acceptodds
Under review as a conference paper at ICLR 2027

Simulating Eutopia: Long-term Fairness with Outcomes, Performativity, and Dynamics

Abstract

As AI-driven Decision Makers (ADMs) influence our socioeconomic reality, their roles in both enhancing efficiency and amplifying the social biases have drawn significant attention. In this paper, we revisit the nuances of *long-term fairness* achievable by an ADM– specifically, through the lens of a loan approver inducing a population-level wealth dynamics. The literature generally (a) considers passive environments, i.e. the decisions of an ADM does not change the population's behaviour, and (b) measures bias in terms of disparity in the instantaneous decisions rather than downstream equity. Modern ADMs challenge both the notions. To address these caveats, we first formalise the wealth dynamics induced by a loan approving ADM interacting with a multi-demographic population as a Performative Markov Decision Process. Then, we mitigate the absence of such a performative test-bed by developing **Eutopia**: a lending-process simulator enabled with a novel performative data generator to learn long-term fair strategies. With **Eutopia**, we test reinforcement learning algorithms with different fairness-aware utilities dependent on approval decisions and downstream wealth. Results show that performative dynamics-aware learning with fairness-aware utility that incorporates the downstream outcomes induce better long-term equity and inclusivity.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Long-term Fairness, Simulator, Markov Decision Process, Performative Optimality, Outcome Fairness

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_5qX8OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper formalizes sequential credit lending as a performative MDP in which the deployed policy shapes both wealth dynamics and future applications. It builds Eutopia, a simulator that joins an RL lender, realized repayments, and social dynamics. Experiments compare policy gradient against its performative counterpart (PePG) across two fairness perspectives, outcome parity versus true-positive-rate parity, and three constraint forms. The paper reports that performative learning performs best, and that opportunity-based utilities reproduce the delayed impact of fairness.

Critical strengths
  1. The testbed is a significant contribution. Existing lending simulators move credit scores in response to decisions but sample arrivals exogenously; here the applicant pool itself reacts to approval history.

  2. The micro-analysis in Section 4.3 and Appendix F.2.4 is the strongest part of the paper. Lorenz curves, reach rates, and never-served shares show that similar macro metrics hide opposite credit allocations, and reproducing the delayed impact of fairness inside a performative environment connects the paper concretely to Liu et al. (2018) and D'Amour et al. (2020).

  3. Reproducibility is strong. Rule-based policies validate the simulator before any learning claim, and code, seeds, and hyperparameters are complete.

Critical weaknesses
  1. Table 2 mixes two training protocols. Five cells learn the multiplier by dual ascent; the cell for FL under Equality of Outcome reports the fixed-multiplier policy of Appendix F.3. Under dual ascent, that cell earns a fifth of the profit with a wider gap (Table 7). The conclusion that FL under Outcome closes the wealth gap while keeping two thirds of the maximal profit (Sections 1 and 4.2) therefore describes the fixed-multiplier policy, while Figure 3 and the surrounding discussion describe the dual-ascent run. The authors ought to report the dual-ascent numbers in Table 2, present the fixed-multiplier policy as a separate result, and restate the conclusion. The fixed-weight result is useful on its own and deserves a named place in the main text.

  2. SAC weakens the claim that the performative agent beats standard RL on all metrics in every scenario (Sections 1, 4.1, and 5). Under FL with Equality of Outcome, SAC earns more profit with a smaller wealth gap and an inequality ratio closer to one than either performative variant (Table 15). Appendix F.4 argues SAC is a fixed cutoff that concentrates loans within each group, but the fixed-multiplier policy shows similar concentration, and SAC's never-served share is not reported. The authors ought to scope the claim to policy-gradient learners and report SAC's micro metrics for this cell.

  3. The comparisons in Table 2 report no significance tests, and several bolded wins are smaller than their standard deviations (Table 7). The wealth gap itself is computed on seed-averaged trajectories, so gaps of opposite sign cancel across seeds, and Appendix F.5 notes the per-seed mean gap is more than three times the value in Table 7. The authors ought to report the per-seed absolute gap and paired per-seed comparisons for the Section 4.1 claims.

Questions for authors
  1. Is the FL-under-Outcome cell in Table 2 the fixed-multiplier run of Appendix F.3 while the other cells use dual ascent, and what are the dual-ascent numbers for that cell?

