acceptodds
Under review as a conference paper at ICLR 2027

Cohesion and Coordination for Strategic Collective Action in Content Recommendation

Abstract

Increasingly, algorithmic systems are mediating which content users are recommended, thus shaping their exposure to creative media. As awareness of this curation process increases, users have begun to organize deliberate, strategic responses in an attempt to influence which content is promoted online. Existing theoretical models of this behavior, spanning both adversarial “attack” and, more recently, algorithmic collective action literature often represent participating users as having no prior histories and adopting a single unified strategy, effectively behaving as a squadron of clones. We argue that this abstraction overlooks two properties central to real human collectives: cohesion, the similarity of members before collective action begins, and coordination, the extent to which members align their behavior during strategic action. We introduce a novel model of collective action that parameterizes cohesion and coordination as continuous, independently controllable quantities. We derive conditions characterizing when a collective's strategic behavior shifts a target item's representation and increases its average predicted relevance, recovering the traditional unified setting as a special case while extending the analysis to partially cohesive and partially coordinated collectives. Controlled synthetic experiments validate these findings and show that successful algorithmic collective action does not require users to behave as clones and, more broadly, that the internal structure of a collective can fundamentally shape its ability to influence algorithmic curation.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
recommender systems, collective action, game theory

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_JLZgOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies collective action in recommender systems, but with real existing users instead of identical fake accounts. Authors describe a collective with two properties, cohesion (how similar members are before acting) and coordination (how much they move together toward a shared strategy). With an ALS recommender, they show how such a collective changes the target item, and that the usual clone setting is a special case. They also show that when members keep their diversity, a larger collective is not always better. Experiments on synthetic data and MovieLens are used to claim that collectives do not need to act like clones, and more coordination is not always helpful.

Critical strengths
  • The decomposition in Theorem 1 is simple but useful. It shows that, for one ALS item update, the collective enters only through its post-action mean and residual covariance, so the classical clone result and the heterogeneous case are put in one formula. The counterexample in Appendix D.2 is also built carefully, with A and b actually realizable, which makes the point that collective size alone does not decide success concrete rather than only intuitive.
Critical weaknesses
  • The paper does not show that the experiments really reflect the parameters of the model. Users can only change their ratings, not their representations, but the paper does not explain how cohesion and coordination are turned into rating changes, and the actually realized values are not reported. So, when the results say more coordination hurts or heterogeneity is not a penalty, we cannot know whether this comes from coordination or cohesion itself, or just from how the ratings were modified. In addition, one main claim is true by construction. The authors claim that relaxing the clone assumption closes the gap between the optimized strategy and simple heuristics, but this result is shown at zero coordination, where in the paper's own theory the chosen strategy has no effect at all. So, all strategies must look the same there, and this does not tell anything about the real advantage of optimization.

  • The measured influence is mostly the collective influencing itself. The downstream metrics are averaged over all users including members, and in the synthetic setting the Top-10 change reaches almost 0.9 when the collective is around 90% of users. This means most of the effect is members getting recommended the item they just rated. But the point of collective action is to change what other users see. Without results on non-members, and a simple control that adds the same amount of ratings without the target, it is hard to say how much real influence the collective has.

Questions for authors
  • Users can only change their ratings, so how are the requested cohesion and coordination levels actually turned into rating changes? Could you report the realized levels measured after this step, for the main settings in Figure 3? Also, when every member simply adds the same ratings to their own history, moving more toward the strategy and keeping less of the original diversity seem to happen together. How are settings with full coordination and full diversity at the same time realized?

  • Could you report the downstream results only on non-members, and compare with a simple control where members add the same amount of ratings without the target item? This would show how much of the effect is real influence on other users, and how much comes from the members themselves or from retraining noise.

  • In the synthetic setup, is the cohesion level set before the ratings are generated, or only on the learned representations? Related to this, at the smallest collective size in Figure 3(b) the collective has only one user, so a non-zero diversity level does not seem possible. How is this point produced?

  • The theory assumes members did not rate the target item before acting. How many users rated the target item in the synthetic data, and how is this assumption kept when the collective becomes very large? Also, which target item is used for MovieLens?

  • In Theorem 1, the effect of the collective depends on how far the shared strategy currently under-predicts the target item. Since each heuristic strategy is built from a profile that already gives the target a high rating, this gap may be small by design.

3 comments
AuthorsOpenReview ↗

Response to Reviewer JLZg

Thank you for your detailed review and helpful comments. We respond below. (W = weakness, Q = question.)

W1: Please see response to Reviewer sAWT. We show in our experiments that our “realized” gamma and beta are relatively close to the hyperparameters theorized by the model. Re: trivial claim, thank you for your observation; we will remove that claim.

