Pith. sign in

REVIEW 5 major objections 5 minor 61 references

The paper argues that controllability — a recommender system's capacity to respond to explicit user guidance — can be measured by an evolutionary multi-agent probing framework, and that applying it reveals a systemic ceiling: long-tail cont

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

2026-08-02 05:13 UTC pith:TANXMVCH

load-bearing objection Worth engaging as a new controllability-evaluation framework, but the headline 'long-tail ceiling' claim measures the probes as much as the recommender, and the reporting needs variance and reconciliation. the 5 major comments →

arxiv 2607.13418 v2 pith:TANXMVCH submitted 2026-07-15 cs.IR cs.AI

Can We Steer the Black-Box? Towards Controllability-Centric Evaluation of Recommender Systems with Collaborative Agents

Jiwen Zhou , Xiang Liu , Mingming Li , Pengbo Mo , Jiao Dai , Honglei Lv , Jizhong Han , Songlin Hu This is my paper
classification cs.IR cs.AI
keywords controllabilityrecommender systemsblack-box evaluationmulti-agent probingpopularity biaslong-tail contentLLM agentsalgorithmic auditing
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.

CtrlBench-Rec is proposed as a standardized way to measure how much a recommender system can be steered by explicit user intent, even when the system's internals are hidden. The framework dispatches a population of LLM-driven agents to interact with the recommender under a strict interaction budget, then evolves them into a compact set of "super probes" through clustering and debate. These probes are deployed on three tasks: finding specific target content, reshaping the user's interest profile, and pulling recommendations away from popular items toward long-tail ones. The central empirical claim is that controllability has a ceiling: coverage saturates near 55% even with 200 interaction rounds, and long-tail genres such as Film-Noir remain almost entirely out of reach. A sympathetic reader would take this as evidence that popularity bias marginalizes long-tail items at the representation level, so probe-level behavior alone cannot fix it.

Core claim

On the paper's own terms, the discovery is twofold. Methodologically, it provides a reusable, budget-aware evaluation protocol that quantifies controllability across five recommender architectures and two datasets. Empirically, it reports a 'strict upper bound on explorable potential': with an extended budget of t=200 rounds, the best-performing controller saturates at roughly 55% coverage, and the long-tail genre Film-Noir (1.1% of items) reaches only about 6.8% coverage. The authors conclude that the black-box system 'cannot establish reliable retrievable associations due to inherent representation sparsity,' and that probe-level optimization alone cannot overcome this barrier. This is pre

What carries the argument

The central object is the 'super probe': an agent profile synthesized by an evolutionary multi-agent fusion pipeline. A population of 100 seed agents, each carrying a persona and a short behavior stream, interacts with the recommender; their trajectories are embedded, clustered with K-means, and agents within each cluster exchange and debate their strategies; a merge step distills each cluster into a single consolidated profile. This profile is then re-deployed for further rounds. The super probes serve as the measuring instrument: by converting evaluation into an optimizable probe-cultivation process, the framework turns controllability assessment into a search for the most effective steeri

Load-bearing premise

The load-bearing premise is that the LLM-driven super probes are a ceiling-reaching controller — that no stronger or simpler probing strategy could do materially better; the paper never validates this, so the 'structural' long-tail barrier could actually be a probe-capability barrier.

What would settle it

Run the identical 200-round budget with a controller that has catalog access and explicitly searches for target items (or with a non-LLM algorithmic controller that greedily clicks any target item that appears). If coverage rises substantially above the reported ~55% saturation, the claimed upper bound is not a property of the recommender. A second check: count how often target items appear in the recommendation lists of unsteered random users; if they appear but agents fail to reach them, the bottleneck is probe steering, not representation sparsity.

