acceptodds
Under review as a conference paper at ICLR 2027

Signed Exposure: Fair Routing of Algorithmic Attention When Attention Can Harm

Abstract

Fairness-of-exposure treats algorithmic attention as a good to be distributed equitably. But when an autonomous agent initiates contact, attention is signed: it delivers value to a willing receiver and imposes burden on an unwilling one. We formalize routing under signed exposure and show that a fair distribution of attention need not be a fair distribution of unwanted attention. Our central result is an incompatibility: within signed-exposure routing, exposure parity (equal contact rates across groups) and burden parity (equal unwanted-contact rates) generically cannot hold at once, and the two are separated by a band that widens as routing grows more selective. A second result shows measurement error is itself a fairness mechanism: group-differential noise in receptivity scores simultaneously inflates a group's exposure and degrades whom it selects, so an apparent exposure-fairness gain is a hidden burden transfer. Calibrating to a public dating-platform survey (𝑛 = 2,499) that, to our knowledge, uniquely measures receive-side receptivity to conversational agents, we find exposure parity costs only 0.2–2.3% of yield yet moves the per-capita burden ratio to 1.7Γ—: the tension is between fairness notions, not between fairness and efficiency. Finally, the burden-parity policy is computable by bisection and learnable online: a plug-in learner recovers it at a 2.3% empirical regret premium. The operative design choice in signed-exposure markets is not efficiency versus fairness but which fairness.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
algorithmic fairness, fairness of exposure, allocation, impossibility results, measurement error, selection, bandits, constrained optimization, recommender systems, two-sided markets, societal considerations, responsible AI

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_r1ceOpenReview β†—
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

The paper is very difficult to follow, but this is what I figured out:

There are users on a platform and they have a certain receptivity to being contacted (by, e.g., an AI agent working for someone else). We can also define a "burden" as the inverse of receptivity. If e∈[0;1]e\in[0;1] is receptivity then b=1βˆ’eb=1-e is the burden. We have a way to estimate the receptivity but there is noise in the estimation. We can divide the users into demographic groups, for example according to gender, as the paper does.

Then we can ask certain fairness-flavored questions:

  • are both groups contacted at the same rate? ("exposure parity")
  • is the burden applied to both groups the same? ("burden parity")

I think the authors want to imply that there is some kind of trade-off between these two things: if the contact is a good thing, we'd want exposure parity, but if the contact is mostly bad or neutral, then we'd want burden parity.

Separately, the authors define the yield. The yield is an average over individual yields. An individual yield is 0 for a person who was not contacted at all, and it is equal to the receptivity for a person who was contacted.

It is assumed that the entity responsible for the contacting is interested in maximizing this yield and it is shown that, for a given budget qq of the number of contacts, the yield is maximized by contacting the people with the highest receptivity in each group.

On a dataset from a dating platform (which contains information about how much each users wants to be contacted), the authors show empirically that enforcing exposure parity does not cost much in terms of lost yield. However, the resulting strategy is then incompatible with burden parity.

Critical strengths

The paper is well-grounded in theorems. Almost any claim is backed by a theorem with a proof or by an experimental result (though only on one dataset).

Critical weaknesses

The main problem of the paper is that it is essentially unreadable for a human reader. The prose is always punchy and sounds good at first (as is typical for LLM-writing; pangram confirms that the paper is almost entirely LLM-written), but it hinders understanding. The text constantly switches up synonyms, I suppose in order to make the text more interesting to read, but when it comes to a scientific text, precise terminology is really important, and this was a constant annoyance while reading.

Take for example the beginning of section 5:

Routing operates on estimates. Let each receiver’s latent receptivity be...

When I read the first sentence, I'm first of all confused. "What is this talking about?" "What routing?" "What estimates?" When you then continue reading, the meaning eventually becomes clear: "routing" refers to the decision on who to contact and "estimates" refers to the receptivity that we need to estimate. In a novel, it's generally good to make your reader think and wonder, but not in a scientific document! Understanding papers is hard enough as it is, don't make it even harder!

