{"id":"d9898eae-112b-49cf-a512-ba8d56073034","arxiv_id":"2505.04331","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"PNBA reports zero-shot cross-subject neural-behavioral alignment in monkey M1/PMd and mouse V1, with preserved representation similarity decaying from trials to sessions to subjects.","lead":"The paper presents PNBA, an algorithm that maps brain activity and behavior into a shared latent space, and uses it to claim that neural representations are preserved across individuals in motor and visual cortex without per-subject calibration. The result matters because cross-subject preservation would support calibration-free brain-computer interfaces, but the similarity may be largely a consequence of the training objective.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cross-subject similarity is an optimization target, not a discovery: PNBA's shared encoder and neural-behavioral matching loss (Eqs.","rationale":"The reader's weakest assumption and mine coincide: high latent similarity is a consequence of the matching objective, not evidence of biological preservation. The paper's baseline and ablation results (Table 1, Fig. 4a) show the method is effective at neural-behavioral alignment, and the code release is a plus, but the scientific claim requires an independent measurement of neural representational preservation. Because the loss explicitly minimizes the distance between neural and behavioral latent distributions, and because identical behavior maps to the same behavioral code, cross-subject neural latent similarity is expected. The absence of a null control is internal to the design, not a disagreement with field consensus. The PMd test set also includes a training subject, which independently weakens the 'across cortices' claim. A subject-specific behavioral encoder control would settle whether the shared behavior code is responsible; consequently the verdict should remain REJECT.","tokens_in":26293,"tokens_out":14132,"duration_ms":158905,"concrete_test":"Train a control PNBA with the same architecture and hyperparameters but replace the shared behavioral encoder gϕ with a separate, subject-specific behavioral encoder for each training subject, keeping the neural encoder fθ shared across subjects. Then evaluate on the same held-out subjects and recompute the cross-subject latent correlations for matched behavior reported in Fig. 5c (M1) and Fig. 6c (V1). If these correlations collapse toward chance, the reported preservation is explained by the shared behavioral target and the central claim fails; if they remain near 0.939/0.892, the shared neural encoder is extracting structure that does not depend on the shared behavior code.","verdict_should_be":"UNCHANGED","load_bearing_attack":"PNBA's central claim is that zero-shot cross-subject correlations (M1 R=0.939, V1 R=0.892) reveal preserved neural representations across cortices and species. The experimental design cannot support that inference because the architecture and loss make high cross-subject similarity a direct training target. Section 3.2.3 states that a single shared network processes all subjects, and Eq. 5 includes L_ProbMatch (Eq. 1), which minimizes d(fθ(x), gϕ(y)) between neural and behavioral latent distributions. For identical behavior y presented to two subjects, the objective pulls fθ(x_A) and fθ(x_B) toward the same gϕ(y); once alignment converges, high cross-subject correlation is expected. The zero-shot numbers therefore demonstrate that a behavior-predicting mapping generalizes to novel subjects, not that the brain's original population representations are preserved independently of the training objective. No null or ablation controls this: the matched-vs-mismatched t-tests (Fig. 3b) compare within a trained alignment, and the reported R values are never compared to a control in which cross-subject similarity is equally enforced but neural-behavioral content is removed. Theorem 3.1 only bounds distances inside the trained model; it does not establish that cross-subject similarity exceeds the loss-imposed ceiling. A compounding concrete flaw: the PMd 'zero-shot' test set (Table 3) includes Monkey M sessions while other Monkey M sessions appear in training, so the PMd zero-shot claim is not strictly valid.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PNBA, a probabilistic neural-behavioral representation alignment framework that uses shared encoders and generative constraints to align neural activity with behavioral variables in a common latent space. The authors evaluate PNBA on monkey primary motor cortex (M1) and dorsal premotor cortex (PMd) during reaching, and on mouse primary visual cortex (V1) during visual stimulation. They report high within-session, cross-session, and zero-shot cross-subject representational correlations (e.g., M1 R=0.939, V1 R=0.892), which they interpret as evidence of preserved neural representations across cortices and species. The paper also demonstrates zero-shot V1-guided movement decoding as a practical application.","tokens_in":26673,"tokens_out":5634,"duration_ms":54387,"significance":"If the central claim were established, this would be a significant advance for systems neuroscience and for calibration-free brain-computer interfaces: it would suggest that population codes are shared across individuals without subject-specific