{"id":"3a4f5e7f-0b68-4069-b07e-eed488b9af92","arxiv_id":"2505.02435","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"BRACE formulates counterfactual explanations as a weighted combination of input-space and latent-space distances under an invertible causal model, claiming to unify several existing counterfactual methods.","lead":"This paper proposes BRACE, a counterfactual explanation method that balances closeness in the input space with closeness in the hidden causal noise space. It shows that the new objective reduces to several earlier methods at different trade-off weights, but its experimental evidence is limited to two hand-picked examples.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 6.1's claim that BRACE strictly outperforms Causal Algorithmic Recourse is falsified by a simple linear example: when cost(A;x)=dX(xCF,x), CAR's full-set intervention already attains the unconstrained CF minimum, so no λ makes BRACE strictly closer.","rationale":"The reader's weakest_assumption focuses on the invertibility of F and unknown causal graphs. That is a real scope limitation, but it is explicitly stated and is a common assumption in this literature. The more decisive problem is internal: Theorem 6.1, which underpins the claimed relationship with Causal Algorithmic Recourse, is false as stated under the theorem's own assumptions. Because the cost function is set to dX(xCF,x) and no restriction is placed on the intervention subset A, CAR can intervene on all variables and therefore reproduce the unconstrained counterfactual optimum. BRACE, which also minimizes dX when λ=0, can at best match that optimum; it cannot be strictly closer. The 2D example with F=I, h(X)=X1-X2, and L1 distances demonstrates this: the minimum distance is 1, CAR achieves it by intervening on both variables, and BRACE achieves exactly the same distance for every λ. Thus the theorem's central inequality is non-strict, and the subsequent Pareto-optimality argument does not establish the claimed superiority. This does not destroy the empirical heuristic or the λ=0 and λ→∞ limits, but it means the conclusion overclaims the theoretical unification with CAR. A conditional acceptance is appropriate, contingent on correcting Theorem 6.1, explicitly restricting the action set or intervention cost, and re-verifying the Pareto-optimality step. The reader did note a 'subtle Pareto-optimality gap' in the same theorem, so there is partial agreement, but the decisive failure identified here is the full-set intervention issue rather than invertibility.","tokens_in":14917,"tokens_out":19290,"duration_ms":268606,"concrete_test":"Check the 2D linear counterexample analytically or with a short script: F is identity, h(x)=x1-x2, x=(0,0), target yCF=1, and distances are L1. Enumerate all intervention subsets A⊆{1,2} for CAR and solve BRACE (8) for λ over a grid such as [0, 0.1, ..., 100]. Confirm that CAR attains dX=1 via A={1,2} and that BRACE's optimal dX is 1 for all λ. This disproves the strict 'closer' statement in Theorem 6.1. If the authors intend a restricted action set, they must restate the theorem with that restriction and redo the proof and experiments.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central unification claim rests on Theorem 6.1, which asserts that BRACE strictly outperforms Causal Algorithmic Recourse (CAR) under convexity and linearity. As stated, the theorem is false. Its assumptions impose no restriction on A, so A may be the full variable set. With cost(A;x)=dX(xCF,x), intervening on all variables lets CAR set xCF to any vector satisfying h(xCF)=yCF; hence CAR's optimum equals the unconstrained counterfactual problem, the same problem BRACE solves at λ=0. BRACE cannot produce an xCF closer than that global minimum. Minimal counterexample within the theorem's assumptions: take X=U (F identity), h(X)=X1-X2, x=(0,0), yCF=1, and dX=dU=L1. The closest feasible point has dX=1 and is achievable by CAR via A={1,2}; BRACE also achieves exactly dX=1 for every λ>0. No λ yields strict inequality, contradicting the theorem's 'closer' claim. The subsequent scalarization/Pareto argument (Eqs. 17-20) cannot repair this because inequality (18) is only non-strict. The claim should be weakened to 'no worse', or the theorem must explicitly exclude full-set interventions and add a non-trivial action cost.