{"id":"b81900de-b5e2-4f6d-8cd9-58c71b7844ef","arxiv_id":"2505.20697","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"REDCLIFF-S models time-varying causal graphs as history-weighted combinations of nonlinear static graphs and reports large F1 gains over baselines in synthetic and brain-data experiments.","lead":"This paper introduces REDCLIFF-S, a method that treats changing causal relationships in brain signals as a weighted blend of several fixed nonlinear graphs, with weights chosen from the recent signal. It reports better recovery of these dynamic causal patterns than current baselines on synthetic data and illustrates the approach on real brain recordings.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Synthetic benchmarks are generated from exactly the model's own factor-superposition class (Eq. 4), so the load-bearing assumption that real neural dynamics are a small weighted sum of static graphs is never tested; the no-assumptions claim in Section 1 is contradicted by the model form.","rationale":"The reader's weakest assumption is indeed the most load-bearing. I considered the paper's own identifiability counterexample (Appendix A.1), but the algebraic identity as written (ReLU(A_t)+B_t = nReLU(-A_t)+B_t) is not correct in general, so I do not build the main critique on it. The superposition assumption is robust and directly testable. The proposed experiment would show whether the method's advantages persist when the data-generating process violates the model's core representational assumption; if they do not, the abstract's 'no assumptions' and 'detection' claims need to be restricted to hypothesis generation within the factor-superposition family. The recommended verdict remains CONDITIONAL, matching the reader, because the method is still a plausible hypothesis-generation tool and the empirical comparisons are reasonable in-distribution.","tokens_in":26556,"tokens_out":8792,"duration_ms":87840,"concrete_test":"Generate a synthetic system with the same nodes, noise, and sample size as Section 4.1 but with continuously time-varying coupling, e.g., edge strength g_ij(t) = beta_0 + beta_1 sin(2 pi t / T) + beta_2 cos(4 pi t / T) and edge functions that change between linear and ReLU-like over time, so the true dynamics are not a finite sum of static factor graphs. Run REDCLIFF-S (with K=3 and with K chosen by their stopping criteria from Eq. 10) and the Section 4.2 baselines on multiple repeats. If REDCLIFF-S's graph recovery (ROC-AUC/F1) collapses toward baseline levels or its selected factors are unstable across repeats, the benchmarks in Section 4 do not support the central detection claim outside the factor-superposition class.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim—that REDCLIFF-S detects complex, time-varying interactions in real neural systems—requires that the underlying dynamics be approximable as Phi(t) = sum_k G(t)_k F_k(t) (Eq. 4) with a small number of reused static factor graphs. The synthetic validation is in-distribution for this assumption: Section 4.1 generates each recording as a time-varying weighted sum of VAR-model outputs (ReLU/nReLU edges), exactly the REDCLIFF-S model class. D4IC is constructed by mixing DREAM4 folds the same way. These benchmarks therefore show recovery when the generative process is inside the hypothesis class; they do not test what happens when true coupling continuously morphs rather than reusing a few discrete factor graphs, when the number of latent states is unbounded, or when edge functions themselves change form over time. The real-data case studies lack ground truth; agreement with a few literature-reported edges cannot establish detection accuracy. Section 1's claim that 'we make no assumptions regarding the underlying generative processes' is contradicted by the factor-superposition model and by the evaluation strategy. Consequently, the abstract's 'enables the detection of complex, time-varying interactions' is not supported for systems outside the superposition family.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes REDCLIFF-S, a factor-based Granger causal discovery method that models dynamic causal graphs as a history-conditioned weighted sum of static factor graphs, with each factor implemented as a cMLP and regularized for sparsity and factor dissimilarity; optional supervised labels are included through an invertible state model. The authors claim substantial improvements over static and linear baselines in detecting nonlinear, state-dependent causal interactions, based on synthetic benchmarks (Synthetic Systems and D4IC) and two real-world LFP case studies (TST and Social Preference). The paper also releases code and reports standard errors and ROC-AUC values alongside F1 scores.","tokens_in":26783,"tokens_out":5891,"duration_ms":62662,"significance":"If the central claims were fully supported, this would be a useful contribution to hypothesis generation in neuroscience: the method is explicitly designed for interpretability (fixed factor graphs, behavior-linked factor scores), combines nonlinear factors with dynamic weighting in a natural way, and is evaluated against multiple baselines with released