The structure of the paper is also confusing and is not suited to being read linearly, the way a human reader usually does. The two last paragraphs of the introduction can't be understood before one has read the rest of the paper. These paragraphs talk about results to algorithms and datasets that haven't been introduced yet. These should go into the discussion section. As another example, on page 3 there is a long paragraph that discusses results from Theorem 1, which is itself on page 4.

Section 4 is (almost) entirely made up of propositions with no explaining text. You can't just drop these propositions without explaining why the reader should care about them or how to interpret their result. I had to make so many notes on that page to understand what was going on. And once you understand the theorems, they're often extremely trivial. For example, why is Proposition 2 interesting? As far as I can tell, it's just a trivial application of the intermediate value theorem. What can we learn from this?

I also don't think the mix of results and theory is helpful for understanding. The experimental results are sprinkled throughout the paper in paragraphs here and there and it's hard to tell what part is general theory and what part is just exploration of the given dataset.

Finally, variables are often suddenly introduced without any explanation. (Another problem I just noticed while trying to explain this problem: there are no equation numbers!) For example, the equation in line 140-141 uses NN as a variable. This is nowhere defined. From context I'm guessing it's the total number of users and if this was a one-off thing, it would be forgivable, but it just adds to the constant obstacles that are put in front of my trying to understand. Line 185-186 use the variable r0r_0, which is nowhere defined. Nor is rβˆ—r^* in line 195.

What's often fundamentally missing from the text is justification or motivation. Lines 140-146 define all these quantities and none of them is properly motivated.

The second main problem is the fact that a lot of claims are made on the basis of only one dataset. This is acknowledged in the text multiple times, but that doesn't erase the problem.

And finally, I'm skeptical about the relevance of this paper to ICLR. It's not clear to me that the scenario the authors present is one that really needs to take into account group fairness. I would think that "individual fairness" (where we only contact people who want to be contacted) is enough.

Questions for authors

These questions should have been answered in section 3 but weren't:

  • How was the definition of the Yield decided?
  • Why is the burden computed "per-capita" instead of per contacted user? It seems you're doing some kind of total utilitarianism here instead of average utilitarianism?
  • Why would we even want exposure parity? What is the problem with maximizing Yield?
2 comments

Thank you for your careful and constructive review. You have understood the paper as intended, and we agree that our presentation made the argument unnecessarily difficult to follow. We revised the paper significantly so readers can follow the argument without reconstructing missing steps. We start with your questions.

1. How is yield defined? Yield YY is the expected number of welcome contacts per contact made. In other words, it measures routing precision. For a platform that pays per contact, maximizing yield is a natural objective. Section 3 now motivates this measure and the other quantities before defining them.

2. Is burden measured per capita or per contacted user? We now distinguish these measures explicitly and report both. Per-contacted burden bgb_g is the probability that a contacted member of group gg finds the contact unwanted. Per-capita burden hg=tgbgh_g=t_gb_g measures the expected unwanted contact received by a random member of that group. The distinction matters because bgb_g describes the experience of those contacted, while hgh_g also accounts for how frequently the group is contacted. We use the latter for burden parity because it captures the group's overall exposure to unwanted contact. Neither is a total utilitarian objective; both are averages over different populations. At q=0.10q=0.10, the ratio bA/bBb_A/b_B is 1.22 under yield maximization and 1.69 under exposure parity. Thus, the per-contacted measure also reveals a shift in burden.

3. Why pursue exposure parity? What is wrong with maximizing yield? Nothing is inherently wrong with maximizing yield, and we do not argue that platforms should pursue exposure parity. Our point is that every routing policy determines who receives unwanted contact, even when fairness is not its objective. In our data, yield maximization produces a female-to-male per-capita burden ratio of 0.66, while exposure parity reverses the imbalance, producing a ratio of 1.69. Under Proposition 2's sign condition, there is a unique burden-parity point rβˆ—r^\ast within group-threshold policies where both groups face equal per-capita burden.

