Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

RTBAgent: A LLM-based Agent System for Real-Time Bidding

T0 review · 4 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read LLM bidding agent beats RL baselines on all three budgets

desk verdict A plausible first LLM-agent-for-RTB framework whose headline gains are too small and too dependent on oracle feedback in the offline evaluation to be believed as-is. read the letter →

arxiv 2502.00792 v1 pith:NJQJY4FY submitted 2025-02-02 cs.AI

Leng Cai , Junxuan He , Yikai Li , Junjie Liang , Yuanping Lin , Ziming Quan , Yawen Zeng , Jin Xu This is my paper · ORCID
classification cs.AI
keywords Real-TimeBiddingLLMagentBidoptimizationMulti-memoryretrievalTwo-stepdecision-makingClick-throughratepredictionSecond-priceauctionInterpretability
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

The paper proposes RTBAgent, an LLM-based agent for real-time bidding in second-price ad auctions. Its central claim is that an agent that decides through a two-step reasoning process — insight then action — and keeps three kinds of memory (environment, bidding, reflection) can adjust a baseline bid factor to win more clicks under a fixed daily budget than standard rule-based and reinforcement-learning bidders. The bid is formed as $b_i^t = v_i^t \cdot \lambda_{base} \cdot (1 + a_t)$, where $v_i^t$ comes from a CTR estimator, $\lambda_{base}$ from an expert strategy, and $a_t$ is the LLM's adjustment. On the iPinYou dataset, the paper reports the highest click counts at 1/2, 1/8, and 1/32 budgets, plus interpretive traces that a panel of advertising experts rated positively. A sympathetic reader should care because this recasts bidding optimization as agent memory-and-reflection design rather than policy training.

What carries the argument

The load-bearing mechanism is a two-step decision loop built around the scaling identity $\lambda_t = \lambda_{base} \cdot (1 + a_t)$. In the first step, insight reasoning, the LLM analyses the current state, summarized memories, and the expert recommendation, and produces reasoning over ten adjustment intervals; in the second step, action making, it selects the adjustment $a_t$ and a rationale. Memory is split into environment memory, bidding memory, and reflection memory, and a daily reflection step compresses the day's experience into summarized memory for later retrieval. The CTR estimator (Factorization Machines) supplies the impression value $v_i^t$, and the expert strategy supplies $\lambda_{base}$, so the LLM only has to choose a bounded correction rather than a bid price from scratch.

What would settle it

Run the same iPinYou replay with feedback masked: hide the market price on lost auctions and delay outcome signals by one or more time steps before they enter memory; if click counts against LP and DRLB drop to overlap, the memory/reflection mechanism is not the source of the gain. A field A/B test with live auction traffic would be the decisive check.

Watch

Extended reading notes

Core claim

On its own terms, the discovery is that equipping an LLM with a CTR predictor, expert bidding strategies, multi-memory retrieval, and a two-step decision loop produces a practical RTB bidder that outperforms both static rule-based methods (MCPC, LIN, LP, ORTB) and RL methods (RLB, DRLB, USCB, DiffBid) in offline replay on iPinYou: 2,281 vs 2,275 clicks at half budget, 1,240 vs 1,235 at eighth budget, and 795 vs 793 at 1/32 budget. The agent never outputs a price directly; it outputs an adjustment in $[-0.5, 0.5]$ to a known bidding factor, thereby anchoring to expert strategy and avoiding the LLM's weakness at producing continuous values. Its multi-memory system records the market state, past bids with reasons, and daily reflections, and the two-step decision process separates market analysis from action selection. Ablations show that adding expert knowledge helps, two-step decision-making beats direct action, and the performance gain persists across several open and closed LLMs.

Load-bearing premise

The offline evaluation lets RTBAgent observe the exact realized outcome of every bid during the test day, while static baselines receive no such feedback; if a real bidder cannot see losing auction prices or gets feedback late, the reported edge may not transfer online.

Editorial extensions

If this is right

  • Every expert strategy tested (MCPC, LIN, LP) gained clicks when wrapped in RTBAgent, with the largest relative gains at the tightest budget, up to +12.75% over MCPC at 1/32.
  • The two-step insight-then-action design beats direct action by the same LLM, so separating market analysis from bid selection improves the final outcome.
  • RTBAgent works across closed and open LLMs of different sizes, so the reported benefit does not depend on one specific model.
  • The agent emits a reason for each bid, giving operations teams a transparent decision trace that rule-based and RL bidders do not provide.
  • The policy is formulated as an MDP whose objective explicitly conditions on a memory-and-reasoning summary $\rho_t$, making the memory loop part of the optimization target.

