Pith. sign in

REVIEW 3 major objections 5 minor 36 references

SR-Agent automates the closed loop from bad-case detection to bounded, verified post-ranking strategy updates, and a one-month A/B test reports higher orders, deeper browsing, and more diverse clicks.

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

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

T0 review · deepseek-v4-flash

2026-08-01 17:10 UTC pith:MFUO3JUO

load-bearing objection Credible system design with a suggestive but not yet rigorous online A/B claim — worth engaging, but don't quote the effect sizes. the 3 major comments →

arxiv 2607.17719 v2 pith:MFUO3JUO submitted 2026-07-20 cs.AI cs.MA

SR-Agent: An Experience-Driven Agentic Framework for Post-Ranking Strategy Refinement in E-Commerce Recommendation

Hanchen Yang , Kaiwen Yang , Junpeng Zhuang , Yang He , Keting Cen , Bochao Liu , Zhongbo Sun , An Liu
show 2 more authors
Zhongteng Han Chenyi Lei
This is my paper
classification cs.AI cs.MA
keywords e-commerce recommendationpost-ranking strategyLLM agentsuser simulationstrategy refinementonline A/B testuser experienceself-evolving system
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Post-ranking rules in industrial recommender systems—diversity thresholds, similarity suppression, exposure caps—grow stale as items and user tastes shift, and refining them today is a slow, manual, multi-role chore. The paper argues this can be automated safely by an agentic loop that inspects displayed lists for user-perceived defects, diagnoses why current strategies let them through, and applies small, typed, reversible configuration updates. SR-Agent implements that loop with a user-simulating inspector, a diagnosis agent, and a constrained refinement harness; deployed on the Kuaishou e-commerce platform, a one-month online A/B test shows order volume up 0.71%, browsing depth up 0.34%, and clicked-category diversity up 0.48%, while the refinement cycle drops from weeks to days. If the results are causal, post-ranking strategy maintenance becomes a continuously self-evolving process rather than a periodic human task.

Core claim

The paper claims that the full cycle of experience inspection, diagnosis, bounded strategy refinement, and guarded validation can be closed by LLM-based agents without changing the base ranking model, and that doing so in production yields consistent gains in both user experience and business metrics. In a one-month A/B test on 2.5% of traffic covering roughly 60 million users, the framework improved clicked categories per user by 0.425–0.482%, browsing depth by 0.342%, total clicks by 0.554%, order volume by 0.707%, and GMV by 0.461%. The system also detected 142 bad cases in 500 sampled recommendation cases with a 95% human validation rate, versus 36 cases found by manual inspection, and i

What carries the argument

The central mechanism is the constrained Strategy Refinement Harness: a fixed, pre-authorized action space of four typed updates—relation-threshold adjustment, redundancy-suppression adjustment, exposure-limit adjustment, and authorized taxonomy correction—each with bounded parameters (e.g., threshold shifts between −0.15 and +0.15 in steps of 0.01). The harness maps diagnoses from an Analysis agent into candidate updates, filters them with a deterministic validator, and admits them only after a four-stage reward pipeline: retesting on labeled bad cases, replay on historical traffic, human verification, and online A/B validation, with rollback on guardrail violations. This decomposition make

Load-bearing premise

The one-month A/B comparison assumes the 2.5% treatment traffic differs from control only in the agent-generated strategy updates; if day-of-week effects, item-supply shifts, or interference between control and treatment users are not balanced, the reported gains could be artifacts rather than causal effects.

What would settle it

Re-analyze the logged one-month experiment with a difference-in-differences or synthetic-control design that adjusts for daily item supply, user composition, and calendar effects; if the point estimates for order volume and browsing depth shrink to near zero or change sign, the central deployment claim is not established.

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

If this is right

  • Post-ranking strategies can remain aligned with a drifting item and user environment automatically, without retraining the base ranker.
  • Human experts shift from case-by-case debugging to auditing a small sample of generated bad cases and diagnoses and approving candidate updates.
  • The refinement cycle shortens from 2–3 weeks to 3–5 elapsed days, allowing faster response to experience regressions.
  • Because each update is bounded and reversible, the framework can serve as a low-risk maintenance layer for always-on production recommenders.
  • The paper states the same guarded loop can be extended to other configurable components, such as cold-start boosting and multi-channel retrieval quotas.

