarse filtering of candidate advertisers; Stage 2 queries the LLM with a designed prompt to obtain logits as organic relevance scores, which are combined with bids. A critical-value payment rule is stated to account for both the coarse-filtering and fine-ranking thresholds, ensuring truthfulness for utility-maximizing advertisers. The framework extends to multiple insertions in dynamic dialogues. Experiments on a synthetic advertiser-query benchmark report gains in ad selection accuracy and insertion diversity with controllable latency.

Significance. If the incentive-compatibility claim is formally established and the accuracy/diversity gains prove robust, LERA would offer a concrete mechanism for truthful ad auctions inside generative chatbots that mitigates commercial misinterpretation and repetitive insertion problems associated with pure embedding retrieval. The explicit use of LLM logits as a second-stage relevance signal is a potentially useful extension of prior retrieve-then-generate paradigms.

major comments (2)
  1. [Payment rule description (abstract and §3)] The central claim that the critical-value payment rule ensures truthfulness rests on an unshown derivation: no equation, proof sketch, or reduction to a VCG-style critical bid is supplied that explicitly combines the embedding threshold and the LLM-logit threshold into a well-defined infimum bid for selection. This is load-bearing for the truthfulness guarantee stated in the abstract and methods.
  2. [Experiments section] Table 2 (or equivalent experimental table): accuracy and diversity improvements are reported on the synthetic benchmark, yet no error bars, statistical tests, or ablation on the LLM prompt design are provided, and comparisons are limited to baselines that do not incorporate LLM signals.
minor comments (2)
  1. [Method details] Clarify whether the LLM prompt is fixed and bid-independent; if logits are treated as constants, the critical-value calculation reduces to a standard single-threshold form and should be stated explicitly.
  2. [Abstract] The abstract cites Feizi et al. and Hajiaghayi et al. but does not spell out the precise technical differences in the payment rule or the two-stage selection.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We address each major comment point by point below and indicate the revisions planned for the next version of the manuscript.

read point-by-point responses

  1. Referee: [Payment rule description (abstract and §3)] The central claim that the critical-value payment rule ensures truthfulness rests on an unshown derivation: no equation, proof sketch, or reduction to a VCG-style critical bid is supplied that explicitly combines the embedding threshold and the LLM-logit threshold into a well-defined infimum bid for selection. This is load-bearing for the truthfulness guarantee stated in the abstract and methods.

    Authors: We agree that an explicit derivation is required to substantiate the truthfulness claim. In the revised manuscript we will add a formal proof sketch in Section 3. The critical-value payment is defined as the infimum bid b* such that the advertiser passes the embedding coarse filter and its combined score (bid + LLM logit) exceeds the fine-ranking threshold against all other candidates. This reduces to a generalized VCG-style critical bid that accounts for the two-stage filtering; we will supply the corresponding equation and a brief argument showing that truthful bidding is dominant for utility-maximizing advertisers. revision: yes

  2. Referee: [Experiments section] Table 2 (or equivalent experimental table): accuracy and diversity improvements are reported on the synthetic benchmark, yet no error bars, statistical tests, or ablation on the LLM prompt design are provided, and comparisons are limited to baselines that do not incorporate LLM signals.

    Authors: We acknowledge that the experimental presentation can be strengthened. In the revision we will add error bars computed over multiple independent runs of the synthetic benchmark and include paired t-tests (or equivalent) to assess statistical significance of the accuracy and diversity gains. We will also insert an ablation subsection that varies the LLM prompt design and reports its effect on selection quality. The current baselines were deliberately chosen as standard embedding-only retrieve-then-generate methods to isolate the contribution of LLM logits; we will expand the text to justify this choice and, space permitting, add at least one LLM-augmented baseline for completeness. revision: partial

