acceptodds
Under review as a conference paper at ICLR 2027

Toward User-centered Agentic Recommendation: When Does Platform Competition Benefit Users?

Abstract

Conventional recommender systems operate within platform-specific catalogs, limiting users' access to potentially valuable resources across platforms. We propose user-centered agentic recommendation, a paradigm in which personal user agents compare recommendations from competing platforms, and formalize their interaction as a repeated game. Platforms can strategically modify item descriptions in order to increase their perceived competitiveness, while user agents maintain each platform's credibility assessment using feedback from consumed items only. Our analysis shows that expanded resource access need not improve user welfare: strategic misrepresentation and limited verification can offset the benefits of platform competition. We quantify when platform credibility assessment allows users to benefit from broader access despite platform's strategic exaggeration. Our analysis also identifies conditions that guarantee higher expected welfare than the original platform-centered recommendation setup, with an interaction budget linear to the number of platforms. Experiments on synthetic and Amazon-grounded markets show that reputation can improve finite-horizon utility, while its benefits depend on market size and mechanism parameters.

Primary area:
learning theory
Keywords:
agentic recommendation, recommender system, welfare analysis, game theory

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_aCtwOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies when user-centered recommendation across competing platforms can improve user welfare. It models the interaction between user agents and strategic platform agents as a repeated game, where platforms can exaggerate item relevance and users update platform-specific credibility estimates using consumption feedback.

The paper derives finite-horizon welfare guarantees under bounded reporting discrepancies and shows that the sufficient interaction budget grows linearly with the number of platforms. A cold-start construction further illustrates why access to more platforms can initially reduce user utility. Experiments on synthetic and Amazon-grounded markets examine how credibility assessment affects user welfare under different market conditions and platform strategies.

Critical strengths
  1. The welfare analysis provides a substantive theoretical contribution.

Theorem 4.3 establishes a finite-horizon welfare bound using feedback from consumed items without requiring minimum exposure to every platform. Corollary 4.4 characterizes a sufficient interaction budget, while Proposition 4.5 provides a complementary cold-start example. Together, these results explain how resource availability and limited verification jointly affect user welfare. This analysis is the strongest reason to consider acceptance.

  1. The evaluation distinguishes different sources of welfare change.

The paper separates resource-access gains from losses caused by strategic reporting and imperfect credibility assessment. It evaluates fixed and adaptive platform strategies, includes alternative credibility updates, and reports cases where reputation improves utility without making users better off than their original platform. This distinction helps avoid overstating the benefits of platform competition.

Critical weaknesses
  1. The central welfare guarantee depends on a restrictive condition that is not established for the main competitive settings.

Theorem 4.3 relies on Assumption 4.2, which bounds the variation of reporting discrepancies across the evaluation horizon. However, strategically adapting platforms are not generally guaranteed to satisfy a sufficiently small bound. The paper also acknowledges that its broad-market experiments fall outside the useful guarantee region. Consequently, the theoretical result does not establish welfare improvement in the main competitive settings used to motivate the work. The conditional guarantee is valuable, but its practical applicability remains insufficiently demonstrated.

  1. The Amazon experiments do not provide convincing evidence about actual user welfare.

The Amazon evaluation constructs virtual platforms and defines consumption value using simulated relevance proxies rather than observed user satisfaction. More importantly, the conclusion about welfare improvement changes with the outcome proxy. Under the default semantic proxy, the reported net gain does not establish an improvement over the original platform, whereas the rating-derived proxy produces positive gains. Appendix E.8 also reports that the rating prediction model does not outperform a bias-only baseline in RMSE. Although the paper acknowledges these limitations, the experiments do not establish which welfare conclusions would hold for actual users. This significantly limits the empirical support for the proposed agentic recommendation paradigm.

  1. The analysis does not establish how strategic platforms behave under the proposed credibility mechanism.

Proposition 4.1 characterizes a weak Nash equilibrium only when credibility assessment is disabled. With credibility assessment enabled, the paper provides welfare bounds for admissible reporting policies but does not characterize equilibrium behavior or the incentives to change reporting strategies. The Exp3 experiments illustrate adaptive competition, but do not establish convergence or equilibrium properties. As a result, it remains unclear whether the strategic conditions assumed in the welfare analysis would arise when platforms actively respond to the proposed mechanism. This limits the game-theoretic interpretation of the results.

  1. The proposed credibility mechanism is not sufficiently compared with alternative user-side learning strategies.

The main experiments compare credibility assessment primarily against uncorrected recommendations and the original platform. Although Appendix E.7 includes an affine calibration baseline, its results show that neither method consistently dominates the other. The paper does not establish whether the proposed exponential averaging rule is competitive with stronger user-side learning strategies operating under the same consumption-feedback constraints. Consequently, the experiments demonstrate the value of using feedback to correct platform reports, but provide limited evidence that the specific proposed mechanism offers a meaningful advantage over alternative approaches.

Questions for authors

Can the authors clarify which practically relevant classes of adaptive platform policies satisfy Assumption 4.2 with a sufficiently small reporting-gap variation? How should Theorem 4.3 be interpreted for the broad-market experiments that fall outside the useful guarantee region?

Given that the sign of the Amazon welfare improvement depends on the outcome proxy, which empirical conclusions remain stable across the reported proxy choices? What evidence supports interpreting either proxy as a meaningful approximation of actual user welfare?

Appendix E.7 compares EMA credibility updates with an affine alternative. Can the authors clarify whether the main theoretical advantage comes from the specific EMA mechanism or from the more general use of consumption-based credibility correction?