Reading between the lines

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

  • An unstated consequence is that this design turns bidding optimization into configuration of prompts, tools, and memory, so future RTB systems could improve by enriching memory structure or adding new tools without retraining a policy.
  • A testable extension would replace the single per-step adjustment with per-impression or per-segment adjustments; the paper's formulation only allows one $a_t$ per time step.
  • If outcome feedback is delayed or censored online, the reported gains may shrink; evaluating RTBAgent under partial observability, such as hiding losing auction prices, would separate the value of reflection from the value of seeing exact results.
  • The competitive framing suggests that multi-agent settings, where several RTBAgents bid against each other, may change behavior qualitatively; the paper names this direction as future work.
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 proposes RTBAgent, an LLM-based agent for real-time bidding in second-price ad auctions. The agent combines a CTR prediction model and rule-based bidding strategies as tools, a two-step (insight then action) decision process, and three types of memory with daily reflection to adjust a multiplicative factor on a base bid. Experiments on the iPinYou dataset compare click counts under three budget levels against rule-based (MCPC, LIN, LP, ORTB), RL (RLB, DRLB, USCB), and diffusion (DiffBid) baselines, reporting the highest click counts for RTBAgent. Ablations show gains from the expert strategy and the two-step process, and a small human evaluation rates the generated decisions positively.

Significance. If the empirical claim were established, the paper would be a useful demonstration that LLM agents can add interpretable, adaptive bid adjustment on top of standard RTB algorithms, and the modular design may extend to other campaign objectives. The public code, the ablation across LLMs, and the decision-trace visualization are strengths that support reproducibility and future work. However, the reported performance advantage is small and the evaluation protocol gives the agent access to feedback that the static baselines do not receive, so the significance of the central claim cannot be assessed from the current evidence.

major comments (4)
  1. [§4.2, Algorithm 1] The offline simulation supplies the realized market price, win/loss outcome, and click label for every impression to RTBAgent as it iterates the test set; Algorithm 1 line 17 stores the resulting state and action in memory, and §3.5.1 states that each action and environmental feedback record is saved in real time. The static baselines (MCPC, LIN, LP, ORTB) and the fixed trained RL/diffusion policies receive no such feedback during evaluation. Because the improvements over the strongest baseline in Table 2 are only 6, 5, and 2 clicks at the three budget levels, the reported gains may be an artifact of test-set outcome feedback rather than of the bidding strategy itself. Please provide an ablation in which memory and reflection are restricted to information observable in a real second-price auction (e.g., own bid and win/loss for won auctions only, no losing clearing price, delayed or aggregated feedback), or use a logged-bandit replay that respects the missing counterfactuals.
  2. [Table 2, §4.4] Table 2 and Table 4 report single click counts without variance, number of seeds, or significance tests. Since the gap to the strongest baseline (DiffBid) is 6, 5, and 2 clicks across the three budget levels, and LLM decoding is stochastic, these differences are within plausible run-to-run noise. Please report means and standard deviations over repeated runs (at least three to five seeds) and a paired significance test (e.g., bootstrap over campaigns/days or a Wilcoxon test) for each budget level. The word "significantly" in the abstract and in §4.4 is currently unsupported.
  3. [Table 2 vs. Tables 3–4] The paper does not specify the LLM and expert-strategy configuration used for the headline RTBAgent row in Table 2. Table 3 shows three expert strategies with Llama-3-8B, and Table 4 shows six LLMs with LP, but the Table 2 row (2,281 / 1,240 / 795) matches only the GPT-3.5-Turbo-1106 with LP row. Please state the exact configuration used for Table 2, and if it is the best configuration obtained by selection, describe a selection procedure that does not use the test set.
  4. [§3.5.3, Algorithm 1 (lines 18–20)] Daily reflection is applied inside the test phase. For the iPinYou test data, which span the final three days of each campaign, reflection on day k updates memory that is used on day k+1. This means decisions on later test days exploit outcomes from earlier test days, while static baselines do not learn across days. Even if within-day feedback is realistic, this cross-day adaptation must be reported separately (e.g., a day-by-day breakdown and a comparison in which the baselines receive the same degree of online adaptation); otherwise the cumulative clicks over the three test days conflate online learning with static bidding ability.
minor comments (5)
  1. [Figure 3 caption] The caption refers to "FP+" and "FP", while the text and the surrounding discussion refer to "LP+" and "LP"; please correct this typo.
  2. [Eqs. (7) and (12)] The summation symbol is used to denote string concatenation, which is nonstandard and confusing; please use an explicit concatenation operator (e.g., ⊕) or describe the operation in prose.
  3. [§3.4.3 and §8.1] The action space is described as a continuous range [-0.5, 0.5], but the prompt template in the appendix discretizes it into ten half-open intervals; please clarify how the LLM's bin selection is mapped to the continuous adjustment factor a_t used in Eq. (11).
  4. [§4.1] The sentence "all monetary values are in RMB, corresponding to the cost-per-thousand-impressions (CPM) pricing model" is confusing because the auction is described as per-impression second-price; please rephrase to avoid implying CPM pricing.
  5. [§4.7 and §4.6] The latency claim in §4.7 is not supported by any measured inference time, and the human evaluation in §4.6 reports a 97% positive rating without specifying how many decisions were rated or how the -1/0/1 scores were aggregated; please provide these details or soften the claims.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: RTBAgent's bid formula is externally anchored and baselines are independent; the offline feedback asymmetry is an evaluation concern, not a derivation-cycle.