W2: We will run experiments and provide an update.

Q1: We distinguish between tr(ΣC)tr(\Sigma_C) , the initial heterogeneity, and γ\gamma the retention of this heterogeneity during strategic action. Recall from Eq. 2 that the post-strategic representation of a user ii can be represented as ui′=(1−β)ui+βuˉC+γϵiu_i' = (1-\beta)u_i + \beta\bar{u}_C + \gamma \epsilon_i. In the case where the collective has complete coordination, this would mean that all the collective members would move towards uhu_h but the persistence in diversity would indicate that their individual deviations from this centroid would be scaled by some factor γ\gamma. Intuitively, we introduce this notion our framework to consider the situation in which users are potentially not able to access all of the exact same strategic items but a more formal characterization of this will be left for future work.

Q2: We will run experiments and provide an update.

Q3: The cohesion level is fully set before the ratings are generated. The full flow is as follows:

seed
│
├── generate items
├── generate users
├── generate popularity
└── choose coalition ordering
       │
       ├── alpha -- proportion of population 
       │    └── select nested coalition
       │
       └── tr(\Sigma_C) -- cohesion (normalized trace of covariance) 
            ├── modify coalition dispersion
            ├── generate historical behavior
            ├── fit baseline ALS -- V_pre, U_pre
            │
            └── \beta -- coordination and \gamma - residual heterogeneity
			           ├── select strategic items and their associated strategic ratings
                 ├── construct strategic vector u_h
                 ├── apply u_i' = (1-\beta)u_i + (\beta)u_h + \epsilon_i
                 ├── translate it into ratings R'
                 ├── retrain ALS from baseline -- V_post, U_post
                 └── evaluate

Q4: In the synthetic data we simulate an environment where the collective members have not rated the target item. However, for each seed we randomly generate the number of non-collective users who have rated each item. For MovieLens, we selected an item with “medium-low” popularity for each seed. We calculated this by applying a log_5 binning procedure to the rate interaction counts associated with each item following [1]. We will happily document and ablate on this protocol for the final camera ready version of our paper.

[1] Salganik, R., Diaz, F., Farnadi, G. (2024). Fairness Through Domain Awareness: Mitigating Popularity Bias for Music Discovery. In: Goharian, N., et al. Advances in Information Retrieval. ECIR 2024. Lecture Notes in Computer Science, vol 14611. Springer, Cham.

Q5: Yes, that is correct. In the future we will augment our experiments with a more diverse set of synthetic settings.

Thank you for the response. Some answers did not fit the paper, so I read the anonymous code linked in the Reproducibility Statement. It disagrees with several rebuttal statements and shows flaws in the experimental design. I list only points that can be checked directly.

Rebuttal vs. released code

  1. Rating protocol. The rebuttal says existing ratings stay unchanged and members have not rated the target. In the code, strategic ratings overwrite a member's existing ratings on those items, and a member's earlier target rating is replaced by 5.
  2. Realized β\beta, γ\gamma. The rebuttal says the experiments show realized values close to the requested ones. The paper reports none. The code already records them in every run, yet your reply to sAWT says these experiments are still to be run.

Experimental design

  1. Item-Shift Success. The code stores only the Theorem 1 formula evaluated at the requested β\beta and γ\gamma. The shift measured after refitting is computed but not stored, although Sec. 5.1 describes it as evaluated after refitting. If the Item-Shift panels plot the stored value, they cannot test Theorem 1.
  2. Definition of uhu_h. uhu_h is fit to ratings that include the target at 5, so the gap (r−uh⊤vj⋆)(r - u_h^\top v_{j^\star}) is small by design, as you confirmed in Q5. Theorem 1 then gives near-zero success at β=1\beta = 1, γ=0\gamma = 0, so the decline with β\beta in Fig. 3(d) follows from the definition of uhu_h.
  3. Top-10 Change. It ranks items a user has already rated and averages over all users, so members who just rated the target 5 are included in the exposure count.
  4. Fig. 3(a) vs. 3(d). Panel (a) at α=0.1\alpha = 0.1 and panel (d) at β=1\beta = 1 are the same setting, yet Item-Shift Success is roughly 0.4–0.5 in (a) and below 0.1 in (d).

I have changed the visibility of all my comments in this thread to Everyone. Several of the authors' replies are currently visible only to the committee and reviewers. For a fair and transparent evaluation by a wider audience, I encourage the authors to set their replies to Everyone as well, so that the full discussion can be read in context.

