acceptodds
Under review as a conference paper at ICLR 2027

When Strategic Adaptation Improves Decisions: Causal Incentive Certificates for Decision-Focused Learning

Abstract

Decision-focused learning optimizes predictions for a downstream decision but normally treats the deployment population as fixed, even when the learned rule changes which actions people take. We prove that two predictors can be identical on all historical observations and have zero historical regret while their post-response deployment losses differ by any prescribed amount. We introduce causal incentive certificates, which map downstream-aware action returns to a strongly monotone strategic equilibrium and then to a lower bound on beneficial effort. A primitive causal-decision identity derives the regret slopes from true-state effects and model-induced decision distortion and gives the smallest beneficial-effort threshold that guarantees improvement uniformly over the stated envelope. Separate smooth and hard-decision branches quantify operator misspecification, anisotropic response covariance, and normal-fan switching, while an independent selected-policy test certifies deployment without a candidate-library penalty. Across 5,000 controlled effort instances, the certificate had no violations, the primitive identity closed with maximum error below . In a UCI-based semi-synthetic study, CS-DFL achieved a worst-environment gain lower bound of and a positive alignment bound under the prespecified response model.

Primary area:
optimization
Keywords:
Decision-Focused Learning, Strategic Behavior, Causal Incentive Alignment, Performative Prediction, Bilevel Optimization, Distribution Shift

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_Q6HLOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

When a model is used to make decisions, people may change their behavior in response to it. Some changes improve real outcomes, while others only change the inputs used by the model. The paper first shows that historical data alone cannot distinguish these effects: two predictors can agree and make optimal decisions on historical data, yet perform very differently once people respond. It then gives conditions under which these responses improve decisions, including a requirement that enough effort goes into beneficial actions. The authors also propose a test using separate data that approves deployment only when there is statistical evidence of improvement. They evaluate the approach through controlled simulations and credit-data experiments in which actions, responses, and causal effects are simulated. The results support the approach in these settings, although the guarantees depend on assumptions about people’s behavior and the effects of their actions.

Critical strengths

The paper connects strategic behavior, causal effects, and decision quality. It clearly distinguishes encouraging useful effort from actually improving decisions. The paper handles smooth and hard decision rules separately and keeps model selection independent of final testing. The controlled experiments check the main formulas, and the authors clearly state the limits of their experiments.

Critical weaknesses

(i) The guarantees depend on assumptions about causal effects and human responses that may be difficult to check. The credit experiments use synthetic actions, responses, and costs, so they do not validate these assumptions for real borrowers.

(ii) The credit experiments mainly compare different selection rules over the same 84 candidates. They therefore provide limited evidence about the proposed training algorithm and its ability to handle interacting agents.

(iii) The gains over Causal SC-DFL are small. In Table 1, the worst-environment gain is 0.0880 versus 0.0875, and several additional results are nearly identical. Repeated data splits and a direct statistical comparison would help establish whether this difference is reliable.

Questions for authors

In Sections 4.2–4.3, can the effort certificate satisfy βcert>βcrit\beta_{\mathrm{cert}} > \beta_{\mathrm{crit}} for the same predictor and response model? Please provide one fully specified example with a nonzero equilibrium response, and compute both thresholds from the agents' utility, causal effects, and downstream optimizer. In particular, verify that γD>δD\gamma_D > \delta_D holds along the entire response path. This would show that the incentive and regret guarantees can jointly certify a strict improvement.

6 comments
AuthorsOpenReview ↗

Response to Critical Weaknesses: Assumption Validity, Training with Interacting Agents, and Statistical Comparisons(1/2)

Critical Weaknesses

W: (i) The guarantees depend on assumptions about causal effects and human responses that may be difficult to check. The credit experiments use synthetic actions, responses, and costs, so they do not validate these assumptions for real borrowers.

A: The guarantees admit explicit corrections for bounded assumption violations. We assume valid causal action maps to evaluate post-response regret from changes in features and true costs. Karimi et al. (2021) use structural interventions for causal recourse. Intervention data covering deployment actions can test predicted effects while historical fit cannot establish off-support effects. Bounded mechanism errors enter our guarantees as loss penalties.

We model agents as maximizing decision benefits minus effort costs, following Hardt et al. (2016). Common separable strictly convex costs yield our explicit effort-share formula and Shavit et al. (2020) provide the quadratic special case. With identified payoffs, new incentive conditions can test response predictions and optimality gaps. Our sensitivity analysis permits heterogeneous, coupled costs.

Strong monotonicity of a continuous game operator on a nonempty compact convex action set gives a unique equilibrium, with perturbation stability governed by its monotonicity modulus. Jacobian bounds can check this property within a specified model. This follows the variational-inequality approach of Parise and Ozdaglar (2019).

For our extension, let the nominal operator have modulus μ>0\mu>0, let its uniform distance from the true operator be at most δ\delta, and let the actual response have true VI dual gap at most ε\varepsilon. Its distance from the nominal equilibrium is bounded by

∣a−a0∣≤δ+δ2+4με2μ.\\|a-a_0\\|\le\frac{\delta+\sqrt{\delta^2+4\mu\varepsilon}}{2\mu}.