code and standard errors. The design principle of reusable factor graphs with state-dependent weights is plausible, and the in-distribution synthetic results are encouraging. However, as detailed below, the external validity of the evaluation and the consistency between the paper's identifiability analysis and its benchmark design require substantial work before the abstract's claims about detecting complex, time-varying interactions can be accepted.","major_comments":[{"comment":"The synthetic evaluation is in-distribution for the model class. Section 4.1 generates each recording as a time-varying weighted superposition of VAR models with ReLU/nReLU edges, exactly the form of Eq. (4), and Section 4.3 constructs D4IC by the same mixing mechanism over DREAM4 folds. The Section 1 statement that 'we make no assumptions regarding the underlying generative processes' is therefore contradicted by both the model form and the evaluation design. Because the real-data case studies in Section 5 lack ground truth, the paper provides no evidence for performance when the true dynamics are not a small weighted superposition of reused static graphs. Please add experiments where the target system is outside the model class (e.g., continuously morphing edge weights, or edge functions that change form over time) or substantially soften the no-assumptions claim and the abstract's characterization of what is supported.","section":"Section 4.1 and Eq. (4)"},{"comment":"The paper's own identifiability result directly weakens the benchmark's validity. Appendix A.1 proves that a two-node system with ReLU and nReLU edges and symmetric innovations has two observationally equivalent state-dependent graphs, and Section 4.1 explicitly uses ReLU and nReLU activations in the synthetic systems. The same sign/flip symmetry may therefore be present in the synthetic ground truth, which would mean that edge-level F1 and ROC-AUC scores in Section 4.2 and Tables 2, 5, and 6 penalize algorithms for recovering observationally equivalent alternatives. Please either establish that the synthetic data generation avoids the A.1 symmetry (e.g., via asymmetric innovation distributions or initial conditions), or evaluate against equivalence classes rather than a single ground-truth graph; at minimum, the relationship between A.1 and the benchmark's ground truth should be discussed explicitly.","section":"Appendix A.1 and Section 4.1"},{"comment":"The behavioral link is partly imposed by the supervised construction and should not be described as an emergent discovery. In Eq. (8), the first B factor weights are assigned to behavioral labels, and Eq. (9) trains the state model to predict those labels through the lambda*MSE term. Consequently, the finding that specific factors 'correspond' to Open Field or Home Cage states in the TST case study reflects the imposed supervised structure as much as an independent property of the data. The real-data analysis should be framed as hypothesis generation with the caveat that the behavior-factor association is built into the loss, and the authors should provide a corresponding unsupervised or held-out validation to support the claim of uncovering behavior-linked relationships.","section":"Section 3.5 and Section 5"}],"minor_comments":[{"comment":"The phrase 'in some of our experiments' makes the headline improvement of 22-28% (and 'well over 60%') untestable; please cite the specific systems and figure numbers (e.g., Figure 6 and Supplementary Figure 9) in the abstract or clarify the qualifier in the main text.","section":"Abstract and Appendix C.1"},{"comment":"The system naming convention '6-2-2', '6-4-2', '12-11-2', '12-11-5' is not defined; please state at first use that the triple denotes (number of nodes, number of inter-variable edges, number of factors).","section":"Section 4.2"},{"comment":"The up/down arrows in Table 3 are ambiguous because the baseline for comparison is not specified; please state that arrows indicate change relative to the full REDCLIFF-S model, and note that for the two systems where the rho=0 ablation improves performance, this should be acknowledged in the main text discussion of the cosine-similarity penalty.","section":"Table 3 and Appendix C.6"},{"comment":"The color scale for 'difference in mean normalized causal estimates' is not defined; please describe how the normalization and difference are computed, since the figure is the primary evidence for the TST case study.","section":"Figure 5"},{"comment":"The proof sketch of Lemma 1 is difficult to follow: the notation p(f_i, X) is used in a way that conflates a joint density and a conditional likelihood, and the step from Eq. (13) to Eq. (14) relies on the uniform-state prior in a way that should be stated explicitly; please rewrite the derivation for clarity.","section":"Appendix A.2.2"},{"comment":"There are minor typographical and formatting issues, including 'it's use' in Section 4.3, inconsistent 'Open Field (OP)' vs 'OF' in Appendix D.1 and Section 5, and the unexpanded 'V AR' spacing in several places.