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes BRACE, a counterfactual explanation method that solves the scalarized objective in Eq. (8): minimize dX(x, xCF) + λ dU(u, uCF) subject to h(xCF) = yCF and xCF = F(uCF), assuming an invertible structural causal model X = F(U). The authors claim that this objective unifies classical Counterfactual Explanations (λ = 0), Deep Backtracking Explanations (λ → ∞), Backtracking Counterfactual Explanations via the Gibbs-type PB in Eq. (14), and Causal Algorithmic Recourse under convexity and linearity assumptions (Theorem 6.1). The paper also proves, under the bijective generation mechanism, that interventional counterfactuals are a special case of backtracking counterfactuals (Theorem 5.1), and evaluates BRACE on a small simulation based on the German Credit dataset.","tokens_in":15232,"tokens_out":15465,"duration_ms":194201,"significance":"The BRACE objective is a natural and computationally attractive formulation, and the reduction of Causal Algorithmic Recourse's combinatorial search to a continuous optimization is a legitimate potential advantage. The derivations in Appendix B and Theorem 5.1 are largely correct: the former shows that BRACE can be recovered as a backtracking counterfactual for a particular choice of PB, and the latter establishes that interventional counterfactuals are among the backtracking solutions when F is invertible. However, the claimed theoretical superiority over Causal Algorithmic Recourse in Theorem 6.1 is false as stated; the counterexample in Major Comment 1 is within the theorem's hypotheses. Since this theorem is one of the paper's advertised unifications, the claim needs substantial repair. If the theorem is corrected to a weaker 'no worse' statement or with additional assumptions on the intervention set, the remaining framework may still be valuable.","major_comments":[{"comment":"The theorem is false as stated. Take the identity SCM X = U (i.e., F is the identity), h(X) = X1 - X2, observed x = (0,0), target yCF = 1, and dX = dU = ℓ1. In CAR (Eq. (5)) with cost(A; x) = dX(xCF, x), intervening on the full set A = {1,2} gives xCF = (1,0), which has dX = 1 and satisfies h(xCF) = 1; this is the unconstrained minimum of dX on the constraint set. For BRACE (Eq. (8)), the objective becomes (1+λ)(|x1|+|x2|) subject to x1 - x2 = 1, and its minimum has dX = 1 for every λ ≥ 0. Therefore no λ yields a BRACE counterfactual with dX strictly smaller than CAR's optimum, contradicting the theorem's claim that BRACE 'outperforms' CAR. The root cause is that Theorem 6.1 imposes no restriction on the intervention set A, so CAR can always choose the full set and match the global CE minimum; a correct theorem would need to exclude full-set interventions, add an action cost, or claim only 'no worse'.","section":"Section 6, Theorem 6.1"},{"comment":"The proof does not establish the conclusion even in a weaker form. Inequality (18) is dX(x*BCF, x) ≤ dX(x*ICF, x), which is non-strict; the statement that 'other feasible values ... may reduce the objective further' does not prove strict improvement. In addition, the assertion that scalarization (20) with λ > 0 yields all Pareto-optimal points of the vector optimization (19) is not generally true: for a non-strictly convex Pareto frontier, points can have supporting normals with zero dU weight and thus are not minima of dX + λ dU for any λ > 0. The argument also does not show that, for the specific α in (16), there exists a λ whose solution of (8) satisfies dU(uCF, u) = α. Convexity and linearity alone do not guarantee this. The proof therefore needs a different argument if the theorem is to be salvaged.","section":"Section 6, proof of Theorem 6.1"},{"comment":"The paper's conclusion and abstract state that BRACE 'generalizes' Causal Algorithmic Recourse and yields counterfactuals closer to the original input than CAR, but the only theoretical support for this claim is Theorem 6.1, which is false as stated. The experiments in Section 8 do not directly compare the dX distances achieved by BRACE and CAR; Table 1 reports feature values and monthly payments, but not the distance to the original input. The conclusion and the paper's claims of outperformance should be rewritten to reflect whatever corrected version of Theorem 6.1 is adopted.","section":"Section 9 and Abstract"}],"minor_comments":[{"comment":"The paper assumes a Bijective Generation Mechanism, which is strong; the authors should state more explicitly that all subsequent results (Theorem 5.1, Appendix B, and the CAR comparison) depend on this invertibility and that the method may not apply to non-invertible SCMs.","section":"Section 2.1, Eq. (3)"},{"comment":"The claim that Theorem 5.1 is 'the first result that relates backtracking and interventional counterfactuals' is misleading, since von Kugelgen et al. (2023) already discuss the relationship between these paradigms; the novelty statement should be softened.","section":"Section 5, after Theorem 5.1"},{"comment":"The method is repeatedly claimed to be computationally more efficient than Causal Algorithmic Recourse, but no runtime or complexity measurements are reported; a comparison of wall-clock times would substantiate the claim.","section":"Section 8"},{"comment":"The optimization problem