This displacement bound also applies to nonunique true responses satisfying the gap condition. Small μ\mu weakens the bound, and hard decisions near ties require separate analysis.

In our single-agent experiment with exact responses and 500 paired parameter settings, we perturb the effort-cost Hessian to Q=I+ρEQ=I+\rho E, where E=E⊤E=E^\top and ∣E∣2=1\\|E\\|_2=1. Counts require positive oracle-regret gain and nonnegative alignment at β=0.6\beta=0.6:

ρ\rhoActually satisfiedCertified after correction
0.03500/500500/500
0.05500/500474/500
0.10500/5000/500
0.20346/5000/500

The counts above evaluate deterministic perturbation bounds across simulated parameter settings. The pathwise regret envelope converts aligned effort into decision improvement. Shavit et al. (2020) distinguish outcome improvement from prediction accuracy; our envelope adds a decision-specific sufficient condition. Given an identified mechanism, pathwise derivative checks can falsify it. On absolutely continuous paths, integrated positive violation SS subtracts SS from the improvement bound. In the two-action example, halving the beneficial causal coefficient from 2 to 1 gave mean gain 0.07833 and mean corrected lower bound 0.04667 across 501 fixed grid points for v∈[1,1.5]v\in[1,1.5].

Mechanism coverage makes the certificate applicable to deployment, paralleling the valid model-error bounds in Petrik et al. (2016). Held-out interventions can expose misspecification, but nonrejection cannot establish coverage. With valid candidate and baseline loss-error bounds e,e0e,e_0, the corrected guarantee is Gtrue≥LCBG−e−e0G_{\mathrm{true}}\ge\mathrm{LCB}_G-e-e_0. These experiments assess conditional robustness; validity for real borrowers still requires intervention and response evidence from that population.

AuthorsOpenReview ↗

Response to Critical Weaknesses: Assumption Validity, Training with Interacting Agents, and Statistical Comparisons(2/2)

Critical Weaknesses

W:(ii) The credit experiments mainly compare different selection rules over the same 84 candidates. They therefore provide limited evidence about the proposed training algorithm and its ability to handle interacting agents.

A: the proposed training algorithm and its ability to handle interacting agents. We added a UCI-based semi-synthetic experiment testing end-to-end policy training with interacting agents. We jointly optimize the covariate scoring parameters and three action coefficients over 160 gradient updates. Each action coefficient enters both the score and its corresponding posted incentive, with a fixed action-specific incentive multiplier.

In each cohort, 20 borrowers compete for six slots. Borrowers choose actions using their expected selection probabilities under independent logistic score noise. Complete solves the coupled response and differentiates through the equilibrium during training. Price-taking uses independent responses during training. Both methods share initialization, training batches, and calibration procedures. Both are evaluated under the same full coupled response model, with each realized allocation selecting exactly the six highest scores after adding noise.

We use 128 training cohorts, 128 calibration cohorts, and 256 held-out test cohorts from one UCI data split. Each reported worst-environment metric is the minimum of its four environment means. Gains are relative to the same status-quo policy.

Worst-environment metricCompletePrice-taking
Raw-cost gain ↑0.0576150.045664
Regret gain ↑0.0112940.009011
Alignment ↑0.0762060.057086

Complete achieves higher empirical values on all three metrics in this experiment. Both methods select calibration-feasible policies. These single-split empirical results demonstrate the feasibility of training through endogenous quota responses. Formal significance testing and a final deployment certificate remain to be established for this run.

W:(iii) The gains over Causal SC-DFL are small. In Table 1, the worst-environment gain is 0.0880 versus 0.0875, and several additional results are nearly identical. Repeated data splits and a direct statistical comparison would help establish whether this difference is reliable.

A: To assess the reliability of the gain difference reported in Table 1 (0.0880 vs. 0.0875) , we added an experiment using the repeated data splits and direct statistical comparison you suggested. Specifically, we used a shared grid of 84 candidates (12 total incentive budgets × 7 allocations between two beneficial actions), four environments and causal-effect drift in the non-nominal environments.

We repeated the experiment with 10 data-split seeds. For each split, we refit the risk model, selected each method’s policy on the calibration set, and directly compared the two policies on the same 5,000 held-out test cohorts using a paired bootstrap.

Worst-environment raw-cost gainResult across 10 splits
CS-DFL, mean0.06468
Causal SC-DFL, mean0.02309
Paired difference, mean+0.04159
Splits with a positive difference10/10
Splits with a paired bootstrap 95% interval entirely above zero10/10

And the results are clear that CS-DFL achieved a higher worst-environment gain in every split, which means CS-DFL achieves better performance than Causal SC-DFL.

AuthorsOpenReview ↗

Response to Questions: Joint Effort and Regret Certification with a Nonzero Equilibrium

Questions For Authors:

A: Yes, they can. We provide a fully specified analytical example showing that the effort and regret certificates can hold simultaneously for the same predictor, response model, and downstream optimizer, with βcert=4/7>βcrit=1/9\beta_{\mathrm{cert}}=4/7>\beta_{\mathrm{crit}}=1/9.