  2. For SAC under FL with Equality of Outcome, what are the never-served share, reach rate, and per-group growth rates?

  3. Do the Section 4.1 results hold under a paired per-seed comparison on the per-seed absolute gap?

5 comments
AuthorsOpenReview ↗

Official Rebuttal – All concerns and questions are addressed in the revised manuscript

We thank the reviewer for their detailed review and appreciating the nuances offered by our proposed simulator. We respond to the concerns raised below.

W1 & Q1. Yes, the FL+Outcome entry of Table 2 corresponds to PeRL-γ⋆\gamma^\star (Appendix F.3), while the other entries use the end-values of learned multiplier. With the end-value of the learned multiplier, this cell yields a profit of 1.304±0.9741.304 \pm 0.974 (×106\times 10^6), a wealth gap of 71.9±298.071.9 \pm 298.0, and ρ=1.059±0.094\rho = 1.059 \pm 0.094 (Table 7).

We highlight that the two policies are two points on the same fairness-utility trade-off of FL+Outcome. Under Outcome, the Matthew effect keeps pushing the wealth gap open, so the constraint keeps binding and dual ascent drives the multiplier towards its fairness-holding extreme: the gap closes and loans are spread evenly, at the cost of profit. Under Equality of Opportunity, the constraint is met early and the multiplier decays towards the profit-earning extreme (Figure 3). γ⋆=1.87\gamma^\star = 1.87 is the strength at which we achieve maximal utility with wealth gap as same as with end-value of the Lagrangian multiplier. PeRL-γ⋆\gamma^\star is, thus, the balanced point learned by the same dual-ascent used for other objectives.

With the end-value of the learned multiplier throughout, FL+Outcome remains the best utility for every learner under relaxed wealth dynamics and on two additional datasets (Appendix G.3 and G.4 in the revised manuscript).

W2 & Q2. We agree with the reviewer on the scope i.e., the claim in Sections 1, 4.1 and 5 concerns PeRL and its passive counterpart, i.e., policy gradient, which is the comparison that isolates the effect of performativity. We have added a detailed comparison with SAC in Appendix G.2 in the revised manuscript, and summarise the observations below.

  • Requested micro metrics (FL+Outcome). Both the PeRL policies reach more of each group and grow both groups faster than SAC.
PolicyNever servedReach rateGrowth rate RgR_g (×10−3\times 10^{-3})
SAC0% / 0%5.3% / 5.0%2.4 / 2.3
PeRL0% / 0%8.6% / 7.4%3.5 / 3.3
PeRL-γ⋆\gamma^\star1% / 1%7.2% / 6.9%2.6 / 2.5
  • SAC responds to only one utility. SAC is a passive learner, i.e., it does not account for how its decisions change who applies next. In four of its six cells, it ends at the same point irrespective of the utility. FL+Outcome is the one utility that reaches parity under SAC.
  • SAC does not improve on PeRL-γ⋆\gamma^\star. The wealth gap of 88 for SAC in Table 15 is the cancellation that the reviewer points out in W3: its seeds split in sign, and its per-seed absolute gap is 7070 against 4848 for PeRL-γ⋆\gamma^\star. When SAC is also given PeRL's pre-training, PeRL-γ⋆\gamma^\star has the smaller wealth gap and ρ\rho closer to 11, SAC earns a slightly higher profit, and in most of the seeds, PeRL-γ⋆\gamma^\star's gap is below 10, whereas SAC's lies between 82 and 94 in every seed.

Thus, SAC reaches parity only under FL+Outcome, there it does not improve on PeRL-γ⋆\gamma^\star, and only PeRL offers the evenly-served end of the trade-off. For the detailed results, we refer to Appendix G.2.

W3 & Q3: Statistical significance. Yes, the results of Section 4.1 hold under paired per-seed comparisons. The standard deviations of Table 7 measure the variation across seeds, whereas the claim of Section 4.1 is about the difference between the two agents on the same seed, which is what a paired test isolates. We ran exact two-sided paired Wilcoxon signed-rank tests over the 10 seeds on the per-seed profit, the per-seed absolute wealth gap, and the per-seed distance ∣ρ−1∣|\rho - 1|, with Holm correction over the 21 comparisons. The table below reports, for PeRL against RL as in Section 4.1, the number of seeds (out of 10) in which PeRL is better, along with the per-seed absolute wealth gap of both the agents. Here, ∗^* denotes a significant difference.

