{"id":"df2d5ad1-6608-4400-8ec1-da35ba40d8e7","arxiv_id":"2511.13899","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"DisRNN, a VAE-based low-rank RNN with a group-wise independence penalty, learns disentangled latent trajectories and interpretable sub-connectivity from neural population recordings.","lead":"Neuroscientists often compress recordings of many neurons into a few latent variables, but those variables are entangled. This paper introduces DisRNN, a low-rank recurrent neural network with a variational-autoencoder twist, which encourages latent variables to split into independent groups and yields a connectivity matrix broken into group-level sub-circuits.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PC penalty (Eq. 10) enforces marginal, not process-level, group independence; lagged cross-group dependencies can persist, so 'independent latent dynamics' and sub-connectivity interpretations are not established.","rationale":"The paper reformulates low-rank RNNs in a VAE framework and adds a partial-correlation penalty to encourage group-wise disentanglement. The central claim is that DisRNN uncovers independent latent dynamics and interpretable sub-connectivities. The reader's conditional verdict focuses on identifiability: the penalty might recover an arbitrary near-independent affine transform rather than the true groups. My stress-test identifies a more specific, internal gap: the penalty in Eq. (10) is applied to the aggregated posterior q(z) = (1/T)Σ_t q(z|x_t), which is a marginal distribution over time. It vanishes when the marginal distribution factorizes, not when the latent stochastic processes are independent. Process independence is strictly stronger and is what the paper's language ('evolve separately', 'independent latent dynamics') promises. The definition in Eq. (9) explicitly uses p(z), the marginal density of the latent vector, not the process distribution p(z_{1:T}), so there is an inconsistency between the stated goal and the formal objective. This is not a mere philosophical concern: a model could have zero KL for the PC term while z_g(t) depends on z_{g'}(t−l), in which case the sub-connectivity decomposition W_g in Eq. (12) does not correspond to independently evolving sub-circuits. The reported evaluations—instantaneous PC, affine alignment R², and rotated-coordinate checks—are all zero-lag and would not detect lagged cross-dependencies. The synthetic experiment uses ground-truth latent processes that are truly independent, so if DisRNN recovers them, lagged independence would hold, but the paper never reports a lagged independence metric. This is directly testable: compute lagged cross-group mutual information on the existing synthetic latents or construct a synthetic dataset with marginal-but-not-process independence. If the concern lands, the main claim is not supported as stated. If the check passes, the paper gains crucial support. I therefore recommend a conditional verdict, consistent with the reader's, but with this sharper condition. The reader's identifiability concern and my temporal-independence concern are related (both concern the gap between penalty and true independence) but distinct, hence 'partial'. No ad hominem; the issue is in the argument and evidence, not the authors.","tokens_in":14208,"tokens_out":9989,"duration_ms":99421,"concrete_test":"On the synthetic dataset of Sec. 4.1, compute for each random seed the lagged cross-group dependence between estimated latent groups, e.g., cross-correlation or mutual information I(z_g(t); z_{g'}(t−l)) for l=1..L, and compare to the ground-truth latent processes. A sharper test: generate a synthetic dataset with two latent groups that are marginally independent but have lagged cross-coupling, e.g., z_1(t)=α z_2(t−1)+ε(t) and z_2(t) i.i.d., generate x(t) via Eq. (6), fit DisRNN with G=2, and measure whether the inferred latent groups retain the lagged coupling. If DisRNN's PC is near zero while lagged cross-dependence remains significant, the penalty does not enforce independent latent dynamics, and the central claim fails.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The disentanglement objective in Eq. (10) is L = (1/T)Σ ELBO − β·KL(q(z) || Π_g q(z_g)), where q(z) = (1/T)Σ_t q(z|x_t) is the aggregated posterior over time. This KL term measures only contemporaneous dependence: it vanishes when the marginal distribution of the latent state (averaged over time) factorizes across groups. It contains no term penalizing cross-group dependencies at nonzero time lags. The Abstract and Sec. 3.2 claim DisRNN recovers 'independent latent dynamics' whose groups 'evolve separately', and Eq. (12) interprets W_g = A_{:,g}B_{g,:} as functionally independent sub-circuits. Process independence requires p(z_{1:T}) = Π_g p(z_{g,1:T}), which is strictly stronger than marginal factorization. A model can achieve KL≈0 while z_g(t) depends on z_{g'}(t−l) for l>0; for example, if z_1(t)=αz_2(t−1)+ε(t) and z_2(t) is i.i.d., then z_1(t) and z_2(t) are marginally independent at each t but the processes are coupled. The evaluations in Sec. 4.1 (PC, instantaneous R² alignment) and Sec. 4.2 (rotated-coordinate check) are zero-lag and cannot detect such temporal cross-talk. Thus the reported evidence does not establish the central claim as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DisRNN, a VAE-based reformulation of low-rank RNNs in which a penalty on the KL divergence