Consider one strategic agent and one passive competitor sharing one allocation unit, with feasible allocation set S=z∈[0,1]2:z1+z2=1\mathcal{S}=\\{z\in[0,1]^2:z_1+z_2=1\\} and quadratic regularizer Ω(z)=∣z∣22/2\Omega(z)=\\|z\\|_2^2/2. For nonnegative actions a=(aD,aU)a=(a_D,a_U), we specify the true costs and predictor as

C(a)=(2−2aD,1),fθ(X(a))=(1−25aD−15aU,1).C(a)=(2-2a_D,1),\qquad f_\theta(X(a))=(1-\frac25a_D-\frac15a_U,1).

The downstream optimizer minimizes predicted allocation cost plus the quadratic regularizer, yielding

z_1(a)=\min\left\\{1,\frac12+\frac15a_D+\frac1{10}a_U\right\\}.

The agent maximizes u(a;v)=vz1(a)−∣a∣22/2u(a;v)=vz_1(a)-\\|a\\|_2^2/2, where v∈[1,3/2]v\in[1,3/2]. On the unsaturated allocation branch, the first-order conditions are

∂u∂aD=v5−aD=0,∂u∂aU=v10−aU=0.\frac{\partial u}{\partial a_D} =\frac{v}{5}-a_D=0, \qquad \frac{\partial u}{\partial a_U} =\frac{v}{10}-a_U=0.

Solving these conditions yields the unique nonzero equilibrium response

a∗(v)=(v5,v10).a^*(v)=\left(\frac{v}{5},\frac{v}{10}\right).

This response lies in the unsaturated region for every v∈[1,3/2]v\in[1,3/2] and is globally optimal because the allocation function is concave in actions and the quadratic effort cost makes the utility strictly concave.

The corresponding effort returns are RD=v/5R_D=v/5 and RU=v/10R_U=v/10. Their uniform bounds yield R‾D=1/5\underline{R}_D=1/5 and R‾U=3/20\overline{R}_U=3/20, hence

βcert=R‾DR‾D+R‾U=1/51/5+3/20=47.\beta_{\mathrm{cert}} =\frac{\underline{R}_D}{\underline{R}_D+\overline{R}_U} =\frac{1/5}{1/5+3/20} =\frac47.

Along every response path a=ta∗(v)a=t a^*(v), t∈[0,1]t\in[0,1], the cost gap g=1−2aDg=1-2a_D lies in [0.4,1][0.4,1], and the true unregularized oracle is z∗=(0,1)z^*=(0,1). Since z1≥1/2z_1\ge1/2, the beneficial-action derivative terms satisfy

(z−z∗)⊤∂DC=−2z1≤−1,C⊤∂Dz=g/5≤1/5.(z-z^*)^\top\partial_D C=-2z_1\le-1, \qquad C^\top\partial_D z=g/5\le1/5.

The corresponding terms for the other action are 00 and g/10≤1/10g/10\le1/10. Thus γD=1>δD=1/5\gamma_D=1>\delta_D=1/5 holds uniformly along the entire response path, with γU=0\gamma_U=0 and δU=1/10\delta_U=1/10. Regret is r(a)=(1−2aD)z1(a)r(a)=(1-2a_D)z_1(a) throughout these paths.

The uniform derivative bounds yield

κD=γD−δD=1−15=45,κU=γU+δU=0+110=110.\kappa_D=\gamma_D-\delta_D =1-\frac15=\frac45, \qquad \kappa_U=\gamma_U+\delta_U =0+\frac1{10}=\frac1{10}.

Therefore, the regret-improvement threshold is

βcrit=κUκD+κU=1/104/5+1/10=19.\beta_{\mathrm{crit}} =\frac{\kappa_U}{\kappa_D+\kappa_U} =\frac{1/10}{4/5+1/10} =\frac19.

Combining this with the effort certificate derived above, both certificates hold simultaneously:

βcert=47>19=βcrit.\beta_{\mathrm{cert}}=\frac{4}{7} > \frac{1}{9}=\beta_{\mathrm{crit}}.

The joint certificate guarantees a strict regret improvement:

r(a∗(v))−r(0)≤−87v700<0.r(a^*(v))-r(0)\le-\frac{87v}{700}<0.

We further verify the construction symbolically and numerically, including 12 independent optimization checks. At v=1v=1, the equilibrium response is (0.2,0.1)(0.2,0.1), and the exact regret decreases from 0.500.50 to 0.330.33.

A fixed status-quo predictor fθ0≡(1,1)f_{\theta_0}\equiv(1,1) induces zero effort and regret 1/21/2, so the calculation also establishes improvement over that baseline. Full derivations are in the appendix, “A Fully Specified Joint Incentive and Regret Certificate.”

We are eager to know whether we have addressed your questions, and we would be happy to discuss further at any time to assist with your re-evaluation.

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

This paper studies decision-focused learning when deployed rules induce strategic user responses. It proposes causal incentive certificates connecting three questions: how rules induce actions, how those actions change downstream decision quality, and whether finite data justify deployment. Under specified response and cost assumptions, an effort-share certificate translates bounds on users’ marginal returns into a lower bound on effort allocated to beneficial actions. The framework is evaluated through controlled simulations and semi-synthetic credit-allocation experiments.