Where Pith is reading between the lines

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

  • The numeric gains are likely environment-specific; the transferable contribution is the closed-loop architecture, and its robustness could be tested by applying the same harness to a different e-commerce platform or domain.
  • The framework's effectiveness hinges on the LLM-based inspector detecting defects that match real user perception; a blind comparison of UserSim's flags against self-reported user satisfaction on a held-out set would sharpen that link.
  • With unbounded runtime, the strategy configuration may drift among locally optimal settings; tracking rollback frequency and update-to-update variance could reveal whether the loop stabilizes or oscillates over months.
  • Because the harness leaves the ranker untouched, it could be paired with any black-box base model, making it a candidate for a standardized post-ranking maintenance service.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. SR-Agent addresses the operational problem of maintaining rule-based post-ranking strategies in a large-scale e-commerce recommender. The paper describes a deployed multi-agent loop: a UserSim agent inspects displayed lists and produces structured bad cases; an Analysis agent consolidates them into diagnoses; a Strategy Refinement Harness maps diagnoses to a typed, bounded action space and admits updates only after bad-case retesting, historical replay, human review, and online A/B validation, with rollback. The empirical section reports a one-month online A/B test on Kuaishou with 2.5% treatment traffic, claiming relative gains of +0.707% order volume, +0.342% browsing depth, +0.554% clicks, +0.461% GMV, and +0.482% 7-day clicked categories, plus a 2–3 week to 3–5 day cycle compression and a ten-update trend in bad-case resolution from 49.2% to 83.1%.

Significance. If the empirical claims withstand scrutiny, the paper would be a meaningful industrial contribution: it is one of the few full closed-loop agentic systems for strategy maintenance, with explicit safety mechanisms (bounded actions, deterministic validation, staged reward pipeline, rollback) and reusable memory. The decomposition of bad-case detection, diagnosis, and bounded refinement is sensible, and the deployment details are concrete. The strongest contribution is the system design and the demonstration that a constrained agent loop can be operated in production. However, the headline quantitative claims are not yet established because the A/B test is reported without confidence intervals, correction for sequential decisions, or details on the adaptive candidate-selection process; and the RQ4 resolution metric is partly circular.

major comments (3)
  1. [§4.1, Table 2; Algorithm 1, §3.5] Table 2 reports one-month relative changes without confidence intervals, p-values, or pre-registered metrics. This is not a fixed-treatment A/B contrast: Algorithm 1 and §3.5 describe a continuously adapting loop in which candidate updates are generated from interim bad-case diagnoses, pass online A/B validation (r_online in Eq. 5) before deployment, and can be rolled back. If only candidates with positive interim effects are kept, the aggregate estimates inherit winner's-curse/selection bias. The paper gives no number of candidates generated, tested, deployed, or rolled back, and no correction for sequential decisions. Please report these counts and provide interval estimates or an alternative design; otherwise the headline 0.707% order-volume and 0.342% browsing-depth gains cannot be distinguished from noise.
  2. [§4.4, Fig. 3(a); §3.3–3.4] The bad-case resolution rate is computed by retesting the targeted bad cases with the same harness and UserSim experience evaluator that produced the labels. Since the defect definition is endogenous, improvement in this metric need not correspond to user-perceived improvement. §3.4 states that human reviewers randomly audit about 1% of bad cases and diagnoses, so the ground-truth check is sparse. RQ4 should report resolution on an independent, human-labeled hold-out set and give audit coverage of the retested cases.
  3. [§4.2, Table 3] Table 3 reports detected cases and validation rates on 500 recommendation cases but gives no sampling details, inter-annotator agreement, confidence intervals, or false-negative rate. The validation rate apparently is the fraction of detected cases confirmed by humans, not the fraction of actual bad cases found. Manual inspection detects 36 cases while UserSim detects 142 at 95% validation, but without a recall estimate relative to a complete label set, 'broader coverage' is not established. Please clarify the annotation protocol and report error bars.
minor comments (5)
  1. [Abstract and §2.3] The claim that SR-Agent is 'the first agentic framework deployed to refine post-ranking strategies in industrial RS' is based on recent preprints and cannot be conclusively verified; consider softening to 'to our knowledge' or adding a more precise prior-art sweep.
  2. [§3.3, Eq. (1)] The symbol u is used both for the user and for 'current intent' in the formal definition. Clarify the notation to avoid ambiguity in Eq. (1).
  3. [§3.5, Table 1] The admissible domains for exposure-limit and taxonomy-correction actions are given only as abstract sets (C_cfg and C_auth_tax). Give representative examples or explicit bounds so readers can assess the actual boundedness claim.
  4. [§4.3, Table 4] The end-to-end iteration times are reported as single numbers without variance or a defined measurement protocol. Report how many loops were measured and the range.
  5. [§4.4, Figure 3] Figure 3(b) shows a trend over time without CI bands or any indication of traffic-composition adjustments. State whether the daily values are raw or debiased and add interval estimates.