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The reader's report captures the central tension well. The in-distribution synthetic evaluation and the paper's own identifiability proof are the two load-bearing concerns; both are fixable within scope by adding out-of-class experiments, adjusting the evaluation to equivalence classes, and calibrating the claims. The real-data case studies are interesting but should be presented as hypothesis-generation illustrations rather than validated discoveries. I would not recommend acceptance in the current form, but I see a clear path to a defensible revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper deserves a serious referee. The core formulation is new: REDCLIFF-S models a dynamic Granger causal graph as a history-conditioned weighted superposition of static nonlinear factor graphs, with optional supervised labels. That combination is absent from the cited literature, which is mostly linear factor models (dCSFA, switching linear dynamical systems) or discrete-state nonlinear trackers (SDCI, SNNTs). The authors did the empirical legwork: Synthetic Systems and D4IC benchmarks show consistent F1 and ROC-AUC gains over static and linear baselines, with standard errors, ablations that demonstrate the multi-factor and supervised components matter, and public code. That is a real contribution for the neuroscience hypothesis-generation niche.\n\nThe soft spots are real but mostly addressable. The paper's own Appendix A.1 proves that even a two-node ReLU/nReLU system is not identifiable. To their credit, the authors are upfront about this and frame the method as hypothesis generation, but the abstract's claim that the method “enables the detection of complex, time-varying interactions” oversells what is possible. More importantly, the synthetic benchmarks are in-distribution: the true generators are weighted sums of static factor graphs, exactly the REDCLIFF-S model class. So the experiments show recovery inside the hypothesis class, not robustness when the true dynamics continuously morph or when a small set of reusable factors does not exist. The Section 1 claim that “we make no assumptions regarding the underlying generative processes” is contradicted by Equation 4 itself; this should be rephrased as a modeling choice.\n\nThe evaluation has two smaller weaknesses. Optimal F1 with oracle thresholds is generous, though the ROC-AUC tables partially compensate; threshold-free or threshold-averaged metrics (AUPRC, average precision) would be more persuasive. Hyperparameters are tuned on a single repeat, which the authors openly acknowledge in B.4. The real-data TST and SP analyses produce plausible behavior-linked connections, including the lateral habenula finding, but without ground truth these are hypothesis illustrations, not validation; the supervised construction does invite circularity between labels and factor weights.\n\nBottom line: this is a solid methodology paper for a within-subfield audience, not a breakthrough. The method is defensible as a hypothesis generator when the superposition assumption holds, and the authors are unusually candid about limitations. With tempered claims, threshold-free metrics, and ideally one synthetic setting outside the model class, it would be a strong contribution. I would send it to peer review and recommend acceptance after revision.","headline":"A genuinely new dynamic-graph factor model with honest limitations; the synthetic wins are real but in-distribution, and the paper oversells the no-assumption framing.","tokens_in":27328,"tokens_out":1642,"would_cite":true,"duration_ms":20058,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62M10","62D20","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Dynamic causal graphs in time series can be recovered as behavior-weighted mixtures of a few static nonlinear factor graphs, and the REDCLIFF-S model recovers them with 22–28% higher F1 scores than baselines on synthetic benchmarks.","keywords":["dynamic causal discovery","Granger causality","nonlinear time series","factor models","hypothesis generation","behavioral state labels","neuroscience","state-dependent dynamics"],"falsifier":"Run the method on a system whose causal edge strengths drift continuously through many values rather than switching between a few patterns; if REDCLIFF-S's fixed factors cannot track the drift at better than static-baseline accuracy, the superposition model is the reason. A sharper test is the paper's own two-node counterexample, in which ReLU and nReLU edges become observationally equivalent: Appendix A.1 argues no method can identify which state generated a sample, and verifying empirically that REDCLIFF-S assigns such samples at chance level would confirm the stated limit rather than a recoverable signal.","tokens_in":26291,"feed_emoji":"🧠","tokens_out":12735,"duration_ms":110316,"temperature":0.7,"pith_summary":"The paper's aim is to show that time-varying causal structure, of the kind that switches as the brain changes behavioral state, can be recovered from observed time series by writing the unknown system as a weighted sum of a few reusable static interaction graphs, each of which may be nonlinear. The proposed model, REDCLIFF-S, learns the interaction graphs by forecasting the next observation and