We do not prescribe this policy either. Our contribution is to make the consequences of these choices explicit, allowing operators to assess the trade-offs rather than assume that equal exposure or maximum yield settles the fairness question. Section 1 now explains this motivation.

Would individual fairness be sufficient? We agree that contacting only willing receivers would be ideal if receptivity were perfectly known. However, receptivity must be estimated, and estimates contain errors that may differ across groups. Even a policy targeting the most willing individuals will therefore generate some unwanted contact. Theorem 2 examines how this remaining burden is distributed across groups and how differences in measurement accuracy change that distribution. Our approach complements individual-level safeguards rather than replacing them. This is also why the work is relevant to fairness and societal considerations in machine learning: it studies the distributional consequences of routing using noisy learned scores.

Presentation and organization We agree with your concerns and have revised the manuscript throughout. Variables are now defined at first use. The unnecessary NN has been removed because the model uses unit mass; r0r_0 and rβˆ—r^\ast are introduced in Section 3; and Ο†\varphi, Ξ¦Λ‰\bar\Phi, and Ξ»\lambda are defined with the noise model. All displayed equations are numbered. We also reorganized the argument. The introduction now gives a short preview instead of referring to unexplained results. The discussion of Theorem 1 follows its statement, and each proposition in Section 4 has a brief explanation of its purpose.

We appreciate your observation about Proposition 2. Its proof is indeed an intermediate value argument. Its importance is establishing a unique burden-parity point under the stated conditions, giving the bisection algorithm and online learner a well-defined target. We now explain this directly. Finally, we have collected the empirical results in one Experiments section and standardized terminology.

Single dataset. We agree that one dataset limits generalizability, and we now state this in the Limitations. The dataset provides stated preferences about receiving contact from agents, directly relevant to our question. However, these are model-derived receptivity estimates from one platform, not observed engagement behavior. The non-Gaussian stress tests in Section 7 examine dependence on distributional assumptions but cannot replace replication on other populations.

We hope these changes, particularly the clearer definitions, revised motivation, and reorganized experiments, address your concerns about readability and presentation.

Dear reviewer r1ce, we have addressed all your questions and concerns and have substantially revised the paper. Please let us know if there are any other questions we can answer, so we have time to respond before the end of the discussion on the 18th.

Reviewer_opfgOpenReview β†—
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper examines fairness in settings in which AI agents decide who should receive machine-initiated contact/attention. In such settings, being contacted is not necessarily beneficial, since some users may perceive the contact as unwanted attention, giving rise to the notion of burden. Using both theoretical results and empirical evidence from survey data collected from 2499 dating platform users, the authors show that fairness in exposure can conflict with fairness in burden (i.e., unwanted contact), such that the two often cannot be satisfied simultaneously. They further show that increased noise in estimated receptivity can make a group appear more fairly exposed while, in fact, increasing the burden experienced by that group. It is also empirically shown that enforcing burden parity achieves only slightly lower yield than a yield-maximizing policy, even when the policy must be learned over time from user feedback.

Critical strengths

The theoretical analysis appears solid, and I appreciate that the authors complement the theoretical results with an empirical analysis using real-world data.

Critical weaknesses

My principle problem with this paper is its premise: I am not sure why fairness is a critical or even relevant metric in this problem setting. If targeting is the objective (in the sense of the dating platform, or in the sense of targeted advertising), then why should attention be distributed equitably, and what does equitable even mean in this case? I would argue that if targeting is the purpose (and is done right), then exposure will inevitably be unequal, with hopefully high yield and low burden. If exposure parity is meant to improve yield, then yield should be the problem to directly address, not fairness. And if exposure parity is the wrong thing to pursue then burden parity is a non-problem.

Some other comments:

Since burden disparity is measured as a ratio (h_A/h_B), it can appear large when the absolute burden is low for both groups. This may partly explain why disparity is larger at small routing budgets, where only the most receptive users are targeted. I wonder whether the results would change if disparity were instead measured using h_A-h_B.

Details regarding the underlying dataset are somewhat sparse. Providing more information about the data (e.g., how is wanted/unwanted attention quantified) would make the paper more self-contained.

