{"id":"af6adf98-c5f5-428e-abef-4dda763ee594","arxiv_id":"2412.01065","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Lookahead counterfactual fairness requires that an individual's future status, not just the current decision, is equal in factual and counterfactual worlds; the paper gives a predictor that achieves this under linear causal models and gradient-based strategic responses.","lead":"The paper introduces a fairness test for the future, not just the present: it asks whether people of different groups end up with the same later outcomes after they adapt to an algorithm's decisions. It then shows a specific predictor that can pass this test when the causal model and human responses have a particular mathematical form.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 5.1 overclaims LCF for multi-valued A: the proof's cancellation works only for the single counterfactual group ˇa used to construct ˇY, not for every ˇa required by Definition 4.1.","rationale":"Reader's weakest assumption (exact gradient response with known η and linear model) is a legitimate limitation about external validity: if the response model is misspecified, LCF is not guaranteed, and the paper's Theorem 4.1 already shows CF predictors can violate LCF. However, the central theorem is a conditional statement; under its stated assumptions it is internally sound for binary A. My concern is more direct: the theorem as written is false for multi-valued A because Definition 4.1 quantifies over all counterfactual groups while the proof only establishes equality for the single ˇa baked into ˇY. The algebra in Appendix A.1 indeed yields exact cancellation for that ˇa; for a second ˇa, the residual is ˇy2 - ˇy1. This is a formal gap, not a question of modeling realism. The paper's experiments all use binary sensitive attributes, so the empirical demonstrations remain valid, but the theoretical contribution needs an explicit binary-A assumption or a multi-valued generalization (Appendix A.2 gives only the Relaxed-LCF version). The proof also contains a sign slip in Theorem 5.2's convexity argument and a T-notation inconsistency in A.1, but those are repairable. Verdict remains conditional acceptance with the added requirement to fix Theorem 5.1's scope.","tokens_in":27264,"tokens_out":17014,"duration_ms":135479,"concrete_test":"Run the synthetic experiment of Section 7.1 with A∈{0,1,2} (e.g., α=β=w=γ=1, η=1), keeping the predictor built from ˇY = Y_{A←1}. For any fixed u, compute y' and ˇy' for counterfactual ˇa=2 using the response model in Eq. 5. The difference will equal ˇy2 - ˇy1 = 1 ≠ 0, demonstrating that Theorem 5.1's LCF guarantee fails for the second counterfactual group. Equivalently, add a third group to the synthetic data and measure AFCE for all pairs; the pair (a=0, ˇa=2) will have AFCE > 0 even at p1 = T/2.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Definition 4.1 requires equality for all ˇa ∈ A, but Theorem 5.1 constructs a single predictor g(ˇY,U) with ˇY = Y_{A←ˇa1}(U) for one fixed ˇa1. The appendix proof shows y' = ˇy' for that ˇa1 only. For any second counterfactual group ˇa2, using the same g, the response in the ˇa2-world uses g(y,u) while the factual response uses g(ˇy1,u). Subtracting the future-status equations (as in Eq. 17) leaves ˇy2' - y' = (ˇy2 - y) + η(||w⊙α||^2+γ^2)[2p1(y - ˇy1)] = ˇy2 - ˇy1 = w^T β (ˇa2 - ˇa1), which is nonzero unless w^Tβ(ˇa2-ˇa1)=0. Thus LCF is violated for ˇa2. Appendix A.2 only repairs the non-binary case for Relaxed LCF via an average construction, not for the perfect-LCF claim of Theorem 5.1. The theorem's statement and Algorithm 1 therefore need an explicit binary-A restriction or a multi-valued construction. The synthetic and law-school experiments use binary A (A∈{0,1} and Sex), so the demonstrated results are unaffected, but the central theorem as stated is false for |A|>2.