{"id":"3f9b0740-8feb-43ba-9c9b-8a11b5a60aad","arxiv_id":"2607.02572","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"ACC-CRL improves multi-source image fusion OOD robustness by shared-anchor response alignment and uncertainty-gated fusion, with gains on ColorMNIST structure metrics and multi-center MVI prediction.","lead":"The paper proposes Additive Causal Construction (ACC) and a trainable version (ACC-CRL) that fuses multi-source images by aligning shared causal content and gating unreliable cross-source links with uncertainty. It targets better out-of-distribution performance in multi-center medical imaging, especially microvascular invasion prediction.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"The load-bearing gap is that e_str/γ is treated as edge credibility without a direct falsification that residual mismatch tracks causal-path reliability rather than residual task noise or incomplete disentanglement.","rationale":"The reader correctly isolates the weakest link: treating sample-level response residuals plus a lightweight uncertainty head as a proxy for cross-system edge credibility (Appendix C; §IV.A–B). That proxy is what converts ACC-CRL from a multi-branch fusion network into a method that “constructs and reconfigures causal graphs.” The paper’s structural diagnostics (h-diff, CF, t-SNE, residual–error correlation in Fig. 6) are supportive but indirect; they do not rule out residual contamination by task noise or incomplete disentanglement, and headline OOD accuracy/AUC is not uniformly superior. I therefore agree with the reader’s weakest_assumption and keep CONDITIONAL: the empirical recipe is interesting and the multi-center MVI setup is valuable, but the causal-regulation interpretation—and thus the strongest claim as stated—remains conditional on a residual-source check of the kind above, plus clearer metric-consistent baselines and artifacts. No stronger internal inconsistency was found; the concern is identification of the gate, not a contradiction in the math.","tokens_in":22241,"tokens_out":715,"duration_ms":7445,"concrete_test":"On ColorMNIST Bias=0.99 (and one MVI OOD fold), ablate the residual source: (i) freeze task heads and recompute e_str after only content/mechanism updates; (ii) inject controlled label noise into one system’s head while holding encoders fixed; (iii) report Spearman(e_str, |Ŷ−Y|) vs Spearman(e_str, known shortcut strength / domain label). If e_str tracks head noise more than shortcut strength, and γ-gated OOD BACC/R@1 does not improve over a non-causal uncertainty gate, the CGR credibility claim fails.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim (Abstract; §V) that ACC-CRL improves OOD via CGT+CGR rests on Appendix C’s identification γ ≈ P(C(e)=1 | e_str, u_unc), with e_str from bidirectional response residuals (Eqs. 24–27) after content–mechanism decoupling. That identification is load-bearing: if residuals mainly reflect task-head noise, incomplete HSIC disentanglement (Eq. 21), or residual correlated non-causal mechanisms (Appendix B’s U(1) ̸⊥ U(2) case), then γ does not selectively preserve stable causal paths and CGR is mis-targeted. ColorMNIST shows reduced h-diff and better R@1/k-NN under Intervention+UAF, but OOD Acc often does not rise (and can fall) relative to Concat/Intervention (Table I), which is consistent with either successful CSE suppression or with over-suppression of useful signal. MVI OOD gains are mixed (IRM higher AUC; ACC-CRL better BACC/ACC; Tables III–VI). Theory (App. A–C) reinterprets the losses as first-order Jacobian / credibility proxies but does not empirically separate residual sources. Without that separation, “mechanism alignment + uncertainty modeling” remains a plausible training recipe whose causal-graph interpretation is under-supported.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper argues that multi-source image fusion under heterogeneous generative mechanisms suffers from cross-system discrepancy (CSD) and cross-system entanglement (CSE), and proposes Additive Causal Construction (ACC) with two principles: causal graph transferability (CGT) via shared anchors and interventional response consistency, and causal graph reconfigurability (CGR) via credibility-aware structural regulation. ACC-CRL instantiates this by content–mechanism decoupling (contrastive + HSIC), bidirectional response alignment under shared anchors, and an uncertainty-gated fusion variable γ that mixes dual-system content with a conservative single-system fallback (plus an observation-level branch). Experiments on ColorMNIST (controlled shortcut bias) and