calibration. The paper has concrete strengths: a detailed architecture description, public code, evaluation on three datasets, and comparisons to several baselines. However, the central inference is not independently tested. As I detail below, the cross-subject similarity is largely a consequence of the training objective rather than a discovery about the brain, and one of the zero-shot test sets is contaminated. The reported correlations therefore do not support the paper's main scientific claim.","major_comments":[{"comment":"The central claim of preserved neural representations is circular because cross-subject similarity is an optimization target by construction. Section 3.2.3 states that a single shared network processes all subjects, and Eq. (5) includes L_ProbMatch (Eq. 1), which minimizes d(f_theta(x), g_phi(y)) between neural and behavioral latent distributions for matched pairs. For identical behavior y presented to two subjects, both f_theta(x_A) and f_theta(x_B) are pulled toward the same g_phi(y), so high cross-subject correlation (Fig. 5c: R=0.939; Fig. 6c: R=0.892) is expected once alignment converges. The matched-vs-mismatched t-tests (Fig. 3b, Supplementary D.3) only demonstrate that the trained latent is behavior-discriminative; they do not show that cross-subject similarity exceeds the level enforced by the loss. A null control is needed, such as shuffling behavior labels across subjects or training without L_ProbMatch, to establish that the observed cross-subject correlation is not a direct consequence of the training objective.","section":"Section 3.2.3, Eq. (5)"},{"comment":"The PMd zero-shot test set is contaminated because it includes sessions from Monkey M, a subject present in the training set. Table 3 lists test sessions M-CO-20150623 and M-CO-20150625 from Monkey M, while training sessions from Monkey M include M-CO-20140203 through M-CO-20150615. Thus the PMd zero-shot result (Fig. 10c: R=0.856) is at best a held-out-session test within a known subject, not a cross-subject generalization. This undermines the abstract's claim of zero-shot validation in PMd and the broader 'across cortices and species' conclusion.","section":"Table 3"},{"comment":"The proofs of the theoretical guarantees are not rigorous and contain unjustified steps. In part (i), the argument that p(x|z_const) must approach zero for all x under a normalization constraint is not valid for continuous densities: a density can be finite pointwise while still integrating to one, so the claimed limit L_total -> +inf is not established. In part (ii), the existence of constants delta_1 < delta_2 is assumed without derivation, so the claimed eta > 0 is not proven. In part (iii), the lower-bound argument asserts that if two distinct inputs map to arbitrarily close latent codes, then log p(x_i|z) + log p(x_j|z) -> -inf, which depends on the specific likelihood parameterization and is not generally true. Since Theorem 3.1 is used to justify the framework's non-degeneracy and representation stability, these proofs need to be corrected or the claims softened.","section":"Supplement B, Theorem B.1 (Theorem 3.1)"}],"minor_comments":[{"comment":"The heading 'CROSS-SUBJECT NETWORK FOR VAIRABLE NEURAL ACTIVTIY' contains typos; it should be 'VARIABLE' and 'ACTIVITY'.","section":"Section 3.2.3 heading"},{"comment":"The reported p-values of exactly 0.0 are not meaningful; the authors should report actual small values or bounds (e.g., p < 1e-300) instead of p=0.","section":"Supplementary D.3"},{"comment":"The phrase 'Codes are availiable' should read 'Codes are available'.","section":"Contributions list"},{"comment":"The definition of the joint approximate posterior q(z|x,y) is unclear as written: the notation 'for x' and 'for y' suggests different distributions depending on the modality being reconstructed, but the derivations in Eqs. (43)-(46) use q(z|x,y) in both directions. Please clarify the exact form used in practice.","section":"Equation (9)"}],"recommendation":"reject","confidential_remarks":"The paper's central claim is circular: cross-subject similarity is directly enforced by the shared-network architecture and the neural-behavioral matching loss, so the reported zero-shot correlations do not constitute evidence for biologically preserved neural representations. The PMd test-set contamination is a separate, clear data-split error. The method may have value as an engineering contribution for zero-shot decoding, but the neuroscience framing and the main claim would need substantial rework, including new control experiments and a reframing of what the results can establish. I recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First: this paper is a solid engineering contribution wrapped in an overclaimed scientific narrative. The claim that zero-shot cross-subject correlations (M1 R=0.939, V1 R=0.892) reveal 'preserved neural representations across cortices and species' is not supported by the design, because the architecture and loss make cross-subject similarity a direct training target. The shared encoder and the probabilistic matching term in Eq. 5 are doing exactly that work. This is a real problem, not a nitpick.