Reviewer_sAWTOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies algorithmic collective action where a group of users tries to promote a target item on a recommender system. The work is motivated by the claim that existing models treat members as identical agents without histories, and it adds two properties: cohesion (how similar members are before acting) and coordination. The model is set only in the latent space of ALS, where member ii moves to ui′=(1−β)uˉC+βuh+γϵiu_i' = (1-\beta)\bar{u}_C + \beta u_h + \gamma \epsilon_i. Here ϵi\epsilon_i is the member’s deviation from the collective mean, β\beta moves the mean toward a point uhu_h given by the strategy, and γ\gamma sets how much of each deviation is kept. For γ=0\gamma = 0, the authors prove monotonicity of success and diminishing returns. In the synthetic validation, the collective size α\alpha, the initial spread tr⁡(ΣC)\operatorname{tr}(\Sigma_C), β\beta, γ\gamma and the strategy are varied independently. Four heuristics (Random Filler, Mainstream Mimic, Target Neighborhood, Target-Supporter Mimic) are compared with an optimized direction under a matched movement budget.

Critical strengths

The paper is well motivated. Members of a real collective already have rating histories and will not all move by the same amount, and models of collective action in recommenders say little about this. Definition 1 separates the shift of the collective mean (β\beta) from the initial spread (ΣC\Sigma_C) and the part of it that survives (γ\gamma). It also recovers the identical-members setting as a special case. Experiments are abundant. Overall, if the model were tied to rating-level actions (see W1), I would expect follow-up work to reuse this parameterization.

Critical weaknesses

W1. The theory is stated in the latent space of ALS, while in reality a collective can only change ratings, and the step between the two is not specified. As far as I can reconstruct it, the pipeline is: (i) a heuristic produces modified ratings R′′R'', from which uhu_h is derived (how is not explained); (ii) Definition 1, with β\beta and γ\gamma given, yields target representations ui′(theory)u_i^{\prime(\mathrm{theory})}; (iii) these targets are reverse-engineered into the ratings R′R' that are used to refit ALS. Appendix E.3 acknowledges the gap: it states that ui′(theory)u_i^{\prime(\mathrm{theory})} is turned into R′R' and that the realized state ui′(real)u_i^{\prime(\mathrm{real})} is measured with the frozen encoder before refitting. How step (iii) works is never described. Appendix E.4 defines each strategy as the target rating plus B=10B = 10 filler items chosen by a rule. None of these rules seem to depend on β\beta, γ\gamma or ϵi\epsilon_i, and the ratings the fillers receive are not stated. The point uhu_h is only described as the direction the fillers are “intended to induce”. I could not find ui′(real)u_i^{\prime(\mathrm{real})} reported anywhere, so the results seem to be indexed by the requested β\beta and γ\gamma, and it is unclear how close the collective gets to them. If my reconstruction is right, the ratings the strategy itself produces (R′′R'') never reach the model, and what is evaluated is R′R', which depends on β\beta and γ\gamma. The optimized strategy, which should be read as an oracle, is also defined in latent space with a latent budget, so comparing it with the heuristics inevitably mixes the quality of a direction with whether ratings can reach it.

Overall, the paper is motivated by users who organize to influence curation, but such a group gets no condition in terms of quantities it controls and no guarantee for any strategy. Users usually cannot see the latent space. And even if they could, they would still have no way to quantify β\beta and γ\gamma. Other work in the literature does give collectives something to compute. Statistical Collusion by Collectives on Learning Platforms (ICML 2025, arXiv:2502.04879) lets a collective infer good strategies and the parameters that drive its success. The bounds in A Statistical Framework for Algorithmic Collective Action with Multiple Collectives (arXiv:2605.06749, which the authors don’t cite) can be computed by each collective with only partial knowledge of the others. What remains for users here is an empirical benchmarking of four heuristics in a fixed setting.

W2. Everything is specific to explicit-feedback ALS. The analysis is built on the closed-form ALS update of the target item, and the model has no user or item features: no xix_i enters the treatment or the experiments. As a stylized model this is fine, and ALS is still a competitive baseline. But the abstract speaks about algorithmic curation in general, and most creative-media platforms learn from implicit feedback. The closest prior work on collective action in recommenders (Baumann and Mendler-Dünner, 2024) already used a transformer-based model from a streaming platform. It is unclear which conclusions carry over, so the reach of the contribution is much narrower than the framing suggests. I would suggest making this clear in the abstract and the contributions, ideally changing the title to “COHESION AND COORDINATION FOR STRATEGIC COLLECTIVE ACTION IN ALS-BASED CONTENT RECOMMENDATION”.

W3. The heterogeneous case that motivates the paper is not covered by the theory. Theorem 2 holds only without heterogeneity, and the counterexamples show that its condition can fail otherwise. Yet the abstract claims an extension to partially cohesive collectives, and Figure 3(c) shows very low success rates with heterogeneity, especially at β=1\beta = 1. I could not find the result that supports the claim that successful collective action does not require clones. If success needs a small γ\gamma, members have to collapse toward each other, which is close to acting as clones.