Critical strengths
  • The analysis distinguishes behavioral alignment from actual decision improvement. It establishes sufficient conditions under which strategic adaptation reduces regret and characterizes a sharp effort-share threshold relative to the stated bounds. Besides, the theory considers equilibrium perturbations and directional response noise, clarifying how modeling errors and behavioral variability consume the improvement margin.
  • Separating training, policy selection, and independent testing provides a principled basis for deploying a candidate or retaining the baseline when evidence is insufficient.
Critical weaknesses
  • The central causal and behavioral mechanisms are prescribed rather than learned in the experiments. The framework assumes access to credible structural relationships or uncertainty sets, but does not establish how these can be identified from available data. Certification therefore remains conditional on their validity.
  • The empirical evidence remains semi-synthetic, and the incremental empirical benefits are not that strong, like the reported worst-environment gain is 0.0880 for CS-DFL versus 0.0875 for Causal SC-DFL.
Questions for authors
  • The implementation is narrower than the framework. The credit study searches an 84-point grid over two policy parameters and uses specified response functions. Is it possiblr to demonstrate the complete learning procedure with endogenous responses to the actual quota-constrained decision rule.
  • Can causal effects and response functions be estimated from intervention or policy-variation data, with estimation uncertainty propagated into the deployment certificate?
4 comments
AuthorsOpenReview ↗

Response to Critical Weaknesses: Prescribed Mechanisms, Conditional Certification, and Small Empirical Gains(1/2)

Critical Weaknesses

W(i): The central causal and behavioral mechanisms are prescribed rather than learned in the experiments. The framework assumes access to credible structural relationships or uncertainty sets, but does not establish how these can be identified from available data. Certification therefore remains conditional on their validity.

A: Mechanism learning is compatible with certification when its uncertainty is quantified. We have derived a learning-error extension and completed a UCI semi-synthetic experiment that estimates hidden causal and behavioral relationships and constructs a statistically calibrated mechanism set.

Let G⋆G_\star denote population loss reduction relative to the baseline and H⋆H_\star denote population alignment. Their fitted-mechanism test averages are G^,H^\widehat G,\widehat H. The bounds eL(θ)e_L(\theta) and eA(θ)e_A(\theta) control absolute population loss and alignment errors caused by mechanism estimation. The experiment below takes loss to be the mean bounded cost of selected borrowers.

Proposition (Certification with learned mechanisms). Freeze the candidate, baseline and learner outputs before independent testing. Suppose nonnegative loss/alignment error bounds hold jointly and uniformly over admissible policies, including the baseline, with probability 1−δ1-\delta. Assume conditionally i.i.d. test scores with the correct population means and sampling corrections rG,rHr_G,r_H having joint failure probability at most α\alpha. Then, with probability at least 1−δ−α1-\delta-\alpha,

G⋆(θ^;θ0)≥G^−rG−eL(θ^)−eL(θ0)G_\star(\widehat\theta;\theta_0) \ge \widehat G-r_G-e_L(\widehat\theta)-e_L(\theta_0)

H⋆(θ^)≥H^−rH−eA(θ^).H_\star(\widehat\theta) \ge \widehat H-r_H-e_A(\widehat\theta).

Alternatively, a frozen, nonempty measurable set covering the complete mechanism with probability 1−δ1-\delta supports certification through empirical worst-case bounds and sampling corrections. Each mechanism must evaluate both policies’ respective responses. Deploying only when the gain bound is positive and the alignment bound nonnegative limits the unconditional probability of unsafe deployment to δ+α\delta+\alpha.

Evidence. DirectLiNGAM (Shimizu et al., 2011) learns the graph and effects without receiving their true values or ordering, assuming linearity, acyclicity, no hidden confounding and independent non-Gaussian errors. Separate randomized probes provide simultaneous regression intervals under known Gaussian measurement noise. UCI covariates are real; discovery and calibration observations are simulated.

Records per learning/calibration taskMean loss-error boundPlug-in coverageError-corrected coverageLearned-set coverage
2560.0154722%100%100%
1,0240.0008346%100%100%
4,0960.0000650%100%100%

Coverage measures whether both lower bounds lie below true gain/alignment over 100 repetitions per row in the beneficial setting. Corrected rules target 95%; all rules include sampling corrections. Plug-in ignores mechanism error. Across both correctly specified settings, parameter-set coverage was 98% to 100% against nominal 97.5%, with no audited error-bound violations among 598 covered repetitions.

Omitting a nonlinear cost term invalidated all 100 learned-set bounds in the misspecification control. The guarantee therefore depends on valid calibration assumptions; discovery alone supplies no such guarantee. This experiment evaluates bounded selected-cost certification and establishes neither real borrower-response identification nor oracle-regret improvement. The theoretical supplement and UCI report document the proof, calibration procedure and scope.

AuthorsOpenReview ↗

Response to Critical Weaknesses: Prescribed Mechanisms, Conditional Certification, and Small Empirical Gains(2/2)

Critical Weaknesses

W(ii): The empirical evidence remains semi-synthetic, and the incremental empirical benefits are not that strong, like the reported worst-environment gain is 0.0880 for CS-DFL versus 0.0875 for Causal SC-DFL.

