acceptodds
Under review as a conference paper at ICLR 2027

Welfare-Enhancing Mechanism Design for Agent Ecosystems

Abstract

The emerging Agentic Web forms a multi-agent ecosystem: a decentralized service market where the user agent selects which content agents to query from their submitted profiles, before seeing any documents. Profiles therefore become a new target of optimization. Existing Generative Engine Optimization studies how a single supplier rewrites documents the engine can already retrieve, to improve their citation. In the Agentic Web, multiple content agents optimize profiles to compete for limited opportunities to deliver documents. To study this repeated interaction, we build WildRegistry, a simulation framework and benchmark based on a real agentic-search request stream, and model it as a dynamic game under private monitoring. Long-horizon experiments show that this competition lowers website-selection quality and stabilizes at a harmful approximate rewriting equilibrium. We therefore design a welfare-enhancing mechanism, DEC (Delivery-Evidence Credit), a platform-side policy that uses recorded delivery outcomes to reallocate later opportunities and reduce information asymmetry. DEC improves this endpoint, raising website-selection quality while preserving most of providers' exposure gains, for a Welfare Gain of points and a healthier long-horizon outcome. More broadly, our findings highlight the importance of mechanism design in agent ecosystems: turning recorded delivery outcomes into evidence that endogenously shapes later user-agent selection and content-agent adaptation can steer long-horizon adaptation toward delivered capabilities.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Agent Ecosystems, Agentic Web, Mechanism Design

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_ESXfOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies providers competing for agentic-search exposure by rewriting website profiles. WildRegistry replays a real request stream, while Delivery-Evidence Credit uses past delivery to reward or withhold future opportunities. The mechanism improves website selection and retains most adaptive providers’ exposure gains in the tested simulations.

Critical strengths
  1. The benchmark captures a useful feedback loop: selection determines the evidence available for later provider adaptation. Credit and Withhold have clear operational roles, supported by component ablations and profile-replacement tests.
  2. Evaluation extends beyond headline scores. The appendix includes human relevance checks, session-level uncertainty estimates, alternative backbones and provider groups, and sensitivity to welfare weights. Offline relevance labels are excluded from the agents’ decisions.
Critical weaknesses
  1. The reported rewriting equilibrium establishes stability of one prescribed updater. The alternative provider strategies are mostly submitted once, whereas the main updater learns continually. These comparisons do not establish resistance to unilateral adaptive deviations. The appendix scopes this limitation; the main text should consistently use the narrower stability claim.
  2. Reported confidence intervals condition on realized trajectories. Backbone and provider-group checks are helpful, but independent repetitions of the full coupled adaptation process are missing. It remains unclear how reliably the long-run gains recur when early selections and provider updates change.
Questions for authors
  1. Can a provider profitably deviate from the settled profiles using a different continual updater while competitors remain fixed? Otherwise, could the paper consistently describe its endpoint as updater stability?
  2. Can the authors repeat the complete adaptation trajectory and report variation in selection quality and provider exposure, rather than only uncertainty over requests within a run?
4 comments

We thank the reviewer for drawing a useful distinction between updater-induced stability and a website-level adaptive deviation, and for examining the reproducibility of the long-horizon result.

In short. The second continual updater reproduces the stable endpoint and DEC's gains under a different update rule, while two full-trajectory replays reproduce DEC's gains under different realized histories. We clarify the definition and answer each empirical question below, with the corresponding additions to the revised PDF.

W1 & Q1: We agree that we should describe more precisely what "equilibrium" means in our empirical setting.

Section 3.3 (pp. 3–4) defines our game with two players: the user agent and the adaptive content agents together forming the provider-side player, whose objective is their summed exposure. The provider-side strategy is the shared updater FPF_P, applied separately to each website's provider-local history. The same section defines the empirical approximate rewriting-equilibrium criterion by stability of deployed profile terms under repeated FPF_P updates, and Appendix D.2 (pp. 25–26) gives the tail-window measurements. Under this model, our evidence identifies an empirical approximation to a performative fixed point of the provider-side dynamics. This follows the fixed-point notion of [1]; multiplayer performative-learning work distinguishes this solution concept from Nash equilibrium [2,3].

The deviation in the Q1 instead changes one website's updater while keeping its competitors fixed, which is a website-level Nash test. Our empirical claim concerns provider-side fixed-point stability. We have therefore revise the endpoint in terms of empirical profile stability under the deployed updater (Abstract, §3.3, §5.5, and Appendix D.2). We call this fixed-point criterion an approximate rewriting equilibrium, and §5.4 now limits the best-response comparison to the evaluated strategies.

