{"id":"ff330411-1471-4cf2-a156-e32c82bf5904","arxiv_id":"2411.17206","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Nonequilibrium thermodynamics applied to a 30-area cortical network model shows that higher global connection strength lowers entropy production at equal memory stability and that hierarchical response times are intrinsic to the network.","lead":"This paper applies a physics-based 'landscape and flux' approach to a computer model of the macaque cortex, suggesting that long-range connections may reduce energy costs for the same memory stability. It also reports that the brain's natural response timing, sensory areas first and frontal areas later, is built into the network structure rather than caused by where a stimulus lands.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Gaussian-mixture closure is the load-bearing approximation: the order-of-magnitude EPR reduction and the temporal-hierarchy indicator are computed from an unvalidated ansatz, so the central claims may be artifacts of the closure rather than properties of the cortical network.","rationale":"The reader's weakest assumption—the Gaussian-mixture self-consistent mean-field closure—is exactly the load-bearing point. The paper's most impressive quantitative claims, especially the order-of-magnitude EPR reduction at equal stability, are derived from an ansatz rather than from direct solution of the 90-dimensional Fokker-Planck equation or direct simulation. The ansatz imposes a locally Gaussian measure with weights determined by initial-condition sampling; it does not capture anharmonic barriers, non-Gaussian tails, or the topology of basin boundaries, all of which are known to control barrier crossing and entropy production in multistable nonequilibrium systems. Therefore the central claim could survive or fail depending on the adequacy of this closure. A concrete numerical check is feasible and would settle the matter. I also note secondary issues: the copy-paste artifact 'savanna and forest ecosystems' in Sec 3.4 and the statement that all data are included, when no code or data are actually provided, reduce confidence but are not the main scientific objection. The verdict stays CONDITIONAL because the paper's modeling framework is standard and the qualitative bifurcation and path results are plausible; however, the new quantitative thermodynamics claims need independent validation before acceptance.","tokens_in":87,"tokens_out":6371,"duration_ms":127026,"concrete_test":"Recompute the central Fig 3 result without the Gaussian-mixture closure at a reduced but faithful version of the model (e.g., 5–10 areas, 15–30 variables) by solving the Fokker-Planck equation with a finite-volume or spectral method in the low-dimensional subspace spanned by the two principal components, or by running long direct SDE simulations with a weighted-ensemble rare-event sampler to estimate the stationary density and current along the principal path. Specifically, pick two points on the white line of Fig 3a that differ in G by a factor of two, compute EPR from the empirical density/current, and compare the ratio to the Gaussian-mixture ratio. If the empirical ratio differs by more than a factor of 2, or if the projected marginal density deviates strongly from the Gaussian mixture (e.g., visible skewness or heavy tails), the order-of-magnitude claim should be withdrawn.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The quantitative headline—EPR decreases by an order of magnitude at fixed stability—is computed, not measured, and every ingredient of that computation (U, Jss, EPR, action) is obtained from the Gaussian-mixture closure in Sec 3.2. There, the steady-state distribution is written as Pss = Σ_l w_l N(μ_l, Σ_l), with μ_l and Σ_l from linearized moment equations and w_l from basin-sampling initial conditions. This ansatz has no error control. In a 30-area (90-dimensional) multistable system with saddle-node bifurcations, the true stationary density is generally non-Gaussian; near the bifurcation line the anharmonic barrier region controls transitions, and the tails of the distribution dominate the entropy production rate. The reported order-of-magnitude EPR variation along the white line in Fig 3a could therefore be a property of the assumed Gaussian shapes and weights rather than of the cortical network. The temporal-hierarchy result in Sec 2.2 is less directly affected because it uses path optimization of the original force, but the path action still depends on the same landscape/flux decomposition. Because no code or simulation data are provided, this cannot be checked from the manuscript.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper applies nonequilibrium landscape-flux theory to a 30-area large-scale cortical network model of working memory. The authors construct a Gaussian-mixture approximation to the steady-state distribution of the 90-dimensional Fokker-Planck equation, compute the nonequilibrium potential landscape, entropy production rate (EPR), and dominant transition paths, and report three main results: (i) the temporal ordering of area responses during memory-state transitions is an intrinsic property independent of stimulus location; (ii) increasing the fraction of global connections reduces EPR by about an order of magnitude while preserving memory-state stability, indicating an energy-efficient evolutionary organization; and (iii) thermodynamic and dynamical indicators (EPR, flux, autocorrelation, cross-correlation asymmetry) show precursor changes before bifurcations, offering early-warning signals. All results are model-derived, with no fitting to external data.","tokens_in":16995,"tokens_out":9112,"duration_ms":77114,"significance":"The framework is a reasonable extension of previous work (Mejias & Wang 2022; Yan & Wang 2020) and addresses a timely question—thermodynamic cost of distributed working memory—with no circular fitting. If the Gaussian-mixture closure is validated, the EPR-stability trade-off result would be an interesting theoretical prediction about global versus local connectivity. The path-optimization treatment of transition dynamics is a strength because it uses the original force rather than the linearized approximation, so the temporal-hierarchy result is less susceptible to the closure error. However, the central quantitative claims are computed from an unvalidated ansatz and the manuscript provides no code or data, so the current evidence is insufficient to establish the conclusions.","major_comments":[{"comment":"The quantitative results—landscape U, probability flux Jss, entropy production rate, and path action—are all computed from the Gaussian-mixture ansatz Pss = Σ_l w_l N(μ_l, Σ_l), where μ_l and Σ_l come from linearized moment equations and w_l from basin-of-attraction sampling. In a strongly nonlinear, multistable system with saddle-node bifurcations, the true stationary density is generally non-Gaussian, and the anharmonic barrier region can dominate transitions and entropy production. No comparison with direct numerical simulation of the 90-dimensional stochastic differential equations, or with a higher-order closure, is provided in the manuscript or in any code/data release. The order-of-magnitude EPR reduction along the white line in Fig. 3a could therefore be an artifact of the closure. The authors should validate the approximation (e.g., by simulating the SDE for selected parameter points and comparing the computed Pss, EPR, and action) before the central claim is accepted.","section":"3.2"},{"comment":"The stability comparison underlying the claim 'a higher fraction of global connections can significantly reduce the entropy production rate over an order of magnitude while maintaining stability' is not well founded. The text states that on the white bifurcation line ΔU → ∞, and then asserts that 'as the ΔU remains the same along the bifurcation line, the stabilities of the memory states are similar.' At a saddle-node bifurcation, the memory state is marginal, not robustly stable; an infinite potential difference does not provide a meaningful common baseline. Moreover, the paper itself admits that relative stability does not exactly pin down global stability. The subsequent action and MFPT analysis in Fig. 3d-f is performed at fixed Jmax = 0.243 while varying G, not along the white line, so it does not directly support the claim. A clear definition of ΔU and a stability comparison along constant-action or constant-MFPT curves is needed.","section":"2.3, Fig. 3a"},{"comment":"The conclusion that 'the temporal scale disparity is an inherent attribute, rather than a consequence of the stimulus input area' is not demonstrated. All path computations use a stimulus applied to V1, which is at the bottom of the hierarchy. To support stimulus-location independence, the authors need to apply the same transition-path analysis with inputs to higher-order areas (or to multiple areas) and show that the ordering of response times is unchanged. Without this comparison, the observed hierarchy may simply reflect the known gradient of connectivity and the chosen input location.","section":"2.2"},{"comment":"The early-warning indicators are presented as 'practical methods for the prediction and detection of state switching,' but the manuscript does not establish that these computed quantities correspond to observable signatures in the actual network. The EPR and average flux values are derived from the Gaussian-mixture closure, while the autocorrelation and ΔCC results are based on simulated trajectories whose relationship to the large-scale model's real stochastic dynamics is not validated. For the claimed practical utility, the authors should demonstrate that the trends persist in direct SDE simulations and, ideally, discuss how noisy, finite-length experimental time series would be processed.","section":"2.4"}],"minor_comments":[{"comment":"Section 3.4 contains a leftover phrase 'governing transitions between savanna and forest ecosystems' that appears to be copied from a different application; it should be removed.","section":"3.4"},{"comment":"Section 3.2 uses 'donates' where 'denotes' is intended (e.g., 'which donates the net flow').","section":"3.2"},{"comment":"The bifurcation threshold for inactivation is given as 0.083 in Sec. 2.1 and as 0.085 in Sec. 2.4; please make the values consistent.","section":"2.1 and 2.4"},{"comment":"The Data Availability statement says 'All data is included in the manuscript and/or supporting information,' but no supporting information or code is provided; please clarify what data are available and how the figures can be reproduced.","section":"Data Availability"},{"comment":"References [31] and [59] are the same reference (Cao et al., Nat. Phys. 2015), as are [32] and [60] (Lan et al., Nat. Phys. 2012); these duplicates should be merged.","section":"References"},{"comment":"In Sec. 3.1, the statistical properties of S_noise are not specified (mean, variance, correlation time), which makes the noise model ambiguous.","section":"3.1"},{"comment":"Figure 3a caption should state that the EPR color scale is logarithmic, since the text refers to a 'logarithmic trend.'","section":"Figure 3a"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of q-bio.NC and addresses an important question, but the central claims rest on an unvalidated Gaussian-mixture closure and a problematic definition of stability along the bifurcation line. The lack of code or simulation data is a reproducibility concern. I recommend major revision, with the expectation that the authors add direct SDE validation, clarify the stability measure, and broaden the stimulus-location test before the claims are published."