Circularity Check

0 steps flagged

No significant circularity; the central A/B result is grounded in an external control-group comparison, and the bounded action space is predefined rather than fitted.

full rationale

The paper's central claim (Section 4.1, Table 2) rests on a one-month online A/B test against a control group that retains the existing manually configured post-ranking strategies, with the base ranking model fixed. The treatment arm applies only typed, bounded strategy updates generated through the predefined action space in Section 3.5/Table 1 and gated by a four-stage validation pipeline. The reported metric gains are measured in live traffic and are not derived algebraically from the model's own labels or from fitted parameters. No load-bearing self-citation or uniqueness-import chain appears in Sections 3-4: the cited prior systems are positioned as context, and the framework's components are described and defined within this paper. The only potentially self-referential element is RQ4's bad-case resolution rate (Section 4.4, Figure 3a), which is defined against UserSim-generated cases and the system's own retesting procedure. However, this is a secondary, internally defined efficiency metric, not the paper's primary evidence; the external online trend and the A/B table provide independent support. The paper also discloses that only approximately 1% of bad cases and diagnoses receive human audit, which is a limitation on internal evaluation quality rather than a circularity in the derivation. The absence of confidence intervals or p-values is a statistical robustness concern, not evidence of circularity. Overall, the central derivation is self-contained and externally benchmarked.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

SR-Agent is a software system, not a newly postulated physical or mathematical entity; no new particles, forces, or conserved quantities are introduced. The central claim depends on the operational assumptions above plus the hand-configured bounds in the action space.

free parameters (3)
  • Bad-case confidence threshold for storing UserSim outputs
    Algorithm 1 line 4 uses 'sufficiently confident' without a stated threshold; this hand-chosen value determines which cases enter the loop and influences all downstream metrics.
  • Bounded action-space parameters in Table 1
    Threshold shift bounds, step sizes, exposure-limit configurations, and taxonomy whitelists are configured by strategy owners; they constrain what the agent can do and therefore shape the online outcomes.
  • Human audit rate of ~1% = 1%
    Section 3.4 states roughly 1% of bad cases and diagnoses are human-audited. This rate is a manual choice that defines the effective quality control of the unsupervised majority.
axioms (4)
  • domain assumption LLM-based UserSim experience judgments approximate real user experience well enough to guide production changes.
    The whole loop relies on UserSim's bad-case flags. Only ~1% are audited in production; the 500-case validation (Section 4.2) is a small sample and uses manual judgment as ground truth.
  • domain assumption Historical traffic replay predicts online safety and effectiveness of candidate strategy updates.
    Section 3.5 admits candidates after replay on D_hist, assuming replay outcomes transfer to live traffic. The paper gives no evidence on replay calibration or false-accept rates.
  • domain assumption The one-month A/B treatment-control comparison is free of interference, seasonality, and selection effects.
    Section 4.1 reports only relative changes with no statistical tests, so the causal interpretation requires this assumption.
  • domain assumption Post-ranking strategies degrade over time and require continuous refinement.
    Challenge 1 motivates the entire paper; if this were false, the deployment would have nothing to fix, but the paper provides no quantitative evidence of degradation over the test period.

pith-pipeline@v1.3.0-alltime-deepseek · 11631 in / 8498 out tokens · 97493 ms · 2026-08-01T17:10:00.706166+00:00 · methodology