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces lookahead counterfactual fairness (LCF), a fairness notion requiring that an individual's future status Y' (after the individual strategically responds to an ML prediction Ŷ) has the same distribution in the factual and counterfactual worlds. Under a linear structural causal model X = α⊙U_X + βA, Y = w^T X + γU_Y and a gradient-ascent response U' = U + η∇_U Ŷ, the paper's central result (Theorem 5.1) constructs a predictor g(Ȳ,U) = p1 Ȳ^2 + p2 Ȳ + p3 + h(U), with p1 = 1/(2η(||w⊙α||_2^2 + γ^2)), which it claims satisfies perfect LCF. The paper also proposes Relaxed LCF (Definition 5.2) with corresponding constructions (Theorems 5.2, 5.3, 5.4), a path-dependent extension (Section 6), and experiments on synthetic and real data showing that the proposed predictor reduces the average future causal effect while maintaining competitive MSE.","tokens_in":27642,"tokens_out":5613,"duration_ms":46623,"significance":"The notion of evaluating counterfactual fairness on downstream status after strategic response is a meaningful and relatively unexplored problem, and the paper provides a constructive, closed-form predictor for a tractable linear model, together with a useful negative result (Theorem 4.1) showing that standard CF predictors can violate LCF. The paper also ships code and validates the method empirically on synthetic and real data. However, the main theorem's claim of perfect LCF is too strong as stated, and a key proof in the appendix contains a sign error; these issues do not destroy the core idea but require a revision to make the scope and proofs precise.","major_comments":[{"comment":"Theorem 5.1 is stated for arbitrary A, but the proof only establishes the cancellation for the single counterfactual group ˇa used to define ˇY. Definition 4.1 requires the equality for every pair (a, ˇa). For a second counterfactual group ˇa2, the same derivation gives ˇy2' - y' = w^T β (ˇa2 - ˇa1), which is generally nonzero. Thus the theorem is false as stated when |A| > 2. Appendix A.2 repairs the non-binary case only for Relaxed LCF through an averaging construction, not for the perfect-LCF claim. The paper should either explicitly restrict Theorem 5.1 to binary A or provide a multi-valued construction; the experiments, which use binary A, are unaffected but the central claim needs a scope correction.","section":"Section 5, Theorem 5.1; Appendix A.1 and A.2"},{"comment":"The proof states that for a strictly convex g, (ˇy - y)(∂g(y)/∂y - ∂g(ˇy)/∂ˇy) > 0. This sign is wrong: for a strictly convex function the derivative is increasing, so the inequality is < 0 when ˇy ≠ y. The main-text proof sketch has the correct sign. The contraction argument in Theorem 5.2 would still work with the corrected sign, but the appendix proof must be fixed.","section":"Appendix A.1, proof of Theorem 5.2"},{"comment":"The proof of Theorem 5.3 is difficult to follow and contains apparent inconsistencies: in Case 1 the conclusion is written as |y' - ˇy'| < |y - y'| rather than |y - ˇy|, and the use of the M-Lipschitz condition in Case 3 is not fully justified in the transition from the bound on |ϕ3 - ϕ4| to the final comparison. Since this theorem is a secondary result, it is less critical than the Theorem 5.1 issue, but the proof should be made rigorous and self-consistent.","section":"Section 5, Theorem 5.3; Appendix A.3"}],"minor_comments":[{"comment":"The symbol T is defined inconsistently: Theorem 5.1 sets T := 1/(η(||w⊙α||_2^2 + γ^2)), while Appendix A.1 sets T = η(||w⊙α||_2^2 + γ^2). This makes the proof harder to check and should be unified.","section":"Section 5, Theorem 5.1 vs. Appendix A.1"},{"comment":"The response is written as u'_X = u_X + ∇u_X g, with the step size η omitted; Eq. (13) then reintroduces η. Add the missing η in the response definition for consistency.","section":"Appendix A.1, proof of Theorem 5.1"},{"comment":"The line \"|y' - ˇy'| = |ˇy - y + (2/T)p1(y - ˇy)| = 0\" should have a strict inequality \"< |ˇy - y|\" rather than equality to zero; as written it incorrectly suggests perfect cancellation for any p1 ∈ (0, T).","section":"Appendix A.1, proof of Corollary 5.1"},{"comment":"The reported MSE for the proposed method (5.298 ± 1.704) is substantially worse than the baselines (UF: 0.012, CF: 0.329); the text says the method \"achieves a large improvement in LCF\" but does not comment on this accuracy cost, which should be discussed.","section":"Appendix D, Table 4"}],"recommendation":"major_revision","confidential_remarks":"The paper has a central theorem that is false as stated for multi-valued protected attributes, but the fix (restricting to binary A) is local and does not affect the empirical demonstrations. The main constructive idea is sound for the binary case. The appendix has several proof-quality issues (sign error, notational inconsistency) that should be corrected before publication. This is a solid but not flawless contribution to the fairness-in-strategic-settings literature; after revision it could be a worthwhile addition."