UtilityProfitPer-seed absolute gap (PeRL vs. RL)∣ρ−1∣\lvert\rho - 1\rvert
FL+Outcome10/10$^*$252 vs. 830, 5/109/10$^*$
RMM+Outcome9/10$^*$3,082 vs. 3,190, 9/10$^*$10/10$^*$
SW+Outcome10/10$^*$2,040 vs. 3,170, 10/10$^*$10/10$^*$
FL+Opportunity10/10$^*$1,087 vs. 3,155, 10/10$^*$10/10$^*$
RMM+Opportunity10/10$^*$1,075 vs. 3,177, 10/10$^*$10/10$^*$
SW+Opportunity10/10$^*$1,089 vs. 3,116, 10/10$^*$10/10$^*$
MaxUtil10/10$^*$1,181 vs. 3,138, 10/10$^*$10/10$^*$

We observe that PeRL is better than RL in 9 or 10 of the 10 seeds on profit and on ρ\rho for nearly every utility, and on the per-seed absolute gap in every utility. These conclusions remain unchanged when RL is additionally given PeRL's pre-training (Appendix G.1 in the revised manuscript).

We hope we have satisfactorily addressed all the concerns. We are eager to discuss further if there are any further questions/concerns.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer 5qX8,

Thank you again for your constructive feedback on our work. We submitted our responses a few days ago but have not yet heard back from you. As the discussion period is coming to an end, we would appreciate it if you could revisit your score if our responses have addressed your concerns.

If you have any further questions, please feel free to let us know. We look forward to hearing your thoughts.

Kind regards, Authors

Thank you for your response and the new appendix experiments. My Q2 and Q3 are sufficiently answered. One question remains on Q1:

How was γ* = 1.87 selected? Appendix F.3 says it is where the Outcome and Opportunity multipliers meet during pre-training; the response describes it as the strength that maximizes utility at the dual-ascent wealth gap, which would use deployment outcomes. Also related, Table 2 and contribution (c) still report PeRL-γ* as the FL/Outcome result.

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

The paper studies long-term fairness under AI decision-making, where decisions change the population it later faces, It formalizes sequential lending as a performative MDP, where repaid loans raise borrower wealth and each group's application rate depends on past approvals via Hawkes process. The paper proposed a lending simulator termed Eutopia, with an initial population statistics derived from UCI Adult Census dataset,. They compare a standard policy gradient learner (RL) with a performative learner (PeRL) under 6 different fairness-aware utility settings - combining equality of outcome and equality of wealth, with 3 ways of turning them into a contraint: social welfare, Rawlsian max-min and a fairness Lagrangian. Empirical results over multiple deployment episodes and seeds report that: PeRL outperforms standard RL which ends up approving almost every applicant, under equal opportunity the fairness constraint is met early and then stops binding, and the Fairness Lagrangian under equality of outcome gives the smallest wealth gap while keeping 2/3rds the maximum profit.

Critical strengths

The paper offers a test bed in which lending decisions affect who applies and how wealth evolves, and in which the population response mechanism can be switched off to recover the non-performative case. This makes it easy to compare learners and fairness objectives on long-term outcomes rather than on one-shot decisions. The within-group analysis is useful. Lorenz curves, reach rates and never-served shares show that the policies with similar group level metrics can distribute credit very differently inside each group, for example leaving a large part of each group unserved under equality of opportiunity. This is a point that aggregate metrics miss.

Critical weaknesses

The participation feedback in the PeMDP, where a group that is consistently rejected applies less often, is similar to some earlier literature on long-term fairness. Hashimoto et al. (2018) and Zhang et al. (2019) model group retention as a function of how the decision rule treats each group. Mouzannar et al. (2019) ["From Fair Decision Making to Social Equality"] and Zhang et al. (2020) ["How Do Fair Decisions Fare in Long-term Qualification?"] model how group qualification evolves under the decision policy in a partially observed MDP setting. Puranik et al. (2022) ["A Dynamic Decision-Making Framework Promoting Long-Term Fairness"] model the applicant pool composition as the MDP state, with a group's future share increasing when its share of admissions exceeds its share of applicants, and further this is extended to multi-agent settings. D'Amour et al. (2020) simulate lending with credit feedback. Many of these work also give formal theoretical results on conditions for equality or growing disparity, characterizations of equilibria, convergence results, bounds on loss and fairness violations for RL. conditions under which Rawlsian policies outperform utilitaian ones (Wu et al, 2025), which is similar to the SW v/s Rawlsian comparison in this paper.