0 comments
read the original abstract

User experience is a first-class objective in industrial e-commerce recommender systems (RS). Post-ranking strategies, which govern diversity, similarity, and exposure over a ranked list, are widely deployed in industrial RS for their simplicity and low serving cost. However, as the online recommendation environment evolves continuously, these statically configured strategies gradually become stale, thereby degrading the user experience. Refining them typically relies on manual inspection, diagnosis, and updates, making it slow, costly, and difficult to scale or reuse. Although recent LLM-based agents (e.g., RecUserSim, SimUSER, and Self-EvolveRec) offer promising directions, none of them close the full loop of automated, self-evolving strategy refinement. To bridge this gap, we introduce SR-Agent, which, to the best of our knowledge, is the first agentic framework deployed to refine post-ranking strategies in industrial RS. SR-Agent unifies three components: (i) a UserSim agent that applies inspection skills to surface user-perceived bad cases; (ii) an Analysis agent that consolidates recurring bad cases into structured, reusable diagnoses; and (iii) a constrained Strategy Refinement Harness that maps diagnoses to typed and bounded actions, gated by a four-stage reward pipeline with reversible rollback. Deployed on the Kuaishou e-commerce platform, SR-Agent continuously runs this refinement loop and, in a one-month online A/B test, increases order volume by 0.71%, browsing depth by 0.34%, and clicked-category diversity by 0.48%, while markedly shortening the refinement cycle and lowering operational cost.

Figures

Figures reproduced from arXiv: 2607.17719 by An Liu, Bochao Liu, Chenyi Lei, Hanchen Yang, Junpeng Zhuang, Kaiwen Yang, Keting Cen, Yang He, Zhongbo Sun, Zhongteng Han.

Figure 1
Figure 1. Figure 1: Overview of the motivation and proposed SR-Agent framework. Existing maintenance of post-ranking strategies relies on static configurations, manual inspection, and delayed metric feedback. Our agentic framework closes the loop from bad-case discovery and diagnosis to bounded strategy updates, offline verification, and controlled online validation. industrial RS. Their low serving cost, operational flexibil… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed agentic framework. The base recommender produces displayed recommen￾dation lists with a fixed ranking model. The agentic loop then performs user-perspective inspection, bad-case diagnosis, bounded strategy refinement, memory update, reward validation, and safe online deployment. similarity suppressions, item/relation penalties, exposure caps, and item groupings. The displayed list … view at source ↗
Figure 3
Figure 3. Figure 3: Results over ten strategy updates. (a) Per-update bad-case resolution rate of the refinement harness. (b) Relative change in number of clicks and browsing depth during the one-month online rollout, versus the pre-rollout baseline. BOUNDED REFINEMENT Apply an authorized taxonomy correction to the cross-category massage-tool group. Five similar pants products occur at positions 6, 8, 10, 12, and 14. Interlea… view at source ↗
Figure 4
Figure 4. Figure 4: Two cases processed by SR-Agent. The first requires cross-category item-relation inference, whereas the second requires session-level exposure analysis. In both cases, UserSim evidence is converted into an actionable diagnosis and mapped to a bounded post-ranking strategy refinement. agent attributes the defect to incomplete relation coverage caused by fragmented item groupings and localizes its scope to t… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

36 extracted references · 5 linked inside Pith

  1. [1]

    Q. Peng, H. Liu, H. Huang, Q. Yang, M. Shao, A survey on llm-powered agents for recommender systems, arXiv preprint arXiv:2502.10050 (2025)

    Pith/arXiv arXiv 2025

  2. [2]

    I. Adaji, Towards improving e-commerce users experience using personalization & persuasive tech- nology, in: Proceedings of the 25th Conference on User Modeling, Adaptation and Personalization, 2017, pp. 318–321

    2017

  3. [3]

    X. Song, User-centric internal tools in e-commerce: Enhancing operational efficiency through ai integration, in: Proceedings of the 2025 International Conference on Economic Management and Big Data Application, 2025, pp. 702–706

    2025

  4. [4]

    P. Pu, L. Chen, R. Hu, Evaluating recommender systems from the user’s perspective: survey of the state of the art, User Modeling and User-Adapted Interaction 22 (2012) 317–355

    2012

  5. [5]

    Duricic, D

    T. Duricic, D. Kowald, E. Lacic, E. Lex, Beyond-accuracy: a review on diversity, serendipity, and fairness in recommender systems based on graph neural networks, Frontiers in big data 6 (2023) 1251072

    2023

  6. [6]

    S. M. McNee, J. Riedl, J. A. Konstan, Being accurate is not enough: how accuracy metrics have hurt recommender systems, in: CHI’06 extended abstracts on Human factors in computing systems, 2006, pp. 1097–1101