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the core idea—counterfactual fairness of the future status after strategic response, rather than of the current decision—is new relative to Hu & Zhang and Tang et al., and the constructive predictor in Theorem 5.1 is a real result for the binary-A linear case. The cancellation is a derivation, not a fit, and the experiments correctly confirm it. Second, the central theorem as stated is false for |A|>2. The proof cancels only for the single counterfactual group used to build ˇY; for a second group ˇa2, the future-status gap is wᵀβ(ˇa2−ˇa1), generically nonzero. The appendix's average construction repairs only the relaxed version, not perfect LCF. This is a genuine theorem-statement error, though the experiments (binary A and Sex) are unaffected.\n\nCredit where due: the paper clearly distinguishes LCF from long-term fairness of decisions and from parity-based long-term fairness, and the related-work placement is honest. Algorithm 1 is straightforward and reproducible. The authors also state Theorem 4.1, which correctly shows that ordinary CF can violate LCF—that is a useful observation.\n\nSoft spots, in proportion. The multi-A gap is the main one. The appendix also has a sign error in the Theorem 5.2 proof (the convexity inequality is written with the wrong direction), and Theorem 5.4's condition a1a2>0 is strange for binary attributes and deserves clarification. The empirical validation is entirely model-internal: the same SCM generates the data and defines the response, so the reported AFCE=0 is a check of the construction rather than evidence about real strategic behavior. The assumption of a known common step size η and exact gradient ascent in exogenous-variable space is narrow, but the paper is upfront about it. None of these are fatal for the binary linear setting that the experiments cover; they do mean the paper's theoretical scope is more limited than the theorem statements claim.\n\nBottom line: this is a paper for researchers in algorithmic fairness and strategic classification who want a concrete, provable first step toward fairness of future outcomes. It deserves a serious referee. If I were refereeing, I would ask for either an explicit binary-A restriction in Theorem 5.1 or a genuinely multi-valued construction, and a fix to the appendix sign error. The notion and the binary construction are worth citing.","headline":"The LCF notion and the binary-A construction are genuinely new and the algebra checks out, but Theorem 5.1 overclaims for multi-valued A and the supporting proofs have warts that need fixing.","tokens_in":28132,"tokens_out":7168,"would_cite":true,"duration_ms":59049,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Under a linear causal model with gradient-ascent responses, a predictor that includes a quadratic term in the counterfactual label equalizes the future status $Y'$ between factual and counterfactual worlds.","keywords":["counterfactual fairness","lookahead fairness","strategic classification","causal fairness","downstream effects","future status","path-dependent fairness"],"falsifier":"Estimate the causal parameters and the response step size $\\eta$ from data, deploy the Algorithm 1 predictor with $p_1 = T/2$ on held-out subjects, and compute the average future causal effect (AFCE). The theorem predicts AFCE = 0; any held-out AFCE that is statistically nonzero, for example outside the sampling noise of the paper's synthetic experiment, would falsify the claim, as would a version of the simulation with heterogeneous individual step sizes $\\eta_i$.","tokens_in":27043,"feed_emoji":"⚖️","tokens_out":7926,"duration_ms":62195,"temperature":0.7,"pith_summary":"The paper argues that counterfactual fairness, which equalizes the prediction an individual receives, does not guarantee fairness in what happens to that individual afterward, because people adapt strategically to the model. It introduces lookahead counterfactual fairness (LCF), which requires the individual's future status $Y'$ to have the same distribution in the factual and counterfactual worlds. The main constructive result is Theorem 5.1: under a linear causal model and a gradient-ascent response $U' = U + \\eta \\nabla_U \\hat{Y}$, the predictor $g(\\check{Y},U) = p_1 \\check{Y}^2 + p_2 \\check{Y} + p_3 + h(U)$ with $p_1 = 1/(2\\eta(\\|w \\odot \\alpha\\|_2^2 + \\gamma^2))$ satisfies LCF exactly. If correct, this is the first algorithm that