Pith. sign in

REVIEW 3 major objections 32 references

Request-level hard selection of one recommender list is too coarse; item-level aggregation is the better action space.

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

2026-07-14 14:40 UTC pith:J33KBXKA

load-bearing objection Honest OOF bake-off: hard agent selection loses to BM25 despite big oracle headroom; learned shortlist stacking recovers signal—but the granularity lesson is still protocol-bound. the 3 major comments →

arxiv 2607.09908 v1 pith:J33KBXKA submitted 2026-07-10 cs.CL cs.IR

RouteRec: Strict Evaluation of Recommender-Agent Selection and Aggregation

Kaiji Zhou , Vladimir Kalmykov , Yue Feng This is my paper
classification cs.CL cs.IR
keywords recommender systemsagent routingLLM rerankingrank aggregationalgorithm selectioncost-aware recommendationshortlist stackingMovieLens
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.

Recommender systems now choose among heterogeneous agents—collaborative filters, sequential models, content retrievers, and LLM rerankers—none of which wins on every request. RouteRec tests that choice under cost constraints by comparing two policies on MovieLens-1M: hard selection of one complete agent list versus learned aggregation over the union of shortlists. A quality oracle shows large headroom (HR@10 = 0.584), so complementary signal exists. Under a leakage-free 5-fold out-of-fold protocol, however, hard selection stays below BM25 and selective LLM escalation does not help, while cheap-only stacking matches BM25 and gated all-agent stacking reaches HR@10 = 0.295. The paper’s lesson is methodological: the unit of the routing decision, not the absence of cross-agent signal, is the bottleneck in this sparse fixed-candidate setting.

Core claim

Despite substantial oracle headroom from complementary recommender agents, request-level hard selection of one complete ranked list fails to beat the best fixed cheap agent under strict out-of-fold evaluation, and selective LLM escalation does not repair it. The same evidence becomes useful when the system scores items from the union of agent shortlists: cheap-only learned stacking matches BM25, and gated all-agent stacking improves both hit rate and ranking quality. The authors therefore conclude that request-level selection is too coarse for this sparse fixed-candidate regime and that item-level aggregation is the more promising action space.

What carries the argument

RouteRec, a framework that holds the same five-agent pool and strict 5-fold out-of-fold protocol fixed while comparing request-level hard selection (cheap selector plus escalation gate driven by probe disagreement) against item-level learned shortlist stacking over deployable rank, score, membership, and popularity features. Changing only the decision unit isolates which action space can recover cross-agent signal.

Load-bearing premise

The SASRec-seeded fixed-candidate MovieLens setup—with one leave-one-out label per user and the LLM limited to the first 20 candidates—is a fair enough testbed to conclude that hard list selection is structurally too coarse, not merely underpowered or protocol-bound.

What would settle it

If, on the same 500-user leakage-free protocol with richer multi-item or free-form feedback, a hard selector clearly beat BM25 while stacking did not, or if stacking gains disappeared under open retrieval or multi-domain data, the claim that request-level selection is the wrong action space would be overturned.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Whole-list routing policies will leave complementary agent signal on the table even when oracles show large headroom.
  • LLM rerankers can still help as shortlist features for stacking even when they are weak as standalone complete lists.
  • Agent-routing evaluations must separately measure oracle complementarity, policy learnability, and decision granularity.
  • Budgeted item-level aggregators can expose a smoother cost–quality frontier than hard escalation gates.
  • Agent-search systems should treat shortlist fusion as a first-class action, not only model selection.

Where Pith is reading between the lines

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

  • The same coarseness may appear in multi-tool agent systems that force a single tool’s full output rather than span- or item-level fusion.
  • Sparse leave-one-out labels may systematically favor stacking over selection; denser multi-item feedback could shift the balance.
  • When an LLM can influence candidate generation rather than only top-20 reranking, the relative value of hard escalation may rise.
  • Open-ended agent pools will likely need sparse or calibrated aggregation rather than denser stacking as the union grows.