between the aggregated posterior and the product of its group marginals encourages group-wise independence among latent dimensions. The resulting model yields a decomposition W=AB=Σ_g A_{:,g}B_{g,:}, whose terms are interpreted as sub-connectivities associated with independent latent groups. The authors evaluate the method on a synthetic Lorenz/Thomas dataset, macaque M1 spiking data, and mouse dorsal cortex voltage imaging, comparing against lrRNN, LINT, SMC, and variants. They report improved latent alignment, better disentanglement, and more interpretable connectivity decompositions. The framework is coherent as a variational inference scheme, and the paper includes useful ablations and a rotated-coordinate check on M1. However, the central claim that DisRNN uncovers 'independent latent dynamics' is not fully supported by the training objective or the evaluations, and the synthetic data generation is under-specified.","tokens_in":14627,"tokens_out":7834,"duration_ms":80265,"significance":"If the central claim holds, the paper would make a useful contribution by extending low-rank RNN decompositions from orthogonal (SVD-based) to independent groups, enabling functional assignment of neural sub-circuits. The VAE reformulation is clean and the idea of partial disentanglement (rather than full dimension-wise independence) is well motivated. The synthetic ground-truth alignment and the rotated-coordinate analysis are constructive, and the β ablation is helpful. However, the manuscript currently does not establish the process-level independence it advertises, and the synthetic experiment—the main external validation—needs clarification. The paper also frankly acknowledges its own limitations (pre-specified group counts, linear connectivity, no external inputs), which is to its credit. Overall, the framework is promising but the evidence does not yet support the strongest claims.","major_comments":[{"comment":"The independence penalty in Eq. (10) is KL(q(z)||Π_g q(z_g)), where q(z) is the time-averaged aggregated posterior. This is a marginal, zero-lag independence criterion: it can vanish even when z_g(t) depends on z_{g'}(t−l) for l>0. Eq. (9) also defines disentanglement through the marginal p(z), not the full process p(z_{1:T}). Therefore the Abstract's claim that DisRNN recovers 'independent latent dynamics' that 'evolve separately' is not supported by the training objective. The evaluations in Sec. 4.1 (PC, R² alignment) and Sec. 4.2 (rotated-coordinate alignment) are all zero-lag and cannot detect temporal cross-talk. Please add a lagged-dependence metric (e.g., lagged cross-correlation or transfer entropy) or modify the penalty/architecture to enforce process-level independence (e.g., a block-diagonal latent transition), and adjust the claims accordingly.","section":"Sec. 3.2, Eq. (10), Eq. (9), Abstract"},{"comment":"The synthetic generation procedure is underspecified and appears contradictory. The text first states 'To generate the latent, we simulate Lorenz and Thomas’ cyclically symmetric dynamics,' then says 'Since z(t) follows the generative process in Eq. (7), we can get the parameters including A, B, and hence generate the observed data x(t) using the fitted A and B.' A Lorenz/Thomas trajectory does not, in general, satisfy Eq. (7) for arbitrary A,B; conversely, if z(t) is generated from Eq. (7), it is unclear how the Lorenz/Thomas dynamics are used. This ambiguity undermines the interpretation of the ground-truth ‘recovery’ results. Please provide the exact generative scheme, including how A and B are chosen and how z(t) is made consistent with Eq. (7), or re-run the synthetic experiment with a properly specified generative process.","section":"Sec. 4.1, Dataset"},{"comment":"The decomposition W=AB=Σ_g A_{:,g}B_{g,:} is an algebraic identity for any A and B; it does not by itself imply that each W_g is a 'sub-circuit' driving an independent source. That interpretation requires process-level independence of the latent groups, which is not guaranteed by the marginal PC penalty in Eq. (10). Consequently, the real-data connectivity interpretations in Secs. 4.2 and 4.3 (e.g., horizontal vs. vertical sub-networks, somatosensory sub-circuits) are post hoc descriptive decompositions rather than established independent sub-circuits. The paper should either enforce process-level independence (e.g., using bdRNN's block-diagonal J on the linear case) or explicitly frame W_g as an algebraic decomposition whose functional independence remains to be verified.","section":"Sec. 3.2, Eq. (11)–(12)"}],"minor_comments":[{"comment":"The quantity KL(q(z)||Π_g q(z_g)) is the total correlation (multi-information) between groups, not a 'partial correlation' as named. Please use terminology that matches the quantity or justify the name.","section":"Eq. (10)"},{"comment":"The text refers to 'our code for details' but no code or data link is provided. Given the complexity of the method and the need for reproducibility, please include a public release or a clear availability statement.","section":"Sec. 4.2"},{"comment":"The paper concedes that the K=2 trajectories 'remain visually poor' despite an alignment R² of about 0.65. Please address the visual quality (e.g., different scaling, plotting choices) and ensure the quantitative claim is not overstated in the abstract or main text.","section":"Sec. 4.2, K=2"},{"comment":"The combinatorial count in Eq. (19) appears off by a factor of G!. For K=GH components partitioned into G unlabeled groups of size H, the standard formula is K!