But the submission offers no theoretical guarantees. Remark 1 which argues that feedback cannot be absorbed into an augmented state seems like a modeling choice rather than a necessity, given the POMDP of Zhang et al (2020) and applicant-share MDP in Puranik et al. (2024). The main contribution of this paper rests on the simulator and the empirical findings. Another concern is whether the emprical results are general enough, as all results come from one simulator and dataset. Only the initial wealth comes from the dataset, while loan sizes, default probabilities, constraints are set by authors. Some of the main findings seem to follow from how the objectives are defined, for eample, the FL under equality of outcome directly penalizes wealth gap that is then used as the evaluation metric.

Questions for authors
  • Given prior work that models participation and qualification dynamics often with formal guarantees, could the authors clarify the main contributions? If it is the simulator, has the application dynamics model been validated against real lending or application data? How easily can the data generator be used with datasets other than UCI Adult?
  • Could the authors clarify what the PeMDP response captures participation and qualification dynamics that earlier models do not, or compare this with prior works?
  • It seems like RL and PeRL are not trained under the same conditions. If I understand the paper correctly, RL is pre-trained in an environment that resets every episode, while PeRL is pre-trained in the continual one. Could the authors compare the two under the same pre-training environment, and test PeRL when its copy of the dynamics is not exact? This would show whether PeRL gains come from being performative or from knowing how simulator works.
5 comments
AuthorsOpenReview ↗

Official Rebuttal – All concerns and questions are addressed in the revised manuscript

We thank the reviewer for their careful review on our work. We respond to the questions and concerns raised below.

Q1--Q2: We thank the reviewer for the detailed list. Each prior model isolates wealth and application dynamics. In performative lending, these two dynamics are coupled: an approval raises wealth, wealth raises the tendency to apply, and who applies decides which cohort the lender learns from next. Our contributions emerge from this coupled regime.

Formalism. The PeMDP lets the transition kernel depend on the deployed policy π\pi, and not only on the realised action. The action-dependent models above are recovered as special cases (see our response to W3).

Test-bed. Eutopia is, to our knowledge, the first lending simulator in which the applicant pool itself reacts to the approval history. The performative response can be switched off to recover the non-performative setting, which makes the effect of the feedback measurable.

Novel findings.

  • A learner that ignores the performative feedback (i.e., standard RL) drifts to approving nearly every applicant, while a performative learner does not (Sec 4.1).

  • Utilities with similar group-level metrics allocate credit in opposite ways inside each group. Under Opportunity, 15−2615-26% of each group never receive a loan (Figure 4) and the female reach rate decays (Figure 14).

  • FL+Outcome is the only fairness-aware objective that keeps the fairness-utility trade-off alive in long-run. A decision-level constraint (TPR parity) is met within the first episodes, its multiplier decays to the floor, and the agent then behaves as a MaxUtil one while the Matthew effect widens the gap (Figure 3). An outcome-level constraint cannot be satisfied while the harm accumulates, so it keeps binding throughout deployment. This is what reproduces, and then overcomes, the delayed impact of fairness (Liu et al., 2018) in a performative environment.

Q1. The findings hold across ablation studies and datasets. We adapted Eutopia to Taiwan Credit (30,000 credit-card clients, groups by sex) and German Credit (1,000 applicants, groups by age); see Appendix G.4 in the revised manuscript. We observe that all our reported results naturally extend to these settings.

Q3: In the paper RL is pre-trained in the static environment, the standard setting for a passive learner, and PeRL in the performative one.