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

If this is right

  • Controllability can be treated as a standard, quantifiable evaluation dimension alongside accuracy and fairness, enabling comparable audits across different recommender systems.
  • Long-tail resistance appears to be a structural property of current recommender representations, not an artifact of a particular probing method; fixing it likely requires architectural or training changes, not just better user behavior.
  • A short seed history of 4-6 interactions is enough to establish effective steering, meaning controllability audits can start with minimal user data.
  • The three-task taxonomy (target discovery, profile shaping, popularity mitigation) can serve as a shared protocol for user-empowerment and regulatory tools.
  • If the 55% saturation generalizes, user-facing steering features will have a hard ceiling on niche content discovery regardless of interface effort.

Where Pith is reading between the lines

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

  • A natural test of the ceiling claim is to compare against an oracle controller that knows the item catalog and can directly issue searches for target items; if such a controller breaks past 55% coverage, the bottleneck is in the probe interface rather than the recommender's representation.
  • The framework's 'controllability' score conflates the recommender's response with the probe's skill; separating these would require calibrating the same probes against a deliberately steerable surrogate recommender.
  • One could extend the audit to measure how controllability changes as the recommender is retrained or as popularity distributions shift, turning a single ceiling measurement into a monitoring signal.
  • The observed 'medium-popularity trap' — popular genres being harder to steer than niche ones due to semantic dispersion — is a testable prediction about embedding geometry: target categories with high internal variance should show lower steerability.

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

5 major / 5 minor

Summary. The paper proposes CtrlBench-Rec, a multi-agent evolutionary framework for evaluating the controllability of black-box recommender systems. It formalizes three tasks — target content discovery, interest profile shaping, and popularity bias mitigation — and evolves a large population of LLM-driven agent probes into a compact set of 'super probes' via clustering, debate, and profile fusion. Experiments on MovieLens-1M and Amazon Toys & Games with five recommender backbones (NARM, SASRec, TwHIN-BERT, BGE, Qwen3.5-4B) compare the framework against random-policy baselines and a KMeans-Fusion ablation. The paper claims that CtrlBench-Rec provides the first standardized controllability toolkit and reports a systemic bottleneck: long-tail items resist steering, with coverage saturating around 55% even at 200 interaction rounds.

Significance. If the framework and the empirical saturation finding were fully validated, the contribution would be substantial: controllability is a real, under-measured dimension of recommender systems, and a reusable evaluation toolkit with released code would benefit algorithmic auditing and user-empowerment research. The paper also deserves credit for formalizing three distinct controllability tasks, evaluating across multiple recommender architectures, and publicly releasing code. However, the headline claims are currently stronger than the evidence. The 55% saturation ceiling and the 'inherent representation sparsity' interpretation are not established because no ceiling controller is tested; Task 2 has no baseline comparison; and no variance or statistical significance is reported anywhere. These are load-bearing issues for both the framework-validation claim and the empirical finding, so the current manuscript is not yet ready for acceptance.