/(G!(H!)^G), not the expression given (which evaluates to K! G!/(H!)^G). Please verify the calculation.","section":"Appendix A.2.1, Eq. (19)"},{"comment":"Interpretations such as 'potentially indicating the formation of episodic memory of receiving the airpuff' are highly speculative. Please phrase these as hypotheses and avoid overstating the evidence from a single voltage-imaging dataset.","section":"Sec. 4.3"}],"recommendation":"major_revision","confidential_remarks":"The central weakness is that the headline claim—'independent latent dynamics'—is stronger than what the objective enforces. The synthetic ground-truth experiment, which could have resolved this, is too vaguely described to be an adequate external check. I believe the framework is promising and the issues are fixable: the authors could add a lagged-dependence analysis, clarify or redo the synthetic generation, and soften the sub-circuit language. If those points are addressed, the paper could become a solid contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a competent methodological paper that does something genuinely useful—reformulates low-rank RNNs in a VAE framework and adds a group-wise partial-correlation penalty to encourage disentangled latent groups, with a clean decomposition of connectivity into sub-connectivities. The synthetic ground-truth experiment is a real check, and the rotated-coordinate alignment on M1 data is a nice touch. But the paper's central claim is stronger than what it actually enforces. The penalty in Eq. (10) acts on the aggregated posterior over time, so it only kills contemporaneous dependence between groups. It does nothing to prevent lagged cross-group coupling. The abstract and Section 3.2 say the groups 'evolve separately' and are 'independent sources,' which is process-level independence—strictly stronger than the marginal factorization the KL term enforces. The stress-test example (z1(t) depending on z2(t−1) while being marginally independent at each time) is a real counterexample. The empirical evaluations—PC, instantaneous R² alignment, the rotated-coordinate check—are all zero-lag, so they wouldn't detect this kind of temporal cross-talk. On the synthetic data the inferred latents align well with true groups, so the issue may not bite there, but the paper doesn't demonstrate it, and on real data the interpretation of sub-connectivities as independent sub-circuits is not warranted by the evidence.\n\nOther soft spots are less severe. No code or data release makes the M1 and mouse results hard to evaluate. The beta and group counts are hand-chosen, and the same PC metric used for training is also used for evaluation, so there's some circularity. The K=2 M1 trajectories are admittedly poor, which the paper acknowledges. The real-data connectivity stories are post hoc. Also the abstract calls the model FacRNN while the body uses DisRNN—sloppy. The reliance on Li et al. (2025) for the penalty is fine but means the novelty is more about the combination than the components.\n\nWhat's genuinely good: the VAE reformulation is clean, the sub-connectivity decomposition W=ΣA_g B_g is useful, the ablation of beta and group ranks is honest, and the conclusion thoughtfully lists limitations. This is a solid contribution for computational neuroscience audiences who want interpretable low-rank RNNs. It deserves a serious referee, but the authors should be asked to either weaken the independence claims or add a lagged-dependence check (e.g., cross-correlation or Granger causality between inferred groups) and tighten the writing.","headline":"Useful VAE-based method for group-wise disentanglement of low-rank RNN latents, but the paper overclaims process-level independence from a marginal penalty.","tokens_in":15127,"tokens_out":2809,"would_cite":true,"duration_ms":25873,"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 a low-rank recurrent neural network, reframed as a variational autoencoder with a partial-correlation penalty, can disentangle neural latent dynamics into independent groups and decompose the connectivity matrix into i","keywords":["low-rank recurrent neural network","latent disentanglement","partial correlation","variational autoencoder","neural population dynamics","connectivity decomposition","motor cortex","voltage imaging"],"falsifier":"A concrete check would be to generate synthetic data from a low-rank RNN with known but entangled latent groups (i.e., no true independent groups) and see if the proposed model still forces a group-wise decomposition with high confidence; if it does, the penalty can create artifacts. Alternatively, on real data, compare the recovered sub-connectivities to known synaptic or anatomical connectivity via independent measurements; a mismatch would challenge the sub-circuit interpretation.","tokens_in":14105,"feed_emoji":"🧠","tokens_out":4514,"duration_ms":37367,"temperature":0.7,"pith_summary":"This paper tries to show that neural population recordings can be explained by a low-rank recurrent network whose latent dynamics split into a few independent groups, each with its own sub-connectivity. The authors reformulate the low-rank RNN as a variational autoencoder and add a partial-correlation penalty that pushes the latent groups toward independence without forcing individual dimensions to be independent. They claim