{"id":"4681762b-07c4-4009-9956-68c6db1e947d","arxiv_id":"2506.19800","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Fast non-local projections strongly shape modeled cortical responses under about 30 ms or for focused inputs, but contribute little to slow spontaneous dynamics, reconciling geometric and connectomic views.","lead":"A computer model of the brain's cortex combines fast long-range shortcut fibers with local wave-like spread, and finds the shortcuts matter for the first few tens of milliseconds of a response but almost not at all for slow spontaneous activity. The result offers a mechanistic explanation for why geometry-only brain models succeed at resting-state fMRI while connectome-based models are needed for fast sensory responses.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (24)'s activity-preserving, zero-delay FNP operator may predetermine the timescale separation: with nonzero conduction delays or net gain, FNP effects could persist beyond 30 ms, and no sensitivity test is reported.","rationale":"The reader's weakest-assumption analysis and mine converge: the model's central result is not robustly separated from the design of the FNP operator. I see this as a correctness-risk concern rather than an internal inconsistency; the model derivation is transparent, the code is open, and the authors candidly flag the assumptions. Those are real strengths. The issue is the absence of a sensitivity test at the one point where the conclusion would change if the assumption failed. A delay/gain sweep is straightforward because the numerical scheme already supports nonzero τ_m in Eq. (24) and would settle whether the >30 ms negligibility holds. Since the reader's verdict is already CONDITIONAL and my analysis reinforces rather than overturns that conditional, I recommend keeping the verdict UNCHANGED.","tokens_in":31364,"tokens_out":9867,"duration_ms":118070,"concrete_test":"Modify Eq. (24) to c_m[α δ(r−b_m)ϕ(a_m,t−τ_m) − δ(r−a_m)ϕ(a_m,t)] with α ∈ {1.0, 1.1, 1.2} and τ_1 ∈ {0, 5, 10, 14} ms for the 14 cm FNP, reducing c_m as needed to keep the linear system stable. Re-run the single-FNP simulations behind Figs. 3 and 5, recording C(t) at t = 30 ms and t = 50 ms and the Pearson correlation r between hybrid and geometric stationary correlation structures. If any nonzero-delay or α > 1 case keeps C(t) above its τ = 0 limiting value or drops r by more than about 0.1, the slow-spontaneous half of the central claim fails outside the conservative, instantaneous operator.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—FNPs dominate fast evoked dynamics but matter little for slow spontaneous fluctuations—depends on how the FNP term is constructed. In Eq. (24), (C◦ϕ) is conservative: each projection removes activity at a_m and deposits the same amount at b_m, and Table 1 fixes τ_m = 0 for all m. The FNP therefore acts as an instantaneous rerouting operator, and the whole-sheet perturbation is naturally transient, decaying after the geometric waves from source and target interfere (Fig. 3B). That transient character is the paper's stated reason for the >30 ms decay. Real long-range projections have finite conduction delays (a 14 cm projection at physiological speeds has roughly 7–14 ms of delay) and nonzero synaptic gain, which can convert the conservative term into delayed feedback or amplification; both can sustain FNP-induced perturbations on longer timescales. The Discussion explicitly lists the activity-preservation property and fixed τ_m values as limitations, but no sensitivity analysis is provided. Because the timescale-separation conclusion is exactly what follows from the conservative, instantaneous construction, the paper's general claim is not yet secured for physiologically nonzero delays or gains.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces a hybrid neural-field/connectome model in which cortical activity propagates both as traveling waves on a two-dimensional sheet and through discrete fast-conducting non-local projections (FNPs). The FNP operator, Eq. (24), is constructed to preserve total space-time integrated activity, and the numerical experiments fix all conduction delays at zero (Table 1). Using single-FNP and multi-FNP simulations, the paper shows that FNPs perturb stimulus-evoked dynamics most strongly on millisecond timescales (C(t) peaks near 8 ms and decays by about 30 ms), that the time-integrated BOLD perturbation is much smaller than the peak perturbation, and that FNP effects on noise-driven activity depend on the spatial precision of the input drive. The paper concludes that geometry-only neural field models are adequate for slow spontaneous fMRI-scale dynamics, while connectome-specific FNP wiring matters for fast, spatially precise evoked responses.","tokens_in":31661,"tokens_out":7101,"duration_ms":85558,"significance":"If the result holds, the paper provides a mechanistic reconciliation of two previously conflicting literatures: geometric neural field models that ignore connectome specificity and connectome-based models that emphasize long-range projections. The manuscript is careful in its derivations (S1), clearly specifies numerical schemes and parameter values, and supports the central mechanism with an analytic estimate of the perturbation peak time. It also ships open code and makes falsifiable predictions, such as the millisecond-scale peak in FNP