major comments (5)
  1. [§4.2, 'Exhaustive exploration' and Fig. 4] The claim that 'a strict upper bound exists on explorable potential' and that saturation near 55% reflects the recommender's 'inherent representation sparsity' is not supported by the experiments. Because the agent can only click items in the returned list Lt (Appendix B), coverage measures the joint probe–recommender system. All compared baselines are either random policies or prompt variants; no ceiling controller is tested, e.g., an agent with access to target item metadata, a search tool over item IDs/titles, or an oracle that can inject arbitrary behavior sequences. A stronger controller could plausibly raise the saturation ceiling, making the result an artifact of probe capability rather than a property of the recommender. The manuscript must either add such a ceiling controller or substantially weaken the 'inherent sparsity' conclusion.
  2. [§4.3, Task 2 (Interest Profile Shaping)] Task 2 reports no baseline comparison at all. Table 2 and Figure 7 show only CtrlBench-Rec's similarity values, yet the text claims the framework achieves the 'steepest initial accuracy slope' and 'highest final similarity' (Level-2). Without comparing against Base-Large, Base-Small, KMeans-Fusion, or a simple random-click controller, these claims are unsubstantiated. The measured plasticity could reflect the recommender's behavior under any reasonable interaction sequence, not the specific contribution of CtrlBench-Rec. At minimum, the random-policy baselines used in Task 1 should be run for Task 2.
  3. [All experiments (Tables 1, 2, 7; Figures 3–5, 8–11)] No error bars, seeds, or variance statistics are reported anywhere. Agents are driven by an LLM at temperature 1.0 (Appendix D.2), so interaction trajectories and outcomes are stochastic. The claimed 'consistent and substantial performance leap' over baselines, the 55% saturation value, and all coverage/efficiency comparisons are point estimates that could change substantially across runs. The manuscript should report means and standard deviations over multiple independent seeds, and ideally significance tests, for all quantitative claims.
  4. [§4.2 vs. Table 1 and Table 7] Several key numbers conflict between the prose and the tables. In §4.2, the BGE Amazon environment is said to give CtrlBench-Rec 'approximately 22.5% coverage at t=20, nearly 1.8 times that of Base-Small (12.5%)', but Table 1 reports CtrlBench-Rec 27.41% and Base-Small 23.35% for that condition. Similarly, Table 7 reports aggregate Amazon coverage for CtrlBench-Rec as 13.85%, which is not reconciled with the per-environment Table 1 value. Also, §4.2 states a NARM Base-Large exploration efficiency of 25.0 versus CtrlBench-Rec's 'stable 13.0', while Table 1 shows 19.42 and 13.50 respectively. These discrepancies undermine the reliability of the reported results and must be corrected.
  5. [§4.1 and 'effectively quantifies controllability'] The evaluation is circular in an important sense: the framework's validity is demonstrated by comparing its own probes against random baselines under task definitions that the authors designed. As §4.1 states, no established external baselines exist. This supports a weak version of the claim — the proposed probes are more effective than random exploration — but not the stronger claim that CtrlBench-Rec provides a standardized, independently meaningful measure of system controllability. The paper should either provide external validation (e.g., known controllable vs. non-controllable systems), or explicitly frame the contribution as a benchmark proposal and temper the abstract/conclusion claims about 'first standardized toolkit'.
minor comments (5)
  1. [Appendix C vs. §4.1/§D.5] The embedding model is referred to as 'Twin-BERT' in Appendix C but 'TwHIN-BERT' elsewhere. Please unify the nomenclature.
  2. [Abstract and §4.1] The code URLs differ: the abstract gives https://github.com/caskcsg/CtrlBenchRec, while §4 gives an anonymous.4open.science link. Use a single consistent URL depending on the intended anonymity stage.
  3. [§4.2, Figure 3] The figure caption and the figures themselves are partially garbled in the provided PDF (e.g., unicode escape sequences). Please ensure the final PDF renders all figures legibly, especially Figure 3 and Figure 4.
  4. [Appendix B, Table 3] Linit is described both as 'seed behavior length' and as the 'strictly budgeted' interaction sequence length. Clarify whether Linit is the pre-interaction seed history or the total evaluation budget, and use distinct notation for the two.
  5. [§6, Conclusion] The phrase 'first standardized toolkit' appears in the Abstract and Conclusion but is not positioned relative to related evaluation frameworks cited in §5.2 (e.g., AgentRecBench, RecBole). A sentence explaining what 'standardized' means in this context, and what the new framework adds over those works, would help the reader calibrate the novelty claim.