Same pre-training (Appendix G.1). We retrained RL with exactly PeRL's pre-training and left everything else unchanged. PeRL keeps its advantage. It earns more in all seven utilities, by a factor of 1.41.4 to 5.25.2 in six of them. Its ρ\rho is significantly closer to 1 in every utility: under Equality of Opportunity and MaxUtil, RL lets the red group's wealth grow 7 to 21 times faster than the blue group's, against 1.21.2 to 4.74.7 times under PeRL. Its wealth gap there is 208 to 1,181, against 2,075 to 2,437 for RL. RL is not significantly better on any metric in any utility. Changing only RL's pre-training environment from static to continual changes none of its end-of-deployment metrics significantly, and the comparison with PeRL is the same. Experiencing the population's response during pre-training is therefore not what RL lacks; what separates the two agents is how the gradient is computed.

An inexact copy of the dynamics. The performative-strength runs of Appendix F.5 contain this test. There the simulator's feedback is scaled by 0.50.5 and by 22 while PeRL's model keeps the default wealth gain per repaid loan, so the model is wrong by a factor of two in each direction. PeRL's ρ\rho is still closer to 1 than RL's in 8 of 10 seeds at ×0.5\times 0.5 and 9 of 10 at ×2\times 2 (9 of 10 with the exact model), and PeRL earns more than RL in 9 of 10 seeds at both scales. PeRL's gains thus do not rest on knowing the simulator exactly.

AuthorsOpenReview ↗

Clarification on weaknesses

Absence of theoretical guarantees. We agree that the paper does not contain new theorems, and we state this as an open direction in Section 5. This is a deliberate choice of scope, for the following reasons:

  1. with participation and wealth coupled, heterogeneous individuals, and an ADM that learns over a long horizon, a closed-form characterisation of equilibria is no longer available. This is the regime where simulation is the established instrument (D'Amour et al., 2020), and where single-step analyses are known to mispredict long-term behaviour;

  2. the learner is the performative policy gradient algorithm, which has convergence guarantees to performative optimality (Basu et al., 2025). The simulator is validated with rule-based policies before any learning claim is made;

  3. the findings are not only consistent with the existing theory but also extend it. Eutopia reproduces the delayed impact of fairness (Liu et al., 2018) and representation disparity (Hashimoto et al., 2018) without forcefully building them in. MaxUtil becomes as harmful as the Equality of Opportunity agent once the applicant pool reacts, and the Rawlsian utility under Outcome over-lends to the female group, which complements the conditions of Wu et al. (2025).

We see the simulator as a tool to find the phenomena that are worth proving for long-term fairness.

Added references. We thank the reviewer for these references. Hashimoto et al., 2018, Zhang et al., 2019 and D'Amour et al., 2020 are already discussed. Now, we also cite and discuss Zhang et al., 2020, Puranik et al., 2022 and Mouzannar et al., 2019, in Section 1 and Appendix B.

Remark 1. We clarify below why Remark 1 states a structural property of our setting and not a modelling preference.

Puranik et al. (2022)'s modelling is a special case of ours. When the response function depends only on realised decisions and the augmented state is observed, as in Puranik et al. (2022), the PeMDP reduces to an augmented MDP. These models are thus special cases of our formalism, and the regime we target is the complementary one. Our experiments show that the distinction matters: standard policy gradient, which ignores the dependence of the kernel on the policy, converges to approving nearly every applicant, whereas the performative learner does not.

Action-dependent versus policy-dependent dynamics. In Zhang et al. (2020), qualification evolves through a fixed kernel conditioned on the current qualification and the realised decision. In Puranik et al. (2022), the applicant-pool composition is a fully observed state with transitions P(st+1∣st,at)P(s_{t+1} \mid s_t, a_t). In both, there is one fixed environment and the policy only selects actions inside it, so state augmentation is exactly the right tool. In our PeMDP, the population responds to the deployed approval rule, pt+1apply=fψ(ptapply,St,Pπ(At∣St))p^{\mathrm{apply}}_{t+1} = f_{\psi}(p^{\mathrm{apply}}_{t}, S_{t}, P_{\pi}(A_{t} \mid S_{t})). The kernel is therefore indexed by the policy, M(π)M(\pi), and this includes the behaviour of π\pi on individuals who do not apply and for whom no action is ever realised. Augmented MDP of Puranik et al. (2022) does not capture this.

"Some of the main findings seem to follow from how the objectives are defined, for eample, the FL+Outcome directly penalizes wealth gap that is then used as the evaluation metric".