learns the per-moment weights from the recent history of the signal, optionally supervised by behavioral labels. This matters because current methods assume either linear relationships or a single fixed graph, leaving hypothesis generation blind to state-dependent neural wiring. On synthetic systems and a multi-state adaptation of DREAM4, the paper reports average F1-score improvements of roughly 22–28% over those baselines, and on two real local-field-potential datasets it produces behavioral-state-specific wiring hypotheses. The authors themselves prove in an appendix that such systems are generically not identifiable, and they argue that hypothesis generation remains useful despite this.","feed_headline":"Dynamic causal graphs recovered 22–28% better than baselines","feed_subtitle":"Mixing a few static nonlinear graphs, weighted by behavior, turns LFP recordings into testable wiring hypotheses.","key_machinery":"The load-bearing identity is the conditionally weighted superposition $\\Phi(t) = G(t) \\odot^\\dagger F(t) = \\sum_k G(t)_k F_k(t)$, where $\\odot^\\dagger$ denotes broadcast multiplication of the score vector $G(t)$ against the tensor of factor outputs. The identity splits estimation into two tasks: each factor $f_{\\phi_k}$ is a cMLP whose first-layer weights are time-lagged adjacency matrices $\\hat{A}_{:,:,t}$, so a factor literally is an inspectable causal graph; and the state model $g_\\theta$ maps a longer history window to nonnegative scores, deciding which factor dominates at each time step. A cosine-similarity penalty between factor graphs enforces the prior that factors should serve different purposes, and in REDCLIFF-S an invertible map lets a subset of scores double as predicted behavioral labels, so the factorization carries the hypothesis directly.","core_discovery":"REDCLIFF-S claims that a dynamic causal graph can be modeled as $\\Phi(t) = \\sum_{k=1}^{K} G(t)_k F_k(t)$: a small number of static factor functions $F_k$, each mapping past variable values to present values and each yielding an explicit time-lagged adjacency matrix of Granger causal weights, combined by history-conditioned scores $G(t)$ produced by a state model $g_\\theta$. Because each factor is a component-wise MLP, the relationships inside a factor can be nonlinear, and because the scores vary with time, the composite graph can change direction and strength without assuming a discrete switch between states. In the supervised variant, behavioral labels are attached to particular factor scores through invertible maps, so each factor's adjacency matrix can be read as the candidate wiring hypothesis for that behavior. The paper reports that this construction outperforms static nonlinear, dynamic-linear, and regime-switching baselines in F1 and ROC-AUC on synthetic data and on D4IC, a five-fold mixture version of DREAM4, and it demonstrates factor graphs for Open Field, Home Cage, Tail Suspension, and Social Preference states that reproduce published findings such as lateral habenula involvement in stress.","pith_inferences":["If real neural dynamics only approximate the superposition assumption, factor graphs may still serve as a low-dimensional grammar of the dominant interaction modes, with the scores acting as a behavioral readout; under misspecification the model would remain a dimension-reduction tool rather than an exact recovery method.","The cosine-similarity penalty is a proxy for functional dissimilarity of graphs, and a testable extension would replace it with a sparsity prior on factor usage to check whether rare behavioral states are recovered better.","The TST hypothesis that the anterior-lateral habenula drives several regions in the stressful Open Field condition is directly interventional: silencing the lateral habenula during tail suspension would test whether that edge carries causal weight.","Because the model needs no interventional data and accepts any scalar time series with global state labels, it could transfer to sleep staging, anesthesia depth monitoring, or economic regime detection, wherever states mark distinct dynamics."],"forward_implications":["Each trained factor yields a fixed, inspectable adjacency matrix, so a REDCLIFF-S run outputs concrete candidate edges, with direction, lag, and strength, that can seed interventional studies such as dynamic causal models.","Attaching behavioral labels to factor scores means the output includes a wiring diagram per behavior rather than one averaged graph; on the TST data this produces distinct Open Field, Home Cage, and Tail Suspension hypotheses from a single training run.","Because factors are nonlinear, the approach beats linear dynamic factor models on systems with ReLU and nReLU edges, where the ground truth is unrepresentable by any linear model.","The ablation results attribute most of the gain to multiple factors, learned scores, and the supervised term, while the cosine-similarity penalty helps on sparse systems and can be dropped on dense ones.","The authors' non-identifiability proof implies that no method of this family can uniquely recover true factor graphs in general, so the practical claim is narrower: the recovered graphs