Circularity Check

0 steps flagged

No significant circularity: the 55% ceiling is an interpretive overreach, not a definitional reduction.

full rationale

The paper's target sets T are defined externally from genre/popularity metadata (Appendix D.1, Table 4), and the coverage metric in Table 5 is the natural normalization of the optimization objective in Eq. (1); no equation in the paper defines the measured quantities in terms of the conclusions drawn from them. The headline saturation result ('strict upper bound ... inherent representation sparsity,' §4.2) is an empirical claim about the CtrlBench-Rec probe population after 200 rounds. It may overstate the recommender's properties because no ceiling controller is tested and actions are restricted to the returned list Lt (Appendix B), but that is a validity/overgeneralization concern, not a circular reduction: the bound is not forced by the definitions, and no fitted parameter is renamed as a prediction. The framework's internal baselines (Base-Large, Base-Small, KMeans-Fusion) are comparisons under the same protocol rather than independent ground truth, which is a standard benchmark limitation. I found no load-bearing self-citation: the framework builds on external systems (Oasis, DeepSeek, TwHIN-BERT, etc.) and no uniqueness theorem from the authors is invoked to rule out alternatives. Accordingly, no step exhibits Eq. X = Eq. Y by construction or a fitted-input-called-prediction pattern.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 1 invented entities

The central measurement relies on several under-specified design choices: agent count, seed history length, embedding blend, and LLM sampling. The framework also assumes that clicking modifies recommender state and that LLM agents are neutral probes. Because no independent ground truth for controllability exists, the ledger is dominated by domain assumptions that the paper does not separately validate.

free parameters (5)
  • initial agent count Kinit = 100
    Chosen by design; all results depend on this population size, which is not derived from first principles.
  • super-probe count Ksuper = ≈26
    Emerges from two halving fusion rounds; comparisons to Base-Small are meant to control for it, but the final size is an implementation choice.
  • seed behavior length Linit = 4–6
    Selected after a sweep on the evaluation datasets (Figure 5), making it a protocol parameter fitted to the testbed.
  • embedding blend coefficient α = not reported
    Vfinal = α·Vprofile + (1−α)·Vmean_likes is defined in Appendix C but α is never given a value, so the clustering/representation step is under-specified.
  • LLM sampling temperature = 1.0
    High stochasticity in the decision engine affects reproducibility of all agent behaviors and therefore all reported metrics.
axioms (5)
  • domain assumption Clicking items in a recommender's ranked list changes its internal user state and can steer future recommendations
    Underlies all three tasks and the entire interaction protocol; never validated independently for the five recommender models used.
  • domain assumption LLM agents with only a short-term memory and a text profile are a valid stand-in for user/regulator steering
    The framework's central instrument is a language model; no evidence links its behavior to real user controllability or regulatory intent.
  • domain assumption K-means on BERT/Twin-BERT embeddings produces behaviorally meaningful clusters
    Stage 3 of the fusion process depends on this clustering without a clustering-quality check or alternative representations.
  • ad hoc to paper The target sets, defined by genre/popularity gradients, are a fair operationalization of controllability goals
    Target sets are constructed by the authors; the long-tail finding depends on this choice, and no sensitivity analysis is provided.
  • domain assumption The recommenders being probed are effectively static during evaluation
    The framework mentions non-stationarity as a challenge but does not model adaptive recommender updates in the experiments.
invented entities (1)
  • Super probe no independent evidence
    purpose: A merged user profile produced from clustered agents via debate and fusion, intended as a compact, high-capability evaluation instrument.
    No external validation that super probes measure controllability better than other strong controllers; their superiority is demonstrated only against internal random baselines.

pith-pipeline@v1.3.0-alltime-deepseek · 28123 in / 9453 out tokens · 100623 ms · 2026-08-02T05:13:13.136757+00:00 · methodology

0 comments
read the original abstract

Recommender systems operate as Black-Boxes, leaving users and regulators unable to steer their outputs toward specific intentions or audit their behavior. This lack of controllability, defined as the system's ability to respond to explicit guidance, remains an unaddressed dimension in existing evaluation paradigms. To fill this gap, we propose CtrlBench-Rec, a collaborative multi-agent framework for systematic assessment of controllability. We formalize three fundamental tasks: target content discovery, interest profile shaping, and popularity bias mitigation, which together measure steerability from explicit commands to implicit representation steering and finally to overcoming algorithmic biases.Extensive experiments on real-world datasets and multiple recommendation models demonstrate that our framework effectively quantifies controllability and exposes critical system bottlenecks, most notably persistent resistance to guiding long tail content. CtrlBench-Rec provides the first standardized toolkit for controllable recommendation research, algorithmic auditing, and user empowerment. Our code is released on https://github.com/caskcsg/CtrlBenchRec.