full rationale

RTBAgent's bid derivation is externally anchored: Eq. (11) sets b_i^t = v_i^t * lambda_base * (1 + a_t), where v_i^t comes from a shared FM CTR estimator, lambda_base comes from expert strategies fitted on the train set (Section 4.3 states that the derivation of lambda_base by RTBAgent "is exclusively derived from the train set"), and a_t is an LLM-chosen adjustment action, not a fitted parameter. No fitted quantity is renamed as a prediction, and no equation reduces the reported click counts to an input quantity by construction. The comparison baselines (MCPC, LIN, LP, ORTB, RLB, DRLB, USCB, DiffBid) are external methods, not outputs of RTBAgent's own framework. The only co-author citation in the text ([16], Keyu Pan and Yawen Zeng) appears in a general list of LLM question-answering work and is not load-bearing. The offline evaluation does feed realized win/loss and click outcomes into the agent's memory during the simulated day, which raises a legitimate evaluation-protocol concern about oracle feedback and online transfer, but that is not a circularity in the derivation chain: the LLM is not fitting a parameter to the test labels, and the bid formula does not collapse into a baseline by definition. The limitation passage ("we have deployed this system into a real-world advertising bidding environment, and we plan to disclose the revenue in the future") is a missing-evidence caveat, not a circular step. Therefore no circularity is identified.

Assumptions & free parameters 5 free parameters · 7 assumptions · 0 invented entities

The central claim rests on standard second-price auction theory, the validity of the iPinYou dataset as a market proxy, the accuracy of the FM CTR predictor, and several ad hoc design choices (adjustment range, time step, prompt templates, two-step structure) that are justified only by ablation or not at all. The offline simulation feedback loop is the most consequential untested assumption.