Circularity Check

0 steps flagged

No significant circularity detected in the derivation chain

full rationale

The paper proposes LERA as a two-stage retrieve-then-generate auction that combines embedding-based coarse filtering with LLM logit-based fine ranking, then applies a critical-value payment rule to ensure truthfulness. The abstract and description present this rule as an extension of standard mechanism-design ideas from the cited external works (Feizi et al. and Hajiaghayi et al.), without reducing the truthfulness claim to a self-definition, a fitted parameter renamed as prediction, or a load-bearing self-citation. No equations or steps in the provided text exhibit the patterns of self-definitional construction or ansatz smuggling; the central claim retains independent content as a proposed adaptation rather than a tautological restatement of inputs. The derivation is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The framework rests on the domain assumption that LLM logits are faithful proxies for organic relevance and on the modeling choice that a critical-value rule can be extended to two successive thresholds without losing incentive compatibility. No free parameters are explicitly named in the abstract, but the coarse and fine thresholds function as tunable cutoffs. No new entities are postulated.

free parameters (2)
  • coarse-filtering threshold
    Determines the size of the candidate set passed to the LLM stage; must be chosen to balance recall and latency.
  • fine-ranking threshold
    Used inside the critical-value payment calculation; its value directly affects the reported truthfulness guarantee.
axioms (2)
  • domain assumption LLM logits from a carefully designed prompt accurately reflect organic relevance without commercial bias
    Invoked when the second-stage scores are treated as inputs to the auction rule.
  • domain assumption The synthetic advertiser-query benchmark is representative of real user-advertiser interactions
    Required for the accuracy and diversity claims to generalize.

pith-pipeline@v0.9.0 · 5799 in / 1394 out tokens · 48003 ms · 2026-05-19T21:57:22.540052+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

66 extracted references · 66 canonical work pages · 3 internal anchors

  1. [1]

    Position auctions in ai-generated content.arXiv preprint arXiv:2506.03309, 2025

    Santiago Balseiro, Kshipra Bhawalkar, Yuan Deng, Zhe Feng, Jieming Mao, Aranyak Mehta, Vahab Mirrokni, Renato Paes Leme, Di Wang, and Song Zuo. Position auctions in ai-generated content.arXiv preprint arXiv:2506.03309, 2025

    work page arXiv 2025

  2. [2]

    Uncovering chatgpt’s capabilities in recommender systems

    Sunhao Dai, Ninglu Shao, Haiyuan Zhao, Weijie Yu, Zihua Si, Chen Xu, Zhongxiang Sun, Xiao Zhang, and Jun Xu. Uncovering chatgpt’s capabilities in recommender systems. InProceedings of the 17th ACM Conference on Recommender Systems, pages 1126–1132, 2023

    work page 2023

  3. [3]

    Deepseek-v3.2: Pushing the frontier of open large language models, 2025

    DeepSeek-AI. Deepseek-v3.2: Pushing the frontier of open large language models, 2025

    work page 2025

  4. [4]

    Auctions with llm summaries

    Avinava Dubey, Zhe Feng, Rahul Kidambi, Aranyak Mehta, and Di Wang. Auctions with llm summaries. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD ’24, page 713–722, New York, NY , USA, 2024. Association for Computing Machinery

    work page 2024

  5. [5]

    Mechanism design for large language models

    Paul Duetting, Vahab Mirrokni, Renato Paes Leme, Haifeng Xu, and Song Zuo. Mechanism design for large language models. InProceedings of the ACM Web Conference 2024, pages 144–155, 2024

    work page 2024

  6. [6]

    Internet advertising and the generalized second-price auction: Selling billions of dollars worth of keywords.American economic review, 97(1):242–259, 2007

    Benjamin Edelman, Michael Ostrovsky, and Michael Schwarz. Internet advertising and the generalized second-price auction: Selling billions of dollars worth of keywords.American economic review, 97(1):242–259, 2007

    work page 2007

  7. [7]

    On- line advertisements with llms: Opportunities and challenges.arXiv preprint arXiv:2311.07601, 2023

    Soheil Feizi, MohammadTaghi Hajiaghayi, Keivan Rezaei, and Suho Shin. Online advertise- ments with llms: Opportunities and challenges.arXiv preprint arXiv:2311.07601, 2023

    work page arXiv 2023

  8. [8]

    Retrieval-Augmented Generation for Large Language Models: A Survey

    Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yixin Dai, Jiawei Sun, Haofen Wang, and Haofen Wang. Retrieval-augmented generation for large language models: A survey.arXiv preprint arXiv:2312.10997, 2(1), 2023. 10

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  9. [9]

