{"id":"2e2eb55a-1356-4292-886f-4a1de4ffe848","arxiv_id":"2508.20961","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Task-specific brain activity patterns correspond to probability fluxes that can be reproduced by weak, task-dependent antisymmetric interactions in an inferred Ising model.","lead":"By analyzing fMRI data from 590 people, the authors find that brain state transitions carry task-specific 'probability fluxes' and that a fitted asymmetric Ising model shows task-dependent antisymmetric interactions between brain regions. The work suggests the brain may switch functions by subtly rewiring directional interactions, which could cost little extra energy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim of task-dependent antisymmetric interactions lacks uncertainty quantification; apparent variability in small antisymmetric J entries may be sampling noise.","rationale":"The reader's weakest_assumption was the Arrhenius-rate extrapolation to asymmetric interactions. That is indeed an acknowledged modeling simplification, but it is not necessarily fatal: Eq. (27) can be read as a direct definition of a kinetic Ising model with asymmetric couplings, and the detailed-balance derivation is only a heuristic motivation. The model is a valid Markov chain even when J is asymmetric, so the fitted J parameters are not logically invalidated by that derivation gap. The more load-bearing weakness is that the paper's headline conclusion is a comparative statement about fitted antisymmetric matrices, and no uncertainty is attached to those fits. The visual difference in Fig. 4c could arise from sampling noise because the antisymmetric entries are small. The Extended Data Fig. 9/10 comparison provides indirect evidence but no statistical test. This gap is addressable with bootstrapping or likelihood-ratio testing, and it directly affects whether the central claim is established. Therefore I agree with the reader's CONDITIONAL verdict, but for a different primary reason: the lack of uncertainty quantification on the inferred interactions rather than the Arrhenius extrapolation. The paper's descriptive findings (task-dependent probability fluxes and plausible fitted models) are likely robust, but the specific mechanistic claim about task-dependent antisymmetric interactions needs statistical support before it can be accepted.","tokens_in":36818,"tokens_out":7437,"duration_ms":91854,"concrete_test":"For each task, generate B=1000 trajectory-bootstrap replicates of the state-transition matrix (the paper already uses trajectory bootstrapping for flux errors in Methods), refit the Ising model (Eq. 29) on each replicate, and compute the bootstrap distribution of the Frobenius norm of the antisymmetric matrix, and of each element. Then test whether the observed cross-task spread of J^(a) (e.g., ||J_a(emotion)-J_a(rest)||_F) exceeds the within-task bootstrap spread. Equivalently, fit a null model with task-independent J but task-specific h and A, and compare per-task fits via a likelihood-ratio or cross-validated log-likelihood test; if the null is not rejected, the central claim fails.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central mechanistic claim—that tasks differ by subtly modifying the antisymmetric part of the inferred interaction matrix (Fig. 4c)—is not backed by any uncertainty estimate on the fitted parameters. The inference is a linear least-squares fit of log transition rates (Eq. 28), so the covariance of βJ is in principle available, but no error bars, confidence intervals, or significance tests are reported for the interaction matrices. The antisymmetric entries are about an order of magnitude smaller than symmetric entries (Fig. 4 color scales: -0.1..0.1 vs -1..1); at such small magnitudes, the apparent cross-task variability could be dominated by finite-sample noise in the transition-rate estimates, especially for rarely visited hypercube states. Extended Data Fig. 9/10, which compares against a task-independent interaction network, is suggestive but is not a formal test of whether the per-task antisymmetric differences are statistically significant. Without quantifying the noise floor, the observation 'symmetric part is similar, antisymmetric part is not'—the load-bearing support for the energy-efficiency narrative—is not established. This is a correctable but central gap: the main conclusion is a comparative claim about fitted matrices, and comparative claims require confidence intervals.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The authors analyze public HCP fMRI data (590 subjects; rest plus seven tasks), coarse-grain 100 cortical parcels into 7 clusters by hierarchical clustering, binarize the time series into Ising states, and estimate hypercubic probability fluxes. They then infer an asymmetric Ising model from the empirical single-spin-flip transition rates using an Arrhenius-type rate formula, decompose the inferred interaction matrix into symmetric and antisymmetric parts, and report that the symmetric part is strong and task-independent while the antisymmetric part is weak and task-dependent (Fig. 4). The fitted model is used to reconstruct probability fluxes; Pearson correlations with the empirical fluxes are r = 0.53–0.96 (Fig. 5). The authors conclude that brain function is carried by task-dependent, subtle modifications of antisymmetric interactions, which may explain the brain's low task-related energy overhead.","tokens_in":37071,"tokens_out":6021,"duration_ms":60552,"significance":"If the central claim is established, the paper would offer a physically interpretable, low-energy mechanism for task switching and would extend probability-flux analysis to whole-cortex human