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 / 0 minor

Summary. The paper studies cost-aware routing over heterogeneous recommender agents (Popularity, BPR-MF, SASRec, BM25, and an LLM reranker) via RouteRec, comparing request-level hard selection (RouteRec-Select: cheap selector plus escalation gate) with item-level learned shortlist aggregation (RouteRec-Stack variants). On MovieLens-1M under a leakage-free 5-fold out-of-fold protocol over 500 users, the full quality oracle has large headroom (HR@10 = 0.584), yet hard selection stays below BM25 (0.223 vs 0.254) and selective LLM escalation does not help, while cheap stacking matches BM25 HR with a higher NDCG point estimate and gated all-agent stacking reaches HR@10 = 0.295 at 70.2% LLM calls. The central claim is methodological: in this sparse fixed-candidate setting, selecting one complete agent list is too coarse, and item-level aggregation is the more promising action space.

Significance. If the result holds, it is a useful corrective for recommender-agent routing and LLM-routing work: oracle complementarity does not imply learnable hard selection, and decision granularity (list-level vs item-level) can dominate architecture. Strengths include a strict OOF protocol, clear separation of deployable features from oracle labels, cost–quality curves, traffic analysis, bootstrap CIs for key stacker gains, and an explicit limitations section. The contribution is primarily empirical and cautionary rather than a new SOTA system, but that caution is timely for agent-search and selective-LLM recommendation papers that lean on oracle headroom alone.

major comments (3)
  1. Experimental Setup and Limitations: the central granularity claim (hard selection too coarse; item-level aggregation more promising) is load-bearing but rests on a SASRec-seeded fixed 50-candidate pool with the held-out target randomly inserted, LLM prompts restricted to the first 20 candidates (target visible in only 36.4% of requests), and one leave-one-out label per 500 users. Under this construction, hard selection must pick an entire list that is often zero/tied, while stacking scores the de-duplicated top-10 union (cheap coverage 50.8%, all-agent 58.4%). The paper already shows Select collapses to a BM25/SASRec switcher (Fig. 3) and flat ablations (Table 4). Please either (i) add at least one less constrained candidate-generation or denser-feedback check that could falsify the structural claim, or (ii) substantially narrow the abstract/conclusion wording from a general action-space
  2. §6.2 / Table 1 and §7.1–7.2: with n=500 sparse LOO labels, hard-selection underperformance and flat feature ablations are consistent with underpowered request-level supervision as much as with inherent coarseness. Subgroup slices are tiny (cold n=26, long-tail n=18), and the gate never finds the sparse LLM-best cases (F1 near 0.1). The manuscript should report power/uncertainty more explicitly for the Select–BM25 gap and avoid reading the flat ablations as evidence that disagreement or request features are intrinsically uninformative; currently the Discussion leans toward a structural interpretation that the data regime cannot cleanly support.
  3. §3 Eq. (1) and §4.6–4.7: policy utility uses fixed α=1, β=0.5, λ=μ=0.01, and stacker model/threshold selection is by mean utility on validation. Main tables emphasize HR/NDCG, and StackGate’s quality gains come with high LLM usage (70.2%). Please show sensitivity of the Select-vs-Stack ranking and the cost–quality frontier to these coefficients (or to pure quality selection), so the claim that aggregation is the more promising action space is not an artifact of one utility parameterization.

Circularity Check

0 steps flagged

Empirical bake-off with OOF evaluation; no derivation reduces to its inputs by construction.

full rationale

RouteRec is a controlled empirical comparison of request-level hard selection versus item-level shortlist stacking on MovieLens-1M under a leakage-free 5-fold out-of-fold protocol. The utility formula (Eq. 1), quality-only ranking labels, escalation-gate BCE targets, and validation-selected thresholds are fixed design choices applied out-of-sample; they do not redefine the reported HR/NDCG outcomes. Oracle rows (CheapOracle, Oracle) are explicit quality upper bounds, not training features or fitted parameters renamed as predictions. Stacker positives are leave-one-out held-out items inside the agent-union shortlist, with only deployable rank/score/popularity features—standard supervised ranking, not a tautology. There is no self-citation chain, uniqueness theorem, or ansatz imported from the authors’ prior work that forces the Select-vs-Stack conclusion. Protocol limitations (SASRec-seeded pool, sparse LOO labels) affect external validity, not circularity of the derivation. Score 0 is therefore the correct finding.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 4 invented entities