perturbation, the small BOLD-level perturbation, and the predicted increase of FNP effects with eigenmode order. These strengths make the paper a useful contribution to the field, provided the load-bearing sensitivity concerns below are addressed.","major_comments":[{"comment":"The central timescale-separation result is obtained with τ_m = 0 and with a conservative FNP operator that preserves total activity. This is acknowledged as a limitation in the Discussion, but no sensitivity analysis is provided. With τ_m > 0, the FNP is no longer an instantaneous rerouting operator: source depletion at a_m and target deposition at b_m are separated in time, so the cancellation mechanism described in Fig. 3C ('the second wavefront arrives and cancels the perturbation') can be delayed; with non-conservative gain it can be absent altogether. Since the 14-cm FNP in Fig. 3 corresponds to physiological conduction delays of roughly 7–14 ms, the claim that FNP influences are confined to <30 ms is not yet secured for realistic parameter settings. Please add simulations with τ_m in this physiological range and with non-conservative gains, or provide an analytic argument that the conservative zero-delay limit is representative.","section":"§3.2, Eq. (24), Table 1"},{"comment":"The main-text evidence that hub and rich-club constraints reduce the aggregate perturbation to spontaneous dynamics is based on Cmax, the maximum over time of the instantaneous cosine distance for impulse stimuli, rather than on direct statistics of noise-driven spontaneous activity. The direct correlation-based check in Fig. S3 uses only 20 connectomes for N = 100 and is reported without the same ensemble sizes or tolerance details used elsewhere. Because the paper's broader conclusion concerns spontaneous fluctuations and resting-state fMRI, please either report the direct correlation-based measures for the full ensembles used in Fig. 7, or justify the Cmax proxy with quantitative evidence that it tracks the correlation-structure perturbation.","section":"§3.4, Fig. 7B, Fig. S3"}],"minor_comments":[{"comment":"The effective non-local gain is written as G_ew(b)G_we(b), but from Eqs. (10)–(13) the product should involve the source-point gain G_we(a). The final reduced equation is unaffected because the product is absorbed into the parameter c_m, but the notation should be corrected.","section":"S1, Eq. (15)"},{"comment":"The text says the distributions are plotted in 'Figs 7B(i) and (ii)', but there are three panels; it should read '(i)–(iii)'.","section":"§3.4, Fig. 7B"},{"comment":"The phrase 'The decreasing in Cmax for increasing λe' should be revised to 'The decrease in Cmax with increasing λe'.","section":"§3.4"},{"comment":"The choice c_m = r^2 is described as equating local and non-local propagation strengths, but the magnitude of C(t) and Cz will scale with c_m; a brief statement that the qualitative conclusions are robust to this choice, or a supporting sensitivity test, would strengthen the presentation.","section":"Table 1, Eq. (24)"},{"comment":"The mollifier weights are normalized to sum to one, but the smoothing scale ϵ = 0.002 m is close to the grid spacing Δx = 0.002 m; a sentence explaining that this choice avoids spurious oscillations without biasing the delta-function approximation would be helpful.","section":"S2, Eq. (28)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is well within scope and the derivations are internally consistent. The main barrier is the untested dependence of the central timescale-separation claim on the conservative, zero-delay construction of the FNP operator; a focused sensitivity analysis would resolve this."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know first: this is a genuine step forward in the geometry-versus-connectome debate. The model is a 2D continuous neural field sheet augmented with multiple fast non-local projections (FNPs), with finite-speed traveling waves and an FNP operator that preserves total activity. That combination is new—prior hybrids were either 1D with a single FNP (Jirsa) or discrete networks with infinite-speed diffusion (Spiegler). The authors show that FNP perturbations dominate stimulus-evoked responses on millisecond timescales (<30 ms) and for spatially localized inputs, but wash out in slower spontaneous dynamics and diffuse drives. If correct, that reconciles apparently conflicting findings.\n\nThe paper does several things well. The S1 derivation is internally consistent; the numerical scheme is clearly specified; the timescale estimate (~8 ms peak, ~28 ms to negligible) matches the simulated cosine dissimilarity; and the results are reproduced with open code. The activity-preserving construction is a clever way to avoid the instability that plagues earlier delayed-feedback FNP models, and the authors are explicit about its cost.\n\nThe main soft spot is the one the stress-test flags: Eq. (24) sets tau_m = 0 and makes the FNP a conservative instantaneous rerouting operator. That means the FNP removes activity at the source and deposits it at the target with no delay and no net gain. The transient character of the perturbation—peaking at ~8 ms and decaying after wave superposition—follows almost directly from that construction. Real long-range projections have conduction delays of order 7–14 ms and can have net synaptic gain; both could sustain FNP influence beyond 30 ms. The authors acknowledge this in the Discussion (\"The validity of the activity-preserving assumption... could be interrogated\") but do not run sensitivity tests. That is a real gap, though not a fatal one: the timescale separation is plausible and the physiological delays fall mostly within the short-timescale window anyway. The bigger worry is that the \"FNPs matter little for slow fluctuations\" conclusion is not yet secured for nonzero gain.