uses normalized features (dividing by σi), but Table 1 reports values in original units; the relationship between the objective in (25) and the reported counterfactuals should be clarified.","section":"Section 8.2, Eq. (25)"},{"comment":"The backtracking conditional distribution is written without its normalizing constant; the authors should note that it is defined up to proportionality.","section":"Eq. (14)"},{"comment":"The proof does not explicitly justify that uICF_A equals u on indices not in A; this follows from the recursive inversion of the structural equations and should be stated to make the proof self-contained.","section":"Proof of Theorem 5.1"}],"recommendation":"major_revision","confidential_remarks":"The main methodological contribution, the BRACE objective and its derivations, is sound and potentially useful, but the false theorem in Section 6 is a serious issue. The theorem is central to the advertised generalization to Causal Algorithmic Recourse, and the counterexample provided in the report falls squarely within the theorem's assumptions. I would ask the authors to either weaken the claim to a non-strict comparison or add explicit restrictions on the intervention set, and to rewrite Section 6 and the conclusion accordingly. The paper could become acceptable after a careful revision; in its current form, the theoretical claim of outperformance is not defensible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: BRACE is a simple and readable weighted combination of Wachter's input-space distance and Kladny's latent-space distance. That is a legitimate and convenient formulation, and the paper explains it well. But the central theorem claiming strict superiority over causal algorithmic recourse is false as stated, and the empirical section is more toy than evaluation.\n\nThe genuinely useful part is Eq (8). It gives a single λ to move between counterfactuals that are close in observation space and ones that are close in latent space. The limits λ=0 and λ→∞ working out to the two existing objectives is correct, and the exponential PB in Eq (14) is a nice bridge to the backtracking literature. Theorem 5.1's claim that backtracking subsumes interventional counterfactuals under invertibility is plausible and worth quoting.\n\nThe trouble is Theorem 6.1. The proof's inequality (18) is only non-strict, but the theorem asserts a strict 'closer' result. The stress-test example shows this is not a minor slip: with F identity and h(x)=x1-x2, intervening on all variables lets CAR reach the global minimum of dX, so no λ can make BRACE strictly better. The honest conclusion is 'no worse' unless you add restrictions on A or a nonzero action cost. That changes the paper's headline claim.\n\nThe experiments are too thin to compensate. Two individuals from one real dataset, a linear causal model and logistic h, no code, and λ chosen by hand. The sensitivity analysis is three noisy draws. It's a demonstration of the concept, not a validation.\n\nThis paper is for readers who want a clean baseline counterfactual objective and aren't looking for deep theory. With the theorem repaired and a real evaluation, it could be a fine workshop or short-conference paper. As it stands, it deserves a serious referee to insist on those fixes, but not a desk reject.\n\nRecommendation: send it to peer review. It's a coherent, well-written idea with a fixable flaw.","headline":"A clean but incrementally novel weighted counterfactual objective, undermined by an overstated theorem and anecdotal experiments.","tokens_in":15717,"tokens_out":4790,"would_cite":false,"duration_ms":56728,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that one objective with a single trade-off parameter λ contains four existing counterfactual explanation approaches as special cases or limits while remaining cheaper to solve than causal recourse.","keywords":["counterfactual explanations","backtracking counterfactuals","causal algorithmic recourse","structural causal models","bijective generation mechanism","model interpretability","actionable recourse","Pareto optimality"],"falsifier":"Take a linear SCM with $F$ and $h$ linear, enumerate the full Pareto frontier of the vector optimization (19), and check whether for every feasible latent deviation $\\alpha$ the scalarized objective (8) reaches the constrained optimum of (17) for some $\\lambda$; any missed point falsifies the existence claim in Theorem 6.1. A simpler check: choose a non-injective $F$ such as $X_1=U_1$, $X_2=U_1^2+U_2$; then $F^{-1}(x)$ is undefined for some $x$, and BRACE cannot be run as written.","tokens_in":14729,"feed_emoji":"🎯","tokens_out":10578,"duration_ms":108529,"temperature":0.7,"pith_summary":"The paper proposes a single optimization objective for generating counterfactual explanations that are both close