Reviewer_BTR6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies whether platform competition benefits users when they use an agent to choose products across recommendation platforms. The method uses feedback after consumption to record how reliable each platform's past recommendations have been. It then adjusts the recommendation scores reported by the platforms. The paper analyzes the conditions under which users are guaranteed to benefit. It also conducts experiments in synthetic markets and simulated markets based on Amazon data.

Critical strengths
  1. The paper considers two practical issues: platforms may exaggerate product value, and users can only provide feedback on products they have consumed.
  2. The method only uses feedback after consumption and does not need to know the true value of every platform's products to the user.
  3. The experiments consider different numbers of platforms and different platform behaviors, with repeated runs using multiple random seeds.
Critical weaknesses
  1. Experimental comparisons with existing methods are limited.

The paper discusses the screening method of Kleine Buening et al., which appears relevant, but does not compare against it experimentally.

  1. Whether users benefit depends heavily on how user utility is calculated in the experiments.

The Amazon experiments mainly use the similarity between users' needs and products to represent user utility. However, this similarity is only weakly related to user ratings in the real rating samples examined in the paper. In some samples, higher similarity is associated with lower user ratings. In addition, using a different model to calculate user utility changes the average gain over the original platform from negative to positive.

  1. The introduction does not sufficiently explain what additional problems the paper solves beyond prior work.

The introduction devotes considerable space to the idea of user-centered recommendation, but does not sufficiently explain the limitations of existing methods or why the proposed method is needed. In addition, the descriptions of the experimental comparisons are scattered. I suggest bringing them together and explaining what each method does and why it was selected as a baseline.

Questions for authors
  1. Why does the paper not compare experimentally with the screening method of Kleine Buening et al.? In settings where both methods apply, what advantages does the proposed method offer?

  2. Changing the model used to calculate user utility changes the average gain from negative to positive. What evidence from real ratings or user preference data can show that users actually benefit?

Reviewer_2j5iOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This is a paper on agentic recommendations. The setting is the following: there is a repeated interaction over TT rounds. At each round tt, a user jj with interests uju_j posts a query qj,t∼Qjq_{j,t} \sim Q_j to their user agent UAjUA_j. UAjUA_j then forwards the query to all available “platform agents” (PAiPA_i for i∈[n]i \in [n]) — these are agents of separate platforms that may have some resources to answer the query. After each platform returns their response Aji,tA_{ji, t}, the UAjUA_j does a cross-platform comparison, and presents to the user their best (agent) choice. The user consumes from this choice and there’s feedback returned to the platform. The paper asks when does competition among PAs improve user utility (compared to a single fixed platform).

Critical strengths

[S1] I think the question addressed in this work is very important in the evolving ecosystem of personal assistants. I agree with the authors that more and more human users will start going either to a chatbot or a user agent / personal assistant for specific recommendations and so then, it is the user agent that will be tasked with ranking / combining / synthesizing the recommendations received from the separate platforms.

Critical weaknesses

[W1] Fundamentally, this is an Applied Modeling paper, so I’m evaluating as such. I do not agree that there is a common “relevance function” that can dictate the responses obtained by the different platforms at each query — note also that there’s no explanation in the writeup as to why this is a reasonable modeling assumption. At a high-level, my criticism of the relevant function is that it imposes a lot of structure on the quality / type of responses obtained by the different platforms, without including enough specific information for each platform. I also don’t agree that the intensity is the same across users, i.e., \lambda_{i,t} should not be the same across jj. Indeed, a user query either reveals some things about the user themselves or the UA_j already has context about this user which they are probably using when assessing the queries. As such, it can be the case that PA_i uses different intensities \lambda_{i,t} based on the specific type of user they are facing.

Separately, I do not agree that the “platform credibility assessment” is the only or even the primary way in which the UA is assessing how to rank / present the results from PA. Same to above, I didn’t find the authors’ explanation of this modeling choice adequate.

[W2] There are some clarity issues: for example, in the intro the paragraphs from lines 57 - 69 are very hard to parse. This is unfortunate because these are the 2 paragraphs that highlight the contributions of the work for the first time. Another example, in the paragraph about “LLM-based recommendation and strategic presentation”: what is the “exaggeration” that the authors refer to? Also at this point of the manuscript, it’s not clear to me (maybe because of the introduction having unclear parts) what the “resource value” corresponds to. The entire Related Work section reads as very AI-generated to me, and I think that it should be completely re-written in a revised version. Note that this is not clearly articulated in the present manuscript’s AI use statement. There are other parts of the paper that read as AI-generated. I would encourage the authors to make a thorough pass and fix this.

Finally, I would strongly suggest to the authors to rewrite the Discussion and Conclusion; they are currently reading as very AI-y. I was expecting at this part of the paper to see more of the authors’ vision of the research area; they have made a lot of important assumptions for their model, and I was hoping that the Discussion would include some discussion on these assumptions, and where the authors think that this research question is heading wrt different modeling assumptions / approaches.

[W3] If i’m not mistaken, then main theoretical result about the platform competition and its effects to welfare could also be proved for a single user. As such, I could not understand why the paper chose to talk about multiple users.

Additional Comments

  • From the abstract, it was unclear what an “Amazon-grounded” market is
  • Typo in line 35-36: grammar “and can interpret personal information need and …” doesn’t parse
Questions for authors

Feel free to address any of the questions from my criticisms above. The only extra question that I would like addressed is: could you be more specific about the way in which you used AI for the paper? For example, what parts of the intro and the Conclusions were AI generated? If the paper were to be accepted, can you provide a concrete roadmap of things you will change to make the writeup clearer?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.