trains for fairness of the downstream status rather than of the decision itself, and it comes with a relaxed version that trades some fairness for accuracy. The paper also extends the construction to path-dependent unfairness and to a class of nonlinear bijective causal models.","feed_headline":"One quadratic term makes future status counterfactually fair","feed_subtitle":"Strategic games break prediction fairness; this predictor equalizes what happens to people, not just the score.","key_machinery":"The load-bearing object is the predictor $g(\\check{Y},U) = p_1 \\check{Y}^2 + p_2 \\check{Y} + p_3 + h(U)$, a quadratic in the counterfactual label $\\check{Y}$ plus an arbitrary function of the exogenous variables. Its derivative with respect to $\\check{Y}$ is $2p_1 \\check{Y} + p_2$, so the response-induced change in future status, which is proportional to $\\eta(\\|w \\odot \\alpha\\|_2^2 + \\gamma^2)(\\partial g/\\partial y - \\partial g/\\partial \\check{y})$, cancels the original gap $\\check{y} - y$ exactly when $2p_1 = 1/(\\eta(\\|w \\odot \\alpha\\|_2^2 + \\gamma^2))$. The counterfactual random variable $\\check{Y} = Y_{A \\leftarrow \\check{a}}(U)$, obtained by abduction--action--prediction, carries the counterfactual information, and the response model $U' = U + \\eta \\nabla_U \\hat{Y}$ is the strategic adaptation that would otherwise create the downstream disparity.","core_discovery":"The paper's central claim is that LCF is achievable in closed form. In the linear causal model $X = \\alpha \\odot U_X + \\beta A$, $Y = w^T X + \\gamma U_Y$, with individuals responding by $U'_i = U_i + \\eta \\nabla_{U_i} \\hat{Y}$, the predictor $g(\\check{Y},U) = p_1 \\check{Y}^2 + p_2 \\check{Y} + p_3 + h(U)$, where $\\check{Y}$ is the counterfactual random variable for $Y$, satisfies LCF when $p_1 = T/2$ with $T = 1/(\\eta(\\|w \\odot \\alpha\\|_2^2 + \\gamma^2))$; $p_2$, $p_3$, and $h$ are free to be trained. The mechanism is a cancellation: the future-status difference becomes $|\\check{y}' - y'| = |\\check{y} - y + (1/T)(\\partial g/\\partial y - \\partial g/\\partial \\check{y})|$, and with $\\partial g/\\partial y = T y$ the difference is exactly zero for every posterior draw of $U$, so the factual and counterfactual future statuses are equalized. With $p_1 \\in (0,T)$, LCF relaxes to the guarantee that the factual--counterfactual gap strictly shrinks. The paper further proves an analogous path-dependent version and a relaxed version for certain nonlinear bijective causal models.","pith_inferences":["Inference: The exact cancellation depends on a single known $\\eta$ shared by all individuals; if response step sizes vary, the theorem suggests a robustified predictor would need to estimate or bound $\\eta$, and the relaxed version might survive under such bounds.","Inference: The same quadratic-in-counterfactual construction could be applied to other dynamic fairness targets, such as equalizing future qualification trajectories over multiple rounds, by iterating the response update and re-deriving the tuning constant.","Inference: If individuals instead manipulate observed features $X$ directly rather than the exogenous variables $U$, the gradient-ascent model changes and the specific $p_1$ formula would need a different derivation; the paper's own Theorem 4.1 already warns that a counterfactually fair predictor can leave the future-status gap unchanged or worse in such settings.","Inference: A testable extension is to measure $\\eta$ from data, for example by observing feature changes after a deployed predictor, and to check whether the AFCE-minimizing $p_1$ tracks the theoretical value across different populations."],"forward_implications":["If Theorem 5.1 holds, an organization that knows the linear causal model and the response step size $\\eta$ can train a predictor achieving an average future causal effect of zero on future status, as the synthetic and law-school experiments illustrate.","The free parameters $p_2$, $p_3$, and $h$ can be fit to data, so LCF is compatible with optimizing prediction accuracy; in the synthetic experiment the LCF predictor reaches an MSE close to the unfair baseline while the counterfactual-fair baseline alone has much larger MSE.","With $p_1 \\in (0,T)$, the same architecture provides a tunable accuracy--fairness trade-off: smaller $p_1$ gives better MSE and weaker fairness, while $p_1 \\to T$ gives the full LCF guarantee.","Theorem 6.1 extends the construction to