    2006

  7. [7]

    Ziegler, S

    C.-N. Ziegler, S. M. McNee, J. A. Konstan, G. Lausen, Improving recommendation lists through topic diversification, in: Proceedings of the 14th international conference on World Wide Web, 2005, pp. 22–32

    2005

  8. [8]

    Kunaver, T

    M. Kunaver, T. Požrl, Diversity in recommender systems–a survey, Knowledge-based systems 123 (2017) 154–162

    2017

  9. [9]

    Lathia, S

    N. Lathia, S. Hailes, L. Capra, X. Amatriain, Temporal diversity in recommender systems, in: Proceedings of the 33rd international ACM SIGIR conference on Research and development in information retrieval, 2010, pp. 210–217

    2010

  10. [10]

    L. Chen, G. Zhang, E. Zhou, Fast greedy map inference for determinantal point process to improve recommendation diversity, Advances in neural information processing systems 31 (2018)

    2018

  11. [11]

    M. Chen, A. Beutel, P. Covington, S. Jain, F. Belletti, E. H. Chi, Top-k off-policy correction for a reinforce recommender system, in: Proceedings of the twelfth ACM international conference on web search and data mining, 2019, pp. 456–464

    2019

  12. [12]

    S. Wang, L. Cao, Y. Wang, Q. Z. Sheng, M. A. Orgun, D. Lian, A survey on session-based recom- mender systems, ACM Computing Surveys (CSUR) 54 (2021) 1–38

    2021

  13. [13]

    Y. Liu, D. Iter, Y. Xu, S. Wang, R. Xu, C. Zhu, G-eval: Nlg evaluation using gpt-4 with better human alignment, in: Proceedings of the 2023 conference on empirical methods in natural language processing, 2023, pp. 2511–2522

    2023

  14. [14]

    Zheng, W.-L

    L. Zheng, W.-L. Chiang, Y. Sheng, S. Zhuang, Z. Wu, Y. Zhuang, Z. Lin, Z. Li, D. Li, E. Xing, et al., Judging llm-as-a-judge with mt-bench and chatbot arena, Advances in neural information processing systems 36 (2023) 46595–46623

    2023

  15. [15]

    Zhang, Y

    A. Zhang, Y. Chen, L. Sheng, X. Wang, T.-S. Chua, On generative agents in recommendation, in: Proceedings of the 47th international ACM SIGIR conference on research and development in Information Retrieval, 2024, pp. 1807–1817

    2024

  16. [16]

    S.-e. Yoon, Z. He, J. Echterhoff, J. McAuley, Evaluating large language models as generative user simulators for conversational recommendation, in: Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), 2024, pp. 1490–1504

    2024

  17. [17]

    Bougie, N

    N. Bougie, N. Watanabe, Simuser: Simulating user behavior with large language models for recommender system evaluation, in: Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track), 2025, pp. 43–60

    2025

  18. [18]

    Wanyan, D

    X. Wanyan, D. Hettiachchi, C. Ma, Z. Xu, J. Chan, Temporal-aware user behaviour simulation with large language models for recommender systems, in: Proceedings of the 34th ACM International Conference on Information and Knowledge Management, 2025, pp. 5335–5339

    2025

  19. [19]

    L. Shi, C. Ma, W. Liang, X. Diao, W. Ma, S. Vosoughi, Judging the judges: A systematic study of position bias in llm-as-a-judge, in: Proceedings of the 14th International Joint Conference on Natural Language Processing and the 4th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics, 2025, pp. 292–314

    2025

  20. [20]

    S. Kim, S. Park, H. Kang, W. Kim, J. Seo, Y. In, K. Yoon, C. Park, Self-evolverec: Self-evolving recommender systems with llm-based directional feedback, arXiv preprint arXiv:2602.12612 (2026)

    Pith/arXiv arXiv 2026

  21. [21]

    H. Wang, Y. Wu, D. Chang, L. Wei, L. Heldt, Self-evolving recommendation system: End-to-end autonomous model optimization with llm agents, arXiv preprint arXiv:2602.10226 (2026)

    Pith/arXiv arXiv 2026

  22. [22]

    C. Lao, F. Pan, G. Ma, H. Li, H. Lin, J. Shi, K. Zhao, K. Gai, M. Zhou, Q. Zhou, et al., Agentx: Towards agent-driven self-iteration of industrial recommender systems, arXiv preprint arXiv:2606.26859 (2026)