Prob 3(ii): It may be worth mentioning that the transformation image.png restores the noiseless marginal distribution, and therefore the noiseless group exposure rates under the Gaussian model. That would further clarify that exposure distortion is correctable while within-group targeting errors from noise are not.

Prop 4: Was there a specific reason that t^βˆ’1/2 was chosen?

Questions for authors

Can you please articulate why one should care about fairness in the stated problem setting? If targeting is the goal, then I understand one can pose the question of whether fairness considerations might harm or help the goal, but fairness itself should not be the goal. If fairness is the goal, then we have a very different problem to address. This paper seems to take it for granted that it is desirable to have exposure parity, but why? If the problem is to maximize yield with some fairness constraint, it'd be good to know what the rationale is.

4 comments

Thank you for your thoughtful review. Your quantitative suggestions helped improve the paper, and we appreciate the opportunity to clarify its motivation. We start with the question we consider the most important.

Why consider fairness if the objective is targeting? We agree that a platform focused solely on maximizing successful contacts has no reason to impose exposure parity. Our original introduction did not make this distinction clear, and we have now revised it.

Our argument is that every routing policy determines who receives unwanted contact, including unconstrained yield maximization. This distributional consequence exists whether or not fairness is an explicit objective. At q=0.10q=0.10, unconstrained routing produces a female-to-male per-capita burden ratio of 0.66. Exposure parity reverses the imbalance, producing a ratio of 1.69. Between these policies lies a unique burden-parity point, rβˆ—r^\ast, where both groups face equal per-capita burden (Proposition 2, under its sign condition). Reaching it increases total unwanted contact by less than 1%. We do not advocate either parity objective. Instead, we show that equalizing exposure can shift unwanted contact between groups rather than eliminate the imbalance, while burden parity offers an alternative at a small efficiency cost. Theorem 2 also shows that differential measurement noise can redistribute burden even when fairness is not an explicit objective.

We now explain this motivation in the Introduction (Section 1). Consent remains a hard constraint: opt-out restrictions take priority, and our analysis concerns only people who may permissibly be contacted but whose receptivity is uncertain (Conclusion, Section 8).

Ratio versus absolute difference. You are right that burden ratios can exaggerate differences when burden rates are small. We therefore also report hAβˆ’hBh_A-h_B, measured in percentage points of per-capita burden.

qqr0r_0rβˆ—r^\astr=1r=1Ratio at r=1r=1
0.10-1.40+2.01.69
0.25-4.60+4.01.29
0.50-5.90+4.81.13

These numbers do not change the conclusion: unconstrained routing and exposure parity produce imbalances in opposite directions, while burden parity eliminates the gap. The point rβˆ—r^\ast and the incompatibility band are unchanged because both measures define parity by hA=hBh_A=h_B. The measures do, however, tell different stories about magnitude. At exposure parity, the absolute burden gap increases with the budget even as the ratio decreases. Table 1 and Section 7.1 now report both measures so readers can assess this distinction.

Dataset and measurement details. We expanded Section 3 (Setup) to explain how receptivity is measured. The dataset asks respondents how they would react if another user's agent contacted them. Under strict coding, only "would engage" counts as positive; under soft coding, "might engage" also counts. The released measurement model estimates each respondent's probability eje_j of a positive response. Thus, 1βˆ’ej1-e_j represents the estimated probability of unwanted contact, not observed behavior. The main analysis uses strict coding; soft coding reproduces the qualitative findings (Appendix A).

Proposition 3(ii): Group-aware score transformations. Thank you for this suggestion. We added the transformation ΞΌg+ρg(ΞΈ^βˆ’ΞΌg)\mu_g+\sqrt{\rho_g}(\hat\theta-\mu_g). Under the Gaussian model, it restores the noiseless marginal distribution N(ΞΌg,sg2)N(\mu_g,s_g^2) and hence the noiseless group exposure rates under the corresponding pooled threshold. However, the transformation cannot recover information lost through measurement noise. It preserves the within-group ranking of receivers, so the targeting attenuation described in part (iii) remains. We now state this distinction explicitly: exposure distortion can be corrected through group-aware transformations, but targeting error caused by noise cannot.