fMRI data. The authors make useful methodological contributions: flux-based rather than static-correlation analysis, robustness checks across cluster numbers, binarization schemes, and Glauber rates, and a task-independent interaction control. The public HCP dataset is appropriate. However, the central mechanistic claims currently rest on an acknowledged extrapolation of the transition-rate formula to asymmetric interactions and on fitted parameters without uncertainty quantification; the validation against empirical fluxes is in-sample. The significance of the work is therefore conditional on these points being resolved.","major_comments":[{"comment":"The Arrhenius transition rate in Eq. (27) is derived from the detailed-balance condition with a symmetric interaction matrix. The text states: 'We assume the symmetric interaction matrix to derive equation (26) but we apply the result to the asymmetric interaction matrix.' This extrapolation is load-bearing for the central claim. For J ≠ J^T, the pseudo-Hamiltonian is not a true energy and the reverse-rate ratio implied by Eq. (27) does not correspond to a canonical stationary distribution; the inferred antisymmetric part βJ^(a) is therefore a direct product of this heuristic. Please derive the asymmetric rate from a consistent nonequilibrium model or validate the inference on synthetic data with a known asymmetric interaction matrix before interpreting βJ^(a) as a brain property.","section":"Methods, 'Infering Ising spin system from transition rates', Eq. (27)"},{"comment":"The central claim that the antisymmetric interaction is task-dependent is made by visual comparison of point estimates. No standard errors, confidence intervals, or significance tests are reported for βJ or βh. Since Eq. (28) defines a linear least-squares problem, the covariance of βJ is in principle available; alternatively, the bootstrap used for the flux analysis could be extended to the inference. This is essential because the antisymmetric entries are an order of magnitude smaller than the symmetric entries (color scales −0.1..0.1 vs −1..1 in Fig. 4), so the apparent cross-task differences may be sampling noise. The central conclusion is a comparative claim about fitted matrices, and comparative claims require uncertainty quantification.","section":"Fig. 4c and Eqs. (28)–(29)"},{"comment":"The reported correlations (r = 0.53–0.96) are in-sample: the probability fluxes are reconstructed from model parameters fitted to the empirical transition rates of the same task. This measures goodness-of-fit, not predictive success. The task-independent interaction control (Extended Data Figs. 9–10) is more informative, but it is not a formal statistical test. Please add out-of-sample evaluation (e.g., split-half or cross-validation) or a null-model significance test of the improvement from task-dependent interactions. The working-memory result (r = 0.53) is acknowledged, but the interpretation of the remaining correlations as validation needs to be calibrated against the in-sample nature of the comparison.","section":"Fig. 5 and 'To validate our inferred model'"},{"comment":"The number of clusters (7) is set by manually choosing a dendrogram threshold, and the binarization is one of three possible transformations. Extended Data Figs. 3–5 show robustness for the probability-flux diagrams, but not for the inferred interaction matrix βJ or external field βh. Because the inference operates on the 2^7 state space, these preprocessing choices are load-bearing for the structural claim. Please report the inferred symmetric/antisymmetric matrices for alternative cluster numbers and binarization schemes, or quantify the sensitivity in some other way.","section":"Methods: 'Spatial coarse-graining...' and 'Temporal coarse graining...'"}],"minor_comments":[{"comment":"Both sections contain placeholder text ('TBA' and 'zenodo.0000000'). Actual data and code links must be provided for reproducibility.","section":"Data availability / Code availability"},{"comment":"The heading 'Infering Ising spin system from transition rates' contains a typo; it should read 'Inferring'.","section":"Methods heading"},{"comment":"The reported correlation (r = 0.691, p = 0.058) is not significant at the 0.05 level; the text should state this explicitly rather than implying a relation.","section":"Supplementary Fig. S1e"}],"recommendation":"major_revision","confidential_remarks":"The manuscript would fit the journal's scope if the methodological gaps are addressed. The novelty relative to Lynn et al. (2021), which already analyzes broken detailed balance in the same HCP data, should be articulated quantitatively. The placeholder data/code DOIs are a reproducibility concern that must be resolved before public release."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper has a real new piece: inferring an asymmetric Ising interaction matrix from empirical transition rates, then decomposing it into symmetric and antisymmetric parts and showing that the symmetric part looks stable across tasks while the antisymmetric part varies. That is a fair and nontrivial empirical observation, and the robustness checks for clustering, binarization, and the Glauber vs. Arrhenius rates are decent. The method is the contribution; the brain story is the hook.\n\nThe soft spots are real, and the reader's CONDITIONAL verdict is fair. The most serious one is not the derivation gap—though that matters too. The Arrhenius rate in Eq. (27) is derived for a symmetric interaction matrix and then applied to an asymmetric one; the authors acknowledge this directly in the Methods. They give a plausible effective-field justification, but they never test whether the approximation biases the inferred antisymmetric part. That is a legitimate concern.