The central claim rests on an empirical protocol and a hand-specified utility, not on new physical entities. Load-bearing free parameters are the utility weights, ranking margin, escalation weight, gate/budget thresholds, and stacker regularizers chosen on validation. Domain assumptions include the MovieLens leave-one-out fixed-candidate setup, the five-agent pool, and the definition of policy utility. Invented constructs are engineering modules (disagreement features, capability vectors, two-stage router, stack gate), not independently evidenced natural kinds.

free parameters (5)
  • utility weights α, β, λ, μ
    Eq. 1 fixes α=1, β=0.5, λ=μ=0.01 to trade NDCG/HR against latency and tokens; gate and stack thresholds optimize this utility on validation.
  • pairwise ranking margin m
    Training uses m=0.3 in the cheap-selector hinge loss (Section 4.6).
  • escalation loss weight η
    Combined loss uses η=1 for the binary escalation term (Section 4.6).
  • escalation threshold τ and stack budget caps
    Gate threshold and StackBudget@25/50/75 caps are selected on validation utility; realized test LLM ratios depend on these choices.
  • stacker regularization C and model family
    Cheap stacker selects logistic regression C∈{0.03,0.1,0.3}; all-agent stacker may also pick histogram gradient boosting via validation utility.
axioms (5)
  • domain assumption No single recommender agent is uniformly best across requests; complementary per-instance winners exist in the pool.
    Premise of Sections 1 and 6.1; supported empirically on this pool but not proved generally.
  • ad hoc to paper Policy utility is α·NDCG+β·HR−λ·lat−μ·tok with the stated fixed coefficients.
    Eq. 1 defines the optimization target for gates, budgets, and stack model selection.
  • domain assumption SASRec-seeded shared candidate pools with held-out target insertion and leave-one-out labels are a valid offline test of routing vs aggregation.
    Experimental Setup; the paper’s conclusions about action-space coarseness are conditioned on this protocol.
  • domain assumption Deployable probe disagreement, request features, and shortlist rank/score features are admissible inputs; oracle quality labels are not.
    Sections 4.4–4.7 separate deployable signals from oracle upper bounds.
  • standard math Standard ranking metrics HR@10, NDCG@10, MRR and 5-fold stratified OOF estimate generalization for trainable routers.
    Common IR/recsys evaluation practice used throughout Sections 5–6.
invented entities (4)
  • RouteRec-Select two-stage router (cheap selector + escalation gate) no independent evidence
    purpose: Request-level hard selection among cheap agents with optional LLM escalation using request, capability, and disagreement features.
    Primary hard-selection method; engineering construct evaluated OOF, not an external natural object.
  • Cheap-probe disagreement vector d_x (15-D) no independent evidence
    purpose: Summarize pairwise overlap/Jaccard/Kendall, agreement, score entropy/margin, and union statistics from SASRec, BPR-MF, BM25 probes.
    Hand-designed feature family for routing; usefulness is tested, not independently measured outside this study.
  • Hand-designed 12-D agent capability vectors no independent evidence
    purpose: Encode signal type, instruction support, mode, scenario strengths, cost so the selector can match request to agent.
    Author-designed embeddings; paper notes one-hot performs comparably on the closed four-agent pool.
  • RouteRec-Stack / StackGate shortlist aggregation policies no independent evidence
    purpose: Item-level learned fusion of agent top-k unions with optional confidence-gated LLM inclusion under budgets.
    Stacking applied to this agent pool; the entity is a method variant, not an external phenomenon.

reviewed 2026-07-14 · how reviews work

0 comments
Cite this review

Pith. "Pith review of RouteRec: Strict Evaluation of Recommender-Agent Selection and Aggregation." pith.science (2026). https://pith.science/paper/J33KBXKA

@misc{pith2026260709908,
  author       = {Pith},
  title        = {Pith review of: RouteRec: Strict Evaluation of Recommender-Agent Selection and Aggregation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J33KBXKA}},
  note         = {Machine review of arXiv:2607.09908}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Recommender systems increasingly face a choice among heterogeneous agents -- collaborative filters, sequential models, content-based retrievers, and LLM-based rerankers -- yet no single agent is uniformly best. We study this choice as task-aware agent ranking under cost constraints using RouteRec, a framework that compares request-level hard selection with item-level learned aggregation over four traditional recommender agents and one LLM reranker agent. On MovieLens-1M, the full quality oracle has substantial headroom (HR@10 = 0.584), confirming that useful cross-agent signal exists. Under a leakage-free 5-fold out-of-fold protocol, however, hard selection remains below BM25 (0.223 vs. 0.254), and selective LLM escalation does not improve it. The same protocol yields a different outcome for learned aggregation: its cheap-only variant matches BM25 in HR and has a higher NDCG point estimate (0.123 vs. 0.114), while gated all-agent aggregation reaches HR@10 = 0.295 with 70.2\% LLM calls. The resulting lesson is not that routing is solved, but that request-level selection of one complete agent list is too coarse for this sparse fixed-candidate setting; item-level aggregation is the more promising action space.