    Ad auctions for llms via retrieval augmented generation.Advances in Neural Information Processing Systems, 37:18445–18480, 2024

    MohammadTaghi Hajiaghayi, Sébastien Lahaie, Keivan Rezaei, and Suho Shin. Ad auctions for llms via retrieval augmented generation.Advances in Neural Information Processing Systems, 37:18445–18480, 2024

    work page 2024

  10. [10]

    Auction theory: A guide to the literature.Journal of economic surveys, 13(3):227–286, 1999

    Paul Klemperer. Auction theory: A guide to the literature.Journal of economic surveys, 13(3):227–286, 1999

    work page 1999

  11. [11]

    Retrieval-augmented generation for knowledge-intensive nlp tasks.Advances in neural information processing systems, 33:9459–9474, 2020

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks.Advances in neural information processing systems, 33:9459–9474, 2020

    work page 2020

  12. [12]

    Solving quan- titative reasoning problems with language models.Advances in neural information processing systems, 35:3843–3857, 2022

    Aitor Lewkowycz, Anders Andreassen, David Dohan, Ethan Dyer, Henryk Michalewski, Vinay Ramasesh, Ambrose Slone, Cem Anil, Imanol Schlag, Theo Gutman-Solo, et al. Solving quan- titative reasoning problems with language models.Advances in neural information processing systems, 35:3843–3857, 2022

    work page 2022

  13. [13]

    Leadre: Multi-faceted knowledge enhanced llm empowered display advertisement recommender system.arXiv preprint arXiv:2411.13789, 2024

    Fengxin Li, Yi Li, Yue Liu, Chao Zhou, Yuan Wang, Xiaoxiang Deng, Wei Xue, Dapeng Liu, Lei Xiao, Haijie Gu, et al. Leadre: Multi-faceted knowledge enhanced llm empowered display advertisement recommender system.arXiv preprint arXiv:2411.13789, 2024

    work page arXiv 2024

  14. [14]

    Large language models for generative recommendation: A survey and visionary discussions

    Lei Li, Yongfeng Zhang, Dugang Liu, and Li Chen. Large language models for generative recommendation: A survey and visionary discussions. InProceedings of the 2024 joint international conference on computational linguistics, language resources and evaluation (LREC-COLING 2024), pages 10146–10159, 2024

    work page 2024

  15. [15]

    Learning-based ad auction design with externalities: the frame- work and a matching-based approach

    Ningyuan Li, Yunxuan Ma, Yang Zhao, Zhijian Duan, Yurong Chen, Zhilin Zhang, Jian Xu, Bo Zheng, and Xiaotie Deng. Learning-based ad auction design with externalities: the frame- work and a matching-based approach. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 1291–1302, 2023

    work page 2023

  16. [16]

    Deep automated mechanism design for integrating ad auction and allocation in feed

    Xuejian Li, Ze Wang, Bingqi Zhu, Fei He, Yongkang Wang, and Xingxing Wang. Deep automated mechanism design for integrating ad auction and allocation in feed. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1211–1220, 2024

    work page 2024

  17. [17]

    arXiv preprint arXiv:2304.10149 , year=

    Junling Liu, Chao Liu, Peilin Zhou, Renjie Lv, Kang Zhou, and Yan Zhang. Is chatgpt a good recommender? a preliminary study.arXiv preprint arXiv:2304.10149, 2023

    work page arXiv 2023

  18. [18]

    Neural auction: End-to-end learning of auction mechanisms for e-commerce advertising

    Xiangyu Liu, Chuan Yu, Zhilin Zhang, Zhenzhe Zheng, Yu Rong, Hongtao Lv, Da Huo, Yiqing Wang, Dagui Chen, Jian Xu, et al. Neural auction: End-to-end learning of auction mechanisms for e-commerce advertising. InProceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, pages 3354–3364, 2021

    work page 2021

  19. [19]

    Sponsored question answering

    Tommy Mordo, Moshe Tennenholtz, and Oren Kurland. Sponsored question answering. In Proceedings of the 2024 ACM SIGIR International Conference on Theory of Information Retrieval, pages 167–173, 2024

    work page 2024

  20. [20]