multi-center MVI prediction (intra- vs peri-tumoral ART MRI) report reduced directional bias / improved retrieval structure on ColorMNIST and a more favorable ID–OOD trade-off on MVI relative to several OOD baselines, with ablations attributing gains to alignment and UAF.","tokens_in":22718,"tokens_out":1705,"duration_ms":19181,"significance":"If the causal-construction reading is correct, the work offers a useful organizing language for multi-source fusion under mechanism shift—treating fusion as prior-guided construction of transferable anchors and reconfigurable edges rather than static multi-input aggregation—and a concrete training recipe (disentanglement + response alignment + uncertainty gating) that is relevant to multi-center medical imaging. Strengths include dual synthetic/clinical evaluation, component ablations tied to CGT/CGR, representation and uncertainty diagnostics (t-SNE, distance gaps, ECE), and appendices that make the first-order Jacobian and gating interpretations explicit. The practical significance for MVI is real if the OOD gains hold under stricter patient-level and multi-sequence settings; the conceptual significance depends on whether γ truly tracks causal-path reliability rather than residual task noise.","major_comments":[{"comment":"Appendix C and §IV.B.1 identify the sample-wise gate γ = 1 − ½(u_str + u_unc) with e_str from bidirectional response residuals (Eqs. 24–27) as γ ≈ P(C(e)=1 | e_str, u_unc) and thus as expected cross-system edge credibility. This identification is load-bearing for the claim that CGR suppresses CSE rather than merely reweighting features. The manuscript does not empirically separate residual sources (task-head noise, incomplete HSIC disentanglement in Eq. 21, residual U(1) ̸⊥ U(2) correlations discussed in Appendix B) from true path unreliability. A falsification or diagnostic is needed—e.g., controlled residual injection, correlation of e_str with known shortcut strength independent of task error, or ablations that break response heads while holding content fixed—before γ can be treated as structural credibility rather than a useful heuristic gate.","section":"§IV.B.1; Eqs. (24)–(33); Appendix C"},{"comment":"The Abstract and §V claim that ACC-CRL “significantly improves OOD generalization while maintaining ID performance.” On ColorMNIST (Table I), Intervention+UAF reduces h-diff and improves R@1/k-NN structure, but OOD Acc is often not better than Concat/Intervention and can be worse at high bias (e.g., Bias=0.99: 67.28±1.81 vs Concat 69.36±2.49). That pattern is consistent with either successful CSE suppression or over-suppression of useful signal. The paper should either (i) reframe the primary ColorMNIST claim around structural metrics (h-diff, CF, R@1) rather than OOD Acc, or (ii) show regimes where OOD Acc also rises, and explain when accuracy is expected to fall under correct causal regulation.","section":"Abstract; §V.B; Table I; Table II"},{"comment":"On multi-center MVI, OOD gains are mixed relative to strong baselines: IRM reports higher external AUC (0.7143±0.0048 vs ACC-CRL 0.6840±0.0296 in Table IV), while ACC-CRL is stronger on ACC/BACC and the ID–OOD BACC trade-off (Fig. 3). The central claim of superior OOD generalization should be stated with metric-specific precision (BACC/ACC vs AUC) and, ideally, patient-level aggregation in addition to slice-level metrics, given clinical decision units are patients. Without that, “significantly improves OOD” overstates a favorable but incomplete trade-off.","section":"§V.C; Tables III–IV; Fig. 3"},{"comment":"CGT/CGR are introduced as graph-level principles (shared anchors, competitive edge selection with threshold δ, candidate system G′ in Eqs. 15–18), but ACC-CRL never constructs or reports an explicit graph, edge set ΔE, or δ-thresholding; regulation is entirely latent via γ and observation fallback (Eqs. 32, 36). The mapping from graph language to representation-space gating is therefore largely definitional (Appendix C). Either provide a structural readout (e.g., estimated edge retention rates, sensitivity to δ, or a simple two-node path recovery experiment) or tone down claims that the method “constructs” and “reconfigures causal diagrams,” framing ACC-CRL as a representation-space approximation with stated limits.","section":"§III.B; §IV.B; Eqs. (10)–(18), (32)–(36); Appendix C"}],"minor_comments":[{"comment":"Notation drift: shared content is X in the main text but occasionally Z in Fig. 5 caption; unify.","section":"Fig. 5"},{"comment":"Fig. 6 caption defines u = 1 − max(p), while the body defines predictive uncertainty via