{"id":"3dcabe98-eb8d-47c0-87e1-3b61fd3c5615","arxiv_id":"2607.10931","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.5,"correctness_risk":"low","formal_verification":"none","parameter_count":3,"one_line_summary":"SpectralOT regularizes entropic optimal transport with the first three Laplace-Beltrami eigenmodes of cortical geometry to produce fast, parsimonious whole-brain functional alignments that improve cross-subject decoding.","lead":"SpectralOT is a fast fMRI alignment method that matches brain activity across people by blending functional similarity with cortical shape via Laplace-Beltrami eigenmodes inside optimal transport. It yields better cross-subject decoding than prior tools while running roughly 30 times faster, making population-level brain models more practical.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified","rationale":"The strongest claim is supported by direct head-to-head numbers (ISC curves, pairwise decoding Table 1, group decoding Figure 5) and a clear complexity advantage (single Sinkhorn pass vs. multi-block FUGW). The three-mode geometric prior is the only modeling choice that could have been under-powered, yet the authors already quantify its adequacy via geodesic error and show that higher modes add little. Sign-flip validity is a mild, standard assumption for template-registered surfaces and does not threaten the empirical results. Minor issues (small N in THINGS, hand-set ε/α) are already reflected in the reader's CONDITIONAL verdict; they do not rise to a load-bearing flaw. Therefore no adjustment is warranted.","tokens_in":18583,"tokens_out":433,"duration_ms":3721,"concrete_test":"Re-run Experiment 4 (IBC RSVPLanguage leave-one-subject-out) with k=5 and k=10 eigenmodes (instead of k=3) at the same α=0.5; if mean accuracy does not rise by more than 1–2 points and remains statistically above anatomical baseline, the three-mode regularizer is confirmed sufficient for the claimed decoding gains.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's weakest assumption (three LBO eigenmodes after gradient sign-flip) is the natural soft spot, but the paper already stress-tests it: Appendix Figure 7 shows mean geodesic error plateaus after k=3, and the sign-flip procedure is valid under the standard fMRI assumption of similarly oriented meshes. Decoding gains (Table 1, Figure 5) and the 30\times speed-up hold under the same regime used for FUGW comparisons. No internal inconsistency or untested premise undermines the central claim that SpectralOT improves ISC and out-of-subject decoding while remaining far cheaper.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The manuscript introduces SpectralOT, a whole-brain functional alignment method for fMRI that constructs a composite cost matrix as a convex combination of a functional L2 dissimilarity and a geometric cost derived from the first three Laplace-Beltrami eigenmodes of the cortical mesh (after a simple gradient-based sign-flip correction). The composite cost is fed to a single entropic Sinkhorn solver to obtain a soft vertex-to-vertex coupling. Four experiments of increasing complexity (purely anatomical color transfer between fsaverage5 and fsLR, ISC on IBC task contrasts, pairwise out-of-subject decoding on Courtois-Neuromod THINGS, and leave-one-subject-out group decoding on IBC RSVPLanguage) compare SpectralOT against FUGW, ProMises and anatomical registration. The central claim is that SpectralOT improves ISC and out-of-subject decoding accuracy while remaining approximately 30\times faster than FUGW and easier to tune.","tokens_in":18736,"tokens_out":888,"duration_ms":8869,"significance":"If the reported gains hold, SpectralOT supplies a practical, geometry-aware alternative to FUGW that removes the need for nested block-coordinate descent and inhomogeneous scaling of Wasserstein versus Gromov-Wasserstein terms. The single free geometric weight α, the linear composite cost, and the public implementation make the method immediately usable for population-scale decoding pipelines and for integration into differentiable deep-learning frameworks. The explicit comparison of blur (Appendix Figure 8), the Nadeau-Bengio-corrected group-level test, and the open code are strengths that raise the bar for subsequent functional-alignment papers.","major_comments":[{"comment":"Table 1 (THINGS pairwise decoding) shows SpectralOT winning on average, yet two of the six source→target pairs fall below the anatomical baseline and the absolute accuracies remain modest (0.14 average). With only three subjects the pairwise design cannot support a claim of consistent domain-shift reduction; the group-level IBC experiment (Figure 5) is therefore load-bearing, yet it reports only a non-significant difference versus FUGW. A power analysis or bootstrap confidence intervals on the accuracy differences would clarify