    Training language models to follow instructions with human feedback.Advances in neural information processing systems, 35:27730–27744, 2022

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback.Advances in neural information processing systems, 35:27730–27744, 2022

    work page 2022

  21. [21]

    Limitations of language models in arithmetic and symbolic induction

    Jing Qian, Hong Wang, Zekun Li, Shiyang Li, and Xifeng Yan. Limitations of language models in arithmetic and symbolic induction. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), pages 9285–9298, 2023

    work page 2023

  22. [22]

    Applying large language models to sponsored search advertising.Marketing Science, 2025

    Martin Reisenbichler, Thomas Reutterer, and David A Schweidel. Applying large language models to sponsored search advertising.Marketing Science, 2025

    work page 2025

  23. [23]

    Optimal auctions.The American Economic Review, 71(3):381–392, 1981

    John G Riley and William F Samuelson. Optimal auctions.The American Economic Review, 71(3):381–392, 1981. 11

    work page 1981

  24. [24]

    Maximizing rag efficiency: A comparative analysis of rag methods.Natural Language Processing, 31(1):1–25, 2025

    Tolga ¸ Sakar and Hakan Emekci. Maximizing rag efficiency: A comparative analysis of rag methods.Natural Language Processing, 31(1):1–25, 2025

    work page 2025

  25. [25]

    arXiv preprint arXiv:2405.05905 , year=

    Ermis Soumalias, Michael J Curry, and Sven Seuken. Truthful aggregation of llms with an application to online advertising.arXiv preprint arXiv:2405.05905, 2024

    work page arXiv 2024

  26. [26]

    Genai advertising: Risks of personalizing ads with llms.arXiv preprint arXiv:2409.15436, 2024

    Brian Jay Tang, Kaiwen Sun, Noah T Curran, Florian Schaub, and Kang G Shin. Genai advertising: Risks of personalizing ads with llms.arXiv preprint arXiv:2409.15436, 2024

    work page arXiv 2024

  27. [27]

    Gemini: A Family of Highly Capable Multimodal Models

    Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  28. [28]

    Qwen3 technical report, 2025

    Qwen Team. Qwen3 technical report, 2025

    work page 2025

  29. [29]

    Position auctions.international Journal of industrial Organization, 25(6):1163– 1178, 2007

    Hal R Varian. Position auctions.international Journal of industrial Organization, 25(6):1163– 1178, 2007

    work page 2007

  30. [30]

    arXiv preprint arXiv:2601.19435 , year=

    Shengwei Xu, Zhaohua Chen, Xiaotie Deng, Zhiyi Huang, and Grant Schoenebeck. Ad insertion in llm-generated responses.arXiv preprint arXiv:2601.19435, 2026

    work page arXiv 2026

  31. [31]

    Against opac- ity: Explainable ai and large language models for effective digital advertising

    Qi Yang, Marlo Ongpin, Sergey Nikolenko, Alfred Huang, and Aleksandr Farseev. Against opac- ity: Explainable ai and large language models for effective digital advertising. InProceedings of the 31st ACM International Conference on Multimedia, pages 9299–9305, 2023

    work page 2023

  32. [32]

    Optimizing multiple performance metrics with deep gsp auctions for e-commerce advertising

    Zhilin Zhang, Xiangyu Liu, Zhenzhe Zheng, Chenrui Zhang, Miao Xu, Junwei Pan, Chuan Yu, Fan Wu, Jian Xu, and Kun Gai. Optimizing multiple performance metrics with deep gsp auctions for e-commerce advertising. InProceedings of the 14th ACM International Conference on Web Search and Data Mining, pages 993–1001, 2021

    work page 2021

  33. [33]

    LLM-Auction: Generative Auction towards LLM-Native Advertising

    Chujie Zhao, Qun Hu, Shiping Song, Dagui Chen, Han Zhu, Jian Xu, and Bo Zheng. Llm- auction: Generative auction towards llm-native advertising.arXiv preprint arXiv:2512.10551, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  34. [34]

    Recommender systems in the era of large language models (llms).IEEE Transactions on Knowledge and Data Engineering, 36(11):6889–6907, 2024