The gap with prior work is also smaller than claimed. The description of prior work as a “squadron of clones” fits the shilling attacks literature, where injected profiles come from one template.

Smaller points. In Definition 1, uhu_h is called a strategic direction but enters as a target point (the direction is uh−uˉCu_h - \bar{u}_C). The definition also says that ϵiϵi⊤\epsilon_i \epsilon_i^\top controls the initial cohesion, where ΣC\Sigma_C or its trace is meant. It describes γ\gamma as the retention of “cohesion”, while the protocol calls it “heterogeneity retention”. Coordination is defined as how much members align their behavior, but with γ=1\gamma = 1 every member moves by the same vector β(uh−uˉC)\beta(u_h - \bar{u}_C) for any β\beta. So β\beta controls how far the collective moves, and alignment between members is set by γ\gamma. Since the strategies are the only place where the paper says how users act, they deserve their own section with the constructions from Appendix E.4 and the missing details listed in W1. Finally, defining RijR_{ij} and ΩU,V\Omega_{U,V} in the main body would help; Ω\Omega takes another meaning in Section 4.3; the dependence on β\beta in Eq. (5) should be explicit; and the second heuristic is called Mainstream Mimic in the main text and Popular-Item Mimic in Appendix E.4. Should [m][m] in the main body be something similar to the budget BB (e.g., the “true” budget)?

Questions for authors

Q1. How is uhu_h computed for each strategy? Appendix E.4 only says the fillers are “intended to induce” it, while ui′(theory)u_i^{\prime(\mathrm{theory})} needs a concrete vector. Is it obtained from the ratings R′′R'' that the heuristic produces, for example through the frozen encoder?

Q2. How exactly is R′R' built from ui′(theory)u_i^{\prime(\mathrm{theory})}? Appendix E.4 fixes the target rating, the budget B=10B = 10 and the rule for choosing fillers, but it does not say what ratings the fillers receive or how β\beta, γ\gamma and ϵi\epsilon_i affect these choices (including which items count as eligible). How does the model perform when refit directly on, say, R′′R'', i.e., the only version users can actually produce?

Q3. The zero-mean assumption in Proposition 1 looks strong to me. If it concerns the learned factors, ALS fit on uncentered ratings usually stores the overall rating level in a shared component, which keeps the mean away from zero. When should the assumption hold in real-world settings?

Q4. Why is success in Definition 2 measured with an average? For a paper about heterogeneity, an average can hide whether the target reaches many users or a few. Would a rank-based measure or other statistics be more suitable?

Q5. I could not follow the interpretation of Figure 3(d) in the text, but I suspect the behaviour is related to W1.

Suggestions.

S1. Please report the realized state that E.3 says you compute. For each strategy and each requested (β,γ)(\beta, \gamma), give the realized values, for example β^\hat{\beta} from projecting the realized mean shift on uh−uˉCu_h - \bar{u}_C (and how much of the shift is orthogonal to it) and γ^=tr⁡ΣC′(real)/tr⁡ΣC\hat{\gamma} = \sqrt{\operatorname{tr}\Sigma_C^{\prime(\mathrm{real})} / \operatorname{tr}\Sigma_C}, and plot the success rate against them. What realized β\beta does a requested β=1\beta = 1 give with B=10B = 10, for members with long histories? If the realized states follow Definition 1 closely, a good part of W1 would be resolved.

Overall, I don't feel like the paper is ready for publication at an A* conference. The idea is well motivated, but the simplicity of the theoretical analysis and the inapplicability of the method severely undermine the impact of this work. I will be happy to engage in a discussion with the authors about how to improve the manuscript. ACA is a very important topic, especially nowadays.

6 comments
AuthorsOpenReview ↗

Response to Reviewer sAWT

Thank you for your detailed review.

W1a/Q1/Q2: We will clarify the protocol in the camera-ready version:

For heuristic strategic interactions, our pipeline is as follows:

Selecting strategic items and ratings. We operationalize each collective strategy hh through a set of BB strategic item–rating pairs, IC,RCI_C, R_C where BB denotes the interaction budget. The selection of these pairs depends on the strategy. For example, under mainstream mimic, we select the (B) most popular items as well as the target item and assign each the maximum possible rating.

Constructing the strategic direction. We translate these actions into a latent strategy vector (uhu_h) by solving the regularized least-squares problem, holding baseline item factors VV fixed: uh=arg⁡min⁡u∑(j,rj)∈Ah(rj−u⊤vj)2+λ∣u∣22.u_h = \arg\min_u\sum_{(j,r_j)\in\mathcal{A}_h}(r_j-u^\top v_j)^2+\lambda|u|_2^2.

