{"id":"df911a47-3ec6-4852-a9cd-0f5ab27f3ecb","arxiv_id":"2505.17542","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"GIST generates graph counterfactuals by overshooting to another class and backtracking with a spectral style loss, reporting higher validity and fidelity on 8 graph classification datasets.","lead":"This paper introduces GIST, a graph counterfactual explainer that first jumps to another graph of a different predicted class, then edits it back while matching the original graph's spectral 'style'. It reports higher validity and fidelity than four existing graph counterfactual methods on eight benchmark datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported counterfactuals are not minimal edits of the input: Eq. (12) optimizes content reconstruction against the randomly chosen opposite-class graph Gε, so the headline gains may reflect near-copies of target-class graphs rather than explanations of G.","rationale":"The reader's REJECT is warranted, but the most load-bearing weakness is not only the padding/commuting assumption. It is that the objective minimizes content loss against Gε, so the thing being explained is not the input. This mismatch is explicit in Definition 4.1, Eq. (12), the overshooting selection in Appendix B, and Figure 3. The padding issue is real and independently damages the theoretical claims (Lemma 4.3 assumes equal-size commuting Laplacians), so I partially agree with the reader's choice of weakest assumption. The proposed intervention isolates whether the empirical gains come from copying Gε; if G* is closer to Gε than to G, the validity/fidelity numbers do not support the counterfactual-explainability claim. The method is clearly presented and the code is released, but the central explanatory claim is unsupported as stated.","tokens_in":36211,"tokens_out":13111,"duration_ms":164350,"concrete_test":"Re-run the AIDS and BBBP experiments with Eq. (12)'s content term set to reconstruct the input G instead of Gε (Lcont = ||X*−X||_1 + BCE(A*, A)), keeping α=0.9, architecture, and all training settings identical. If validity/fidelity stay at reported levels while GED drops toward iRand, the current objective unnecessarily replaces input content; if validity/fidelity collapse, the reported gains are artifacts of copying Gε rather than explaining G.","verdict_should_be":"UNCHANGED","load_bearing_attack":"GIST's central claim is that it produces valid counterfactuals as interpolations between the input's style and counterfactual content, satisfying the minimal-change requirement of Eq. (3). The optimization, however, defines content faithfulness w.r.t. Gε: Eq. (12) minimizes ||X*−Xε||_1 + BCE(A*, Aε), and only the style term involves G. Since Gε is chosen in Appendix B as a shuffled opposite-class graph from the dataset (Eq. 26) and α=0.9 makes content dominate, G* is pulled toward Gε, not toward the input G. The paper's own Figure 3 shows G* moving closer to Gε as α→1, and Table 5 reports AIDS GED(G,G*) = 52.0, more than twenty times the iRand baseline's 2.3. Thus the generated graph is not a minimal perturbation of G; for molecules, atom/bond content can be replaced by that of another molecule while only the Laplacian spectrum is matched. The spectral theory (Lemma 4.3, Theorems 4.4-4.5, Corollary 4.6) does not repair this: it concerns the matrix αL(Gε)+(1−α)L(G), not the network's sampled output, and requires same-size commuting Laplacians, whereas the benchmarks are variable-size and padded (acknowledged in §5.3). Consequently, the validity/fidelity gains in Tables 1-2 may measure how well GIST copies target-class content, not how faithfully it explains the specific input.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GIST (Graph Inverse Style Transfer), a method for graph counterfactual explainability that first overshoots the decision boundary of a black-box oracle by selecting a random opposite-class graph Gε and then learns a backtracking mechanism that generates a graph G* whose normalized Laplacian spectrum is aligned with the input G while its local content is aligned with Gε. The method is formalized through a loss function (Eq. 12) that combines a content reconstruction term with respect to Gε and a spectral style term with respect to G. The authors provide theoretical results for convex combinations of Laplacians (spectral interpolation, connectivity preservation, spectral-gap bounds, and a Frobenius-norm identity) and evaluate GIST on eight graph classification benchmarks against CF-GNNExp., CF2, CLEAR, and RSGG-CE, reporting improved validity and fidelity.","tokens_in":36517,"tokens_out":5338,"duration_ms":61079,"significance":"The idea of re-framing graph counterfactual generation as a spectral