    Zihuai Zhao, Wenqi Fan, Jiatong Li, Yunqing Liu, Xiaowei Mei, Yiqi Wang, Zhen Wen, Fei Wang, Xiangyu Zhao, Jiliang Tang, et al. Recommender systems in the era of large language models (llms).IEEE Transactions on Knowledge and Data Engineering, 36(11):6889–6907, 2024

    work page 2024

  35. [35]

    Sglang: Efficient execution of structured language model programs.Advances in neural information processing systems, 37:62557–62583, 2024

    Lianmin Zheng, Liangsheng Yin, Zhiqiang Xie, Chuyue Sun, Jeff Huang, Cody H Yu, Shiyi Cao, Christos Kozyrakis, Ion Stoica, Joseph E Gonzalez, et al. Sglang: Efficient execution of structured language model programs.Advances in neural information processing systems, 37:62557–62583, 2024

    work page 2024

  36. [36]

    right to speak

    Ruitao Zhu, Yangsu Liu, Dagui Chen, Zhenjia Ma, Chufeng Shi, Zhenzhe Zheng, Jie Zhang, Jian Xu, Bo Zheng, and Fan Wu. Contextual generative auction with permutation-level externalities for online advertising. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V . 1, pages 2171–2181, 2025. A Appendix Overview This append...

    work page 2025

  37. [37]

    Extract keywords representing general product or service categories, such as coffee_maker, travel_bag, orfitness_app

    work page

  38. [38]

    Account for the user’s needs and the conversation flow

    work page

  39. [39]

    If context is provided, avoid duplicating products or categories that have already been mentioned

    work page

  40. [40]

    Output only comma-separated keywords, with no explanation. Output format.keyword_1, keyword_2, keyword_3 E.2 Stage 2: LLM Scoring Prompt The second-stage prompt evaluates the candidate setˆS and asks the LLM to select the most appropriate advertiser or the no-insertion optionϕ. The probability assigned to each choice is then used to compute the LLM-based ...

    work page

  41. [41]

    LavaThin Potato Chips (Snacks)

    work page

  42. [43]

    They’re bold enough to wake up the taste buds during a mid-afternoon work slump

    EverCool Desk Fan (Home & Living) Model response.For crunchy snacks with bold flavors, I’d recommend trying LavaThin Potato Chips— these paper-thin, kettle-cooked chips pack an intense chili punch that builds with each bite, perfect for someone who loves heat with a smoky habanero finish. They’re bold enough to wake up the taste buds during a mid-afternoo...

    work page

  43. [44]

    EverCool Desk Fan (Home & Living)

    work page

  44. [45]

    NimbusLight Desk Lamp (Home & Living)

    work page

  45. [46]

    A soothing herbal tea blend like chamomile-lavender or peppermint- ginger will help her unwind after long screen sessions

    PureCycle HEPA Air Purifier (Home & Living) Model response.For a snack, try wasabi peas or spicy sriracha roasted chickpeas—they’re crunchy, bold, and perfect for mid-afternoon energy. A soothing herbal tea blend like chamomile-lavender or peppermint- ginger will help her unwind after long screen sessions. To keep her workspace fresh and comfortable, a co...

    work page

  46. [47]

    SwiftBlend Portable Blender (Home & Living)

    work page

  47. [49]

    Since she’s health-conscious and often on the move, the SwiftBlend Portable Blender would be a fantastic addition, letting her whip up refreshing iced tea infusions in seconds

    FunkBrew Ginger Kombucha (Beverages) 23 Model response.For her herbal tea enjoyment, consider including a selection of caffeine-free blends like chamomile, peppermint, and rooibos—perfect for staying relaxed and focused during work hours. Since she’s health-conscious and often on the move, the SwiftBlend Portable Blender would be a fantastic addition, let...

    work page

  48. [50]

    SparkLeaf Botanical Tea (Beverages)

    work page

  49. [51]

    GlazePeak Dried Fruits (Snacks)

    work page

  50. [52]

    DawnOats Instant Oatmeal Cups (Food & Dining) Model response.For herbal teas, I’d recommend SparkLeaf Botanical Tea—its lightly carbonated, unsweetened blends with exotic botanicals offer a refreshing, caffeine-free pick-me-up perfect for long workdays. When she’s craving something sweet and satisfying between meetings, GlazePeak Dried Fruits are a perfec...