Intuitively uhu_h represents the latent profile of a hypothetical user whose preferences are determined exclusively by the selected strategic interactions.

Coordinating collective members. Given uhu_h, we construct each collective member's desired post-intervention representation according to Eq. (2), where β\beta controls movement toward the strategic direction and γ\gamma controls the retention of within-collective heterogeneity.

Realizing strategic interactions. Finally, we translate these desired representations into observable ratings. For each collective member (i) and strategic item (j), we compute the predicted rating (ui′(theory))⊤vj,(u_i'^{\mathrm{(theory)}})^\top v_j, clipped to the permissible rating range. We consider only additive interventions: existing historical ratings remain unchanged, and collective members can only introduce new interactions. The resulting augmented interaction matrix R′R' is used to retrain ALS.

For optimal strategic interactions, we have a slightly different pipeline: rather than explicitly deriving a set of item/rating pairs we use a linear solver to directly compute some optimal direction vector uh∗u_{h^*} Then, the next two steps detailed above (i.e. coordinating coalition members and realizing strategic interactions) hold for the optimal case.

W1b: The suggested methodologies would be interesting to evaluate. Our goal is to understand how a collective’s characteristics affect success, rather than identify ideal strategies. We will cite the suggested work and leave further experiments to future work.

W2: We thank the reviewer for this feedback. We will make this clear in the abstract and contributions, and change the title accordingly.

W3: Our main theoretical contribution characterizes success in terms of cohesion and coordination and extends previous work, depicting a generalization of the setting beyond the unified-clone case. In Theorem 2, we find that the properties of monotonicity and diminishing returns (properties that always hold in the unified-clone case) do not hold unless there is no heterogeneity, but we do characterize general success as a function of heterogeneity, and examine its role in success. Further, in Figure 3(c), we find that varying initial heterogeneity when the heterogeneity retention is 1 yields low but relatively consistent values of success — this points to the tradeoff between initial heterogeneity and heterogeneity retention (i.e., full heterogeneity retention results in low success rates across different initial dispersion values). Finally, we emphasize that we characterize a generalization of the unified clones setting and show across Figure 3(a-b), that we can reconstruct the patterns of success in the unified clone setting with respect to size α\alpha even with limited coordination and full heterogeneity.

Gap with prior work: We will remove this overstatement from the abstract. In the paper, we present a generalized formulation of collective action that can account for both unified and heterogeneous behaviors and prior states, which previous literature has only partially addressed.

Smaller points: Thank you; we will fix.

Q3: We appreciate the opportunity to clarify this point. We aren’t arguing that all the user representations UU are centered around the 0 mean. But rather we are arguing that given some collective mean uˉC\bar{u}_C for a collective CC we can rewrite all the participants i∈Ci \in C as ui=uˉC+ϵiu_i = \bar{u}_C + \epsilon_i. This is by construction of the collective mean uˉC=1∣C∣∑i∈Cui\bar{u}_C = \frac{1}{|C|} \sum_{i \in C} u_i so this shouldn’t be an unreasonable assumption.

Q4: We agree that, in principle, averages can hide imbalanced allocations. However, we do try to mitigate this by normalizing over the users. We hope to extend to more complex metrics in the future.

Q5: Figure 3(d) shows that full coordination (β=1\beta = 1) is not optimal. This suggests that over-committing to a strategy can be harmful, but we have not fully characterized the role of β\beta in success. We leave this to future work.

S1: We will run these experiments and update you with results.

Thank you for the reply. Overall, I don't think my main concerns have been adequately addressed, and I remain of the idea that this paper is interesting, but still not ready for publication at an A* conference.

Considering also the number of elements deferred to future work, I think this is currently more of a workshop paper, and I am happy to iterate more to help the authors improve the manuscript.

W1a/Q1/Q2/S1. Thank you for clarifying the protocol and for committing to make it clearer in the manuscript and integrate the results about the realized state.

W1b. I kindly disagree with the authors. (Algorithmic, Collective) Strategic Action is about action. Although the theoretical exercise of understanding how a collective’s characteristics affect success is necessary, it should be justified by immediate and subsequent methodological work to translate the theory into practical (actionable, inferential) insights. This is almost always the case in the ACA literature (especially the recent one), not only in the cited works. Even assuming that this is the first theoretical step towards more actionable progress, such theoretical analysis is anyway restricted to a single specific (somehow obsolete) model, and I really struggle to see how strategies (regardless of their optimality) can ever be implemented: again, users cannot see the latent space, and even if they could, they would still have no way to quantify β\beta and γ\gamma.

W2. Thank you.