Robustness across continual updaters. We agree that robustness beyond the original FPF_P is best tested with another continual updater. We therefore reran the complete 60-day PA and DEC trajectories after replacing the shared FPF_P for the full adaptive provider side with a less restricted continual updater. This experiment tests whether DEC's empirical benefit depends on the particular continual updater used by the modeled provider side.

This updater removes the fixed eight slots and lets each content agent freely use a 75-word profile budget. Under PA and DEC, 18.1% and 4.7% of profiles contain more than eight terms, so the added rewriting space changes realized profiles. Over the final 20 active days, 91.5% and 91.0% of profiles keep the same term set, reproducing the long-horizon stability in the original paper.

SystemW-nDCG@5D-nDCG@5Provider GEOUnbacked claims
PA49.6838.986.8930.8%
DEC51.2140.457.0915.7%
DEC − PA+1.53 [1.02, 2.03]+1.46 [0.89, 2.04]+0.20−49%
  • Under this less restricted updater, DEC recovers most of PA's user-side quality loss, reduces unbacked claims, and keeps Provider GEO above Original.
  • The result therefore establishes updater robustness at the modeled provider-side level: DEC's benefit is not tied to the original prescribed continual updater.

Added in the revised PDF: The Abstract (p. 1), §3.3 (pp. 3–4), §5.5 (p. 9), and Appendix D.2 (pp. 25–26) clarify the endpoint terminology; §5.4 limits the best-response comparison to the evaluated strategies and Table 4 summarizes the updater-robustness result (pp. 8–9), with the full experiment in Appendix F.2 (p. 31).

W2 & Q2. Two independent full replays reproduce DEC's gains.

Prior work shows that hosted LLM services may change over time [4] and that LLM-agent systems can vary across runs [5]. We therefore ran paired PA and DEC experiments in two different periods, keeping the request stream and all other settings fixed, to test the results under different runtime conditions. The two replays produce different early histories: in the first ten active days, 36.6% of requests select different adaptive websites.

ReplaySystemW-nDCG@5D-nDCG@5Provider GEO
1PA51.5140.757.39
DEC52.8942.067.01
2PA50.7640.567.89
DEC52.7641.826.87
  • User-side result: In the two replays, DEC improves W-nDCG@5 over PA by 1.38 and 2.00 points, and D-nDCG@5 by 1.31 and 1.26 points.
  • Provider-side result: DEC's Provider GEO ranges from 6.87 to 7.01, remaining above the corresponding Original and retaining 63.3%–83.3% of PA's exposure gain.

The experiments show that the variation across periods does not change the main result. The two full runs produce different histories and absolute results, but both reproduce DEC's user-side quality gains and preservation of most provider-exposure gains.

Added in the revised PDF: §5.4 and Table 4 summarize both complete trajectories (pp. 8–9); Appendix F.1 reports their cross-run allocation variation and full outcomes in Tables 14–15 (pp. 30–31).

We hope the clarified equilibrium definition, second-updater experiment, and independent full replays address your questions about long-horizon stability and reproducibility. If any specific concern remains, we would be glad to address it during the discussion.

References

[1] J. Perdomo, T. Zrnic, C. Mendler-Dünner, and M. Hardt. "Performative Prediction." ICML 2020.

[2] A. Narang, E. Faulkner, D. Drusvyatskiy, M. Fazel, and J. L. Ratliff. "Multiplayer Performative Prediction: Learning in Decision-Dependent Games." JMLR 2023.

[3] X. Wang, C.-Y. Yau, and H. T. Wai. "Network Effects in Performative Prediction Games." ICML 2023.

[4] L. Chen, M. Zaharia, and J. Zou. "How Is ChatGPT's Behavior Changing over Time?" 

[5] H. Zhu et al. "OAgents: An Empirical Study of Building Effective Agents." Findings of EMNLP 2025.

Thanks for the clarification and additional experiments. The distinction between profile stability and Nash stability addresses my concerns, and the 60-day replays in Appendix F.1 confirm that the quality gains hold across different histories. My concerns are resolved, and I’m sticking with my weak-acceptance recommendation.

Thank you for your thoughtful feedback and encouraging assessment. We are glad that the clarified equilibrium definition and additional full-trajectory experiments have addressed your concerns. We appreciate your recognition of these revisions and your continued support!