free parameters (5)
  • adjustment_range = [-0.5, 0.5] (10 bins)
    The action space of the LLM, chosen by hand, constrains the maximum bid adjustment to ±50% of the expert strategy baseline (Section 3.4.3).
  • time_steps_per_day = 24 (hourly)
    The MDP time step is set to one hour, a modeling choice that determines the frequency of memory updates and adjustments (Section 4.2).
  • budget_fractions = 1/2, 1/8, 1/32 of total historical test cost
    Evaluation budgets chosen to test different budget constraints; these affect the click counts and comparison (Section 4.2).
  • llm_temperature
    The LLM sampling temperature is not reported, so the stochasticity of the action selection is uncontrolled, affecting reproducibility (Section 4.3).
  • prompt_templates
    The behavior of the agent is largely determined by the hand-crafted prompts, which are only partially shown in the appendix (Section 8.1).
assumptions (7)
  • standard math Optimal bid factor b = lambda * v in second-price auctions (Zhang et al. [30]).
    Invoked in Eq. (2) and used as the basis for the bid formula throughout the paper.
  • domain assumption Clicks are the objective KPI; conversions could be used but are not.
    Section 3.1.1 states the optimization is for clicks, which simplifies the evaluation to click counts rather than a business-validated metric.
  • domain assumption iPinYou dataset (2013, 9 campaigns, 10 days) is a valid proxy for real RTB market dynamics.
    Section 4.1 relies on this dataset without discussing its age or representativeness for current RTB markets.
  • domain assumption FM-based CTR estimate v_i is an accurate estimate of impression value for bidding.
    Section 3.4.2 uses FM for all methods; no calibration or uncertainty of the CTR estimate is analyzed.
  • ad hoc to paper The adjustment range [-0.5, 0.5] and the 10 bins are sufficient to capture optimal bid adjustments.
    Section 3.4.3 defines this action space without justification or sensitivity analysis.
  • ad hoc to paper The two-step decision making (insight then action) improves performance.
    Section 3.5.2 proposes it and Table 5 gives an ablation, but there is no theoretical or mechanistic explanation for why the decomposition helps.
  • ad hoc to paper Offline simulation with logged test data and immediate win/loss/click feedback is a faithful approximation of online bidding.
    Section 4.2 uses the test set as the environment, feeding outcomes back to the agent during the simulated day. This is a strong assumption given the known online-offline discrepancy in RTB.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RTBAgent: A LLM-based Agent System for Real-Time Bidding." pith.science (2026). https://pith.science/paper/NJQJY4FY

@misc{pith2026250200792,
  author       = {Pith},
  title        = {Pith review of: RTBAgent: A LLM-based Agent System for Real-Time Bidding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NJQJY4FY}},
  note         = {Machine review of arXiv:2502.00792}
}
read the original abstract

Real-Time Bidding (RTB) enables advertisers to place competitive bids on impression opportunities instantaneously, striving for cost-effectiveness in a highly competitive landscape. Although RTB has widely benefited from the utilization of technologies such as deep learning and reinforcement learning, the reliability of related methods often encounters challenges due to the discrepancies between online and offline environments and the rapid fluctuations of online bidding. To handle these challenges, RTBAgent is proposed as the first RTB agent system based on large language models (LLMs), which synchronizes real competitive advertising bidding environments and obtains bidding prices through an integrated decision-making process. Specifically, obtaining reasoning ability through LLMs, RTBAgent is further tailored to be more professional for RTB via involved auxiliary modules, i.e., click-through rate estimation model, expert strategy knowledge, and daily reflection. In addition, we propose a two-step decision-making process and multi-memory retrieval mechanism, which enables RTBAgent to review historical decisions and transaction records and subsequently make decisions more adaptive to market changes in real-time bidding. Empirical testing with real advertising datasets demonstrates that RTBAgent significantly enhances profitability. The RTBAgent code will be publicly accessible at: https://github.com/CaiLeng/RTBAgent.

Figures

Figures reproduced from arXiv: 2502.00792 by the authors.

Figure 1
Figure 1. The workflow of our RTBAgent, which is equipped with 4 tools, 3 types of memory, two-step decision-making to [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. An example of RTBAgent using LLMs to assist reasoning process. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Comparison of the remaining budget and CPC be [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. VLM-3D:End-to-End Vision-Language Models for Open-World 3D Perception

    cs.CV 2025-08 reject novelty 4.0 of 10

    The paper promises VLM-3D but the body text is entirely the MVISU-Bench mobile-agent benchmark paper, so the stated result is unsupported.

Reference graph

Works this paper leans on

34 extracted references · 17 canonical work pages · cited by 1 Pith paper

  1. [1]

    Kareem Amin, Michael Kearns, Peter Key, and Anton Schwaighofer. 2012. Budget optimization for sponsored search: Censored learning in mdps. arXiv preprint arXiv:1210.4847 (2012)

    arXiv 2012

  2. [2]

    Han Cai, Kan Ren, Weinan Zhang, Kleanthis Malialis, Jun Wang, Yong Yu, and Defeng Guo. 2017. Real-time bidding by reinforcement learning in display adver- tising. In Proceedings of the tenth ACM international conference on web search and data mining. 661–670

    work page 2017

  3. [3]