Three results do not follow from the definition (Appendix G.2 to G.4).

  1. RMM and SW under Outcome are built on the same group-wealth statistic, and their gaps are larger than FL's for every learner in every one of the six settings; on Adult they are PeRL's two largest. What matters is that FL penalises the difference between the groups.

  2. PeRL, RL and SAC trained on the identical FL/Outcome utility differ on quantities it does not contain: with PeRL's pre-training, RL keeps the gap small by nearly ceasing to lend (it reaches 2% of each group per episode against 7 to 9% for PeRL), and SAC concentrates its loans (Gini of 0.420.42 / 0.390.39 against 0.080.08 / 0.090.09).

  3. FL+Outcome is the only utility under which no individual goes unserved and loans are spread almost uniformly within each group (Table 12); none of these quantities appears in any utility.

We hope we have sufficiently resolved all the raised concerns/question raised by the reviewer. We are happy to discuss if there are any further questions or suggestions.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer fbKw,

Thank you again for your constructive feedback on our work. We submitted our responses a few days ago but have not yet heard back from you. As the discussion period is coming to an end, we would appreciate it if you could revisit your score if our responses have addressed your concerns.

If you have any further questions, please feel free to let us know. We look forward to hearing your thoughts.

Kind regards, Authors

Reviewer_xeEsOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper examines the long-term fairness of AI-driven decision makers (ADMs) when their decisions influence future human behaviour, using credit lending as the main setting. The authors introduce Eutopia, a simulator that models lending decisions, repayment outcomes, changes in wealth, and long-term effects of the ADM’s decisions on future participation and the group wealth gap. Experiments suggest that performative RL consistently outperforms RL in both profit and fairness, with standard RL often converging to a policy that approves nearly every loan. It is also shown that fairness based on outcomes (rather than opportunity) can lead to better long-term equity.

Critical strengths

The paper’s main contribution is the formalization of a performative MDP in the credit lending setting through a simulator, together with an empirical comparison of different learning algorithms and fairness objectives.

Critical weaknesses

While the problem formulation is valid, the scope is a bit narrow. Furthermore, while some of the findings are interesting, they depend substantially on the specific modeling assumptions in Eutopia. Specifically, there are some questionable assumptions underlying the simulator. For instance, repayment outcomes and their effects on wealth are immediate rather than delayed. More importantly, both loan denial and loan default leave wealth unchanged, whereas successful repayment increases wealth, meaning wealth can never decrease. I'm not sure these assumptions make sense: a loan denial could lead the prospective borrower to turn to a different investment with their downpayment; a loan default generally results in negative consequences for the borrower. In short, in leaving out some of these key possibilities, it becomes less clear what part of the problem the simulator is capturing.

On a related note, the results of Table 2 are not very surprising. The reported winner is determined by the wealth gap, which strongly aligns with the specific objective that Outcome+FL is trained on. It is not surprising that it performs better on this metric than policies trained using different fairness criteria.

This paper is part of a rapidly growing literature on placing algorithmic decision systems in a feedback loop, where individuals in a population react to decision outcomes thereby changing the statistics of the population, necessitating evolution on the part of the decision system. Not all these studies focus on RL as the learning architecture, even fewer solely focus on the lending problem. Nevertheless, there are many missed relevant references. For instance, the issue of outcome parity vs. decision-level parity was studied in depth in this paper (albeit in a different application context): X. Zhang et al, “How do fair decisions fare in long-term qualification,” NeurIPS, 2020.

Questions for authors

How would the key results change if some of the underlying assumptions are revised or relaxed?

5 comments
AuthorsOpenReview ↗

Official Rebuttal – All concerns and questions are addressed in the revised manuscript

We thank the reviewer for their valuable feedback on our manuscript. We respond to the concerns raised below.

W1 & Q1: Scope of Eutopia-- revised/relaxed assumptions. Eutopia is a flexible simulator of long-term wealth dynamics induced by an algorithmic lender. To reinforce this claim, we have conducted the following ablation studies across modelling choices and datasets that retain the reported conclusions across settings.