to the observed input and consistent with a known causal model. The objective minimizes the sum of a distance in input space and a weighted distance in the latent-noise space, subject to the counterfactual flipping the model's prediction. The paper argues that this objective, with one trade-off parameter $\\lambda$, contains four existing counterfactual approaches as special cases or limits: classical counterfactual explanations ($\\lambda=0$), deep backtracking explanations ($\\lambda\\to\\infty$), backtracking counterfactual explanations under a specific backtracking distribution, and causal algorithmic recourse under convexity and linearity assumptions. If true, this gives a computationally cheaper way to produce causally grounded, actionable explanations without the combinatorial subset search or latent-distribution integration of prior methods.","feed_headline":"One objective unifies four counterfactual explanation methods","feed_subtitle":"A single parameter λ recovers classical, backtracking, deep-backtracking, and causal-recourse counterfactuals in one objective.","key_machinery":"The load-bearing object is the invertible reduced-form map $X=F(U)$ between independent latent noises and observed variables, which makes $U=F^{-1}(X)$ deterministic and lets the method treat a counterfactual as a nearby latent vector rather than a modified causal graph. On top of this sits the BRACE objective with two distances: observed-space fidelity $d_X$ and latent-space plausibility $d_U$, balanced by $\\lambda$. The proof machinery is Theorem 5.1's identity $u^{ICF}_A=F^{-1}(F_A(F^{-1}(x)))$, which rewrites interventional counterfactuals as backtracking counterfactuals, and the Pareto scalarization argument used in Theorem 6.1 to show that varying $\\lambda$ can reach the point where a constrained latent deviation is traded optimally.","core_discovery":"Under a known structural causal model with invertible reduced form $X=F(U)$, BRACE solves $$\\min_{$x^{{CF}}$,$u^{{CF}}$} d_X(x,$x^{{CF}}$)+\\$\\lambda$ d_U(u,$u^{{CF}}$) \\quad \\text{s.t.}\\quad h($x^{{CF}}$)=$y^{{CF}}$,\\ $x^{{CF}}$=F($u^{{CF}}$),\\ x=F(u).$$ The paper's central claim is that this scalarized two-distance objective unifies four existing paradigms: at $\\lambda=0$ it reduces to classical counterfactual explanations, as $\\lambda\\to\\infty$ it reduces to deep backtracking explanations, with the backtracking kernel $P_B(u^{CF}\\mid u)\\propto\\exp(-d_X(F(u),F(u^{CF}))-\\lambda d_U(u,u^{CF}))$ it coincides with backtracking counterfactual explanations, and under convex distances and linear $F,h$ it yields counterfactuals at least as close to $x$ as causal algorithmic recourse for the same latent deviation. The paper further establishes Theorem 5.1, that under the bijective generation mechanism backtracking counterfactuals contain interventional counterfactuals as a special case, via the construction $u^{ICF}_A=F^{-1}(x^{ICF})$.","pith_inferences":["Editorial extension: the theory only needs invertibility of $F$, not linearity, so BRACE should be testable on nonlinear invertible generative models such as normalizing flows; if the dominance over causal algorithmic recourse persists there, the method extends well beyond the paper's linear experiments.","The exponential backtracking kernel in Eq. (14) is a substantive modeling commitment, not a neutral choice; a user whose true backtracking distribution differs would obtain different counterfactuals from the same objective.","The parameter $\\lambda$ can be read as a Lagrange multiplier on latent deviation, which suggests selecting it empirically by downstream actionability measures (such as effort or cost of the recommended change) rather than by default values.","Theorem 6.1's argument implies a general comparison recipe: any recourse method that fixes a latent-space deviation can be dominated by re-optimizing observed-space distance inside that deviation; a direct test would compare BRACE with causal algorithmic recourse on a nonlinear SCM using a sweep over $\\lambda$."],"forward_implications":["Setting $\\lambda=0$ makes BRACE exactly the classical counterfactual explanation objective, and letting $\\lambda\\to\\infty$ makes it exactly deep backtracking explanations; the same implementation covers both extremes.","BRACE is equivalent to backtracking counterfactual explanations with the exponential kernel in Eq. (14), so it gives a tractable route to backtracking counterfactuals that avoids marginalizing over the backtracking distribution.","Under convex distances and linear $F$ and $h$, for every fixed latent deviation $\\alpha$ there is a $\\lambda$ whose BRACE counterfactual is no farther from the observed input than causal algorithmic recourse's counterfactual; because BRACE avoids the exponential subset search, it is also cheaper to