style-transfer backtracking process is novel and could open a useful new direction if the generated graphs were faithful minimal edits of the input. The paper also provides an unusually detailed appendix, a public code link, and extensive ablations, which are commendable. However, the central claim is not supported: the loss function preserves content of a randomly selected opposite-class graph rather than of the input, so the generated graphs are not minimal counterfactual perturbations of G under the paper's own definition (Eq. 3). The theoretical guarantees apply only to equal-size commuting Laplacians, while the benchmarks are variable-size and require padding, a limitation acknowledged in Section 5.3. The reported improvements in validity and fidelity may therefore reflect copying target-class structure rather than explaining the specific input. Given that this issue is load-bearing and cannot be repaired without redefining the objective, the paper's main contribution as a counterfactual explainer is not established.","major_comments":[{"comment":"The optimization objective minimizes content reconstruction against the randomly selected opposite-class graph Gε, not against the input G. The content term is ||X* - Xε||1 + BCE(A*, Aε), and with α=0.9 it dominates the loss, pulling G* toward Gε. This is confirmed by Figure 3, which shows GED(G,G*) increasing as α→1, and by Table 5, where GIST's GED between G and G* on AIDS is 52.0 versus 2.3 for iRand. This contradicts the counterfactual definition in Eq. (3) and the abstract's claim of 'minimal changes' to the input. The generated graphs are interpolations between G's spectrum and Gε's content, so the reported validity/fidelity gains in Tables 1-2 may measure how well GIST copies target-class structure rather than how faithfully it explains the specific input.","section":"Section 4.2.2, Eq. (12)"},{"comment":"The theoretical results require L~(G) and L~(Gε) to be real, symmetric, commuting matrices of the same dimension. The benchmark graphs have different numbers of nodes (Table 3), and Section 5.3 explicitly states that the zeros in Figures 4-6 are due to padding adjacency and degree matrices to a common dimension. Padding appends zero eigenvalues and alters the spectrum, and the convex combination in Eq. (7) is undefined for matrices of different sizes. Consequently, Lemma 4.3, Theorems 4.4-4.5, and Corollary 4.6 do not apply to the reported experiments. Moreover, the spectral 'validation' in Figures 4-6 checks quantities that are directly minimized by Lstyle in Eq. (12), so it is an optimization sanity check rather than evidence of semantic or structural preservation.","section":"Section 4.1, Definition 4.2 vs. Section 5.3"},{"comment":"The experimental results are internally inconsistent. For AIDS, Table 1 reports GIST validity of 0.969 and Table 2 reports fidelity of 0.957, while Table 5 reports validity of 0.940 and fidelity of 0.928 under the same 5-fold cross-validation protocol. No standard deviations or confidence intervals are reported for any of the tables, so the headline improvements of +7.6% in validity and +45.5% in fidelity cannot be assessed for statistical significance. At minimum, the AIDS discrepancy must be explained, and all aggregate claims need variance estimates.","section":"Tables 1, 2, and 5"},{"comment":"The paper concedes that there is no guarantee of class maintenance during backtracking in multiclass scenarios. This is not a minor caveat: for the multiclass datasets ENZYMES and COLORS-3, a generated graph that changes the oracle's prediction to any of the other classes counts as valid in Table 1, even if it does not land in the intended counterfactual class. Without per-class validity results or a constraint that Φ(G*) equals the target class, the multiclass validity numbers conflate arbitrary class flips with meaningful counterfactual explanations.","section":"Section 6, Limitations"}],"minor_comments":[{"comment":"The sentence 'if G and Gε are connected graphs, then also Gε is connected' should read 'then also G* is connected'; as written it is a tautology and inconsistent with the surrounding discussion.","section":"Section 4.1"},{"comment":"The text says 'we expect that the GED between G and Gε decreases when α→1 (see the red curve)', but Figure 3's red curve is described as the distance between G* and Gε; this should be GED(G*,Gε).","section":"Section 5.3"},{"comment":"The abbreviation 'RSSG-CE' appears once; it should be 'RSGG-CE' for consistency with the rest of the paper.","section":"Section 5.2"},{"comment":"The word 'explainations' should be 'explanations'.","section":"Section 5.1"},{"comment":"Since all results are averages over 5-fold cross-validation, reporting