Why use tβˆ’1/2t^{-1/2} for exploration? We chose this rate because it provides continued exploration while allowing the fraction of exploratory contacts to decline. Under Proposition 4's assumptions, including uniform within-group exploration and uniformly positive expected group allocations, each receiver receives an expected Ξ©(T)\Omega(\sqrt T) exploratory selections over TT rounds. This supports consistent learning while exploration contributes only o(T)o(T) regret. The exponent is not essential. Any rate tβˆ’at^{-a} with 0<a<10<a<1 provides consistency under the same assumptions. We now explain this alongside Proposition 4.

Again, thank you for your constructive comments. We hope that you find that your suggestions, among others, have significantly improved the paper.

Dear reviewer opfg, we have addressed all of your concerns in the updated paper. Please let us know if you have any remaining questions, so we can respond to them by the close of discussion on the 18th. Thank you

I understand the point "unconstrained routing and exposure parity produce imbalances in opposite directions", which frankly speaking is hardly surprising. My point is that if there is a definitive goal (as all these systems must have, such as high yield), then routing should never have been unconstrained to begin with, and that the real question is how to do routing property to maximize yield. I don't think the comments have addressed my key concern, which is why looking at exposure and burden parity as the issue to fix, while the real problem is how to do routing to optimize whatever the objective may be. Put it differently, had the gender (using the example shown in the response) distribution (or other demographic statistics) being different on the platform, you may well not see the burden disparity issue, while the yield may still be poor depending on the targeting criteria.

Thank you for following up. We think we know where the misunderstanding comes from, and we would appreciate the opportunity to resolve the issue.

First, a clarification about "unconstrained" routing. In our paper, "unconstrained" means yield-maximizing. It appears that the way we've used the term leads to an interpretation that it's "not optimized." This is not the case: baseline is the policy that achieves the highest possible yield (Proposition 1), which is represented by r0r_0 in Table 1. We will replace "unconstrained" with "yield-maximizing" throughout the paper to avoid misunderstanding by the reader.

Why does fairness matter here? Consider a platform that works as you appear to suggest: it contacts the people who are most likely to respond, without imposing any fairness constraints. We agree that this is a reasonable objective. However, targeting is based on estimates, not perfect knowledge. Some contacts will inevitably be unwanted, and their distribution across groups depends on the routing policy.

This is exactly what we observe in our data. Under yield-maximizing routing, women receive about two-thirds as many unwanted contacts per person as men. Under exposure parity, which contacts men and women at equal rates, women receive 1.69 times as many. Neither policy produces equal burden. Between them, there is a unique group-threshold policy that equalizes per-capita burden, at an increase of less than 1% in total unwanted contact.

We are not arguing that platforms must choose burden parity. Our point is that every routing policy determines how unwanted contact is distributed, whether or not the platform considers fairness. We provide a way to measure these consequences and compare the costs of different choices.

On targeting accuracy. We agree that improving targeting is important. But targeting accuracy and the distribution of unwanted contact are connected. Theorem 2 shows that, under its stated conditions, greater measurement noise for one group can increase both its exposure and its unwanted-contact burden, while reducing targeting accuracy within that group. Improving measurement quality can help with both problems, although correcting exposure differences alone does not recover information lost to noise.

On other platforms. You are right that the results depend on differences in receptivity between groups. If the groups had identical receptivity distributions, the incompatibility would disappear, as Theorem 1(iii) states. In our dataset, however, the groups differ by 0.30 standard deviations, making the trade-off relevant.

On whether the findings are surprising. We agree that different routing objectives can favor different groups, and that this is intuitive. What is less obvious is that exposure parity and burden parity generally cannot be achieved simultaneously within group-threshold routing, that the resulting trade-off can be quantified using real data, and that measurement noise can increase both exposure and unwanted-contact burden.