g_unc and u = 1 − γ; clarify which quantity is plotted.","section":"Fig. 6; §V.E"},{"comment":"Loss weights λ_con, λ_dis, λ_ali, λ_uaf and temperatures τ are free parameters but not reported with values or sensitivity; a short hyperparameter table would aid reproducibility.","section":"§IV.C; Eq. (37)"},{"comment":"arXiv date stamp “30 Jun 2026” and journal header “VOL. 14, NO. 8, AUGUST 2015” look like template leftovers; clean for submission.","section":"Header / arXiv line"},{"comment":"Table III header says “INTERNAL ODD-DATA COHORT” (likely “ID”); fix typo.","section":"Table III"},{"comment":"Positive-class augmentation count in Cohort 1 and exact ROI cropping protocol should be stated more precisely for external reproducibility.","section":"§V.C.1–2"}],"recommendation":"major_revision","confidential_remarks":"The empirical recipe (disentanglement + response alignment + uncertainty gating) is plausible and the medical setting is appropriate for this venue, but the manuscript currently over-identifies a useful training procedure with a full causal-graph construction theory. I would accept after major revision if the authors either (a) add diagnostics that e_str tracks path reliability, or (b) substantially soften the causal-graph claims and restate results with metric-specific OOD language. Scope fit for a CV/medical imaging journal is fine; novelty relative to IRM/contrastive multi-modal OOD work is moderate and depends on the CGT/CGR framing holding up under the requested checks."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful part of this paper is not a new causal theorem. It is a concrete dual-region ART MRI setup for multi-center MVI, plus a training recipe that couples content–mechanism split, bidirectional response alignment, and an uncertainty gate, with ColorMNIST diagnostics that actually look at directional bias and representation structure rather than accuracy alone.\n\nWhat is new is the packaging and the medical instantiation: shared-anchor CGT, credibility-regulated CGR, ACC-CRL, and the dual-region (intra/peri-tumoral) cross-center protocol. Empirically they do some real work. On ColorMNIST, Intervention+UAF cuts h-diff and lifts R@1/k-NN structure under strong color shortcuts. On MVI, ACC-CRL sits closer to the ID–OOD diagonal on BACC than MultiOOD/VREx-style baselines and keeps competitive ID numbers while improving external ACC/BACC relative to several strong OOD methods. Ablations and the uncertainty plots (error correlation, OOD shift of u, ECE after temperature scaling) are honest enough that you can see what each piece is doing. Citations cover IRM, DRO, contrastive, and recent multi-modal OOD work without obvious gaps.\n\nThe soft spots are real but not catastrophic. The abstract overclaims “significantly improves OOD”: at high ColorMNIST bias, OOD Acc can be worse than plain concat; on external MVI, IRM still wins AUC. Appendix C’s identification of γ with cross-system edge credibility is definitional more than falsified—e_str could be task-head noise or incomplete HSIC disentanglement rather than causal-path reliability, and they never separate those residual sources. Theory (App. A–C) is a clean first-order reinterpretation of the losses, not independent structure recovery. No code or data release. Free parameters (λ’s, τ, δ, augmentation) are standard but unreported in detail.\n\nThis is for people who train multi-source medical models under center shift and care about fusion gating, not for pure causal-ID theorists. The central argument holds as an engineering claim with mixed metric support; the causal-graph story is optional scaffolding. I would send it to referees. Engage if you work multi-center fusion or MVI; skim the ColorMNIST structural tables and the MVI trade-off figure if you only need the empirical takeaway.","headline":"Solid dual-region multi-center MVI recipe with useful structural diagnostics; the ACC/CGT/CGR packaging is mostly a re-label of disentanglement + response alignment + uncertainty gating, and the load-bearing γ≈edge-credibility claim is under-tested.","tokens_in":23336,"tokens_out":652,"would_cite":false,"duration_ms":8117,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Multi-source image fusion works better out-of-distribution when shared causal anchors are built first and unreliable cross-system links are gated by uncertainty.","keywords":["multi-source image fusion","causal representation learning","out-of-distribution generalization","cross-system discrepancy","cross-system entanglement","uncertainty-aware fusion","microvascular invasion","content-mechanism decoupling"],"falsifier":"On