Reviewer_p1tzOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies repeated competition among content-provider agents that rewrite their profiles to increase exposure in an Agentic Web setting. It introduces WildRegistry, a simulation framework and benchmark for repeated competition in the agentic Web, combining real-search logs with a set of 100 websites. The authors find that profile adaptation increases provider exposure while reducing user-side quality and eventually stabilizing in a harmful state under the deployed rewriting policy.

To address this behavior, the paper proposes Delivery-Evidence Credit (DEC), a platform-side mechanism that rewards websites with relevant past delivery evidence and withholds opportunities after repeated selection and non-delivery. Validation includes two protocols (30-day and 60-day) for baseline comparison, component ablations, multiple LLM backbones, and an alternative set of adaptive providers. The reported results show improved user-side quality while preserving much of the exposure (provider GEO), yielding a positive welfare gain.

Critical strengths

A timely problem and a distinctive mechanism for an emerging ecosystem - The paper addresses strategic competition in the Agentic Web, where providers can influence access to their content through self-authored profiles before their documents are observed. This is an emerging setting in which the stakeholders, interactions, and downstream outcomes are unclear. The paper introduces a framework, based on real-world data, as a starting point for strategic competition research in the Agentic Web. Its main contribution (INO) is a mechanism that connects past delivery outcomes to both future selection opportunities and the feedback shaping subsequent profile updates. DEC provides a concrete intervention through complementary credit and withholding rules. The reported improvements in user-side quality, while retaining much of providers’ exposure gains, demonstrate the potential for mutually beneficial outcomes.

Comprehensive evaluation - The evaluation combines a stream of real search requests with a substantial web corpus. Although provider adaptation is simulated, this grounding makes the experimental setting more informative than one based entirely on synthetic data. The authors evaluate DEC through multiple-baseline comparisons, an ablation study, multiple LLM backbones, and several adaptive providers. They also study why DEC works, helping explain the observed underlying effects beyond just performance.

Critical weaknesses

Limited alignment between the simulated ecosystem and the real-world Agentic Web setting - In my opinion, three modeling choices restrict the conclusions: (i) providers can revise only their profiles, while documents remain fixed. In a real ecosystem, providers could also improve, expand, or manipulate their content in response to platform incentives. The experiments therefore leave unexplored whether DEC encourages substantive content improvements or remains effective under joint profile-and-content optimization; (ii) the user agent directly outputs a ranked document list, with selection treated as citation, rather than a generated answer with citations. This omits the interaction between retrieval, answer generation, and attribution (citation) assignment. Notably, there is an interplay between the three. Consequently, the user-side metrics measure document relevance, without establishing improvements in final-answer correctness, completeness, or groundedness; and (iii) only some of the hundred providers modify profiles, while the remaining remain static. Testing several profile groups addresses provider identity but does not establish robustness when profile modifications become widespread.

Baseline implementation - Full-history ICL originates from a general continual-learning setting, whereas VCR addresses document-level optimization competitions for generative search engines; neither was originally designed for competition through website profiles. These gaps make it difficult to determine how much of DEC’s advantage reflects the proposed mechanism versus limitations of the baseline implementations in this setting. For example, one may wonder how VCR performs in an Agentic Web setting in which provider-agents modify documents and profiles, and the mechanism rewards document modifications as well, as originally designed. Hence, I find the comparison problematic and requiring further discussion and details at the least.

Questions for authors
  1. May you please provide statistical test results for the reported results in the paper?

  2. What are the results of + Withhold alone (ablation study)?

6 comments

We thank the reviewer for asking how well the conclusions extend beyond the main profile-only setting and whether the comparisons and ablations establish the source of DEC's gains.

In short. The questions concern two issues: whether DEC generalizes beyond the main profile-only setting, and whether the evaluation isolates its benefit. Our matched extensions show that DEC retains its two-sided benefit in broader Agentic Web settings, while the added baselines, uncertainty estimates, and ablations strengthen the evidence for its mechanism.

1. Scope and generalization (W1).

W1.1. Clarification of our study scope: We focus on website-selection competition before retrieval in the Agentic Web, which is important and remains largely unexplored.

Content agents compete for limited selection opportunities through their profiles before returning the documents. This is a different stage from the document optimization studied in existing GEO work. We therefore fix documents in the main experiments to study this important upstream competition and its long-horizon effects. The scope clarification appears in §1 (p. 1).