A: Regarding the semi-synthetic setting, we use real borrower characteristics and observed default outcomes from the UCI dataset, while specifying the causal intervention effects, effort costs, and strategic response mechanisms synthetically. These components cannot be jointly identified from the available observational data, which lack individual counterfactual outcomes and responses to alternative incentives. As discussed in our responses to Critical Weakness 1 (assumptions underlying the causal and behavioral mechanisms) and Question 2 (estimation of causal effects and response functions), these assumptions are motivated by relevant empirical evidence, while our guarantees remain conditional on the specified models.

Regarding the incremental empirical benefits, following the suggestion from reviewer Q6HL, we conducted repeated data splits and direct paired statistical comparisons to further investigate the small performance gap in Table 1. Specifically, we used a shared grid of 84 candidates across four environments and causal-effect drift in the non-nominal environments.

This repeated-split experiment changes the policy class and environments of Table 1, including an 80% attenuation of one beneficial action’s non-nominal effects chosen after exploring milder drift levels (Appendix I.4); its results concern this modified setting rather than a same-protocol replication of the original Table 1 comparison.

We repeated the experiment across 10 data splits, refitting the risk model and selecting policies on the calibration set for each split. Both methods were evaluated on the same 5,000 held-out test cohorts per split using paired bootstrap comparisons.

Worst-environment raw-cost gainAcross 10 splits
CS-DFL, mean0.06468
Causal SC-DFL, mean0.02309
Mean paired difference+0.04159
Splits with positive differences10/10
Splits with positive paired bootstrap 95% lower bounds10/10

These results show that CS-DFL consistently achieves higher worst-environment gains, demonstrating a substantially larger empirical advantage.

AuthorsOpenReview ↗

Response to Questions: Learning beyond the 84-Point Grid and Propagating Estimation Uncertainty into Certification(1/2)

Questions For Authors

Q(i): The implementation is narrower than the framework. The credit study searches an 84-point grid over two policy parameters and uses specified response functions. Is it possiblr to demonstrate the complete learning procedure with endogenous responses to the actual quota-constrained decision rule.

A: Motivation for the restricted implementation. The original 84-point grid search is motivated by computational tractability and controlled comparisons of policy-selection criteria. Related decision-aware model-selection approaches have also been studied in robust optimization (e.g., Bao et al., 2025).

Consistency with the general framework. The original implementation is a natural special case of our general framework in two respects:

  • Restricted policy class: The finite candidate set satisfies Θgrid⊆Θ\Theta_{\mathrm{grid}}\subseteq\Theta, with ∣Θgrid∣=84|\Theta_{\mathrm{grid}}|=84.

  • Separable strategic responses: The deterministic independent responses correspond to a degenerate response kernel,

    Rθ(da∣x)=∏i=1Nδai∗(θ,xi)(dai).\mathcal R_\theta(da\mid x) =\prod_{i=1}^{N}\delta_{a_i^*(\theta,x_i)}(da_i).

Thus, the original implementation is consistent with the general framework, which also accommodates coupled response equilibria and gradient-based policy optimization.

End-to-end learning with endogenous quota-constrained responses.

We added a UCI-based semi-synthetic experiment testing end-to-end policy training with interacting agents. We jointly optimize the covariate scoring parameters and three action coefficients over 160 gradient updates. Each action coefficient enters both the score and its corresponding posted incentive, with a fixed action-specific incentive multiplier.

In each cohort, 20 borrowers compete for six slots. Borrowers choose actions using their expected selection probabilities under independent logistic score noise. Complete solves the coupled response and differentiates through the equilibrium during training. Price-taking uses independent responses during training. Both methods share initialization, training batches, and calibration procedures. Both are evaluated under the same full coupled response model, with each realized allocation selecting exactly the six highest scores after adding noise.

We use 128 training cohorts, 128 calibration cohorts, and 256 held-out test cohorts from one UCI data split. Each reported worst-environment metric is the minimum of its four environment means. Gains are relative to the same status-quo policy.

Worst-environment metricCompletePrice-taking
Raw-cost gain ↑0.0576150.045664
Regret gain ↑0.0112940.009011
Alignment ↑0.0762060.057086

Complete achieves higher empirical values on all three metrics in this experiment. Both methods select calibration-feasible policies. These single-split empirical results demonstrate the feasibility of training through endogenous quota responses. Formal significance testing and a final deployment certificate remain to be established for this run.

AuthorsOpenReview ↗

Response to Questions: Learning beyond the 84-Point Grid and Propagating Estimation Uncertainty into Certification(2/2)

Questions for Authors

Q: (ii) Can causal effects and response functions be estimated from intervention or policy-variation data, with estimation uncertainty propagated into the deployment certificate?

Yes, both can be estimated under an appropriate intervention design. In a UCI-based semi-synthetic experiment, we compare four standard estimators: difference-in-means, inverse probability weighting (IPW), stratified standardization, and augmented IPW (AIPW) (Neyman, 1923/1990; Hernán and Robins, 2020; Bang and Robins, 2005).