Figures

Figures reproduced from arXiv: 2607.13418 by Honglei Lv, Jiao Dai, Jiwen Zhou, Jizhong Han, Mingming Li, Pengbo Mo, Songlin Hu, Xiang Liu.

Figure 1
Figure 1. Figure 1: A taxonomy of controllability tasks for Black [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The overview of the evolutionary training process for the [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 5
Figure 5. Figure 5: Impact of initial behavior flow on perfor [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 4
Figure 4. Figure 4: Extended interaction coverage curves. 22.5% coverage at t = 20, nearly 1.8 times that of Base-Small (12.5%). This demon￾strates that our LLM-driven evolutionary fu￾sion mechanism can capture system-specific retrieval signals with far greater precision than random exploration, even with an identical number of agents. • CtrlBench-Rec delivers vastly superior ef￾ficiency and economic feasibility compared to t… view at source ↗
Figure 6
Figure 6. Figure 6: Interest profile shaping task. t0 t3 t6 t9 t12 t15 t18 t20 Interaction Rounds 0.35 0.40 0.45 0.50 0.55 0.60 0.65 0.70 Similarity Score (%) Level-1 Level-2 Level-3 Old Interest [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Evolution of target-profile similarity under [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Newly discovered target items across interac [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: Coverage comparison across different recommendation architectures. [PITH_FULL_IMAGE:figures/full_fig_p016_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Coverage evolution over interaction rounds. [PITH_FULL_IMAGE:figures/full_fig_p017_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: illustrates the interactive discussion pro￾cess during the agent fusion phase. The user debate process presented below mainly adopts a two-round debate format. The first round is for independent statements. During this stage, each Agent is estab￾lishing their own personality profile, identifying commonalities, and initially constructing a consen￾sus around the core preferences of deep plot and character-d… 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

61 extracted references · 9 linked inside Pith

  1. [1]

    proceedings of the 30th acm international conference on information & knowledge management , pages=

    Recbole: Towards a unified, comprehensive and efficient framework for recommendation algorithms , author=. proceedings of the 30th acm international conference on information & knowledge management , pages=

  2. [2]

    International Conference on Computation of Artificial Intelligence & Machine Learning , pages=

    A comprehensive survey of evaluation techniques for recommendation systems , author=. International Conference on Computation of Artificial Intelligence & Machine Learning , pages=. 2024 , organization=

    2024

  3. [3]

    2018 IEEE international conference on data mining (ICDM) , pages=

    Self-attentive sequential recommendation , author=. 2018 IEEE international conference on data mining (ICDM) , pages=. 2018 , organization=

    2018

  4. [4]

    arXiv preprint arXiv:2302.04556 , year=

    Auditing Recommender Systems--Putting the DSA into practice with a risk-scenario-based approach , author=. arXiv preprint arXiv:2302.04556 , year=

    Pith/arXiv arXiv

  5. [5]

    ACM Transactions on Information Systems , volume=

    Bias and debias in recommender system: A survey and future directions , author=. ACM Transactions on Information Systems , volume=. 2023 , publisher=

    2023

  6. [6]

    ACM Transactions on Information Systems , volume=

    Recommender ai agent: Integrating large language models for interactive recommendations , author=. ACM Transactions on Information Systems , volume=. 2025 , publisher=

    2025

  7. [7]

    arXiv preprint arXiv:2510.12742 , year=

    CTRL-Rec: Controlling recommender systems with natural language , author=. arXiv preprint arXiv:2510.12742 , year=

    arXiv

  8. [8]

    Proceedings of the 47th international ACM SIGIR conference on research and development in Information Retrieval , pages=