a held-out multi-center or multi-region fusion task, if removing the response-alignment residual from the gate (or replacing it with random/noise scores) leaves OOD metrics unchanged while ID performance stays similar, the claim that structural mismatch correctly targets unstable cross-system edges would fail.","tokens_in":23116,"feed_emoji":"🧠","tokens_out":635,"duration_ms":6258,"temperature":0.7,"pith_summary":"When images come from different sources or regions, they are generated by different mechanisms. Simply concatenating them can create mismatched meanings and unstable shortcut links, which collapse under distribution shift. This paper argues that fusion should be treated as building a joint causal graph: first find shared causal anchors by making content comparable and interventional responses consistent across systems, then reconfigure the graph by keeping only high-credibility cross-system paths and suppressing the rest with structural uncertainty. The learnable version, ACC-CRL, disentangles shared content from system-specific mechanisms, aligns bidirectional responses under those anchors, and uses a sample-level gate that falls back to a conservative representation when mismatch or predictive risk is high. On ColorMNIST shortcut settings and multi-center MRI prediction of microvascular invasion, the method keeps competitive in-distribution accuracy while cutting the drop on external data, supporting the claim that mechanism alignment plus uncertainty-aware reconfiguration yields more transferable fusion.","feed_headline":"Shared causal anchors make multi-source fusion travel","feed_subtitle":"Align mechanisms first, then gate unstable cross-system links with uncertainty to cut OOD drop","key_machinery":"Additive causal construction (ACC), realized as ACC-CRL: content–mechanism decoupling yields shared anchors; bidirectional response alignment produces a structural-mismatch residual; that residual plus predictive uncertainty define a gate γ that strengthens dual-system fusion when reliable and falls back to a conservative representation when not.","core_discovery":"The paper claims that multi-source fusion under heterogeneous generative mechanisms is best cast as additive causal construction: establish transferable shared anchors via semantic comparability and interventional response consistency (CGT), then reconfigure candidate causal edges by credibility and structural uncertainty (CGR). Instantiated as ACC-CRL, content–mechanism decoupling plus bidirectional response alignment and uncertainty-aware fusion suppress cross-system discrepancy and entanglement, improving OOD generalization while preserving ID performance on ColorMNIST and multi-center MVI prediction.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Shared causal anchors transfer multi-source image fusion","Additive causal construction curbs cross-system fusion gaps","Content-mechanism decoupling plus uncertainty reconfigures fusion","Interventional anchors enable reconfigurable multi-source fusion","ACC-CRL aligns mechanisms to cut OOD drop in fusion"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The method assumes that mismatch in bidirectional task responses under shared anchors, together with a lightweight uncertainty score, is a faithful proxy for whether a cross-system causal link is stable enough to keep.","fun_headline_variants_meta":{"raw":{"variants":["Shared causal anchors transfer multi-source image fusion","Additive causal construction curbs cross-system fusion gaps","Content-mechanism decoupling plus uncertainty reconfigures fusion","Interventional anchors enable reconfigurable multi-source fusion","ACC-CRL aligns mechanisms to cut OOD drop in fusion"]},"model":"grok-4.5","effort":"low","cost_usd":0.005768,"raw_usage":{"total_tokens":1598,"prompt_tokens":859,"num_sources_used":0,"completion_tokens":79,"cost_in_usd_ticks":57680000,"prompt_tokens_details":{"text_tokens":859,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":660,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":859,"tokens_out":79,"duration_ms":6905,"temperature":1.0,"reasoning_tokens":660,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T10:07:03.055024+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On a held-out multi-center or multi-region fusion task, if removing the response-alignment residual from the gate (or replacing it with random/noise scores) leaves OOD metrics unchanged while ID performance stays similar, the claim that structural mismatch correctly targets unstable cross-system edges would fail.","supporting_citations":[],"review_version":1}