standard deviations or individual fold results would substantially improve interpretability, especially for the claimed average gains.","section":"Tables 1-2"}],"recommendation":"reject","confidential_remarks":"The manuscript appears to be a camera-ready ICML version with a self-corrected figure; that provenance is not a scientific issue. The more serious editorial concern is that the method's objective is conceptually misaligned with counterfactual explainability: the content-preservation term targets a random opposite-class graph rather than the input, so the generated graphs are not minimal edits of G. Because the problem is in the core formulation rather than in a local technical detail, it cannot be fixed by a revision within the manuscript's current scope. If the authors resubmit, they should either reformulate the method as a graph style-transfer approach without counterfactual claims, or redefine the content term with respect to G and re-evaluate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know about this one. First, the core idea is new: instead of forward-perturbing the input across the oracle's boundary, GIST overshoots to a shuffled opposite-class graph Gε, then learns to backtrack, using a Laplacian spectral 'style' term that pulls the generated graph toward the original G's eigenvalues. That is a genuine departure from the cited GCE baselines, and the paper is clearly written. Second, the central claim collapses on inspection: the content term in Eq. (12) is ||X* − Xε||₁ + BCE(A*, Aε) — faithfulness to Gε, not to the input. With α = 0.9, content dominates, so G* ends up copying another graph's atoms and bonds while matching G's spectrum. That is not a minimal edit of the input. The paper's own numbers say so: GED(G, G*) on AIDS is about 52 versus 2.3 for iRand, and Figure 3 shows validity rising as the distance from G grows. For a paper whose definition of counterfactual (Eq. 3) is maximal similarity to G, this is load-bearing.\n\nCredit where earned: the commuting-case spectral facts are assembled cleanly, with an honest non-commuting Weyl appendix, and the limitations section is candid about the lack of class-maintenance guarantees. The authors also report failed attempts to reproduce RCExplainer and CLEAR, which is good practice. Code is public.\n\nSoft spots, in proportion. The padding of variable-size graphs is acknowledged but real: Definition 4.2 and the theorems require same-size commuting Laplacians, and the benchmarks don't satisfy that. The proof of Theorem 4.5 in Appendix A.3 looks wrong: the displayed interval is misderived, and the min/max conclusion doesn't follow from Weyl. There are no error bars; the AIDS validity is 0.969 in Table 1 but 0.940 in Table 5; and the +7.6%/+45.5% headline averages aren't reproducible from Tables 1–2. Figures 4–6 check the objective GIST was trained to minimize, so they are optimization sanity checks, not independent validation.\n\nWho this is for: anyone in graph counterfactual explainability. The backtracking direction is worth knowing, and the flaw is a useful caution about content-matching to a class exemplar being sold as minimal-change explanation. I would send it to referees. The idea is novel enough that serious review could either reshape the method or dispatch it quickly, and the paper needs major revision either way.","headline":"The backtracking idea is genuinely new and the writing is honest, but GIST's content loss is pinned to a random other-class graph, not the input, so the generated graphs are not minimal counterfactuals of anything.","tokens_in":37076,"tokens_out":7463,"would_cite":true,"duration_ms":75089,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["05C50","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Graph counterfactuals improve when generated by overshooting the decision boundary and then backtracking via spectral style transfer: +7.6% validity and +45.5% fidelity over the second-best baselines on 8 benchmarks.","keywords":["graph counterfactual explainability","spectral style transfer","normalized Laplacian","backtracking","graph neural networks","counterfactual validity","spectral gap","Gumbel-Softmax edge sampling"],"falsifier":"Set $\\alpha = 0.5$ instead of 0.9 and test the paper's central in-between prediction, $\\|L(G_\\varepsilon) - L(G^*)\\|_F \\approx (1-\\alpha)\\|L(G_\\varepsilon) - L(G)\\|_F$: on AIDS the paper's own experiments report an error of 0.537 at $\\alpha = 0.5$, against errors of order $10^{-3}$ at $\\alpha = 0.9$, so the interpolation story must be checked across the whole $\\alpha$-range rather than at the extreme where it was validated. A second check: on a multi-class benchmark, count how often the produced counterfactual