W3. I appreciate the effort, but I remain not convinced by the evidence. The theory in practice provides a negative result, that is, again, appreciated but somehow does not support the claims of the paper. Moreover, 3(c) shows that the success is almost one order of magnitude lower than in settings with limited heterogeneity.

Q3. I agree, sorry for the oversight, and thank you for the clarification.

I want to reiterate again, closing this reply, that I think this type of work is timely, and I encourage the authors to continue working on it. I just believe it does not yet meet the bar of a conference like this, and it could be a much more impactful work when properly strengthened.

First, we want to once again thank the reviewer for their incredibly thoughtful engagement with our work. We sincerely appreciate their effort and interest.

W1b.

First, we wish to highlight the value of our mathematical framework and then gently push back on the reviewer’s comparison with the work presented in Statistical Collusion by Collectives on Learning Platforms.

A prominent paradigm in the study of algorithmic collective action, originating from the canonical formulation of Hardt et al's Algorithmic Collective Action in Machine Learning, examines how coordinated changes to the distribution of training data can influence learned systems. We adopt a complementary perspective, formalizing algorithmic collective action through the geometry of learned representations. This perspective offers several advantages. First, by explicitly characterizing how collective participation changes user and item representations, our framework provides a mechanistic account of collective influence that can inform the analysis of multiple downstream outcomes, including predicted relevance, ranking, and exposure. Second, rather than requiring assumptions about the full distribution of participating users, our theoretical formulation characterizes their influence through the first- and second-order moments of their learned representations, which are sufficient to determine the immediate target-item update under matrix factorization.

Like other existing theoretical formulations of algorithmic collective action, our framework introduces idealizations that may not be fully realizable in practice. We acknowledge that our coordination and residual heterogeneity parameters β\beta and γ\gamma are not necessarily completely controllable by a collective. However, we argue that simply by identifying coordination and residual heterogeneity as distinct determinants of collective success, our mathematical framework provides a foundation for designing future strategies that explicitly seek to manipulate these quantities. For example, a collective could develop strategies that increase coordination β\beta or preserve some degree of individual preference heterogeneity γ\gamma, subject to the constraints imposed by its members' initial representations and feasible actions. We see our work as an important stepping stone towards achieving this kind of group-aware success (similar to some of the motivations behind the contributions in Statistical Collusion by Collectives on Learning Platforms). In this sense, our work shares the broader motivation of Gauthier et al. to develop principled, implementable strategies for collective influence, while addressing what we believe is a complementary, yet extremely important, theoretical question: how do the geometric properties of a heterogeneous collective determine its potential influence on a learned recommender system?

Finally, we respectfully disagree that Statistical Collusion by Collectives on Learning Platforms constitutes a directly applicable baseline for our setting. While both works investigate algorithmic collective action, they consider fundamentally different learning environments and different action spaces. Specifically, Gauthier et al. study classification under strategies that transform feature-label pairs, including strategies that explicitly modify both features and labels. In contrast, our work considers collaborative filtering, where users can strategically modify their interactions with items but cannot directly manipulate their own learned representations or the representations of target items. These representations are instead determined by the platform's learning process. Consequently, the strategies proposed by Gauthier et al. cannot be directly implemented within our setting without introducing additional assumptions about the collective's capabilities. We therefore view their work as complementary rather than as a directly comparable experimental baseline.

W1a. Additional results for Success Metrics on Non-Collective Members

We wish to respond to the reviewer’s previous request to see the downstream metrics calculated over the general population (excluding members of the collective). We present those results in a small markdown write up which can be accessed via this link. [https://anonymous.4open.science/r/CollectivesNotClones-090D/rebuttal_plots/NonCollectiveMetrics/explanation.md]

Reviewer_LPXTOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper analized the impact of users'strategic behaviour in order to game the recommender system algorithm to obtain a more desired recommendation. The user dynamics is driven by two key components: cohesion, namely the similarity of members before collective action begins and coordination, namely the extent to which members align their behaviour for collective actions. The authors find that perfect coordination an perfect cohesion are not necessary, and perfect coordination some times detrimental for success Therefore, groups of real users acting collectively but imperfectly can have big impact in steering the outcome of the recommendation.

Critical strengths

I found the paper very informative and isightful when it comes to say that perfect alignment among the users' features vectors is not needed and that alignment is not monotonic with the impact users have in steering the algorithm outcome. I also appreciate the model validation on both controlled synthetic and real-world experiments.

Critical weaknesses

None

Questions for authors

Q1 How should we read Figure 3(b), given that β = 0 removes the strategy from Eq. (2)? Under Definition 1, setting β = 0 gives u′ᵢ = ū_C + γεᵢ. That expression does not depend on u_h at all. Yet Figure 3(b) compares four heuristics and a “movement-matched optimal” strategy in exactly this setting. Section 5.2.1 then uses it to claim that relaxing the clone assumptions “narrows the gap” between optimized and heuristic strategies. What actually differs between the strategies when β = 0? Is it only the filler ratings added to R′, which the theory says should have no effect? What does the optimizer in Appendix E.5 optimize when the objective is independent of u_h?