whether the claimed superiority over FUGW is supported or whether the methods are statistically equivalent.","section":null},{"comment":"Methods, Geometric Descriptors and Appendix Figure 7: the decision to truncate to the first three eigenmodes is justified solely by geodesic-error saturation under pure anatomical alignment (α=1). It remains untested whether the same truncation remains optimal once functional cost is present (0<α<1) or on individual (non-template) meshes whose higher-frequency geometry may matter for fine functional topography. A short ablation of k under the ISC or decoding protocols would close this gap.","section":null}],"minor_comments":[{"comment":"Equation (5) and the subsequent simplification assume equal-mass marginals; the text should state explicitly that the Sinkhorn solver is always initialized with uniform marginals, otherwise the row-normalization step is required.","section":null},{"comment":"Figure 4B reports wall-clock times on a single GPU; stating the number of vertices per hemisphere and whether the cost matrices are pre-computed or recomputed would make the 30× claim fully reproducible.","section":null},{"comment":"The ProMises low-rank critique in the Appendix is clear, yet the main text still includes ProMises in every comparison; a single sentence noting that the model is retained only for completeness would avoid reader confusion.","section":null},{"comment":"Typographical inconsistencies appear throughout (e.g., “computationalefficiency”, “thepredictive”, missing spaces after periods). A careful copy-edit pass is needed.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The work is a solid incremental contribution that cleanly addresses the practical bottlenecks of FUGW. Fit for a methods-oriented cognitive-neuroscience or neuroimaging journal is good; the main risk is over-claiming superiority over FUGW on the basis of three-subject pairwise numbers. Minor revision should suffice."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"SpectralOT is a practical methods paper that does what it claims. The new piece is the linear composite cost: functional L2 plus a three-mode Laplace-Beltrami geometric cost, fed to a single Sinkhorn solve, plus a simple gradient-based sign-flip to align the eigenmodes. That combination is not in FUGW, ProMises, or the classical functional-maps literature, and it is easy to tune because alpha is linear rather than the awkward nonlinear trade-off in FUGW.\n\nWhat they do well is the evaluation. Four experiments of increasing difficulty on public data (IBC, Courtois-Neuromod THINGS), fair parameter matching for blur, Nadeau-Bengio corrected tests on the group decoding, and an explicit appendix check that geodesic error plateaus after k=3. They beat anatomical registration and ProMises on ISC and decoding, match or slightly beat FUGW on decoding while running ~30× faster, and release code. The low-rank pathology of ProMises is correctly diagnosed in the appendix. Citations are appropriate and self-citation is limited.\n\nSoft spots are real but minor. The three-mode geometric regularizer is the modeling choice that could have been fragile; they stress-test it and it holds under the usual fMRI assumption of similarly oriented meshes. Hyper-parameters are still hand-set rather than nested-CV’d, and one decoding experiment has only three subjects. None of that undercuts the central claim.\n\nThis is for people who actually run multi-subject surface decoding and need something faster and simpler than FUGW. It deserves a serious referee. I would engage with it and expect to cite the method when I next need whole-brain functional alignment.","headline":"Clean, fast geometry-aware OT alignment that actually improves ISC and out-of-subject decoding while being ~30\times cheaper than FUGW.","tokens_in":19338,"tokens_out":447,"would_cite":true,"duration_ms":4860,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"SpectralOT aligns whole-brain fMRI across people by mixing functional signal with three Laplace-Beltrami eigenmodes, improving cross-subject decoding while running far faster than prior optimal-transport methods.","keywords":["functional alignment","fMRI decoding","optimal transport","Laplace-Beltrami eigenmodes","cross-subject generalization","cortical geometry","SpectralOT"],"falsifier":"On a held-out multi-subject fMRI dataset, replace the three-eigenmode geometric cost with pure functional cost (α=0) or pure Euclidean distance; if cross-subject decoding accuracy then falls below the anatomical baseline while SpectralOT stays above it, the geometric embedding is doing essential work; if accuracy stays the same, the claim collapses.","tokens_in":19492,"feed_emoji":"🧠","tokens_out":886,"duration_ms":8836,"temperature":0.7,"pith_summary":"Brain-activity decoders fail to generalize across people because both anatomy