    Pith/arXiv arXiv 2026

  23. [23]

    J. Mao, L. Li, Y. Gao, Z. Peng, S. He, C. Zhang, S. Qin, S. Khalid, Q. Lin, S. Rajmohan, et al., Stepfly: Agentic troubleshooting guide automation for incident diagnosis, Proceedings of the ACM on Software Engineering 3 (2026) 3070–3092

    2026

  24. [24]

    P. Pu, L. Chen, R. Hu, A user-centric evaluation framework for recommender systems, in: Proceedings of the fifth ACM conference on Recommender systems, 2011, pp. 157–164

    2011

  25. [25]

    Adamopoulos, A

    P. Adamopoulos, A. Tuzhilin, On unexpectedness in recommender systems: Or how to better expect the unexpected, ACM Transactions on Intelligent Systems and Technology (TIST) 5 (2014) 1–32

    2014

  26. [26]

    Carbonell, J

    J. Carbonell, J. Goldstein, The use of mmr, diversity-based reranking for reordering documents and producing summaries, in: Proceedings of the 21st annual international ACM SIGIR conference on Research and development in information retrieval, 1998, pp. 335–336

    1998

  27. [27]

    Y. Liu, C. Walder, L. Xie, Determinantal point process likelihoods for sequential recommendation, in: Proceedings of the 45th international ACM SIGIR conference on research and development in information retrieval, 2022, pp. 1653–1663

    2022

  28. [28]

    Wilhelm, A

    M. Wilhelm, A. Ramanathan, A. Bonomo, S. Jain, E. H. Chi, J. Gillenwater, Practical diversified recommendations on youtube with determinantal point processes, in: Proceedings of the 27th ACM International Conference on Information and Knowledge Management, 2018, pp. 2165–2173

    2018

  29. [29]

    Zhang, S

    Z. Zhang, S. Liu, Z. Liu, R. Zhong, Q. Cai, X. Zhao, C. Zhang, Q. Liu, P. Jiang, Llm-powered user simulator for recommender system, in: Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, 2025, pp. 13339–13347

    2025

  30. [30]

    L. Chen, Q. Dai, Z. Zhang, X. Feng, M. Zhang, P. Tang, X. Chen, Y. Zhu, Z. Dong, Recusersim: A realistic and diverse user simulator for evaluating conversational recommender systems, in: Companion Proceedings of the ACM on Web Conference 2025, 2025, pp. 133–142

    2025

  31. [31]

    Huang, J

    X. Huang, J. Lian, Y. Lei, J. Yao, D. Lian, X. Xie, Recommender ai agent: Integrating large language models for interactive recommendations, ACM Transactions on Information Systems 43 (2025) 1–33

    2025

  32. [32]

    Zhang, Y

    J. Zhang, Y. Hou, R. Xie, W. Sun, J. McAuley, W. X. Zhao, L. Lin, J.-R. Wen, Agentcf: Collaborative learning with autonomous language agents for recommender systems, in: Proceedings of the ACM Web Conference 2024, 2024, pp. 3679–3689

    2024

  33. [33]

    Y. Wang, Z. Jiang, Z. Chen, F. Yang, Y. Zhou, E. Cho, X. Fan, Y. Lu, X. Huang, Y. Yang, Recmind: Large language model powered agent for recommendation, in: Findings of the Association for Computational Linguistics: NAACL 2024, 2024, pp. 4351–4364

    2024

  34. [34]

    B. Ma, H. Li, Z. Hu, X. Gui, L. Liu, S. Lau, Agentrec: Next-generation llm-powered multi-agent collaborative recommendation with adaptive intelligence, arXiv preprint arXiv:2510.01609 (2025)

    arXiv 2025

  35. [35]

    J. Park, M. Kim, Madrec: A multi-aspect driven llm agent for explainable and adaptive recommen- dation, arXiv preprint arXiv:2510.13371 (2025)

    arXiv 2025

  36. [36]

    X. Lin, Y. Deldjoo, S. Dai, H. Bao, X. Ye, F. Nazary, W. Wang, T. Di Noia, J. Xu, T.-S. Chua, Autonomous information seeking: A roadmap for agentic recommender systems, arXiv preprint arXiv:2607.04433 (2026)

    Pith/arXiv arXiv 2026