this yields better disentanglement of neural trajectories and interpretable connectivity decompositions on synthetic, monkey motor cortex, and mouse voltage-imaging data. The importance is that it offers a way to assign distinct computational roles to separate neural sub-circuits.","feed_headline":"Low-rank RNN with partial correlation penalty reveals neural sub-circuits","feed_subtitle":"Group-wise independence in latent dynamics improves interpretability and matches brain data better than standard low-rank RNNs.","key_machinery":"The central mechanism is the partial-correlation penalty added to the evidence lower bound objective of a variational autoencoder reformulation of the low-rank RNN. The penalty, scaled by a hyperparameter beta, penalizes the KL divergence between the aggregated posterior q(z) and the product of group marginals prod_g q(z_g). This encourages group-wise independence without forcing individual dimensions within a group to be independent. The second key piece is the decomposition of the learned low-rank connectivity into group-specific sub-connectivities W_g = A_{:,g}B_{g,:}, which gives the interpretation of separate neural sub-circuits.","core_discovery":"The central claim is that group-wise independence, rather than full component-wise independence, is the right structural assumption for latent dynamics in low-rank RNNs. By expressing the low-rank RNN in a variational autoencoder framework, the paper introduces a partial-correlation penalty on the aggregated posterior that encourages independence between groups of latent dimensions while allowing flexible within-group entanglement. This allows the rank-K connectivity W = AB to be decomposed into sub-connectivities W_g = A_{:,g}B_{g,:}, each associated with a distinct latent group, which the paper interprets as neural sub-circuits. Experiments on synthetic data, macaque M1 recordings, and mou","pith_inferences":["If group-wise independence holds, the partial-correlation penalty might be seen as a form of identifiable factor analysis for dynamical systems, but identifiability beyond linear-Gaussian cases remains an open question the paper does not settle.","The paper's evidence for biological plausibility (preferred cardinal axes in M1) suggests a testable prediction: the recovered sub-connectivities should align with known anatomical or functional connectivity, which could be validated with optogenetic or pharmacological perturbations.","The method might be extended to incorporate task inputs or stimulus labels, which the paper notes as a limitation; such extensions could reveal input-driven changes in sub-circuit connectivity."],"forward_implications":["Neural circuits can be decomposed into independent sub-circuits that drive separate behavioral or sensory variables, enabling circuit-level hypotheses from population recordings.","Low-rank RNN models can be made more interpretable without sacrificing reconstruction accuracy, since group-wise independence allows higher within-group rank.","The approach provides a way to identify which latent groups correspond to task-relevant variables (e.g., x/y movement coordinates) even without explicit labels.","The framework extends to nonlinear encoders/decoders and general nonlinear dynamical systems, suggesting broad applicability beyond linear connectivity."],"fun_headline_variants":["FacRNN splits latent dynamics into independent neural circuits","Partial correlation penalty reveals independent latent groups in RNNs","Group-wise independence uncovers neural sub-circuits in low-rank RNNs","Factored RNN: independent latent groups for interpretable brain dynamics","Low-rank RNN with independent groups reveals sub-circuit connectivity"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the true neural population actually has separable latent groups—that minimizing the partial-correlation penalty recovers the true independent groups rather than an arbitrary near-independent transformation of them, and that the model's group structure matches a real biological decomposition.","fun_headline_variants_meta":{"raw":{"variants":["FacRNN splits latent dynamics into independent neural circuits","Partial correlation penalty reveals independent latent groups in RNNs","Group-wise independence uncovers neural sub-circuits in low-rank RNNs","Factored RNN: independent latent groups for interpretable brain dynamics","Low-rank RNN with independent groups reveals sub-circuit connectivity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000713,"raw_usage":{"total_tokens":3019,"prompt_tokens":697,"completion_tokens":2322,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":441,"completion_tokens_details":{"reasoning_tokens":2236}},"tokens_in":441,"tokens_out":2322,"duration_ms":15366,"temperature":1.0,"reasoning_tokens":2236,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T21:41:51.907226+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete check would be to generate synthetic data from a low-rank RNN with known but entangled latent groups (i.e., no true independent groups) and see if the proposed model still forces a group-wise decomposition with high confidence; if it does, the penalty can create artifacts. Alternatively, on real data, compare the recovered sub-connectivities to known synaptic or anatomical connectivity via independent measurements; a mismatch would challenge the sub-circuit interpretation.","supporting_citations":[],"review_version":1}