Q2 I missed the intuition on why greater coordination can sometimes reduce a collective’s success under realizable strategies. What mechanisms explain this counterintuitive result, and under what conditions would increasing coordination improve outcomes instead?

2 comments
AuthorsOpenReview ↗

Response to Reviewer LPXT

Thank you for your review. We are glad that you recognize the paper’s contribution. We address your two questions below.

Q1: Thank you for the observation; you are correct. We will explicitly make this acknowledgment and remove that claim. Setting β=0\beta = 0 means that the only differentiation between a clone and collective setting is the initial and residual heterogeneity.

Q2: Since the strategies that we evaluate in our work are purely heuristic and not directly optimized for the metrics, it’s possible that the reason that complete coordination decreases success is because the collective is “overcommitting” to a slightly suboptimal strategy. For example, as stated by Reviewer JLZg, if the model’s rating for a target item is already high, pushing it too high might move the item too far towards the collective and away from the general population, leading to a drop in immediate success. We hope to continue exploring this phenomenon in future work.

Thank you

Comments on the paper

4 comments

Thank you for the answers and for the updated Fig. 3(a). Having read them, I believe the paper's claims about coordination and heterogeneity are not yet established by the experiments, for three reasons.

  1. The results are indexed by requested, not realized, parameters. You confirm that the requested latent states are realized imperfectly. Every β\beta and γ\gamma axis in Fig. 3 shows requested values, so the findings stated for settings such as β=1\beta = 1, γ=0\gamma = 0 are findings about requests. The trends cannot yet be attributed to coordination or heterogeneity retention as defined; at most their direction can. Please state the realized β\beta and γ\gamma for the Fig. 3 settings at B=10B = 10, in particular the range of realized γ\gamma in Fig. 3(e).

  2. Theorem 1 is not compared with a measured item shift. You confirm that Item-Shift Success is the formula evaluated at requested values. The item shift measured after refitting, which Sec. 5.1 describes, is not reported. The abstract's validation claim therefore rests only on similar trends between the formula and the downstream metrics, and for the heuristics the updated Fig. 3(a) does not show that similarity.

  3. The corrected Fig. 3(a) does not match the text. The updated subfigure differs substantially from the submitted one. For the heuristics, Item-Shift Success is at most about 0.1 up to α=0.3\alpha = 0.3 and below 0.5 at α=0.7\alpha = 0.7, while for α≥0.3\alpha \ge 0.3 Refit grows roughly as 2α2\alpha and Top-10 Change roughly as α\alpha. Growth in proportion to the member share with almost no predicted item shift is what one would expect if most of the measured effect at these sizes came from the members' own scores, which the pending non-member results can settle. The optimum's Item-Shift is at least six times larger, yet its Refit and Top-10 stay within about 0.2 of the heuristics for α\alpha between 0.3 and 0.7. The first two statements of Sec. 5.2.1 therefore do not hold in Fig. 3(a): Item-Shift does not follow the downstream metrics, and for the heuristics the size effect appears in it only between α=0.7\alpha = 0.7 and 0.90.9.

During this discussion, the claim about the gap at β=0\beta = 0 was withdrawn, the statement that realized parameters are close to the requested ones was withdrawn as too broad, the coordination finding was qualified, and Fig. 3(a), the unified clone setting, was replaced by a different version. If the paper's claims are revised to fit these changes and the three points above, I would like to understand what remains. Which finding is attributable to cohesion or coordination as defined, and which experiment tests Theorem 1 against a measured item shift? My concern is that it is unclear how much of the contribution on coordination and heterogeneity is left once the claims are corrected.

Thank you very much for the new experiments. Within a few days you added the non-member metrics, the target-only baseline, and the measured item shift, and you released the CSV (in the repository) for anyone to check. I sincerely appreciate that effort and the open way you have answered every point. Because the data are public, I could check the numbers myself. Let me start with one fact from the CSV.

1. The Item-Shift panels plot a prediction from the requested state. In your target-only baseline, the submitted ratings are identical in all nine (β,γ)(\beta, \gamma) settings, and every measured outcome is identical as well. Its Item-Shift Success still ranges from 0.690.69 at (β,γ)=(0,1)(\beta, \gamma) = (0, 1) to −0.002-0.002 at (β,γ)=(1,0)(\beta, \gamma) = (1, 0), with α=0.1\alpha = 0.1 and tr(ΣC)=0\mathrm{tr}(\Sigma_C) = 0. The 6.3×10−116.3 \times 10^{-11} agreement compares two ways of computing the same formula from the same requested vectors. Your new plot holds the real comparison. For the heuristics in the clone setting at α=0.5\alpha = 0.5, the formula gives about 0.10.1 and the measured shift is about 0.80.8. At α=0.1\alpha = 0.1, a requested β=1\beta = 1 is realized as β^=0.47\hat\beta = 0.47, and the formula is still evaluated at β=1\beta = 1. The experiments do not validate the theory once β>0\beta > 0.