  • Relaxed Assumptions: penalty on default, and outside options. With a default penalty, a borrower who defaults loses 0.3 li0.3\,l_i. With an outside option, a denied applicant gains 0.07 li0.07\,l_i, the return of investing a down payment of 0.2 li0.2\,l_i elsewhere. The third setting applies both. All seven utilities were re-run with the protocol of the paper (10 seeds). The findings remain unchanged: (i) FL+Outcome gives the smallest wealth gap of the seven utilities under each rule for PeRL and for RL, significantly so against every other utility; with the default penalty PeRL's gap is 15 with ρ=0.96\rho = 0.96. (ii) PeRL is significantly better than RL on profit, wealth gap, and inequality ratio ρ\rho under the three rules, and earns 1.91.9 to 5.95.9 times RL's profit in the two settings with an outside option. Thus, our earlier findings do not depend on wealth being non-decreasing. For more detailed discussion, we refer to Appendix G.3 in the revised manuscript.

  • Additional datasets. We adapted Eutopia to Taiwan Credit (30,000 credit-card clients, groups by sex) and German Credit (1,000 applicants, groups by age); see Appendix G.4 in the revised manuscript. We observe that the results extend: (i) Under group-neutral rules the group that starts ahead pulls further ahead on every dataset, in the order of its initial lead: ρ\rho ends at 8.88.8 to 10.210.2 on Adult (27% ahead), 2.62.6 to 4.94.9 on German Credit (12% ahead), and 0.750.75 to 0.860.86 on Taiwan Credit, where the other group starts 2% ahead. (ii) FL+Outcome gives the smallest wealth gap for every learner on all three datasets. (iii) PeRL is significantly better than RL in all comparisons on German Credit and intermediately on Taiwan Credit, where the two groups start almost equal and there is little to amplify or to correct.

All these results validate the correctness of the reported conclusions corresponding to PeRL as an algorithm, and FL+Outcome as a fairness-aware utility.

AuthorsOpenReview ↗

Clarification to weaknesses

Results of Table 2. While we agree with the reviewer on the narrow point, i.e., the wealth gap is the quantity that FL+Outcome penalises, so its win on that one column is expected, we would like to highlight that Table 2 is not the only evidence on which our recommendation rests.

We would like to draw the reviewer's attention to three results that do not follow from the definition of the objective:

  1. Standard RL trained on exactly the same objective ends with a wealth gap of 745 and ρ=0.19\rho = 0.19, and the female group grows five times faster. RMM and SW under Equality of Outcome are built on the same group-wealth statistics, yet they end up with gaps of 3,0823,082 and 1,9561,956 and ρ=0.10\rho = 0.10 and 0.140.14. Among the outcome-based utilities, only FL combined with a performative learner closes the gap with ρ\rho close to 11 (gap 71.971.9, ρ=1.06\rho = 1.06 with the learned multiplier, Table 7).

  2. The micro metrics are not part of the objective, and FL+Outcome is the only utility that is equitable on them as well (Section 4.3, Table 12). The utility only sees the two group means. It could be satisfied by lending repeatedly to a small subset of the disadvantaged group, and nothing in it rewards reaching every individual. Yet under FL with Equality of Outcome, no individual of either group goes unserved, loans are spread almost uniformly within each group, and the two groups receive the same volume (243243 and 227227 loans per individual, reach rates of 8.6%8.6\% and 7.4%7.4\%).

  3. FL+Outcome is the only fairness-aware objective that keeps the fairness-utility trade-off alive in long-run. A decision-level constraint (TPR parity) is met within the first episodes, its multiplier decays to the floor, and the agent then behaves as a MaxUtil one while the Matthew effect widens the gap (Figure 3). An outcome-level constraint cannot be satisfied while the harm accumulates, so it keeps binding throughout deployment. This is what reproduces, and then overcomes, the delayed impact of fairness (Liu et al., 2018) in a performative environment.

Hence, FL+Outcome is the only utility that is equitable between groups, within groups, and over time. None of these three properties is in the training criterion.

Added References. We thank the reviewer for pointing us to Zhang et al. (2020). We cite and discuss it in Section 1 and Appendix B in the revised draft. We see it as complementary to our work in three aspects:

  1. Zhang et al. (2020) characterise the equilibria of group qualification rates when a myopic decision maker imposes static, decision-level constraints (demographic parity, equal opportunity). We instead learn long-horizon policies and compare an outcome-level objective (the wealth gap) against an opportunity-level one (TPR parity), tracking wealth, participation and within-group allocation (micro metrics) along the whole trajectory.