    On generative agents in recommendation , author=. Proceedings of the 47th international ACM SIGIR conference on research and development in Information Retrieval , pages=

  9. [9]

    Proceedings of the ACM Web Conference 2024 , pages=

    Agentcf: Collaborative learning with autonomous language agents for recommender systems , author=. Proceedings of the ACM Web Conference 2024 , pages=

    2024

  10. [10]

    Proceedings of the 48th International ACM SIGIR conference on Research and Development in Information Retrieval , pages=

    Agentic feedback loop modeling improves recommendation and user simulation , author=. Proceedings of the 48th International ACM SIGIR conference on Research and Development in Information Retrieval , pages=

  11. [11]

    ACM Transactions on Information Systems , volume=

    User behavior simulation with large language model-based agents , author=. ACM Transactions on Information Systems , volume=. 2025 , publisher=

    2025

  12. [12]

    arXiv preprint arXiv:2505.16429 , year=

    Beyond Static Testbeds: An Interaction-Centric Agent Simulation Platform for Dynamic Recommender Systems , author=. arXiv preprint arXiv:2505.16429 , year=

    arXiv

  13. [13]

    CEUR Workshop Proceedings , volume=

    Librec: A java library for recommender systems , author=. CEUR Workshop Proceedings , volume=. 2015 , organization=

    2015

  14. [14]

    Proceedings of the 45th international ACM SIGIR conference on research and development in information retrieval , pages=

    User-controllable recommendation against filter bubbles , author=. Proceedings of the 45th international ACM SIGIR conference on research and development in information retrieval , pages=

  15. [15]

    Foundations and Trends

    Explainable recommendation: A survey and new perspectives , author=. Foundations and Trends. 2020 , publisher=

    2020

  16. [16]

    Explainable Recommendation: A Survey and New Perspectives , volume=

    Zhang, Yongfeng and Chen, Xu , year=. Explainable Recommendation: A Survey and New Perspectives , volume=. Foundations and Trends® in Information Retrieval , publisher=. doi:10.1561/1500000066 , number=

    doi:10.1561/1500000066

  17. [17]

    arXiv preprint arXiv:2407.06083 , year=

    A survey of controllable learning: Methods and applications in information retrieval , author=. arXiv preprint arXiv:2407.06083 , year=

    Pith/arXiv arXiv

  18. [18]

    Proceedings of the 46th international ACM SIGIR conference on research and development in information retrieval , pages=

    When search meets recommendation: Learning disentangled search representation for recommendation , author=. Proceedings of the 46th international ACM SIGIR conference on research and development in information retrieval , pages=

  19. [19]

    arXiv preprint arXiv:2411.11581 , year=

    Oasis: Open agent social interaction simulations with one million agents , author=. arXiv preprint arXiv:2411.11581 , year=

    Pith/arXiv arXiv

  20. [20]

    2023 , eprint=

    C-Pack: Packaged Resources To Advance General Chinese Embedding , author=. 2023 , eprint=

    2023

  21. [21]

    arXiv preprint arXiv:2209.07562 , year=

    TwHIN-BERT: A Socially-Enriched Pre-trained Language Model for Multilingual Tweet Representations , author=. arXiv preprint arXiv:2209.07562 , year=

    Pith/arXiv arXiv

  22. [22]

    Proceedings of the 2017 ACM on Conference on Information and Knowledge Management , pages=

    Neural attentive session-based recommendation , author=. Proceedings of the 2017 ACM on Conference on Information and Knowledge Management , pages=

    2017

  23. [23]

    European Conference on Information Retrieval , pages=

    Large language models are zero-shot rankers for recommender systems , author=. European Conference on Information Retrieval , pages=. 2024 , organization=

    2024

  24. [24]

    ACM Transactions on Information Systems , volume=

    Multi-agent attacks for black-box social recommendations , author=. ACM Transactions on Information Systems , volume=. 2024 , publisher=

    2024

  25. [25]