2. β\beta changes the level of the shared filler ratings. With γ=0\gamma = 0, Eq. (2) gives every member the same vector, ui′=(1−β)uˉC+βuhu_i' = (1-\beta) \bar u_C + \beta u_h, and all members submit identical ratings on the same items at every β\beta. Increasing β\beta only moves those ratings toward 5. In the Fig. 3(d) setting, the heuristics minus target-only in non-member Top-10 is +0.03+0.03 at β=0\beta = 0 and −0.16-0.16 at β=1\beta = 1. Fig. 3(d) and the 3(a)/3(b) comparison therefore show that five-star fillers hurt. In 3(b) every member still rates the same items and gives the target 5, which is hard to call uncoordinated.

3. Realized γ\gamma covers a narrow range. In Fig. 3(e), a requested γ\gamma from 00 to 11 is realized as γ^\hat\gamma from 0.570.57 to 0.700.70 in the ratings, and from 0.730.73 to 0.780.78 after refitting. The figure cannot support a conclusion about γ\gamma between 00 and 11.

One correction on my side. Non-members are affected at β=0\beta = 0, with a Top-10 change of 0.330.33 to 0.770.77 in the 3(b) setting. In the clone setting the four-heuristic mean stays below 0.060.06 at every α\alpha. My W2 holds for the clone setting only.

The β\beta trend is real. The validation claim, the reading of β\beta as coordination, and the conclusion on γ\gamma are not supported, and my assessment stays the same.

Thank you again for the care you have put into this discussion. You carried out every analysis I asked for and conceded points openly, and I respect that a great deal. Please tell me if I misread the CSV.

I respond briefly to the three points.

First, Pangram flags all three responses they have posted as 100% AI-generated.

-> I wrote my review and comments in my native language and used an AI tool to translate and condense them into English. ICLR 2027 permits AI assistance in writing reviews provided it is disclosed together with the original self-written text, and I did exactly that in the AI Documentation field (visible to the SAC and PC) before the discussion period began. An AI detector cannot distinguish a translation of self-written text from AI-generated text, so the Pangram score reflects the translation step and says nothing about who did the analysis. The content of my comments is based on your paper, code, and released repo and CSV, and can be checked against them.

Second, we find their responses to be increasingly incoherent and illogical in relation to the flow of the discussion.

-> This reflects the authors' view, and I leave it to the AC.

Third, we can see that our anonymous repository has been accessed over 300 times despite the fact that only three reviewers and ourselves should be able to find it.

-> ICLR 2027 submissions are publicly visible on OpenReview (please log out of OpenReview and open your paper's page to check this yourself), and the Reproducibility Statement of your paper gives the repository URL. Also, your own comments of 16 Nov (to Reviewer sAWT), which are visible to everyone, also link to it. Anyone can therefore reach the repository.

  1. "Therefore, item_shift_success as defined in the code is a quantity without any real-world implications."

-> item_shift_success is Definition 3 applied to the V′ actually measured after refitting. Like Definition 3 and your Item-Shift Success, it is evaluated against the pre-action users.

  1. "We are not sure how the reviewer would know that the submitted ratings are identical as we don't believe that they are tracked in the csv."

-> In target_only, every member submits only the target at 5 in all nine (β, γ) settings. The CSV you uploaded to the public repository confirms it. The realized β and γ computed from the submitted ratings and every post-refit outcome recorded there are identical across the nine settings.

  1. "Again, this value was provided to show that our theory_success (which is the exact formulation of success in our theorem) is aligned with the empirical counterpart. These vectors are not the same."

-> Both values are computed from the requested u′ᵢ before any rating is submitted. Their agreement confirms the algebra, not the theory against measured data.

  1. "We disagree with this claim that our parameter values aren't "fully" realizable is a sign that our theory does not hold."

-> I did not say the theory does not hold. I wrote, "The experiments do not validate the theory."

  1. "It is unclear which plot they're referring to… in our new results, there are no updated figures which are labeled as Figure 3(b)"

-> I mean Figure 3(b) of the submitted paper (β = 0, tr(Σ_C) = γ = 1), which you compared with the updated 3(a) in your 16 Nov response. In that setting, every member rates the same strategic items and gives the target 5.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.