and functional response patterns differ. SpectralOT solves this by building a single composite cost that linearly interpolates functional dissimilarity with geometric dissimilarity derived from the first three Laplace-Beltrami eigenmodes of each cortical mesh. An entropic optimal-transport solver then yields a soft vertex-to-vertex map that can be applied to new data. The method improves inter-subject correlation and out-of-subject decoding accuracy relative to pure anatomical registration and two strong baselines, while requiring only one Sinkhorn pass and running roughly thirty times faster than the previous state-of-the-art. The practical payoff is that population-level decoders become feasible on modest datasets without sacrificing anatomical fidelity.","feed_headline":"Three eigenmodes align brains 30\times faster for better decoding","feed_subtitle":"SpectralOT mixes functional signal with cortical geometry, lifting cross-subject accuracy while staying cheap to run.","key_machinery":"SpectralOT: the composite cost C = (1-α)C_func + α C_geom, where C_geom is the squared Euclidean distance between the first three sign-aligned Laplace-Beltrami eigenmodes of the source and target meshes; this cost is fed once to a Sinkhorn solver to obtain the soft correspondence matrix used for signal transfer.","core_discovery":"Embedding cortical geometry via the first three Laplace-Beltrami eigenmodes into a linearly weighted functional-geometric cost, then solving a single entropic optimal-transport problem, produces whole-brain alignments that raise both inter-subject correlation and cross-subject decoding accuracy while remaining orders of magnitude faster than fused Gromov-Wasserstein approaches.","pith_inferences":["The method supplies a natural building block for a multi-subject functional template: each new subject can be aligned once to a growing average rather than to every other subject.","Because the transport plan is differentiable with respect to the cost, SpectralOT can be inserted as a layer inside an end-to-end neural decoder that jointly optimizes alignment and classification.","If the three-mode geometric regularizer proves sufficient, many other surface-matching problems outside neuroimaging can replace expensive geodesic distances with the same cheap spectral embedding."],"forward_implications":["Population-level decoders can be trained after a single, fast alignment step rather than after expensive pairwise Gromov-Wasserstein solves.","Because the geometric prior is mesh-intrinsic, alignments can be computed between individual cortical surfaces without first warping them to a common template.","The linear α parameter and single Sinkhorn pass make nested cross-validation and integration into deep-learning pipelines practical.","The same eigenmode cost extends immediately to volumetric data, opening whole-brain (surface-plus-volume) functional templates."],"fun_headline_variants":["Three eigenmodes align brains 30× faster for better decoding","SpectralOT embeds geometry for whole-brain functional alignment","Geometry-aware OT raises cross-subject decoding accuracy fast","First three eigenmodes regularize fMRI alignment via SpectralOT","SpectralOT aligns brains faster while preserving cortical structure"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The first three Laplace-Beltrami eigenmodes, after a simple gradient-based sign correction, already capture enough geometry to keep distant brain regions from matching, provided the two meshes are roughly isometric and similarly oriented.","fun_headline_variants_meta":{"raw":{"variants":["Three eigenmodes align brains 30× faster for better decoding","SpectralOT embeds geometry for whole-brain functional alignment","Geometry-aware OT raises cross-subject decoding accuracy fast","First three eigenmodes regularize fMRI alignment via SpectralOT","SpectralOT aligns brains faster while preserving cortical structure"]},"model":"grok-4.5","effort":"low","cost_usd":0.005642,"raw_usage":{"total_tokens":1429,"prompt_tokens":637,"num_sources_used":0,"completion_tokens":82,"cost_in_usd_ticks":56420000,"prompt_tokens_details":{"text_tokens":637,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":710,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":637,"tokens_out":82,"duration_ms":7618,"temperature":1.0,"reasoning_tokens":710,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T08:13:03.900402+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On a held-out multi-subject fMRI dataset, replace the three-eigenmode geometric cost with pure functional cost (α=0) or pure Euclidean distance; if cross-subject decoding accuracy then falls below the anatomical baseline while SpectralOT stays above it, the geometric embedding is doing essential work; if accuracy stays the same, the claim collapses.","supporting_citations":[],"review_version":1}