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Thanks for passing this along. Short version: this is a competent application of the landscape-flux machinery to a 30-area macaque cortical model, and it produces two genuinely new results worth thinking about — an intrinsic temporal hierarchy that does not depend on stimulus location, and a trade-off where higher global connection strength lowers the entropy production rate at roughly fixed stability. The temporal hierarchy result is the more robust of the two, since it comes from path optimization on the original force rather than from the distributional ansatz.\n\nWhat the paper does well: the narrative is clear, the methods are standard for this group, and the authors are honest about the limits of their stability measure — they explicitly say relative stability is not global stability and then supplement it with action and mean first passage time. The citation pattern looks fine; they are building on Mejias–Wang and their own prior work, which is appropriate.\n\nNow the soft spots, in proportion. The stress-test note is on target for the EPR claim. Everything quantitative in Sec 2.3 — the landscape U, the flux J, the entropy production rate — is computed from the Gaussian-mixture approximation in Sec 3.2. That ansatz has no error control, and in a 90-dimensional multistable system near saddle-node bifurcations the true stationary density is likely non-Gaussian in the barrier region. The order-of-magnitude reduction in EPR along the white line in Fig 3a could easily be a property of the assumed Gaussian shapes and weights. The paper does not validate the ansatz against direct simulation of the original SDEs, and no code or data are provided to allow an independent check. The 'Data Availability' line is not credible for a computational paper.\n\nThere is also a copy-paste error in Sec 3.4: the text mentions transitions 'between savanna and forest ecosystems,' which is clearly from another manuscript. That is minor but suggests the final proofreading was rushed.\n\nThe abstract and significance statement overreach a bit: the 'practical methods for the prediction and detection of state switching' are computed from a model, not demonstrated on real neural time series. The early-warning indicators section inherits the same closure issue, so I would temper the clinical language.\n\nNet: the core results may be correct, but the load-bearing approximation is unvalidated. I would send this to peer review — it deserves referee time — but with the clear expectation that the authors either release code and data, validate the Gaussian-mixture closure against direct stochastic simulation in a few parameter regimes, or soften the EPR claim. The temporal hierarchy result is likely to survive that scrutiny; the order-of-magnitude energy claim needs the check first.\n\nIf you are looking for a paper to assign, this is a legitimate one. I would not cite it myself until the closure is checked. For a reading group, maybe — it is a good case study in how a standard approximation can drive a headline.","headline":"Solid but unvalidated modeling application; the order-of-magnitude EPR claim rests on a Gaussian-mixture closure that needs direct simulation checks before I'd trust it.","tokens_in":17510,"tokens_out":3212,"would_cite":false,"duration_ms":29361,"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":"In a 30-area cortical model, a larger share of global connections cuts the energy cost of working memory by an order of magnitude while preserving stability, and the lower-to-higher-area response order is intrinsic to the network.","keywords":["cortical working memory","nonequilibrium landscape","entropy production rate","large-scale brain network","temporal hierarchy","critical slowing down","time irreversibility","energy-stability trade-off"],"falsifier":"Run direct stochastic simulations of the same 30-area network with small diffusion, estimate the entropy production rate from long trajectories (for example via time-irreversibility of cross-correlations or by measuring the flux), and check whether increasing global coupling G still produces a tenfold drop in EPR while the memory states stay equally stable; if the drop disappears or reverses under this non-Gaussian sampling, the central energy-efficiency claim is refuted.","tokens_in":16538,"feed_emoji":"🧠","tokens_out":9446,"duration_ms":76436,"temperature":0.7,"pith_summary":"This paper tries to establish that two organizational features of the macaque cortex — the ordering of response times across cortical areas and the energetic cost of maintaining working memory — can be explained from the nonequilibrium dynamics