Figures

Figures reproduced from arXiv: 2607.09908 by Kaiji Zhou, Vladimir Kalmykov, Yue Feng.

Figure 1
Figure 1. Figure 1: RouteRec-Select architecture. A request is encoded and combined with cheap-probe disagreement features to select the best cheap agent (Stage 1) and decide whether to escalate to the LLM reranker (Stage 2). disagreement, a deployable measure of cheap-agent output diver￾gence, to choose one cheap agent and decide whether to escalate to the LLM ( [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Distribution of the quality-best agent per user. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 5
Figure 5. Figure 5: Strict OOF ablation impact relative to full [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 4
Figure 4. Figure 4: Strict subgroup HR@10 comparison across BM25, [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 7
Figure 7. Figure 7: Strict selection and aggregation versus quality [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] 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

32 extracted references · 1 canonical work pages

  1. [1]

    Keqin Bao, Jizhi Zhang, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He. 2023. TALLRec: An Effective and Efficient Tuning Framework to Align Large Language Model with Recommendation. InProc. of the 17th ACM Conf. on Recommender Systems (RecSys). ACM, New York, NY, USA, 1007–1014. https: //doi.org/10.1145/3604915.3608857

    doi:10.1145/3604915.3608857 2023

  2. [2]

    Robin Burke. 2002. Hybrid Recommender Systems: Survey and Experiments. User Modeling and User-Adapted Interaction12, 4 (2002), 331–370. https://doi. org/10.1023/A:1021240730564

    doi:10.1023/a:1021240730564 2002

  3. [3]

    Lingjiao Chen, Matei Zaharia, and James Zou. 2023. FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance. arXiv:2305.05176 [cs.LG]

    Pith/arXiv arXiv 2023

  4. [4]

    Andrew Collins, Dominika Tkaczyk, and Joeran Beel. 2018. One-at-a-time: A Meta-Learning Recommender-System for Recommendation-Algorithm Selection on Micro Level. arXiv:1805.12118 [cs.IR]

    Pith/arXiv arXiv 2018

  5. [5]

    Cormack, Charles L

    Gordon V. Cormack, Charles L. A. Clarke, and Stefan Büttcher. 2009. Reciprocal Rank Fusion Outperforms Condorcet and Individual Rank Learning Methods. InProc. of the 32nd Int. ACM SIGIR Conf.ACM, New York, NY, USA, 758–759. https://doi.org/10.1145/1571941.1572114

    doi:10.1145/1571941.1572114 2009

  6. [6]

    Tiago Cunha, Carlos Soares, and André C. P. L. F. de Carvalho. 2018. Metalearning and Recommender Systems: A Literature Review and Empirical Study on the Algorithm Selection Problem for Collaborative Filtering.Information Sciences 423 (2018), 128–144. https://doi.org/10.1016/j.ins.2017.09.050

    doi:10.1016/j.ins.2017.09.050 2018

  7. [7]

    Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. 2022. Recommendation as Language Processing (RLP): A Unified Pretrain, Personalized Prompt & Predict Paradigm (P5). InProc. of the 16th ACM Conf. on Recommender Systems (RecSys). ACM, New York, NY, USA, 299–315. https://doi.org/10.1145/ 3523227.3546767

    arXiv 2022

  8. [8]

    Maxwell Harper and Joseph A

    F. Maxwell Harper and Joseph A. Konstan. 2015. The MovieLens Datasets: History and Context.ACM Transactions on Interactive Intelligent Systems5, 4 (2015), 19:1–19:19. https://doi.org/10.1145/2827872

    doi:10.1145/2827872 2015

  9. [9]