    IEEE Transactions on Knowledge and Data Engineering , volume=

    Adversarial attacks for black-box recommender systems via copying transferable cross-domain user profiles , author=. IEEE Transactions on Knowledge and Data Engineering , volume=. 2023 , publisher=

    2023

  26. [26]

    Neural Computing and Applications , volume=

    A novel shilling attack on black-box recommendation systems for multiple targets , author=. Neural Computing and Applications , volume=. 2025 , publisher=

    2025

  27. [27]

    arXiv preprint arXiv:2512.11426 , year=

    AgentBalance: Backbone-then-Topology Design for Cost-Effective Multi-Agent Systems under Budget Constraints , author=. arXiv preprint arXiv:2512.11426 , year=

    arXiv

  28. [28]

    ACM Transactions on Information Systems , volume=

    A thorough performance benchmarking on lightweight embedding-based recommender systems , author=. ACM Transactions on Information Systems , volume=. 2025 , publisher=

    2025

  29. [29]

    ACM Transactions on Intelligent Systems and Technology , volume=

    Fairness and diversity in recommender systems: a survey , author=. ACM Transactions on Intelligent Systems and Technology , volume=. 2025 , publisher=

    2025

  30. [30]

    arXiv preprint arXiv:2308.00894 , year=

    User-controllable recommendation via counterfactual retrospective and prospective explanations , author=. arXiv preprint arXiv:2308.00894 , year=

    Pith/arXiv arXiv

  31. [31]

    2020 , publisher=

    Bandit algorithms , author=. 2020 , publisher=

    2020

  32. [32]

    ACM Computing Surveys (CSUR) , volume=

    A survey on conversational recommender systems , author=. ACM Computing Surveys (CSUR) , volume=. 2021 , publisher=

    2021

  33. [33]

    Artificial Intelligence Review , volume=

    Multi-agent deep reinforcement learning: a survey , author=. Artificial Intelligence Review , volume=. 2022 , publisher=

    2022

  34. [34]

    Journal of Machine Learning Research , volume=

    F2a2: Flexible fully-decentralized approximate actor-critic for cooperative multi-agent reinforcement learning , author=. Journal of Machine Learning Research , volume=

  35. [35]

    International conference on machine learning , pages=

    Cooperative exploration for multi-agent deep reinforcement learning , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  36. [36]

    The 41st international acm sigir conference on research & development in information retrieval , pages=

    Conversational recommender system , author=. The 41st international acm sigir conference on research & development in information retrieval , pages=

  37. [37]

    Artificial Intelligence Review , volume=

    Evaluating conversational recommender systems: A landscape of research , author=. Artificial Intelligence Review , volume=. 2023 , publisher=

    2023

  38. [38]

    Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining , pages=

    Model-agnostic counterfactual reasoning for eliminating popularity bias in recommender system , author=. Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining , pages=

  39. [39]

    Vicinagearth , volume=

    A survey on LLM-based multi-agent systems: workflow, infrastructure, and challenges , author=. Vicinagearth , volume=. 2024 , publisher=

    2024

  40. [40]

    arXiv preprint arXiv:2602.03794 , year=

    Understanding Agent Scaling in LLM-Based Multi-Agent Systems via Diversity , author=. arXiv preprint arXiv:2602.03794 , year=

    arXiv

  41. [41]

    arXiv preprint arXiv:2510.03215 , year=

    Cache-to-cache: Direct semantic communication between large language models , author=. arXiv preprint arXiv:2510.03215 , year=

    arXiv

  42. [42]

    arXiv preprint arXiv:2511.20639 , year=

    Latent collaboration in multi-agent systems , author=. arXiv preprint arXiv:2511.20639 , year=

    Pith/arXiv arXiv

  43. [43]

    arXiv preprint arXiv:2511.09149 , year=

    Enabling Agents to Communicate Entirely in Latent Space , author=. arXiv preprint arXiv:2511.09149 , year=