of a 30-area network model. When the network transitions between a resting state and a selective memory state, the dominant transition path activates lower hierarchical areas before higher ones, and this ordering persists regardless of where the stimulus is applied; the temporal hierarchy is therefore an intrinsic property of the connectivity, not a consequence of sensory input location. The paper further claims that increasing the global fraction of long-range connections reduces the entropy production rate, the thermodynamic cost of sustaining the steady state, by more than an order of magnitude while keeping memory states equally stable. If these claims are right, they connect the brain's hierarchical organization and its long-range wiring to a quantitative energy-stability trade-off, and they supply early-warning indicators for cortical state transitions.","feed_headline":"Global brain wiring cuts working-memory energy tenfold","feed_subtitle":"In a 30-area model, more global connections preserve memory stability while cutting energy cost tenfold.","key_machinery":"The central object is the nonequilibrium potential landscape $U = -\\ln P_{\\mathrm{ss}}$ together with the steady-state probability flux $J_{\\mathrm{ss}}$, defined on the 90-dimensional state space of 30 cortical areas, each with three populations. The argument is carried by the force decomposition $F = -D G \\cdot \\nabla U + J_{\\mathrm{ss}}/P_{\\mathrm{ss}} + D\\nabla \\cdot G$, the entropy production rate $\\dot S = \\int J \\cdot (DG)^{-1}\\cdot J / P \\, dx$, and the Onsager-Machlup path integral whose minimized action selects the dominant transition path between attractors. All of these quantities are computed under a self-consistent mean-field approximation that represents the steady-state distribution as a weighted sum of Gaussian distributions around each attractor, with mean and variance evolved through linearized moment equations.","core_discovery":"The central discovery, stated on the paper's own terms, is that the thermodynamic cost and response-time structure of distributed working memory are controlled by the balance between global and local connectivity. In the 30-area Wong-Wang model of the macaque cortex, the authors compute the nonequilibrium steady-state probability distribution approximately, decompose the driving force into a landscape gradient and a rotational probability flux, and identify dominant transition paths by minimizing the Onsager-Machlup action. Along the path from resting to memory state, primary sensory areas such as V1 and MT rise first, while higher-order areas ramp later; the same order appears when the stimulus is delivered to different areas, so the temporal scale disparity is an inherent attribute rather than an effect of the stimulus location. Along the bifurcation curve where the resting state disappears, memory states have equal relative stability, yet increasing global coupling strength lowers the entropy production rate by more than an order of magnitude. Since the entropy production rate measures the thermodynamic cost of maintaining the steady state, the paper concludes that a more globally connected network achieves the same memory stability at lower energy dissipation, which it interprets as an evolutionary advantage of long-distance cortical connections.","pith_inferences":["The Gaussian-mixture closure is load-bearing for the quantitative factor of ten; if the true steady-state distribution is non-Gaussian, the qualitative direction of the result might survive but the exact magnitude could change. This is an editorial caution, not a claim of the paper.","A direct experimental test of the temporal-hierarchy claim: optogenetically stimulate a high-order association area and record lower-order sensory areas during working memory; the intrinsic ordering predicts lower-order areas still ramp earlier during memory encoding.","The same landscape-flux machinery could be applied to other multistable biological networks with tunable global coupling, such as gene regulatory circuits or ecological systems, to see whether global coupling similarly lowers thermodynamic cost while preserving multistability.","One could estimate EPR, flux, and time-irreversibility indicators from empirical calcium or electrode recordings during gradual pharmacological perturbation, comparing the predicted early rise before state transitions against measured autocorrelation and forward-backward cross-correlation asymmetry."],"forward_implications":["If the temporal hierarchy is intrinsic, then observed latency differences between sensory and association areas do not require stimulus-specific routing explanations; they follow from the network's hierarchical connectivity alone.","If global connections lower energy cost at equal stability, then the evolutionary expansion of long-range cortical connections may be driven at least in part by thermodynamic efficiency rather than only by robustness.","Entropy production rate and average probability flux change sharply and early as the network approaches a bifurcation, making them candidate early-warning indicators for imminent cortical state transitions.","The time-irreversibility measure (forward-backward cross-correlation difference) and critical slowing down (autocorrelation relaxation) are obtainable from time-series data, offering experimentally accessible ways to detect latent state transitions.","The energy-stability-flexibility analysis points