lands in a third class rather than the overshoot class — the paper explicitly disclaims any guarantee that $\\Phi(G^*) = \\Phi(G_\\varepsilon)$.","tokens_in":35925,"feed_emoji":"🕸️","tokens_out":11275,"duration_ms":75468,"temperature":0.7,"pith_summary":"The paper tries to establish that graph counterfactual explanations are better produced backwards than forwards. Instead of nudging an input graph until the underlying model changes its verdict, GIST first overshoots the decision boundary to a different-class graph and then learns a reverse transformation that pulls the candidate back toward the input's global spectral structure while keeping the overshoot graph's local content. The authors argue this makes the overshooting problem controllable through a single interpolation factor and produces counterfactuals that are more often valid and more faithful to the true class distribution. If correct, the claim matters because it reframes a search problem that tends to over-edit or destroy structure as a style-transfer problem with explicit control over the validity-proximity trade-off.","feed_headline":"Backtracking lifts graph counterfactual validity by 7.6%","feed_subtitle":"GIST reframes counterfactual generation as spectral style transfer, gaining 45.5% in fidelity across 8 benchmarks.","key_machinery":"The load-bearing object is the convex combination of normalized Laplacians, $\\tilde{L}(G^*) = \\alpha \\tilde{L}(G_\\varepsilon) + (1-\\alpha)\\tilde{L}(G)$, together with the spectral style loss it justifies: the $\\ell^1$ distance between the sorted eigenvalues of the input and the counterfactual. When the two Laplacians commute and share the same size, the eigenvalues of $G^*$ are exactly the convex combination of the endpoints' eigenvalues (Lemma 4.3), which yields connectivity preservation (Theorem 4.4), a between-bounds spectral gap (Theorem 4.5), and the Frobenius identity $\\|L(G_\\varepsilon) - L(G^*)\\|_F = (1-\\alpha)\\|L(G_\\varepsilon) - L(G)\\|_F$ (Corollary 4.6). The learned backtracking network — transformer convolution layers on the overshoot graph, an MLP edge-probability head, and Gumbel-Softmax edge sampling — is the mechanism that realizes this interpolation on real graphs.","core_discovery":"On the paper's own terms, GIST's central claim is that a counterfactual graph $G^*$ can be obtained as a spectral interpolation between the input graph $G$ and a different-class overshoot graph $G_\\varepsilon$: the normalized Laplacian of $G^*$ is modeled as $\\tilde{L}(G^*) = \\alpha \\tilde{L}(G_\\varepsilon) + (1-\\alpha)\\tilde{L}(G)$, so its eigenvalues, spectral gap, and Frobenius deviation from $G_\\varepsilon$ all inherit bounds from the two endpoint graphs. The learning objective realizes this interpolation by minimizing $\\alpha$ times a content-preservation term (node-feature and edge reconstruction against $G_\\varepsilon$) plus $(1-\\alpha)$ times a style term, namely the $\\ell^1$ distance between the sorted normalized-Laplacian eigenvalues of $G$ and $G^*$. Empirically the paper reports that this backtracking procedure beats forward-perturbation baselines on the benchmarks used, with the largest validity gains on multi-class tasks and a reported +45.5% improvement in fidelity.","pith_inferences":["A testable corollary the paper does not draw: if spectral alignment is the effective mechanism, GIST-type methods should dominate on benchmarks where classes differ by global connectivity (trees versus cycles) and lose their edge on motif-localized classes, where local edits suffice.","The exact theory needs equal-size commuting Laplacians; since real benchmarks vary in size, the sorted-eigenvalue $\\ell^1$ comparison rests on zero padding. Swapping in a Wasserstein distance on spectra — which the paper names as future work — would show whether the interpolation story survives variable-size graphs.","The random overshoot selection leaves performance hostage to which different-class graph is drawn; the paper sketches but does not run a proximity-conditioned variant (Equation 27). If that variant materially improves convergence, much of the reported gain may come from spectral alignment rather than the backtracking mechanism itself.","The reported degradation at $\\alpha = 0.5$ suggests practitioners should treat the in-between interpolation guarantee as valid only in the content-dominant regime."],"forward_implications":["If GIST is right, forward perturbation is not the only workable paradigm for graph counterfactuals: crossing the boundary first and then refining gives higher validity on 4 of 8 benchmarks, parity on the rest, and higher fidelity almost everywhere.","The interpolation factor $\\alpha$ becomes a direct control dial: moving $\\alpha$ toward 1 pushes counterfactuals closer to the overshoot graph and raises validity, at the cost of larger spectral distance from the input (Figure 3).","In the idealized equal-size setting, produced counterfactuals inherit connectivity and a spectral gap bounded by the two endpoint graphs, so structural coherence is guaranteed rather than tuned.","Spectral distance (sorted-eigenvalue $\\ell^1$ distance) is proposed as a more principled proximity measure than Graph Edit Distance, since edge edits that do not move the spectrum are treated as structurally neutral.","The framework requires no paired counterfactual data: the overshoot graph is any different-class graph from the dataset, so the method applies wherever an oracle and a dataset exist."],"supporting_citations":[{"why":"Supplies the counterfactual formalization (minimal change that flips the outcome) and the regulation-motivated framing that GIST builds on.","marker":"(Wachter et al., 2017)"},{"why":"Provides the graph-counterfactual definition used in Equation (3), the iRand baseline, and the benchmarking methodology behind Tables 1 and 2.","marker":"(Prado-Romero et al., 2023)"},{"why":"Introduces content/style separation in neural style transfer, the conceptual template for splitting GIST's loss into content and style terms.","marker":"(Gatys et al., 2016b)"},{"why":"CycleGAN's cycle-consistent reverse mapping is the precedent for the overshoot-then-backtrack generation strategy.","marker":"(Zhu et al., 2017)"},{"why":"Gumbel-Softmax makes binary edge decisions differentiable, allowing end-to-end training of the backtracking network.","marker":"(Jang et al., 2017)"},{"why":"Transformer convolution layers provide the node-embedding backbone that processes the noisy overshoot graph.","marker":"(Shi et al., 2021)"},{"why":"CF-GNNExplainer is the strongest forward-perturbation baseline that GIST must beat in Tables 1 and 2.","marker":"(Lucic et al., 2022)"}],"fun_headline_variants":["Spectral backtracking lifts counterfactual validity by 7.6%","GIST: inverse style transfer for graph counterfactuals gains 45.5% fidelity","Backtracking beats forward perturbation in graph counterfactual explainability","Graph counterfactuals via spectral style transfer: +7.6% validity","First backtracking framework for graph counterfactuals, +7.6% validity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The theory assumes that Laplacians of graphs with different node counts can be compared by padding adjacency and degree matrices to a common size, so that sorted eigenvalues can be matched entry-by-entry; the exact interpolation theorems additionally require equal-size, commuting Laplacians.","fun_headline_variants_meta":{"raw":{"variants":["Spectral backtracking lifts counterfactual validity by 7.6%","GIST: inverse style transfer for graph counterfactuals gains 45.5% fidelity","Backtracking beats forward perturbation in graph counterfactual explainability","Graph counterfactuals via spectral style transfer: +7.6% validity","First backtracking framework for graph counterfactuals, +7.6% validity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000836,"raw_usage":{"total_tokens":3656,"prompt_tokens":962,"completion_tokens":2694,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":578,"completion_tokens_details":{"reasoning_tokens":2591}},"tokens_in":578,"tokens_out":2694,"duration_ms":15482,"temperature":1.0,"reasoning_tokens":2591,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:46:00.215406+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Set $\\alpha = 0.5$ instead of 0.9 and test the paper's central in-between prediction, $\\|L(G_\\varepsilon) - L(G^*)\\|_F \\approx (1-\\alpha)\\|L(G_\\varepsilon) - L(G)\\|_F$: on AIDS the paper's own experiments report an error of 0.537 at $\\alpha = 0.5$, against errors of order $10^{-3}$ at $\\alpha = 0.9$, so the interpolation story must be checked across the whole $\\alpha$-range rather than at the extreme where it was validated. A second check: on a multi-class benchmark, count how often the produced counterfactual lands in a third class rather than the overshoot class — the paper explicitly disclaims any guarantee that $\\Phi(G^*) = \\Phi(G_\\varepsilon)$.","supporting_citations":[{"cited_title":"A., Prenkaj, B., and Stilo, G","cited_arxiv_id":null,"evidence_quote":"Provides the graph-counterfactual definition used in Equation (3), the iRand baseline, and the benchmarking methodology behind Tables 1 and 2."},{"cited_title":"Categorical reparametrization with gumble-softmax","cited_arxiv_id":null,"evidence_quote":"Gumbel-Softmax makes binary edge decisions differentiable, allowing end-to-end training of the backtracking network."}],"review_version":1}