We use two separate simulated pilots. The first varies the action dose and measures the downstream outcome to estimate the causal coefficient κ\kappa. The second varies the incentive rule and observes borrowers’ actions to estimate vv, the parameter governing the response function. These pilots estimate two scalar parameters within the specified causal and response models, using known assignment probabilities.

The table reports results for the beneficial-policy setting, with 2,048 estimation observations per pilot and 400 replications. AIPW additionally uses 256 independent observations per pilot to fit its outcome regressions. Final evaluation uses 2,048 independent test draws. Regret change is defined as candidate policy minus baseline policy. The upper bounds are averaged across replications, and certification is assessed separately in each replication.

EstimatorMean regret upper bound (uniform assignment)Mean regret upper bound (covariate-dependent assignment)Certification rate (covariate-dependent assignment)
Difference-in-means−0.07581——
IPW−0.05348−0.0103090.75%
Stratified standardization−0.026950.0023752.50%
AIPW−0.11544−0.07450100%

A dash indicates that the estimator was not evaluated under that assignment design.

For each estimator, we propagate uncertainty in both κ^\widehat\kappa and v^\widehat v by maximizing regret change over their joint confidence set and then adding a held-out sampling correction. Under covariate-dependent assignment, AIPW and stratified standardization had similar RMSE for κ\kappa (0.13224 and 0.13231), yet their final regret bounds differed substantially. For deployment, the relevant comparison is the bound after both sources of estimation uncertainty have been propagated.

We construct simultaneous confidence intervals across both parameters and all candidate estimators, allocating a total failure probability of 0.025 to mechanism estimation. This simultaneous coverage permits estimator selection on the pilot data. Independent testing receives an additional failure budget of 0.025, giving overall validity of at least 95% under the stated assumptions. Among estimators with a positive alignment lower bound, we select the one with the lowest pilot regret upper bound. We then use independent test data to compute the final regret upper bound. The policy is certified if that upper bound is negative and the alignment lower bound is positive. Under covariate-dependent assignment, the pilot selected AIPW in all 400 beneficial-policy replications, and all 400 passed the independent test. Across the simulated harmful-policy settings, we observed no false certifications.

We are eager to know whether we have addressed your questions, and we would be happy to discuss further at any time to assist with your re-evaluation.

Reviewer_n2RBOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies learning with strategic agents. Agents respond to a deployed prediction rule by choosing costly actions that can change both observable features and true decision costs. This strategic behavior makes the evaluation of the predictor policy-dependent. The objective is to improve worst-case regret relative to a status quo while encouraging beneficial effort.

The paper proposes a certificate that guarantees regret reduction when the minimum benefit from beneficial effort exceeds the maximum harm from other effort, accounting for uncertainty. Its proposed learning algorithm trains predictors using simulated strategic responses and downstream optimization, then uses independent testing to decide whether to deploy a selected policy. The experiments include controlled simulations and semi-synthetic credit studies using real covariates with assumed behavioral and causal relationships; the credit studies report reductions in raw decision cost rather than demonstrating the central regret-reduction guarantee.

Critical strengths

The paper studies when predictions can elicit beneficial strategic behavior from agents. It uses connections between strategic behavior, causal changes in true costs, and the regret for decision making. The certificate provides a useful perspective for thinking about policy from predictions. I think many mathematical tools are established, but the integration still feels useful here.

Critical weaknesses

The main weakness is that the paper assumes knowledge of the cost function. This assumption seems important for the result, but I'm not sure how realistic the assumption is. The paper would benefit from more discussion around this. Also, it would be helpful if the simulations also tested the proposed training algorithm.

Questions for authors

Can the authors elaborate on how to obtain a cost estimate from a heterogenous population? Will it be without loss to consider a population average cost function?

2 comments
AuthorsOpenReview ↗

Response to Critical Weaknesses: Realism of Cost Assumptions and Evaluation of the Training Algorithm

Critical Weaknesses

W(i): The main weakness is that the paper assumes knowledge of the cost function. This assumption seems important for the result, but I'm not sure how realistic the assumption is. The paper would benefit from more discussion around this.

A: Our framework distinguishes two cost functions, agents’ effort costs function and downstream decision cost function.

Effort costs.
Theoretical scope of the effort-cost model. Our uniform effort-share theorem applies to separable costs g(a)=∑mψ(am)g(a)=\sum_m\psi(a_m) with a common continuously differentiable, strictly convex ψ\psi satisfying ψ′(0)=0\psi'(0)=0 and ψ′(t)→∞\psi'(t)\to\infty. At a selected optimum, each action satisfies am∗=(ψ′)−1([Rm(a∗)]+)a_m^*=(\psi')^{-1}([R_m(a^*)]_+), where Rm(a∗)R_m(a^*) is its downstream-aware marginal return. The inverse marginal-cost map converts a lower bound on beneficial returns and an upper bound on undesirable returns into a lower bound on the beneficial effort share. For example, ψ(t)=tp/p\psi(t)=t^p/p for p>1p>1 gives am∗=[Rm(a∗)]+1/(p−1)a_m^*=[R_m(a^*)]_+^{1/(p-1)}, with the quadratic case recovered at p=2p=2.