    Yirong Chen, Xiaofen Xing, Jingkai Lin, Huimin Zheng, Zhenyu Wang, Qi Liu, and Xiangmin Xu. 2023. Soulchat: Improving llms’ empathy, listening, and comfort abilities through fine-tuning with multi-turn empathy conversations. In Findings of the Association for Computational Linguistics: EMNLP 2023. 1170–1183

    work page 2023

  4. [4]

    George B Dantzig. 2002. Linear programming. Operations research 50, 1 (2002), 42–47

    work page 2002

  5. [5]

    Zihan Ding and Hao Dong. 2020. Challenges of reinforcement learning. Deep Reinforcement Learning: Fundamentals, Research and Applications(2020), 249–272

    work page 2020

  6. [6]

    Jiayan Guo, Yusen Huo, Zhilin Zhang, Tianyu Wang, Chuan Yu, Jian Xu, Bo Zheng, and Yan Zhang. 2024. Generative Auto-bidding via Conditional Diffusion Modeling. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Dis- covery and Data Mining (Barcelona, Spain) (KDD ’24). Association for Computing Machinery, New York, NY, USA, 5038–5049. https://do...

    doi:10.1145/3637528 2024

  7. [7]

    Izzeddin Gur, Hiroki Furuta, Austin Huang, Mustafa Safdari, Yutaka Matsuo, Dou- glas Eck, and Aleksandra Faust. 2023. A real-world webagent with planning, long context understanding, and program synthesis. arXiv preprint arXiv:2307.12856 (2023)

    arXiv 2023

  8. [8]

    Yue He, Xiujun Chen, Di Wu, Junwei Pan, Qing Tan, Chuan Yu, Jian Xu, and Xiaoqiang Zhu. 2021. A unified solution to constrained bidding in online display advertising. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining . 2993–3001

    work page 2021

Show all 34 references
  1. [9]

    Sirui Hong, Xiawu Zheng, Jonathan Chen, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, et al . 2023. Metagpt: Meta programming for multi-agent collaborative framework. arXiv preprint arXiv:2308.00352 (2023)

    2023 arXiv

  2. [10]

    Dong Huang, Qingwen Bu, Jie M Zhang, Michael Luck, and Heming Cui. 2023. Agentcoder: Multi-agent-based code generation with iterative testing and opti- misation. arXiv preprint arXiv:2312.13010 (2023)

    2023 arXiv

  3. [11]

    IAB and PwC. 2023. Internet Advertising Revenue Report 2022. In Interactive Ad- vertising Bureau and PwC. IAB and PwC, 1–31. https://www.iab.com/wp-content/ uploads/2023/04/IAB_PwC_Internet_Advertising_Revenue_Report_2022.pdf

    2023

  4. [12]

    Kuang-chih Lee, Burkay Birant Orten, Ali Dasdan, and Wentong Li. 2018. Esti- mating conversion rate in display advertising from past performance data. US Patent 10,037,543

    2018

  5. [13]

    Junkai Li, Siyu Wang, Meng Zhang, Weitao Li, Yunghwei Lai, Xinhui Kang, Weizhi Ma, and Yang Liu. 2024. Agent hospital: A simulacrum of hospital with evolvable medical agents. arXiv preprint arXiv:2405.02957 (2024)

    2024 arXiv

  6. [14]

    Hairen Liao, Lingxiao Peng, Zhenchuan Liu, and Xuehua Shen. 2014. iPinYou global rtb bidding algorithm competition dataset. In Proceedings of the Eighth International Workshop on Data Mining for Online Advertising . 1–6

    2014

  7. [15]

    Xiao-Yang Liu, Guoxuan Wang, and Daochen Zha. 2023. Fingpt: Democra- tizing internet-scale data for financial large language models. arXiv preprint arXiv:2307.10485 (2023)

    2023 arXiv

  8. [16]

    Keyu Pan and Yawen Zeng. 2023. Do LLMs Possess a Personality? Mak- ing the MBTI Test an Amazing Evaluation for Large Language Models. arXiv:2307.16180 [cs.CL]

    2023 arXiv

  9. [17]

    Joon Sung Park, Joseph O’Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. 2023. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th annual acm symposium on user interface software and technology . 1–22

    2023

  10. [18]

    Claudia Perlich, Brian Dalessandro, Rod Hook, Ori Stitelman, Troy Raeder, and Foster Provost. 2012. Bid optimizing and inventory scoring in targeted online advertising. In Proceedings of the 18th ACM SIGKDD international conference on Knowledge discovery and data mining . 804–812