compute.","The identity $u^{ICF}_A=F^{-1}(F_A(F^{-1}(x)))$ implies that any interventional counterfactual can be realized as a backtracking counterfactual once the model is invertible.","In the loan-approval simulations, BRACE changes both the loan amount and the repayment duration, keeping the implied monthly payment closer to the original than the baselines, which is the sense in which the explanations are actionable."],"supporting_citations":[{"why":"Supplies the classical counterfactual explanation objective (4) that BRACE recovers at $\\lambda=0$ and that the paper criticizes for ignoring causal dependencies.","marker":"Wachter et al., 2017"},{"why":"Defines Causal Algorithmic Recourse (5), the combinatorial baseline that Theorem 6.1 reformulates and that BRACE is claimed to dominate.","marker":"Karimi et al., 2021"},{"why":"Introduces backtracking counterfactuals and the backtracking conditional distribution $P_B$; Appendix B derives BRACE's equivalence to that framework.","marker":"Von Kügelgen et al., 2023"},{"why":"Defines Deep Backtracking Explanations (7), which BRACE recovers as $\\lambda\\to\\infty$.","marker":"Kladny et al., 2024"},{"why":"Provides the Bijective Generation Mechanism assumption, used to justify the inverse mapping $U=F^{-1}(X)$.","marker":"Nasr-Esfahany et al., 2023"},{"why":"Gives the SCM background and the regression-based causal coefficient estimation used in the experiments.","marker":"Peters et al., 2017"},{"why":"Supplies the Pareto scalarization result used in the proof of Theorem 6.1.","marker":"Boyd & Vandenberghe, 2004"},{"why":"Provides the German Credit dataset used in the empirical evaluation.","marker":"Hofmann, 1994"}],"fun_headline_variants":["Single objective unifies four counterfactual methods","One formula covers four counterfactual explanation paradigms","Unified objective for backtracking and causal counterfactuals","Four counterfactual methods, one unified objective","Efficient causal counterfactuals via one scalarized objective"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole construction assumes a known structural causal model with independent noise and no hidden confounders, whose reduced-form map $F$ is invertible; if $F^{-1}$ is not well-defined or the causal graph is misspecified, the latent vector $u$ cannot be recovered and the proofs of Theorem 5.1 and the equivalence to backtracking explanations collapse.","fun_headline_variants_meta":{"raw":{"variants":["Single objective unifies four counterfactual methods","One formula covers four counterfactual explanation paradigms","Unified objective for backtracking and causal counterfactuals","Four counterfactual methods, one unified objective","Efficient causal counterfactuals via one scalarized objective"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000704,"raw_usage":{"total_tokens":3161,"prompt_tokens":918,"completion_tokens":2243,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":2178}},"tokens_in":534,"tokens_out":2243,"duration_ms":18257,"temperature":1.0,"reasoning_tokens":2178,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:52:16.135970+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a linear SCM with $F$ and $h$ linear, enumerate the full Pareto frontier of the vector optimization (19), and check whether for every feasible latent deviation $\\alpha$ the scalarized objective (8) reaches the constrained optimum of (17) for some $\\lambda$; any missed point falsifies the existence claim in Theorem 6.1. A simpler check: choose a non-injective $F$ such as $X_1=U_1$, $X_2=U_1^2+U_2$; then $F^{-1}(x)$ is undefined for some $x$, and BRACE cannot be run as written.","supporting_citations":[{"cited_title":"Algorithmic recourse: from counterfactual explanations to interventions","cited_arxiv_id":null,"evidence_quote":"Defines Causal Algorithmic Recourse (5), the combinatorial baseline that Theorem 6.1 reformulates and that BRACE is claimed to dominate."},{"cited_title":"Backtracking counterfactuals","cited_arxiv_id":null,"evidence_quote":"Introduces backtracking counterfactuals and the backtracking conditional distribution $P_B$; Appendix B derives BRACE's equivalence to that framework."},{"cited_title":"u gelgen, J., Sch \\","cited_arxiv_id":null,"evidence_quote":"Defines Deep Backtracking Explanations (7), which BRACE recovers as $\\lambda\\to\\infty$."},{"cited_title":"Counterfactual identifiability of bijective causal models","cited_arxiv_id":null,"evidence_quote":"Provides the Bijective Generation Mechanism assumption, used to justify the inverse mapping $U=F^{-1}(X)$."},{"cited_title":"Elements of causal inference: foundations and learning algorithms","cited_arxiv_id":null,"evidence_quote":"Gives the SCM background and the regression-based causal coefficient estimation used in the experiments."}],"review_version":1}