In semi-synthetic experiments, we model the cost of borrower ii’s actions in environment η\eta as

gi,η(ai)=qiqη2(a1i2+a2i2+aUi2).g_{i,\eta}(a_i)=\frac{q_iq_\eta}{2}\left(a_{1i}^2+a_{2i}^2+a_{Ui}^2\right).

This has the same quadratic form as a special case of the norm-based cost family in Rosenfeld and Rosenfeld (2024), obtained by choosing the Euclidean norm and ϕ(t)=t2/2\phi(t)=t^2/2. Our appendix considers the more general action cost g(a)=12a⊤Qag(a)=\tfrac12 a^\top Qa. Setting Q=IQ=I recovers the action cost 12∥a∥22\tfrac12\|a\|_2^2 used by Shavit, Edelman, and Axelrod (2020).

Also, the setting still has the reality meanings. The quadratic form captures increasing marginal effort costs; To introduce borrower heterogeneity, we specify qiq_i as an increasing function of the observed bill amount. This modeling choice is motivated by the possibility that additional repayment is more burdensome when liquidity is limited; The multiplier qηq_\eta takes values 1.001.00, 1.051.05, 1.251.25, and 1.001.00 across the four environments, allowing us to evaluate the policy under different overall effort-cost levels.

Downstream default costs. We define

Ci,η(ai)=clip⁡[0,1](pi−g1ηa1i−g2ηa2i+gUηaUi),C_{i,\eta}(a_i)=\operatorname{clip}_{[0,1]} \left(p_i-g_{1\eta}a_{1i}-g_{2\eta}a_{2i}+g_{U\eta}a_{Ui}\right),

where pip_i is the baseline default risk fitted using the UCI training data. The coefficients g1ηg_{1\eta}, g2ηg_{2\eta}, and gUηg_{U\eta} specify how the two repayment actions and the debt-reshuffling action affect default risk. Like the effort-cost multipliers qηq_\eta, these coefficients are prespecified for each environment.

In practice, recorded expenditures and responses to known incentives can inform effort costs, while randomized interventions or policy variation with appropriate identification conditions can identify action effects on downstream costs.

W(ii): Also, it would be helpful if the simulations also tested the proposed training algorithm.

A: We added a UCI-based semi-synthetic experiment testing end-to-end policy training with interacting agents. We jointly optimize the covariate scoring parameters and three action coefficients over 160 gradient updates. Each action coefficient enters both the score and its corresponding posted incentive, with a fixed action-specific incentive multiplier.

In each cohort, 20 borrowers compete for six slots. Borrowers choose actions using their expected selection probabilities under independent logistic score noise. Complete solves the coupled response and differentiates through the equilibrium during training. Price-taking uses independent responses during training. Both methods share initialization, training batches, and calibration procedures. Both are evaluated under the same full coupled response model, with each realized allocation selecting exactly the six highest scores after adding noise.

We use 128 training cohorts, 128 calibration cohorts, and 256 held-out test cohorts from one UCI data split. Each reported worst-environment metric is the minimum of its four environment means. Gains are relative to the same status-quo policy.

Worst-environment metricCompletePrice-taking
Raw-cost gain ↑0.0576150.045664
Regret gain ↑0.0112940.009011
Alignment ↑0.0762060.057086

Complete achieves higher empirical values on all three metrics in this experiment. Both methods select calibration-feasible policies. These single-split empirical results demonstrate the feasibility of training through endogenous quota responses. Formal significance testing and a final deployment certificate remain to be established for this run.

AuthorsOpenReview ↗

Response to Questions: Estimating Heterogeneous Costs and the Validity of Population Averaging

Questions For Authors:

Q:(i) Can the authors elaborate on how to obtain a cost estimate from a heterogenous population?

A: In our original UCI semi-synthetic experiment, borrower-specific effort costs follow the model above, using observed bill amounts and a prespecified mapping.

To introduce borrower heterogeneity, we let qiq_i increase with the observed bill amount:

\right),$$ where $\sigma$ is the logistic function and $m$ and $s$ are the median and standard deviation of the transformed bill amounts in the training set. We added a semi-synthetic pilot to illustrate how borrower-specific effort costs can be estimated from incentive-response data. In the reference environment with $q_\eta=1$, we simulate repeated observations of each borrower’s response to two known marginal incentives, $R_L$ and $R_H$. Under the quadratic cost model, an interior response satisfies $a_i(R)=R/q_i$. We estimate the response slope by **within-borrower difference-in-means** and take its reciprocal to obtain the cost coefficient: $$\widehat s_i=\frac{\bar a_{i,H}-\bar a_{i,L}}{R_H-R_L}, \qquad \widehat q_i=\frac{1}{\widehat s_i}.$$ The table reports RMSE across 100 replications with 9,000 borrowers. Increasing the number of observations per borrower at each incentive level from 1 to 16 reduced individual cost RMSE from 0.0526 to 0.0131. | Observations per borrower at each incentive level | Response slope RMSE | Individual cost RMSE | | ------------------------------------------------: | ------------------: | -------------------: | | 1 | 0.0408 | 0.0526 | | 4 | 0.0204 | 0.0261 | | 16 | 0.0102 | 0.0131 | >Q: (ii) Will it be without loss to consider a population average cost function? A: **In general, averaging costs changes the induced responses.** Consider two equally prevalent borrower types with $q_i=1$ and $q_i=3$. Under the same marginal incentive $R>0$ and interior responses, averaging their individual responses gives $(R+R/3)/2=2R/3$. Using the population-average cost coefficient $q=2$ gives $R/2$. **Extension to heterogeneous separable costs.** The same optimality argument gives a conservative certificate when marginal costs differ across borrowers, environments, and action coordinates. On the nonnegative orthant, suppose each coordinate cost satisfies $\psi_m'(0)=0$ and $0<\underline{k}_m\le\psi_m''(t)\le\overline{k}_m$ over the relevant action range, uniformly across borrowers and environments. Then lower return bounds on beneficial actions give $a_j^*\ge[\underline R_j]_+/\overline{k}_j$, and upper return bounds on undesirable actions give $a_k^*\le[\overline R_k]_+/\underline{k}_k$. Writing $L_D$ and $U_U$ for the sums of these respective bounds, the beneficial effort share is at least $L_D/(L_D+U_U)$ when $L_D>0$. Thus uniform curvature bounds suffice for this certificate. If the resulting share bound exceeds $\beta_{\mathrm{crit}}$ and the regret-envelope conditions hold, the same regret-improvement result applies. In the same pilot, using 16 observations per borrower at each incentive level, we evaluated two fixed policies on the same 5,000 cohorts, either retaining each borrower’s $\widehat q_i$ or using the arithmetic mean of those same estimates for everyone. The table reports mean absolute errors across 100 replications relative to evaluation with the known individual costs. **Retaining individual estimates reduced both errors for both policies in this semi-synthetic setting.** | Fixed policy | Cost used in evaluation | Worst-environment raw-cost gain error | Worst-environment alignment error | | -------------- | --------------------------- | ------------------------------------: | --------------------------------: | | CS-DFL | Individual estimates | **0.0000072** | **0.0000257** | | CS-DFL | Population-average estimate | 0.0001907 | 0.0005477 | | Responsive DFL | Individual estimates | **0.0000677** | **0.0000191** | | Responsive DFL | Population-average estimate | 0.0006796 | 0.0004073 |

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to the Area Chair and Reviewers

A Kind General Response to All Reviewers and Dear AC

We sincerely thank the reviewers for their generous praise (provides a useful perspective[n2RB], establishes sufficient conditions [uEZE], and connects strategic behavior, causal effects, and decision quality [Q6HL]). We also deeply appreciate the valuable suggestions for improvement. In the revised manuscript, we have specifically addressed the following points (also please refer to our Supplementary Material in our submission, containing the complete appendix mentioned):

Summary of Clarifications and Additions

  1. Clarification of Framework Assumptions and Estimation

    • Clarifying how effort cost functions are specified and how bounds on cost curvature and marginal returns yield a conservative effort-share certificate under heterogeneous separable costs (n2RB–W1,Q2).
    • Explaining the assumptions on causal and behavioral mechanisms, and how bounded deviations affect the guarantees (Q6HL–W1, uEZE–W1).
    • Explaining how to estimate causal effects and response functions and propagate their uncertainty into the deployment certificate (uEZE–W1,Q2).
  2. Clarification of the Learning Algorithm and Its Guarantees

    • Explaining how the original 84-candidate implementation fits within the general learning framework (Q6HL–W2, uEZE–Q1, n2RB–W2).
    • Providing a fully specified analytical example in which the effort and regret certificates jointly establish strict improvement for the same predictor, with βcert=4/7>1/9=βcrit\beta_{\mathrm{cert}}=4/7>1/9=\beta_{\mathrm{crit}} and γD>δD\gamma_D>\delta_D along the entire response path (Q6HL–Q1).
  3. End-to-End Training with Strategic Interactions

    • Implementing end-to-end policy training that computes the coupled equilibrium of borrowers competing (Q6HL–W2, uEZE–Q1, n2RB–W2).
  4. Addition of Empirical Evidence

    • Repeated splits and direct statistical comparison. Comparing CS-DFL and Causal SC-DFL using paired bootstrap across ten UCI splits(Q6HL–W3, uEZE–W2).
    • Sensitivity to assumption violations. Testing how perturbations to effort costs and causal effects affect the certified improvement margin in controlled examples (Q6HL–W1).
    • Certification with estimated mechanisms. Comparing plug-in and uncertainty-aware certificates in a semi-synthetic UCI study (Q6HL–W1, uEZE–W1,W2).
    • Estimation of causal effects and response functions. Using four estimators to estimate these functions (uEZE–Q2).
    • Estimation of borrower-specific effort costs. Using a simulated incentive-response pilot to estimate individual effort costs and comparing the resulting policy evaluation(n2RB–Q1,Q2).

In our revised version, we have fully prepared and will comprehensively and rigorously implement all reviewers' suggestions. We are eager to know if the esteemed reviewers have any further questions or concerns. We will promptly address them comprehensively and clearly. It is our great pleasure to contribute together to the ICLR community!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.