DEC's two-sided gains also hold under joint profile–document optimization. We agree that joint profile–document optimization is also worth studying. However, our 100 websites contain about 18.43 million documents. Rewriting the full corpus would require many LM calls and index updates [1,2]. Profile adaptation is a less expensive provider action that already brings substantial exposure gains. Existing GEO studies also usually limit each optimization to a small set of candidate documents, such as 5–10 documents per query [3–5].

We therefore use a local rewriting strategy that follows the provider-local history constraint: across rounds, each adaptive content agent can revise its profile and multiple documents previously retrieved from its own corpus and returned in past interactions. Our experiment shows that this broader action space raises PA's Provider GEO by 33.7% over profile-only adaptation.

We implemented PA-Joint, VCR-Joint, and DEC-Joint under the same provider action space and update budget, also addressing W2. VCR-Joint applies VCR to successive profile and document versions. DEC-Joint extends Credit and Withhold to revised documents using past delivery evidence:

SystemW@3W@5W@10D@3D@5D@10Provider GEO
PA-Joint50.7050.9447.5645.6340.6732.246.94
VCR-Joint51.7251.4248.0046.2141.0732.824.66
DEC-Joint52.2451.9748.5147.1241.8433.345.42
  • Best user-side quality: DEC-Joint achieves the best user-side quality. It leads on all six quality measures.
  • Provider exposure remains substantial: DEC-Joint retains 78.1% of PA-Joint's GEO.

These experiments show that DEC can be readily extended to broader provider action spaces, maintaining higher user-side quality with substantial provider exposure.

Added in the revised PDF: the scope clarification appears in §1 (p. 1); the joint-setting discussion and results appear in §5.7 and Table 5 (p. 10), with the joint-action protocol in Appendix C.4 (p. 25).

W1.2. The quality gains also carry over to generated answers.

Our paper focuses on website selection and the quality of the ranked document sources in the final answer. To test whether these gains carry over to generated answers, we follow AutoGEO [3] to generate natural-language answers with citations. We then use rubric-based LLM evaluation [7], with GPT-4o-mini as the judge, to assess correctness, completeness, and citation groundedness. Citation groundedness evaluates whether answer claims are supported by their assigned citations, and citation-based Provider GEO is computed from those citations.

System / StatisticCorrectnessCompletenessCitation groundednessProvider GEO
PA4.5544.5484.5296.18
DEC4.5714.5664.5485.83
Answer change per +0.1 ΔD-nDCG@5+0.119+0.120+0.122—
  • Generated-answer quality improves under DEC: DEC is higher on all three answer-quality measures. The 95% paired bootstrap CIs are [0.0095, 0.0255], [0.0114, 0.0232], [0.0107, 0.0269].
  • The source quality gains also carry over to generated answers: The paired analysis shows that when DEC's D-nDCG@5 improvement over PA increases by 0.1 on the same requests, each answer-quality score rises by about 0.12 points on average. The 95% paired bootstrap CIs are [0.0976, 0.1403], [0.0994, 0.1417], [0.1012, 0.1434]
  • Provider exposure remains: DEC also retains 94.4% of PA's citation-based Provider GEO.

This directly evaluates the interaction in the question: the source-quality improvement carries through answer generation and citation assignment, while citation-based provider exposure remains substantial.

Added in the revised PDF: §5.2 points to this validation (p. 7), and Appendix A reports the protocol and statistical results under "Generated-answer evaluation," including Table 6 (p. 18).

W1.3. DEC's gains hold with 30% adaptive websites.

We further tested broader profile adaptation by increasing the number of adaptive websites from 10 to 30 and ran Original, PA, and DEC to completion.

Adaptive settingΔW@5 (DEC−PA)ΔD@5 (DEC−PA)ΔGEO (DEC−Original)Welfare Gain
Main 10 sites+2.26+1.76+1.82+3.51
Alternative 10 sites+2.03+2.04+0.72+1.90
30 sites+3.34+0.47+2.08+1.91
  • Quality gains remain: With 30 adaptive websites, DEC still improves website-selection and final-document quality.
  • Exposure remains positive: Adaptive providers retain a positive exposure gain over Original.

This confirms that the main results also hold in larger adaptive populations, and DEC's advantage is robust in broader adaptation.