to an optimal balance between global and local connectivity: near the tangent of EPR and relative-stability isolines, the network gets the most stability per unit of thermodynamic cost."],"supporting_citations":[{"why":"Supplies the 30-area macaque cortical network model of distributed working memory that the paper analyzes.","marker":"[22]"},{"why":"Provides the Wong-Wang local circuit dynamics used for each cortical area.","marker":"[26]"},{"why":"Documents the temporal hierarchy in large-scale cortical models and the earlier assumption that stimulus enters lower areas, which the paper tests by relaxing the assumption.","marker":"[15]"},{"why":"Underpins the nonequilibrium landscape-flux theory, including the potential landscape U=-ln P and probability flux decomposition.","marker":"[40]"},{"why":"Supplies the entropy production rate definition used to quantify thermodynamic cost of the steady state.","marker":"[56]"},{"why":"Provides the path-integral action minimization method used to find the dominant transition paths and their temporal ordering.","marker":"[75]"},{"why":"Establishes critical slowing down and autocorrelation as early-warning indicators before bifurcations.","marker":"[62]"},{"why":"Supports the forward-backward cross-correlation difference as a measure of time irreversibility and nonequilibriumness.","marker":"[41]"},{"why":"Gives the marmoset cortical connectivity scaling evidence that motivates the evolutionary energy-efficiency interpretation.","marker":"[35]"},{"why":"Relates entropy production and functional stability trade-offs in neural circuits, framing the energy-stability-flexibility balance.","marker":"[27]"}],"fun_headline_variants":["Global brain wiring cuts memory energy tenfold","Cortical connectivity lowers working-memory energy cost","Global connections slash energy cost of memory states","Memory stability at lower cost via global wiring","Energy-efficient memory from global cortical links"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole quantitative framework rests on the assumption that the steady-state probability distribution of the 90-dimensional network is well captured by a weighted sum of Gaussians around each attractor; if the true distribution is significantly non-Gaussian in the strongly nonlinear regime, the computed entropy production rates and the order-of-magnitude reduction could be numerical artifacts.","fun_headline_variants_meta":{"raw":{"variants":["Global brain wiring cuts memory energy tenfold","Cortical connectivity lowers working-memory energy cost","Global connections slash energy cost of memory states","Memory stability at lower cost via global wiring","Energy-efficient memory from global cortical links"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000194,"raw_usage":{"total_tokens":1391,"prompt_tokens":1019,"completion_tokens":372,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":635,"completion_tokens_details":{"reasoning_tokens":306}},"tokens_in":635,"tokens_out":372,"duration_ms":3967,"temperature":1.0,"reasoning_tokens":306,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:22:24.863764+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run direct stochastic simulations of the same 30-area network with small diffusion, estimate the entropy production rate from long trajectories (for example via time-irreversibility of cross-correlations or by measuring the flux), and check whether increasing global coupling G still produces a tenfold drop in EPR while the memory states stay equally stable; if the drop disappears or reverses under this non-Gaussian sampling, the central energy-efficiency claim is refuted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Wong-Wang local circuit dynamics used for each cortical area."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the 30-area macaque cortical network model of distributed working memory that the paper analyzes."},{"cited_title":"Chaudhuri, K","cited_arxiv_id":null,"evidence_quote":"Documents the temporal hierarchy in large-scale cortical models and the earlier assumption that stimulus enters lower areas, which the paper tests by relaxing the assumption."},{"cited_title":"Wang, Advances in Physics 64, 1 (2015)","cited_arxiv_id":null,"evidence_quote":"Underpins the nonequilibrium landscape-flux theory, including the potential landscape U=-ln P and probability flux decomposition."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the entropy production rate definition used to quantify thermodynamic cost of the steady state."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the path-integral action minimization method used to find the dominant transition paths and their temporal ordering."},{"cited_title":"Scheffer, J","cited_arxiv_id":null,"evidence_quote":"Establishes critical slowing down and autocorrelation as early-warning indicators before bifurcations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supports the forward-backward cross-correlation difference as a measure of time irreversibility and nonequilibriumness."},{"cited_title":"Theodoni, P","cited_arxiv_id":null,"evidence_quote":"Gives the marmoset cortical connectivity scaling evidence that motivates the evolutionary energy-efficiency interpretation."},{"cited_title":"Yan and J","cited_arxiv_id":null,"evidence_quote":"Relates entropy production and functional stability trade-offs in neural circuits, framing the energy-stability-flexibility balance."}],"review_version":1}