We hope this clarifies our central argument. We are not questioning yield maximization as an objective. We are showing that even a yield-maximizing policy has distributional consequences, and that these consequences can be measured and compared.

Reviewer_p8g6OpenReview β†—
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

This paper studies the fair allocation of algorithmically initiated messages, considering both wanted and unwanted attention. They state that a greedy selection of the most receptive users within each group does not generally produce equalised outcomes in terms of rates of unwanted contact. Using survey-based estimated and simulations of message allocation, the authors characterise the trade-off between equalising contact rates and unwanted-contact rate. They investigate how prediction errors can increase unwanted contact, and propose algorithms for equalising unwanted contact rates across groups.

Critical strengths

S1. Problem analysis.
In the age of automated algorithmic spam, soliciting and computerised interaction, the authors have identified a concrete and important problem to investigate in modern media. The paper establishes that equal contact rates does not imply equal unwanted-contact rates, which is a case against treating attention as uniformly beneficial especially since it is a limited resource.

S2. Experimentation.
The authors utilise survey-based simulations which are corroborated via robustness checks, to support some of their conceptual claims. They are helpful for visualising the aforementioned trade-off and how routing policies might be one way of addressing it.

Critical weaknesses

W1. Clarity.
This paper has been extremely cumbersome to read due to the usage of overly complex, non-standard phrases, here are just a few examples:

  • Lines 441-442: "the exposure-parity contour [...] is reachable by noise alone [...] (the masquerade of Theorem 2)" β†’ (disambiguated) measurement noise can make contact rates appear equal;
  • Lines 470-471: "the band … is a menu, not a verdict" β†’ the model describes policy options though it does not specify which is the fairest;
  • Lines 107-108: "the shrinkage escape route is closed" β†’ adjusting estimates toward group averages does not resolve the problem;
  • Lines 124-125: "whose released propensities and measurement diagnostics supply our calibration" β†’ we set the model's parameters using wilingness estimates and uncertainty measures. The usage of underspecified, imprecise and overly metaphorical language is very counterproductive for conveying/understanding technical methdology and formal results. In several places, the framing also makes the results appear more general than is soundly supported by assumptions made, and the restricted class of routing policies under which results are proven. See the soundness concerns below for details.

W2. Soundness.
There is an incongruence between the framing of this work and the actual formal results. As acknowledged by the authors, Thm 1's impossibility result only applies to optimal within-group threshold policies; it is note proven generally for other selection rules, such as uniform random selection or random selection without replacement. The abstract and introduction suggest a broader, unsubstantiated incompatibility. Furthermore, the proof relies on an additional sign condition to establish that burden parity lies between the unconstrained optimum and exposure parity. It should be clarified that this ordering is conditional on that assumption. Finally, the widening incompatibility gap (under greater selectivity) is extensively discussed as a general property, but lacks concrete theoretical support. While some experimental evidence has been put forth, they fall short of certifying that this behaviour holds generally acorss distributions considered.

Questions for authors

Q1. Have the authors considered the effect of limited recipient attention on the simulation and trade-offs? E.g., the effects of repeated unwanted contact on each recepient's willingness rate? Do the authors think this rate would realistically be constant or diminish over exposure and time?

Q2. Would equalising unwanted contact rates also tend to increase the total volume of unwanted contact relative to unconstrained routing? It might be helpful to report both the distribution and total burden, since greater equity does not imply greater overall "utility" (ie. less "harm") overall.

3 comments

Thank you for your careful and constructive review. We agree with your concerns about clarity and the scope of our claims, and have substantially revised the manuscript to address these concerns. We start with your two questions, and then address the weaknesses you have identified.

Q2: Does equalizing burden increase total burden? Yes, but only slightly. Total unwanted contact equals total contact minus welcome contact. Since the selected mass is fixed at qq, total burden per capita is q(1βˆ’Y)q(1-Y). Thus, any reduction in welcome contact increases unwanted contact by the same amount.