    Pith/arXiv arXiv

  44. [44]

    adaptable commitment interfaces

    Switchtube: a proof-of-concept system introducing “adaptable commitment interfaces” as a tool for digital wellbeing , author=. Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems , pages=

    2023

  45. [45]

    Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems , pages=

    How the design of youtube influences user sense of agency , author=. Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems , pages=

    2021

  46. [46]

    2023 2nd International Conference on Applied Artificial Intelligence and Computing (ICAAIC) , pages=

    Multi-agent personalized recommendation system in e-commerce based on user , author=. 2023 2nd International Conference on Applied Artificial Intelligence and Computing (ICAAIC) , pages=. 2023 , organization=

    2023

  47. [47]

    Advances in Neural Information Processing Systems , volume=

    Multi-agent reinforcement learning is a sequence modeling problem , author=. Advances in Neural Information Processing Systems , volume=

  48. [48]

    International Conference on Machine Learning , pages=

    Off-belief learning , author=. International Conference on Machine Learning , pages=. 2021 , organization=

    2021

  49. [49]

    Proceedings of the 13th international conference on web search and data mining , pages=

    Estimation-action-reflection: Towards deep interaction between conversational and recommender systems , author=. Proceedings of the 13th international conference on web search and data mining , pages=

  50. [50]

    The twelfth international conference on learning representations , year=

    MetaGPT: Meta programming for a multi-agent collaborative framework , author=. The twelfth international conference on learning representations , year=

  51. [51]

    Acm transactions on interactive intelligent systems (tiis) , volume=

    The movielens datasets: History and context , author=. Acm transactions on interactive intelligent systems (tiis) , volume=. 2015 , publisher=

    2015

  52. [52]

    Justifying recommendations using distantly-labeled reviews and fine-grained aspects , author=. Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP) , pages=

    2019

  53. [53]

    arXiv preprint arXiv:2510.11062 , year=

    Stronger-MAS: Multi-Agent Reinforcement Learning for Collaborative LLMs , author=. arXiv preprint arXiv:2510.11062 , year=

    arXiv

  54. [54]

    Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    ConvApparel: A Benchmark Dataset and Validation Framework for User Simulators in Conversational Recommenders , author=. Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  55. [55]

    Findings of the Association for Computational Linguistics: EACL 2026 , pages=

    Toward safe and human-aligned game conversational recommendation via multi-agent decomposition , author=. Findings of the Association for Computational Linguistics: EACL 2026 , pages=

    2026

  56. [56]

    arXiv preprint arXiv:2509.10397 , year=

    Recoworld: Building simulated environments for agentic recommender systems , author=. arXiv preprint arXiv:2509.10397 , year=

    Pith/arXiv arXiv

  57. [57]

    arXiv preprint arXiv:2604.17459 , year=

    Transparent and Controllable Recommendation Filtering via Multimodal Multi-Agent Collaboration , author=. arXiv preprint arXiv:2604.17459 , year=

    Pith/arXiv arXiv

  58. [58]

    Advances in Neural Information Processing Systems , volume=

    Agentrecbench: Benchmarking llm agent-based personalized recommender systems , author=. Advances in Neural Information Processing Systems , volume=

  59. [59]

    Proceedings of the Nineteenth ACM International Conference on Web Search and Data Mining , pages=

    Multi-Agent Video Recommenders: Evolution, Patterns, and Open Challenges , author=. Proceedings of the Nineteenth ACM International Conference on Web Search and Data Mining , pages=

  60. [60]

    arXiv preprint arXiv:2601.04554 , year=

    Exploring Recommender System Evaluation: A Multi-Modal User Agent Framework for A/B Testing , author=. arXiv preprint arXiv:2601.04554 , year=

    arXiv

  61. [61]

    Proceedings of the ACM Web Conference 2026 , pages=

    Multi-agent collaborative filtering: Orchestrating users and items for agentic recommendations , author=. Proceedings of the ACM Web Conference 2026 , pages=

    2026