are useful hypotheses even when not unique."],"supporting_citations":[{"why":"Supplies the cMLP architecture whose first-layer weights become time-lagged adjacency matrices, used as each factor's base unit and as static nonlinear baselines.","marker":"Tank et al. (2021)"},{"why":"Regime-PCMCI, the regime-dependent causal discovery method that serves as the principal baseline for dynamic graph recovery.","marker":"Saggioro et al. (2020)"},{"why":"DREAM4, the benchmark adapted into the D4IC multi-state dataset used to test recovery under noise.","marker":"Marbach et al. (2009)"},{"why":"The multi-region local-field-potential Tail Suspension Test dataset used for the main real-data case study.","marker":"Carlson et al. (2023)"},{"why":"The Social Preference dataset and the earlier factor-model hypothesis generation whose behavioral-label convention the paper extends.","marker":"Mague et al. (2022)"},{"why":"Supervised autoencoders for behavior-linked brain networks, providing the supervised-label strategy and the published anxiety-network result the TST findings are compared with.","marker":"Talbot et al. (2023)"},{"why":"The tidybench regression-based baselines (LASAR, QRBS, SLARAC) used in the supervised edge-detection comparison.","marker":"Weichwald et al. (2020)"},{"why":"SDCI, the source of the state-determined setting that motivates conditioning factor weights on history alone.","marker":"Balsells-Rodas et al. (2022)"},{"why":"Supplies the identifiability definition used in the paper's own proof that nonlinear dynamic causal graphs are generically unidentifiable.","marker":"Maclaren & Nicholson (2019)"}],"fun_headline_variants":["Dynamic causal graphs decoded via nonlinear superposition","State-weighted graph soup beats causal baselines by 28%","From LFP to wiring hypotheses with factored graphs","Nonlinear graph mixture lifts causal F1 by up to 60%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The recorded system must actually be expressible as a small set of reusable interaction patterns whose relative importance at each moment is set by recent history; if the true interactions instead morph continuously into new forms, the factor graphs and scores have no faithful target to recover.","fun_headline_variants_meta":{"raw":{"variants":["Dynamic causal graphs decoded via nonlinear superposition","State-weighted graph soup beats causal baselines by 28%","From LFP to wiring hypotheses with factored graphs","Nonlinear graph mixture lifts causal F1 by up to 60%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000494,"raw_usage":{"total_tokens":2438,"prompt_tokens":973,"completion_tokens":1465,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":589,"completion_tokens_details":{"reasoning_tokens":1400}},"tokens_in":589,"tokens_out":1465,"duration_ms":12779,"temperature":1.0,"reasoning_tokens":1400,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:48:53.870830+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the method on a system whose causal edge strengths drift continuously through many values rather than switching between a few patterns; if REDCLIFF-S's fixed factors cannot track the drift at better than static-baseline accuracy, the superposition model is the reason. A sharper test is the paper's own two-node counterexample, in which ReLU and nReLU edges become observationally equivalent: Appendix A.1 argues no method can identify which state generated a sample, and verifying empirically that REDCLIFF-S assigns such samples at chance level would confirm the stated limit rather than a recoverable signal.","supporting_citations":[{"cited_title":"Reconstructing regime-dependent causal relationships from observational time series","cited_arxiv_id":null,"evidence_quote":"Regime-PCMCI, the regime-dependent causal discovery method that serves as the principal baseline for dynamic graph recovery."},{"cited_title":"Generating realistic in silico gene networks for performance assessment of reverse engineering methods","cited_arxiv_id":null,"evidence_quote":"DREAM4, the benchmark adapted into the D4IC multi-state dataset used to test recovery under noise."},{"cited_title":"Multi-region local field potential recordings during a tail-suspension test","cited_arxiv_id":null,"evidence_quote":"The multi-region local-field-potential Tail Suspension Test dataset used for the main real-data case study."},{"cited_title":"D., Talbot, A., Blount, C., Walder-Christensen, K","cited_arxiv_id":null,"evidence_quote":"The Social Preference dataset and the earlier factor-model hypothesis generation whose behavioral-label convention the paper extends."},{"cited_title":"Estimating a brain network predictive of stress and genotype with supervised autoencoders","cited_arxiv_id":null,"evidence_quote":"Supervised autoencoders for behavior-linked brain networks, providing the supervised-label strategy and the published anxiety-network result the TST findings are compared with."},{"cited_title":"Causal discovery from conditionally stationary time series","cited_arxiv_id":null,"evidence_quote":"SDCI, the source of the state-determined setting that motivates conditioning factor weights on history alone."}],"review_version":1}