The percentages below measure increases relative to total burden, rather than the yield losses reported in Table 1, which are relative to YY.

qqunconstrained r0r_0burden parity rβˆ—r^\astexposure parity r=1r=1
0.100.03510.0354 (+0.9%)0.0366 (+4.2%)
0.250.15140.1516 (+0.1%)0.1520 (+0.4%)
0.500.38650.3866 (+0.02%)0.3867 (+0.06%)

At all three budgets, burden parity increases total unwanted contact by less than 1%, substantially less than exposure parity. Following your suggestion, Table 1 (Section 7.1) now reports total burden alongside its distribution across groups. The underlying identity is given in Equation (4) of Section 3.

Q1: What if receptivity declines with repeated contact? Our model treats receptivity eje_j as fixed and therefore does not capture this possibility. We agree that repeated unwanted contact may reduce willingness to engage. We now explicitly acknowledge this limitation and identify a model in which receptivity changes with accumulated burden as a natural extension. Our current results do not apply to that setting.

W2: Scope of the incompatibility results. We agree with your three points and have revised the relevant claims.

(a) Scope of Theorem 1. The incompatibility result applies to within-group threshold policies, not to all possible selection rules. We now state this restriction in the theorem, abstract, and introduction.

(b) Existence of the burden-parity point. The ordering r0<rβˆ—<1r_0<r^\ast<1 requires the sign condition Ξ²(r0)<1<Ξ²(1)\beta(r_0)<1<\beta(1). This condition is now explicit in Proposition 2. The proof of Theorem 1 refers to that proposition rather than assuming the existence of rβˆ—r^\ast.

(c) Widening of the incompatibility band. We now describe this as an empirical comparative static, state the sufficient tail condition, and no longer present widening as a general theoretical result.

We also tightened four related claims: Proposition 4, the generalized-utility remark, the noise range in Theorem 2, and the comparison baseline and regime in Proposition 3(ii). Specifically,

  • Theorem 1, abstract and introduction are scoped to within-group threshold policies;
  • The sign condition is now a hypothesis of Proposition 2, and the widening band is labelled an empirical comparative static only;
  • Proposition 4 now concerns expected rather than realized burden, and specifies randomized rounding of the fractional group allocation;
  • Section 7.3 identifies the configuration in Figure 3 (deterministic rounding, no forced exploration) as the simpler variant and reports the learner exactly as specified in Proposition 4 alongside it;
  • The remark on non-complementary benefit and harm now says that the incompatibility extends under group-constant harm intensity;
  • Proposition 1's threshold structure needs expected contact utility nondecreasing in receptivity;
  • Theorem 2 is stated over the noise range on which its regime conditions hold, and part (iii) is claimed for the probit link only;
  • Proposition 3(ii)'s raw-versus-shrinkage comparison is stated against a noiseless comparison group, in the scarce regime along the whole comparison;
  • Proposition 4's coverage assumption is now a fixed bisection range containing rβˆ—r^\ast, which bounds the expected group allocations below in every round, with an endpoint fallback for rounds in which the plug-in root is not bracketed; and
  • Section 7.2 identifies the logistic noise sweep as a numerical validation of the probit result in Theorem 2(iii).

W1: Clarity and presentation

We removed the expressions you identified and replaced them with more direct explanations. We also substantially revised the wording of the manuscript to make it easier to read (we hope you agree). In addition, we made the presentation changes requested by Reviewer r1ce: defining variables at first use, numbering displayed equations, and clearly distinguishing theoretical results from empirical findings. The new PDF is uploaded.

We hope these revisions address your main concerns, particularly the scope of the incompatibility theorem, the conditions for burden parity, and the previously unreported effect on total unwanted contact.

Dear Reviewer p8g6, the revised manuscript, which contains all the updates you requested, is uploaded. Please let us know if there are any questions or concerns that we can still address before the end of the discussion period. Thank you.

I appreciate the authors' clarifications of the assumptions and the new numeric results on total unwanted contact. However, my concerns about the scope of contributions remain: the impossibility result applies only to within-group threshold policies; the evidence that the gap widens with greater selectivity is still empirical. I maintain my score of 1 (clear rejection).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.