- 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 is receptivity then 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 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 receiverβs 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 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 , which is nowhere defined. Nor is 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
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.
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 Y 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 bgβ is the probability that a contacted member of group g finds the contact unwanted. Per-capita burden hgβ=tgβbgβ measures the expected unwanted contact received by a random member of that group. The distinction matters because bgβ describes the experience of those contacted, while hgβ 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=0.10, the ratio bAβ/bBβ 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 rβ 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 N has been removed because the model uses unit mass; r0β and rβ are introduced in Section 3; and Ο, Ξ¦Λ, and Ξ» 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.