    Yupeng Hou, Junjie Zhang, Zihan Lin, Hongyu Lu, Ruobing Xie, Julian McAuley, and Wayne Xin Zhao. 2024. Large Language Models are Zero-Shot Rankers for Recommender Systems. InProc. of the 46th European Conf. on Information Retrieval (ECIR). Springer, Cham, 364–381. https://doi.org/10.1007/978-3-031- 56060-6_24

    doi:10.1007/978-3-031- 2024

  10. [10]

    Jacobs, Michael I

    Robert A. Jacobs, Michael I. Jordan, Steven J. Nowlan, and Geoffrey E. Hinton

  11. [11]

    https://doi.org/10.1162/neco.1991.3.1.79

    Adaptive Mixtures of Local Experts.Neural Computation3, 1 (1991), 79–87. https://doi.org/10.1162/neco.1991.3.1.79

    doi:10.1162/neco.1991.3.1.79 1991

  12. [12]

    Dongfu Jiang, Xiang Ren, and Bill Yuchen Lin. 2023. LLM-Blender: Ensembling Large Language Models with Pairwise Ranking and Generative Fusion. InProc. of the 61st Annual Meeting of the Association for Computational Linguistics (ACL). Association for Computational Linguistics, Toronto, Canada, 14165–14178. https: //doi.org/10.18653/v1/2023.acl-long.792

    doi:10.18653/v1/2023.acl-long.792 2023

  13. [13]

    Wang-Cheng Kang and Julian McAuley. 2018. Self-Attentive Sequential Recom- mendation. InProc. of the IEEE Int. Conf. on Data Mining (ICDM). IEEE, Piscataway, NJ, USA, 197–206. https://doi.org/10.1109/ICDM.2018.00035

    doi:10.1109/icdm.2018.00035 2018

  14. [14]

    Sein Kim, Hongseok Kang, Seungyoon Choi, Donghyun Kim, Minchul Yang, and Chanyoung Park. 2024. Large Language Models meet Collaborative Filtering: An Efficient All-round LLM-based Recommender System. InProc. of the 30th ACM SIGKDD Conf. on Knowledge Discovery and Data Mining. ACM, New York, NY, USA, 1395–1406. https://doi.org/10.1145/3637528.3671931

    doi:10.1145/3637528.3671931 2024

  15. [15]

    Lei Li, Yongfeng Zhang, Dugang Liu, and Li Chen. 2024. Large Language Models for Generative Recommendation: A Survey and Visionary Discussions. InProc. of the 2024 Joint Int. Conf. on Computational Linguistics, Language Resources and Evaluation (LREC-COLING). ELRA and ICCL, Torino, Italy, 10146–10159. https://aclanthology.org/2024.lrec-main.886/

    2024

  16. [16]

    Junling Liu, Chao Liu, Peilin Zhou, Renjie Lv, Kang Zhou, and Yan Zhang. 2023. Is ChatGPT a Good Recommender? A Preliminary Study. arXiv:2304.10149 [cs.IR]

    Pith/arXiv arXiv 2023

  17. [17]

    Tie-Yan Liu. 2009. Learning to Rank for Information Retrieval.Foundations and Trends in Information Retrieval3, 3 (2009), 225–331. https://doi.org/10.1561/ 1500000016

    2009

  18. [18]

    Hanjia Lyu, Song Jiang, Hanqing Zeng, Yinglong Xia, Qifan Wang, Si Zhang, Ren Chen, Chris Leung, Jiajie Tang, and Jiebo Luo. 2024. LLM-Rec: Per- sonalized Recommendation via Prompting Large Language Models. InFind- ings of the Association for Computational Linguistics: NAACL 2024. Associa- tion for Computational Linguistics, Mexico City, Mexico, 583–612. ...

    doi:10.18653/v1/2024.findings-naacl.39 2024

  19. [19]

    Gonzalez, M

    Isaac Ong, Amjad Almahairi, Vincent Wu, Wei-Lin Chiang, Tianhao Wu, Joseph E. Gonzalez, M. Waleed Kadous, and Ion Stoica. 2024. RouteLLM: Learning to Route LLMs with Preference Data. arXiv:2406.18665 [cs.LG]

    Pith/arXiv arXiv 2024

  20. [20]

    Qwen Team. 2024. Qwen2.5 Technical Report. arXiv:2412.15115 [cs.CL]

    Pith/arXiv arXiv 2024

  21. [21]

    Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt- Thieme. 2009. BPR: Bayesian Personalized Ranking from Implicit Feedback. InProc. of the 25th Conf. on Uncertainty in Artificial Intelligence (UAI). AUAI Press, Arlington, VA, USA, 452–461

    2009

  22. [22]

    John R. Rice. 1976. The Algorithm Selection Problem.Advances in Computers15 (1976), 65–118. https://doi.org/10.1016/S0065-2458(08)60520-3

    doi:10.1016/s0065-2458(08)60520-3 1976

  23. [23]

    Stephen Robertson and Hugo Zaragoza. 2009. The Probabilistic Relevance Frame- work: BM25 and Beyond.Foundations and Trends in Information Retrieval3, 4 (2009), 333–389. https://doi.org/10.1561/1500000019

    doi:10.1561/1500000019 2009

  24. [24]

    Marlesson R. O. Santana, Luckeciano C. Melo, Fernando H. F. Camargo, Bruno Brandão, Anderson Soares, Renan M. Oliveira, and Sandor Caetano. 2020. Con- textual Meta-Bandit for Recommender Systems Selection. InProc. of the 14th ACM Conf. on Recommender Systems (RecSys). ACM, New York, NY, USA, 444–449. https://doi.org/10.1145/3383313.3412209

    work page doi:10.1145/3383313.3412209 2020

  25. [25]

    Le, Geoffrey E

    Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc V. Le, Geoffrey E. Hinton, and Jeff Dean. 2017. Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer. InProc. of the 5th Int. Conf. on Learning Representations (ICLR). OpenReview.net. https://openreview.net/forum? id=B1ckMDqlg

    2017

  26. [26]

    Tal Shnitzer, Anthony Ou, Mírian Silva, Kate Soule, Yuekai Sun, Justin Solomon, Neil Thompson, and Mikhail Yurochkin. 2023. Large Language Model Routing with Benchmark Datasets. arXiv:2309.15789 [cs.LG]

    Pith/arXiv arXiv 2023

  27. [27]

    Lukas Wegmeth, Tobias Vente, and Joeran Beel. 2024. Recommender Systems Algorithm Selection for Ranking Prediction on Implicit Feedback Datasets. In Proc. of the 18th ACM Conf. on Recommender Systems (RecSys). ACM, New York, NY, USA, 1163–1167. https://doi.org/10.1145/3640457.3691718

    doi:10.1145/3640457.3691718 2024

  28. [28]

    David H. Wolpert. 1992. Stacked Generalization.Neural Networks5, 2 (1992), 241–259. https://doi.org/10.1016/S0893-6080(05)80023-1

    doi:10.1016/s0893-6080(05)80023-1 1992

  29. [29]

    Likang Wu, Zhi Zheng, Zhaopeng Qiu, Hao Wang, Hongchao Gu, Tingjia Shen, Chuan Qin, Chen Zhu, Hengshu Zhu, Qi Liu, Hui Xiong, and Enhong Chen. 2024. A Survey on Large Language Models for Recommendation.World Wide Web27 (2024), 60. https://doi.org/10.1007/s11280-024-01291-2

    doi:10.1007/s11280-024-01291-2 2024

  30. [30]

    xAI. 2025. Grok 4.1 Fast and Agent Tools API. https://x.ai/news/grok-4-1-fast Accessed July 10, 2026

    2025

  31. [31]

    Junjie Zhang, Ruobing Xie, Yupeng Hou, Wayne Xin Zhao, Leyu Lin, and Ji-Rong Wen. 2025. Recommendation as Instruction Following: A Large Language Model Empowered Recommendation Approach.ACM Transactions on Information Systems43, 5, Article 114 (2025), 114:1–114:37 pages. https://doi.org/10.1145/ 3708882

    2025

  32. [32]

    Yuyue Zhao, Jiancan Wu, Xiang Wang, Wei Tang, Dingxian Wang, and Maarten de Rijke. 2024. Let Me Do It For You: Towards LLM Empowered Recommendation via Tool Learning. InProc. of the 47th Int. ACM SIGIR Conf.ACM, New York, NY, USA, 1796–1806. https://doi.org/10.1145/3626772.3657828 8

    doi:10.1145/3626772.3657828 2024

This paper was first reviewed by grok-4.5 on July 14, 2026.