\n\nThe bigger problem is that the interaction matrices are fitted without any uncertainty quantification. The antisymmetric entries are roughly an order of magnitude smaller than the symmetric ones (color scales around 0.1 vs. 1), and the central claim is that these small entries differ across tasks. Without confidence intervals, that apparent variability could be finite-sample noise in the transition rate estimates, especially for rarely visited states. The stress-test note is exactly right here: this is a comparative claim about fitted matrices, and comparative claims need intervals. The task-independent control in Extended Data Figs. 9-10 is suggestive but not a formal test.\n\nThe flux reproduction in Fig. 5 is also in-sample: the model is fit to the same transition rates used to compute the fluxes it is compared with. Correlations of 0.53–0.96 show goodness of fit, not predictive success, and working memory at r=0.53 is a visible failure. Finally, the data and code links are placeholders (Zenodo with TBA), which blocks independent verification.\n\nThat said, the paper is coherent and the authors are upfront about several limitations. The derivation gap and the missing error bars are correctable. For a serious referee, this deserves engagement rather than a desk reject; the inference method is interesting enough to warrant the effort. My recommendation: send it to review, but the reviewers should be told to focus on whether the antisymmetric interaction differences survive a bootstrap of the transition rates and a test of the extrapolation to asymmetric interactions.","headline":"A genuinely new inference method with a plausible empirical story, but the central claim about task-dependent antisymmetric interactions is not yet established because the fitted matrices come without error bars.","tokens_in":37549,"tokens_out":1677,"would_cite":false,"duration_ms":21269,"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":"The paper claims that the human brain switches between tasks by making small, task-specific changes to the directional (antisymmetric) part of the couplings between brain regions, while the strong mutual (symmetric) couplings stay essential","keywords":["probability flux","nonequilibrium steady state","asymmetric Ising model","broken detailed balance","fMRI","brain dynamics","energy-efficient computation","task-dependent connectivity"],"falsifier":"Generate synthetic time series from a 7-spin asymmetric Ising model with a known antisymmetric interaction matrix, run the paper's full inference pipeline on those series, and check whether the recovered antisymmetric part matches the known matrix. A mismatch would show that the task-dependent antisymmetric couplings inferred from fMRI are artifacts of applying a symmetric-derivation transition rate to asymmetric interactions.","tokens_in":36666,"feed_emoji":"🧠","tokens_out":9658,"duration_ms":94292,"temperature":0.7,"pith_summary":"The paper argues that the human brain performs its many cognitive and motor functions not by raising its energy consumption, but by subtly modifying the directional part of the interaction network among brain regions. Analysing fMRI data from 590 adults at rest and during seven tasks, the authors find that the probability flux—the imbalance between forward and backward state transitions—forms a distinct pattern for each task, with circulating fluxes that indicate a nonequilibrium steady state. They fit an asymmetric Ising spin model to these fluxes and report that the symmetric part of the inferred interaction matrix is strong and similar across tasks, while the antisymmetric part is weak and task-dependent. The fitted model reproduces the observed flux patterns for most tasks (Pearson correlation above 0.7, working memory at 0.53), and fixing the interaction across tasks destroys the reconstruction. The claim is therefore that human brain function is a sequence of state transitions controlled by a small set of directional couplings.","feed_headline":"Task switches trace to weak asymmetric brain couplings","feed_subtitle":"Fitting fMRI fluxes to an Ising model shows symmetric wiring stays put while antisymmetric couplings change per task.","key_machinery":"The key object is the probability flux on the hypercube of binarized brain states, defined as the difference between the forward and backward joint transition rates; its nonzero value is the microscopic signature of broken detailed balance. The inference machinery is an asymmetric Ising spin model, in which each of seven coarse-grained cortical clusters is a binary variable and the transition rate between states differing by one spin flip is taken to be the Arrhenius rate k exp[-beta sigma_i (sum_j J_ij sigma_j + h_i)]. The interaction matrix is decomposed into symmetric and antisymmetric parts, J_s = (J + J^T)/2 and J_a = (J - J^T)/2; the antisymmetric part is what generates the circulating","core_discovery":"The central discovery claimed is that the functional identity of a brain state is carried by the antisymmetric part of the interaction matrix between seven coarse-grained cortical clusters. The symmetric (mutual) couplings are strong and nearly task-invariant; the antisymmetric (directional) couplings are visibly smaller and differ from task to task. This decomposition is obtained by inferring an asymmetric Ising spin system from the estimated transition rates and then writing the interaction matrix as J = J_s + J_a. The reconstructed probability fluxes match the empirical ones with r > 0.7 for seven of the eight conditions, and the match is lost when the interaction is forced to be