    2012

  11. [19]

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

    2010

  12. [20]

    Tim Roughgarden. 2010. Algorithmic game theory. Commun. ACM 53, 7 (2010), 78–86

    2010

  13. [21]

    Jun Wang and Shuai Yuan. 2015. Real-time bidding: A new frontier of com- putational advertising research. In Proceedings of the Eighth ACM International Conference on Web Search and Data Mining . 415–416

    2015

  14. [22]

    Jun Wang, Weinan Zhang, Shuai Yuan, et al . 2017. Display advertising with real-time bidding (RTB) and behavioural targeting. Foundations and Trends® in Information Retrieval 11, 4-5 (2017), 297–435

    2017

  15. [23]

    Di Wu, Xiujun Chen, Xun Yang, Hao Wang, Qing Tan, Xiaoxun Zhang, Jian Xu, and Kun Gai. 2018. Budget constrained bidding by model-free reinforcement learning in display advertising. In Proceedings of the 27th ACM International Conference on Information and Knowledge Management ...

    2018

  16. [24]

    Songlin Xu, Xinyu Zhang, and Lianhui Qin. 2024. EduAgent: Generative Student Agents in Learning. arXiv preprint arXiv:2404.07963 (2024)

    2024 arXiv

  17. [25]

    Hui Yang, Sifu Yue, and Yunzhong He. 2023. Auto-gpt for online decision making: Benchmarks and additional opinions. arXiv preprint arXiv:2306.02224 (2023)

    2023 arXiv

  18. [26]

    Hao Yu, Michael Neely, and Xiaohan Wei. 2017. Online convex optimization with stochastic constraints. Advances in Neural Information Processing Systems 30 (2017)

    2017

  19. [27]

    Hao Yu and Michael J Neely. 2020. A low complexity algorithm with 𝑂( √ 𝑇) regret and𝑂(1) constraint violations for online convex optimization with long term constraints. Journal of Machine Learning Research 21, 1 (2020), 1–24

    2020

  20. [28]

    Weizhe Yuan, Richard Yuanzhe Pang, Kyunghyun Cho, Sainbayar Sukhbaatar, Jing Xu, and Jason Weston. 2024. Self-rewarding language models. arXiv preprint arXiv:2401.10020 (2024)

    2024 arXiv

  21. [29]

    Hongbo Zhang, Junying Chen, Feng Jiang, Fei Yu, Zhihong Chen, Jianquan Li, Guiming Chen, Xiangbo Wu, Zhiyi Zhang, Qingying Xiao, et al. 2023. Huatuogpt, towards taming language model to be a doctor. arXiv preprint arXiv:2305.15075 (2023)

    2023 arXiv

  22. [30]

    Weinan Zhang, Kan Ren, and Jun Wang. 2016. Optimal real-time bidding frame- works discussion. arXiv preprint arXiv:1602.01007 (2016)

    2016 arXiv

  23. [31]

    Weinan Zhang, Shuai Yuan, and Jun Wang. 2014. Optimal real-time bidding for display advertising. In Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining . 1077–1086

    2014

  24. [32]

    Wentao Zhang, Lingxuan Zhao, Haochong Xia, Shuo Sun, Jiaze Sun, Molei Qin, Xinyi Li, Yuqing Zhao, Yilei Zhao, Xinyu Cai, et al . 2024. FinAgent: A Multi- modal Foundation Agent for Financial Trading: Tool-Augmented, Diversified, and Generalist. arXiv preprint arXiv:2402.18485 (2024)

    2024 arXiv

  25. [33]

    Jun Zhao, Guang Qiu, Ziyu Guan, Wei Zhao, and Xiaofei He. 2018. Deep rein- forcement learning for sponsored search real-time bidding. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining . 1021–1030

    2018

  26. [34]

    adjustment range

    Zhi Zhou, Jiang-Xin Shi, Peng-Xiao Song, Xiao-Wen Yang, Yi-Xuan Jin, Lan-Zhe Guo, and Yu-Feng Li. 2024. LawGPT: A Chinese Legal Knowledge-Enhanced Large Language Model. arXiv preprint arXiv:2406.04614 (2024). 8 APPENDIX 8.1 Partial Display of Prompt Template for RTBAgent Due t...

    2024 arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.