\n\nMinor quibbles: the torus geometry, the random rather than connectome-based FNP placements, and the BOLD-as-zero-frequency approximation are all stylized but reasonable for a first demonstration. The paper is honest about these.\n\nBottom line: this deserves a serious referee. I'd send it out, and ask for sensitivity analyses with nonzero delays/gains and, if feasible, one comparison to a real evoked-response dataset. The central idea is likely to survive, but the strongest version of the claim needs those robustness checks.","headline":"A careful modeling paper that gives a plausible timescale- and input-precision-dependent resolution to the geometry-vs-connectome debate, but the central separation rests on untested zero-delay, activity-preserving assumptions.","tokens_in":32124,"tokens_out":2587,"would_cite":true,"duration_ms":26838,"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":"The paper argues that fast non-local projections shape cortical dynamics mainly on timescales below about 30 milliseconds and for spatially precise inputs, while slower spontaneous activity is captured by geometry alone.","keywords":["cortical dynamics","neural field models","connectome","fast non-local projections","traveling waves","resting-state fMRI","stimulus-evoked responses","timescales"],"falsifier":"Record stimulus-evoked and resting-state dynamics in the same animal while varying whether long-range tracts are intact versus silenced, and check whether differences between geometric and connectome models persist beyond 30 ms in slow fluctuations; alternatively, rerun the paper's connectome simulations with physiologically nonzero $\\tau_m$ and with $c_m$ values that add net gain and see whether the slow-timescale dominance of geometry survives.","tokens_in":31176,"feed_emoji":"🧠","tokens_out":5428,"duration_ms":56281,"temperature":0.7,"pith_summary":"The paper tries to resolve a puzzle: rapid, long-range axonal projections are clearly important for brain function, yet models built purely from cortical geometry predict many observed dynamics. It builds a single model in which populations interact through a continuous sheet (traveling waves) and through fast, non-local projections wired independently of geometry. Simulating the model shows that the projections strongly alter responses to precise, localized stimuli in the first roughly 30 ms, but contribute little to slower spontaneous fluctuations, where the geometry-only model is increasingly accurate. If right, the apparent contradiction between connectome-based and geometry-based models disappears: each is correct for a different regime, defined by timescale and input spatial precision.","feed_headline":"Fast brain links shape only the first 30 ms","feed_subtitle":"Long-range wiring drives fast stimulus responses, while geometry governs slow spontaneous activity.","key_machinery":"The load-bearing object is the FNP operator $(C\\circ\\phi)(r,t)=\\sum_m c_m[\\delta(r-b_m)\\phi(a_m,t-\\tau_m)-\\delta(r-a_m)\\phi(a_m,t)]$ appended to a two-dimensional neural field equation. Each FNP removes activity at its source and deposits it at its target, which preserves total space-time integrated activity and prevents the instabilities that arise when projections add gain; the paper sets conduction times $\\tau_m=0$. This operator is what makes FNPs rerouting shortcuts rather than amplifiers, so their influence is transient and spatially local, and it is the mechanism behind the timescale dependence the paper reports.","core_discovery":"The central claim is that fast-conducting non-local projections (FNPs) act as brief, spatially specific perturbations to geometrically constrained cortical dynamics rather than as a global wiring correction. In the model, a single FNP creates a second wavefront at its target, but once the source and target wavefronts superimpose, the perturbation mostly cancels, so its largest effect is concentrated before about 30 ms; the integrated BOLD-like response captures less than a fifth of the maximum transient perturbation. With multiple FNPs, the perturbation grows with connection number but still peaks between 10 and 20 ms and decays at longer times. For noise-driven spontaneous activity, the perturbation shrinks as the input becomes spatially uniform, and hub or rich-club organization concentrates the effect near hubs without raising the spatial average, and for 50–100 FNPs it lowers it. The paper concludes that connectome specificity is needed for rapid processing of spatially targeted inputs, while geometric mean-field connectivity is a valid approximation for resting-state fMRI.","pith_inferences":["If conduction delays are nonzero, the 30 ms boundary is likely to shift and possibly smear: delayed FNP arrivals might re-excite activity after the initial wavefronts pass, extending the connectome's influence into slower dynamics. This is a testable sensitivity analysis the paper does not perform.","The activity-preserving