task-ind","pith_inferences":["A testable extension is that targeted perturbation of a single directional coupling should alter the corresponding flux cycle without changing the symmetric backbone; this follows from the paper's mechanism but goes beyond its correlational analysis.","The working-memory failure (r = 0.53) hints that pairwise asymmetric interactions are not enough for at least one task; higher-order interactions or non-Markovian transitions might be needed, which the paper acknowledges as a future direction.","If the energy cost of changing couplings scales with their magnitude, the small antisymmetric values imply that the brain can store a large repertoire of task-specific directional modifications at near-zero metabolic overhead—a quantitative version of the paper's energy-efficiency intuition.","Because the transition-rate formula used for inference was derived under symmetric detailed balance, a rigorous nonequilibrium derivation of the same rate is the cleanest way to confirm whether the inferred antisymmetric structure is real; until then it should be read as conditional on that extrapolation."],"forward_implications":["If the claim holds, task switching in the brain is a change in directional couplings, not in overall activation cost, giving a concrete mechanism for energy-efficient computation.","The probability-flux pattern becomes a functional observable: distinct tasks map to distinct flux cycles, so flux diagrams could be used to identify which cognitive operation is being performed.","The symmetric interaction scaffold can be treated as a fixed backbone, and only a small antisymmetric component needs to vary to produce task-specific dynamics; the model's success with task-dependent J and failure with task-independent J is direct evidence for this separation.","The same probability-flux inference method can be applied to other high-dimensional many-body time series to expose asymmetric interactions and nonequilibrium structure beyond the brain."],"supporting_citations":[{"why":"Supplies the resting-state and task fMRI data from 590 adults used for all analyses.","marker":"23"},{"why":"Documents the seven cognitive and motor task conditions in the fMRI dataset.","marker":"24"},{"why":"Provides the preprocessed BOLD time series and the earlier broken-detailed-balance analysis that motivates the nonequilibrium approach.","marker":"16"},{"why":"Defines the 100-region cortical parcellation into which the fMRI data are divided.","marker":"40"},{"why":"Supplies the probability-flux estimation and trajectory-bootstrap procedure used to measure state transitions.","marker":"43"},{"why":"Motivates coarse-graining from empirical correlation when the underlying interaction network is unknown.","marker":"45"},{"why":"Defines the seven known functional networks used to check that the correlation-based clusters align with brain function.","marker":"47"},{"why":"Provides the orthogonal projection of hypercubes used to draw the probability-flux diagrams.","marker":"49"},{"why":"Supplies the master equation and the Glauber transition rate used as an alternative inference scheme.","marker":"53"},{"why":"Supplies the bootstrapping method used to assess the variability of the estimated probability fluxes.","marker":"68"}],"fun_headline_variants":["Weak directional brain links carry task-specific signals","Brain's subtle asymmetric wiring drives task-specific activity","Probability fluxes reveal task-dependent asymmetric brain interactions","Asymmetric couplings, not strong symmetric ones, shape brain function","Task-specific brain states arise from weak asymmetric interactions"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that the transition-rate formula used to infer the interactions—derived under the assumption of symmetric, reversible (detailed-balance) couplings—remains valid when the couplings are allowed to be asymmetric; if that extrapolation is wrong, the task-dependent directional couplings could be artifacts of the fitting procedure rather than properties of the brain.","fun_headline_variants_meta":{"raw":{"variants":["Weak directional brain links carry task-specific signals","Brain's subtle asymmetric wiring drives task-specific activity","Probability fluxes reveal task-dependent asymmetric brain interactions","Asymmetric couplings, not strong symmetric ones, shape brain function","Task-specific brain states arise from weak asymmetric interactions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00069,"raw_usage":{"total_tokens":2986,"prompt_tokens":789,"completion_tokens":2197,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":533,"completion_tokens_details":{"reasoning_tokens":2139}},"tokens_in":533,"tokens_out":2197,"duration_ms":13443,"temperature":1.0,"reasoning_tokens":2139,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T14:39:15.496752+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate synthetic time series from a 7-spin asymmetric Ising model with a known antisymmetric interaction matrix, run the paper's full inference pipeline on those series, and check whether the recovered antisymmetric part matches the known matrix. A mismatch would show that the task-dependent antisymmetric couplings inferred from fMRI are artifacts of applying a symmetric-derivation transition rate to asymmetric interactions.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the preprocessed BOLD time series and the earlier broken-detailed-balance analysis that motivates the nonequilibrium approach."},{"cited_title":"Horiike \\ and\\ author S","cited_arxiv_id":null,"evidence_quote":"Provides the orthogonal projection of hypercubes used to draw the probability-flux diagrams."}],"review_version":1}