  2. In Zhang et al. (2020), the qualification of an individual evolves through a fixed transition kernel conditioned on the current qualification and the realised decision. The environment is, therefore, a single (partially observed) Markov model, and the policy only selects actions inside it. In our PeMDP, the population responds to the deployed approval rule itself, pt+1apply=fψ(ptapply,St,Pπ(At∣St))p^{\mathrm{apply}}_{t+1} = f_{\psi}(p^{\mathrm{apply}}_{t}, S_{t}, P_{\pi}(A_{t} \mid S_{t})), so the transition kernel is a function of the policy, M(π)M(\pi), and changes each time the policy is updated.

  3. As in Remark 1, this dependence cannot be removed just by enlarging the state. Augmenting with ptapplyp^{\mathrm{apply}}_t gives a partially observed problem, and under sustained denial a group stops applying and generates no data, although the dynamics still depend on how π\pi would treat it. Two policies that agree on every visited state can then induce different wealth trajectories. The only statistic that restores the Markov property is π\pi itself, and a state that contains the policy is not a well-posed RL problem.

I hope we have sufficiently addressed all the concerns raised. We are eager to discuss if there are any further questions/concerns.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer xeEs,

Thank you again for your constructive feedback on our work. We submitted our responses a few days ago but have not yet heard back from you. As the discussion period is coming to an end, we would appreciate it if you could revisit your score if our responses have addressed your concerns.

If you have any further questions, please feel free to let us know. We look forward to hearing your thoughts.

Kind regards, Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to reviewers' comments

We thank all reviewers for their constructive feedback. The reviews agree on the value of the test-bed and of the within-group analysis, and the concerns raised were concrete and testable. We have addressed each of them with new experiments in the revised manuscript (Appendix G). We summarise the main concerns and the corresponding changes below.

1. Do the findings depend on Eutopia's modelling assumptions and on one dataset? (xeEs, fbKw) No. We re-ran all seven utilities under relaxed wealth dynamics (a penalty on default, an outside option for denied applicants, and both together; Appendix G.3) and adapted Eutopia to Taiwan Credit and German Credit datasets (Appendix G.4). The conclusions are unchanged: FL+Outcome gives the smallest wealth gap for every learner in every setting, and PeRL retains its advantage over RL.

2. Is FL+Outcome the most effective utility in long-term only because it penalises the evaluation metric? (xeEs, fbKw) No. RMM and SW under Outcome are built on the same group-wealth statistic and end with far larger gaps. FL+Outcome is also the only utility that is equitable within groups (no individual unserved, loans spread almost uniformly) and whose constraint keeps binding throughout deployment. None of these properties appears in the training objective.

3. Is the PeRL vs. RL comparison fair and statistically significant? (fbKw, 5qX8) Yes. (i) With identical pre-training, PeRL still earns more in all seven utilities and RL is not significantly better on any metric (Appendix G.1). (ii) PeRL still earns more than RL in 9 of 10 seeds. (iii) Paired per-seed Wilcoxon tests with Holm correction show that PeRL is significantly better on profit and on the inequality ratio in every utility, and on the per-seed absolute wealth gap mostly everywhere.

4. Table 2 protocol and SAC. (5qX8) We now report the FL+Outcome results for both the end-value learned and the fixed multiplier, and clarify that they are two points on the same fairness-utility trade-off. We have scoped our claim to policy-gradient learners and added the requested SAC micro metrics (Appendix G.2). Once sign cancellation across seeds is removed, SAC does not improve on PeRL.

5. Further related work. (xeEs, fbKw) We now cite and discuss all suggested references (Section 1, Appendix B). Earlier action-dependent models are recovered as special cases of the proposed PeMDP. Our contribution lies in the complementary, policy-dependent regime, where coupled wealth and application dynamics leave no closed-form characterisation and simulation is the established instrument.

We believe these additions directly resolve the concerns behind the current ratings, and that every conclusion of the paper now rests on substantially stronger evidence than at submission. We therefore kindly ask the reviewers to reconsider their scores in light of the revision. We would be glad to address any remaining concern during the discussion period.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.