design implies FNP effects are redistributive; in nonlinear or gain-amplifying regimes, repeated circulation through long-range loops could build persistent slow correlations, which would narrow the regime where geometry suffices.","A direct extension would predict that task-evoked fMRI, with spatially targeted stimuli, should show more connectome sensitivity than resting-state fMRI; comparing task versus rest eigenmode reconstruction accuracy would test this.","The concentration of FNP influence near hubs suggests that brain stimulation targeted at hub sites should produce more non-local, connectome-dependent spread than stimulation elsewhere, which is measurable with simultaneous stimulation and wide-field imaging."],"forward_implications":["Stimulus-evoked responses measured with voltage-sensitive dye, EEG, or calcium imaging should show FNP-mediated non-local activations within tens of milliseconds; geometry-only models will miss these.","Resting-state fMRI correlations over seconds can be approximated by geometric neural field models; adding connectome specificity may improve but is not required for the dominant pattern.","BOLD integration suppresses the FNP signature, so fMRI studies will tend to underestimate the role of long-range projections in fast processing.","Hub and rich-club connectivity matter mainly for inputs that arrive at hub regions; they do not increase and can reduce the spatial average of connectomic perturbation.","The same structural connectome supports different dynamical regimes depending on the spatial precision of the driving input, so structural weight alone does not fix functional influence."],"supporting_citations":[{"why":"Supplies the two-dimensional neural field propagator and parameter values for geometric propagation on the cortical sheet.","marker":"[33]"},{"why":"Provides the noise-driven, spectral framework used to model spontaneous dynamics and justify space-time white noise inputs.","marker":"[27]"},{"why":"Empirical evidence that geometric eigenmodes capture resting-state MEG and fMRI dynamics, the slow-timescale result the paper must explain.","marker":"[28]"},{"why":"Shows geometric constraints predict human brain function, the geometry-only success that motivates the reconciliation.","marker":"[29]"},{"why":"Prior neural field model with heterogeneous connection topology whose instability the paper's design overcomes.","marker":"[42]"},{"why":"Prior model with local and global connectivity and time delay, background for the FNP stability discussion and the choice of zero time delay.","marker":"[43]"},{"why":"Voltage-sensitive dye evidence that stimulus-evoked responses show rapid non-local activation via FNPs, supporting the fast-timescale setting.","marker":"[15]"},{"why":"Wide-field calcium imaging of long-range projections in whisker-related somatosensory cortices, supporting the fast non-local activation setting.","marker":"[18]"}],"fun_headline_variants":["Fast wiring matters only for the first 30 ms","Long-range links: brief flash, then geometry wins","Connectome shapes quick responses, not slow waves","30 ms cutoff: where non-local wiring loses sway"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The assumption that FNPs only move activity one-for-one and with zero conduction delay is load-bearing: if real long-range projections add gain or arrive after delays, their slow-timescale influence could persist, and the claim that geometry dominates after 30 ms would weaken.","fun_headline_variants_meta":{"raw":{"variants":["Fast wiring matters only for the first 30 ms","Long-range links: brief flash, then geometry wins","Connectome shapes quick responses, not slow waves","30 ms cutoff: where non-local wiring loses sway"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000253,"raw_usage":{"total_tokens":1619,"prompt_tokens":1056,"completion_tokens":563,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":672,"completion_tokens_details":{"reasoning_tokens":501}},"tokens_in":672,"tokens_out":563,"duration_ms":6477,"temperature":1.0,"reasoning_tokens":501,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:24:58.035691+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record stimulus-evoked and resting-state dynamics in the same animal while varying whether long-range tracts are intact versus silenced, and check whether differences between geometric and connectome models persist beyond 30 ms in slow fluctuations; alternatively, rerun the paper's connectome simulations with physiologically nonzero $\\tau_m$ and with $c_m$ values that add net gain and see whether the slow-timescale dominance of geometry survives.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the noise-driven, spectral framework used to model spontaneous dynamics and justify space-time white noise inputs."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prior neural field model with heterogeneous connection topology whose instability the paper's design overcomes."},{"cited_title":"Ferezou, F","cited_arxiv_id":null,"evidence_quote":"Voltage-sensitive dye evidence that stimulus-evoked responses show rapid non-local activation via FNPs, supporting the fast-timescale setting."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Wide-field calcium imaging of long-range projections in whisker-related somatosensory cortices, supporting the fast non-local activation setting."}],"review_version":2}