Added in the revised PDF: §5.4 summarizes the result (p. 8), and Appendix F.6 reports the broader 30-site setting, including Table 19 (pp. 32–33).

2. Evaluation and mechanism evidence (W2, Q1–Q2).

W2. DEC still leads with three profile-level baselines.

We add three methods that better match our setting. Global Beta Reputation uses each website's full delivery history. Query-conditioned Beta Reputation uses only history related to the current request. Both follow the Beta Reputation System [8]. Relevant-History ICL directly provides these related service records to the Selector.

SystemW-nDCG@5D-nDCG@5Provider GEOWelfare Gain
PA51.0141.975.190.00
Global Beta Reputation41.9539.222.48−13.31
Query-conditioned Beta Reputation49.1741.017.14+1.20
Relevant-History ICL52.2342.054.25−0.25
DEC53.2743.745.99+3.51
  • DEC leads at the same interface: DEC achieves the highest user-side quality and Welfare Gain.
  • Scalar reputation loses information: In Appendix E.1.2 (p. 27), reducing related outcomes to one reputation score lowers W-nDCG@5 by 2.24 points on the PA trajectory and by 3.68 points on the DEC trajectory.
  • History alone is insufficient: Related history is therefore useful, but neither raw records nor a scalar reputation reproduces DEC's combination of evidence construction and opportunity actions.

The original VCR adaptation also works as intended. VCR [6] originally compares successive document versions. We apply the same comparison to successive profile versions while keeping its judgment and ranking rules. Its warnings cover 81.3% of Selector calls and change candidate ordering in 38.2% of calls. This shows that VCR actively affects website selection, although its user-side quality gain is limited. W1.1 further restores VCR's document-level mechanism. Under joint optimization, DEC still outperforms VCR on user-side quality and Provider GEO. We clarify the baseline implementations in Appendix C.3, and report their results in Tables 1 and 8 (pp. 6 and 24).

Q1. The paper already reports paired 95% confidence intervals.

The 60-day main comparisons are in Appendix D.1, and the 30-day baseline and backbone comparisons are in Appendix F.5. The central comparisons are:

Protocol and comparisonMetricPaired difference95% CI
60-day DEC − PAW-nDCG@5+2.14[1.44, 2.87]
60-day DEC − PAD-nDCG@5+0.70[0.10, 1.33]
60-day DEC − OriginalProvider GEO+2.23[1.53, 3.00]
60-day DEC (PA reference)Welfare Gain+2.01[1.32, 2.74]

Added in the revised PDF: §5.1 now points readers to these existing intervals (p. 6) and to the resampling procedure at the beginning of Appendix F (p. 30).

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

The paper studies competition among content agents that repeatedly rewrite their website profiles to gain exposure. It proposes DEC, which uses past delivery outcomes to improve website selection while retaining most of the exposure gains from profile rewriting.

Critical strengths
  • The paper addresses an important problem: websites compete for selection through their profiles before document retrieval. It examines how providers’ efforts to gain exposure can hurt website-selection quality.

  • The experiments cover multiple model backbones, a second group of adaptive websites, and component ablations. These provide useful support for DEC’s effectiveness in the tested settings.

Critical weaknesses
  1. DEC uses past successful deliveries to guide later selection. Early chance differences could therefore create lasting exposure gaps between similarly capable websites. The market-level concentration analysis does not directly test this possibility.

Could DEC amplify chance differences in early selection opportunities? Among similarly capable websites, could a few early successful deliveries give one site a lasting exposure advantage? Would the long-run results change if early opportunities were varied?

  1. The updater can add at most eight candidate terms. It is unclear how much this constraint contributes to late-stage profile stability and DEC’s effect on term choices.

Does DEC still improve profile composition with a different limit or a less restricted updater?

  1. Withhold only applies when the current profile differs from the original. Consider a website that rewrites its profile, then reverts after repeated failures to obtain citations. Using only its own feedback, could this strategy outperform FPF_P in cumulative exposure over the full feedback process?
Questions for authors

Please refer to the questions in the Weaknesses

4 comments

We thank the reviewer for examining DEC's long-horizon behavior from three important angles: sensitivity to early histories, robustness to the provider updater, and the return-to-Original strategy.

In short. The added experiments show that DEC's result is robust across different histories and a second updater, while the action analysis shows that reversion cannot reset delivery evidence. We answer each point below and identify the corresponding additions to the revised PDF.