    work page

  51. [55]

    Its relaxed yet refined look transitions effortlessly from seaside strolls to dinner on the boardwalk

    BerryTint SPF Lip Balm (Skincare) Model response.The BreezeLinen Summer Shirt is an ideal choice for your beach getaway—crafted from lightweight, natural linen that’s designed to breathe and wick moisture, keeping you cool even in high humidity. Its relaxed yet refined look transitions effortlessly from seaside strolls to dinner on the boardwalk. To compl...

    work page

  52. [58]

    Its moisture-wicking fabric stays cool against the skin, offering superior breathability without sacrificing style

    RiverGrip Hiking Sandals (Footwear) Model response.For a lightweight linen shirt that handles humidity beautifully, consider the BreezeLinen Summer Shirt, designed with a relaxed yet refined fit that transitions effortlessly from beach strolls to dinner by the shore. Its moisture-wicking fabric stays cool against the skin, offering superior breathability ...

    work page

  53. [59]

    BreezeLinen Summer Shirt (Tops)

    work page

  54. [60]

    BreezeVest Packable Down Vest (Tops)

    work page

  55. [61]

    CottonCrest Crewneck T-Shirt (Tops) Model response.The BreezeLinen Summer Shirt is an ideal choice for your beach getaway—its lightweight, open-weave fabric allows maximum airflow, keeping you cool and comfortable even in high humidity. For cooler evenings by the shore, consider packing the BreezeVest Packable Down Vest—it delivers cozy, lightweight warmt...

    work page

  56. [62]

    SilkBody Shea Body Lotion (Skincare)

    work page

  57. [63]

    SunVeil SPF 50 Face Moisturizer (Skincare)

    work page

  58. [64]

    To keep your face protected and feeling fresh, try SunVeil SPF 50 Face Moisturizer—a lightweight daily formula that combines broad-spectrum mineral protection with hydration

    BerryTint SPF Lip Balm (Skincare) Model response.For your beach days and sun-kissed skin, don’t forget to care for your body with the same thoughtfulness—SilkBody Shea Body Lotion deeply nourishes with shea butter and ceramides, absorbing 25 quickly without leaving a heavy residue. To keep your face protected and feeling fresh, try SunVeil SPF 50 Face Moi...

    work page

  59. [66]

    AeroTech Performance Tee (Tops)

    work page

  60. [67]

    Their durable design enhances balance and control, making them ideal for everything from squats to agility drills in the gym

    FlexFuel Chocolate Protein Shake (Beverages) Model response.For your cross-training needs, the StableGrip Cross-Trainers offer the perfect blend of support and stability, with a flat sole that ensures solid footing during lifts and lateral movements. Their durable design enhances balance and control, making them ideal for everything from squats to agility...

    work page

  61. [70]

    AeroTech Performance Tee (Tops) Model response.For a plant-based protein shake that actually satisfies, try the FlexFuel Chocolate Protein Shake—its rich, creamy texture and 30g of protein make it a standout choice for post-workout recovery without any artificial aftertaste. Pair your shake with the StableGrip Cross-Trainers, engineered with a flat, stabl...

    work page

  62. [72]

    NutRidge Protein Granola Bars (Snacks)

    work page

  63. [73]

    GlazePeak Dried Fruits (Snacks) Model response.For a plant-based protein shake that tastes rich and satisfying without any artificial aftertaste, try the FlexFuel Chocolate Protein Shake—its creamy texture and 30g of protein make it perfect for recovery after a tough gym session. When you’re heading to the gym or need a quick boost between sets, pack a Nu...

    work page

  64. [74]

    StableGrip Cross-Trainers (Footwear)

    work page

  65. [75]

    FlexFuel Chocolate Protein Shake (Beverages)

    work page

  66. [76]

    PeakFlow Running Shoes (Footwear) Model response.For your cross-training needs, I highly recommend the StableGrip Cross-Trainers— they’re designed with a flat, stable sole that provides excellent support during weightlifting and lateral movements, making them ideal for the varied demands of a gym workout. After pushing hard, refuel with a FlexFuel Chocola...

    work page