path-dependent LCF, allowing only specified unfair causal paths to be equalized, and Theorem 5.3 transfers the relaxed guarantee to a class of nonlinear bijective causal models with monotone concave link functions."],"supporting_citations":[{"why":"Defines counterfactual fairness and the abduction--action--prediction procedure used to compute $\\check{Y}$ and estimate the conditional distribution of $U$.","marker":"Kusner et al. (2017)"},{"why":"Supplies the gradient-based strategic response model $U' = U + \\eta \\nabla_U \\hat{Y}$ that the paper adopts.","marker":"Rosenfeld et al. (2020)"},{"why":"Establishes the strategic classification setting where individuals move features to increase favorable predictions.","marker":"Hardt et al. (2016a)"},{"why":"Motivates LCF by showing that fair lending decisions can harm future credit scores, the downstream disparity LCF targets.","marker":"Liu et al. (2018)"},{"why":"The most related sequential path-specific counterfactual fairness work; the paper positions LCF as focusing on future status instead of decisions.","marker":"Hu & Zhang (2022)"},{"why":"Prior work on downstream effects of counterfactually fair predictions with parity-based goals, which LCF extends to individual counterfactual fairness of status.","marker":"Tang et al. (2023)"},{"why":"Justifies modeling strategic classification as a causal problem with individual responses in exogenous-variable space.","marker":"Miller et al. (2020)"},{"why":"Provides the law-school dataset used for real-data validation of the LCF algorithm.","marker":"Wightman (1998)"}],"fun_headline_variants":["Fairness beyond scores: equalize future status","New LCF measure targets post-prediction outcomes","Quadratic term yields lookahead counterfactual fairness","Strategic response? Fairness must look ahead","Closed form makes future outcomes counterfactually fair"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The exact guarantee depends on every individual responding with the same known gradient-ascent step size $\\eta$ in the exogenous-variable space and on the structural equations being exactly the linear form in Theorem 5.1 with known parameters; if responses vary, act on observed features instead, or the causal functions are nonlinear, LCF is no longer guaranteed, and the paper's own Theorem 4.1 shows a counterfactually fair predictor can then leave the future-status gap unchanged or worse.","fun_headline_variants_meta":{"raw":{"variants":["Fairness beyond scores: equalize future status","New LCF measure targets post-prediction outcomes","Quadratic term yields lookahead counterfactual fairness","Strategic response? Fairness must look ahead","Closed form makes future outcomes counterfactually fair"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00049,"raw_usage":{"total_tokens":2473,"prompt_tokens":1073,"completion_tokens":1400,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":689,"completion_tokens_details":{"reasoning_tokens":1327}},"tokens_in":689,"tokens_out":1400,"duration_ms":10610,"temperature":1.0,"reasoning_tokens":1327,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:43:50.177123+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Estimate the causal parameters and the response step size $\\eta$ from data, deploy the Algorithm 1 predictor with $p_1 = T/2$ on held-out subjects, and compute the average future causal effect (AFCE). The theorem predicts AFCE = 0; any held-out AFCE that is statistically nonzero, for example outside the sampling noise of the paper's synthetic experiment, would falsify the claim, as would a version of the simulation with heterogeneous individual step sizes $\\eta_i$.","supporting_citations":[{"cited_title":"From predictions to decisions: Using lookahead regularization","cited_arxiv_id":null,"evidence_quote":"Supplies the gradient-based strategic response model $U' = U + \\eta \\nabla_U \\hat{Y}$ that the paper adopts."},{"cited_title":"Tier Balancing: Towards Dynamic Fairness over Underlying Causal Factors","cited_arxiv_id":"2301.08987","evidence_quote":"Prior work on downstream effects of counterfactually fair predictions with parity-based goals, which LCF extends to individual counterfactual fairness of status."},{"cited_title":"Strategic classification is causal modeling in disguise","cited_arxiv_id":null,"evidence_quote":"Justifies modeling strategic classification as a causal problem with individual responses in exogenous-variable space."}],"review_version":1}