W1 & Q1: DEC is robust to different realized early histories. Early differences do not lead to lasting exposure gaps. This is because DEC does not build a cumulative global reputation: later outcomes continually update the evidence used by both actions.

We test this with two matched, independent 60-day replays, making each website its own capability-matched comparison. Due to runtime variation in the LLM-agent chain, the two replays produced different early histories. In the first ten active days, 36.6% of requests selected different adaptive websites, and 19.5% used different adaptive websites in the final answer. However, the allocation gaps between the two DEC replays became much smaller:

Allocation gap between DEC replaysFirst 10 daysFinal 10 days
Website opportunities25.7%4.3%
Final use20.0%3.9%
Provider GEO21.7%3.0%

We further compare the same website across the two replays to hold website capability fixed. For example, Answers.com:

  • first 10 days: 30 early successful deliveries in Replay 2, higher than 13 in Replay 1
  • final 10 days: 3.3% exposure share in Replay 2, close to 4.5% in Replay 1

This comparison shows that more early successful deliveries did not persist as a larger late-run exposure share.

Long-run results are robust. Both complete replays also reproduce DEC's user-side quality gains:

ReplaySystemW-nDCG@5D-nDCG@5Provider GEO
1PA51.5140.757.39
DEC52.8942.067.01
2PA50.7640.567.89
DEC52.7641.826.87
  • Across the two replays, DEC improves W-nDCG@5 over PA by 1.38/2.00 points and D-nDCG@5 by 1.31/1.26 points, while preserving most of the provider GEO.

Together, the two complete replays show that DEC's quality benefit persists under different early histories while late provider allocations remain closely aligned.

Added in the revised PDF: §5.4 and Table 4 summarize the replay results (pp. 8–9); Appendix F.1 reports the early-versus-late allocation analysis and complete outcomes, including Tables 14–15 (pp. 30–31).

W2 & Q2: The DEC gains hold with a less restricted updater. We ran another 60-day experiment, removing the eight-term limit while keeping the 75-word profile budget. Providers can now use this budget to rewrite more freely.

Under PA and DEC, 18.1% and 4.7% of daily deployed profiles used more than eight terms, showing that providers used the larger rewriting space. Over the final 20 active days, 91.5% and 91.0% of profiles kept the same term set, consistent with the long-horizon stability reported in the paper.

SystemW-nDCG@5D-nDCG@5Provider GEOUnbacked claims
PA49.6838.986.8930.8%
DEC51.2140.457.0915.7%
DEC − PA+1.53 [1.02, 2.03]+1.46 [0.89, 2.04]+0.20−49%
  • Fewer unbacked claims: DEC still reduces unbacked claims by 49% compared with PA, improving profile composition.
  • Quality improves without losing provider gain: It also raises W-nDCG@5 and D-nDCG@5 by 1.53 and 1.46 points, while both systems remain above Original in Provider GEO.

These results confirm that the main findings on profile composition, user-side quality, and long-horizon stability hold with a less restricted updater.

Added in the revised PDF: §5.4 and Table 4 summarize the comparison (pp. 8–9); Appendix F.2 and Table 16 report the complete setting and outcomes (p. 31).

W3 & Q3. Reverting to Original does not increase the exposure. We agree that returning to the Original profile temporarily removes the profile-change condition for Withhold, so we implemented returning to Original as a provider action. At every eligible update, each content agent used only its own service history to choose among returning to Original, keeping its current profile, and continuing to rewrite. Across all observed decisions, no content agent selected returning to Original. The proposed response was therefore available endogenously throughout the repeated interaction.

The state transition explains this choice. Returning to Original suspends the Withhold, but this also removes the added claims that created the extra matching opportunities, so the website cannot keep the additional exposure while suspending Withhold. If the website later adds those claims again, the records can still trigger Withhold; waiting for their expiry also gives up those matching opportunities in the meantime. This supports the experiment result that reverting was not an attractive action under the current provider policy.

Added in the revised PDF: §4.2 clarifies the state transition (p. 5), and Appendix C.1 reports the return-to-Original path and provider action audit (p. 23).

We hope these replay, updater, and action analyses address the reviewer's questions about DEC's robustness over repeated interaction. If any specific concern remains, we would be glad to address it during the discussion.

Thanks for the additional experiments. Since my rating was already positive, I think it is still appropriate, so I will maintain it. Good luck with the paper!

Thank you for your thoughtful feedback and positive assessment. We sincerely appreciate your recognition of our work and your continued support!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.