\n\nWhat is genuinely new: PNBA is a clean way to align neural and behavioral latents with generative constraints, and the paper demonstrates that the alignment generalizes to held-out subjects—and even to a different cortical area (V1) and recording modality (calcium imaging) than the motor cortex data it was designed for. The zero-shot V1 decoding (R^2 ~0.88) is a useful existence proof for calibration-free decoding. The authors also ship code. Those are real contributions.\n\nThe soft spots, in order of severity. First, the circularity: Eq. 1-5 explicitly pull f(x_A) and f(x_B) toward the same g(y) for matched behavior y; the high cross-subject R is the expected outcome, not a discovery. No null control (e.g., shuffled pair labels, or a behavior-free version of the same architecture) is reported that would show the similarity exceeds what the loss enforces. Second, the PMd 'zero-shot' test set is not fully held out: the supplement admits that two Monkey M test sessions come from a monkey whose sessions were used in training. That undermines the claim of cross-subject generalization for PMd, and the main text does not flag it. Third, the proofs in Supplement B are informal and contain questionable steps; the mutual information claim is asserted rather than derived. These are not fatal to the method, but they are fatal to the paper's central scientific inference.\n\nWho is this for? A reader who wants a practical recipe for neural-behavioral alignment and robust zero-shot decoding will get value. A reader looking for evidence about the organization of neural codes across individuals should read Safaie et al. 2023 and treat this paper's interpretation with caution. The citation practice is fine—the relevant prior work is engaged, and the comparison baselines are reasonable.\n\nI would send this to a serious referee: the method deserves scrutiny and the claims need to be reined in, but there is enough substance here for a productive review.","headline":"Solid neural-behavioral alignment method whose headline 'preserved representations' claim is largely an artifact of the training objective; worth engaging for the engineering, not for the neuroscience conclusion.","tokens_in":27148,"tokens_out":2590,"would_cite":true,"duration_ms":25829,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A probabilistic alignment of neural activity to behavior reveals preserved neural representations across animals, cortices, and species, without per-subject calibration.","keywords":["neural representation alignment","zero-shot cross-subject generalization","probabilistic representation learning","motor cortex","visual cortex","brain-computer interface","variational autoencoder","population coding"],"falsifier":"Train PNBA on a dataset where the pairing between neural activity and behavior is permuted, replacing each trial's behavior with a randomly chosen behavior from another trial of the same subject, then measure cross-subject latent correlation under zero-shot testing; if the correlation stays near the intact-data values (e.g., $R \\approx 0.9$ in M1 or V1), the reported preservation is imposed by the objective rather than discovered in the brain.","tokens_in":26099,"feed_emoji":"🧠","tokens_out":10328,"duration_ms":95777,"temperature":0.7,"pith_summary":"The paper claims that a shared low-dimensional neural code survives across individuals, sessions, and even cortical areas and species, once neural activity is aligned to behavior in a probabilistic latent space. If correct, this would resolve the paradox of heterogeneous brains producing consistent function and would let brain-computer interfaces decode from a new person without per-user calibration. The paper demonstrates the claim by training a probabilistic dual-encoder model on monkey motor cortex (M1, PMd) and mouse visual cortex (V1) and reporting high neural-behavioral correlation on held-out subjects, with cross-subject latent correlations near $R = 0.9$ under zero-shot testing. The interpretive step the paper makes is that these correlations are evidence of preserved neural representations, not just of the alignment objective.","feed_headline":"Brain representations survive across animals once aligned to behavior","feed_subtitle":"Preserved neural codes across individuals could let brain-computer interfaces skip per-user calibration.","key_machinery":"The machine is PNBA's probabilistic dual-encoder alignment. A neural encoder $f_\\theta$ maps a spike matrix into a Gaussian in a shared latent space $\\mathcal{Z}$, and a behavioral encoder $g_\\phi$ maps kinematics or visual stimuli into a Gaussian there too; a sigmoid probabilistic matching loss with learnable temperature and bias pulls matched neural-behavioral pairs together and pushes mismatched pairs apart, while the distributional distance includes both mean and variance terms. Generative ELBO constraints—reconstruction of spikes with a Poisson likelihood and of behavior with a Gaussian/MSE likelihood—prevent the trivial constant-code solution, and a shared network with adaptive pooling standardizes the neuron dimension so the same parameters serve all subjects. The preservation claim is measured by Pearson correlation between latent codes across trials, sessions, and subjects, with matched behavioral conditions defined by kinematic correlation above 0.9 in motor cortex and by identical visual stimuli in V1.","core_discovery":"On the paper's own terms, the discovery is that preserved neural representations are not confined to motor cortex or to post-hoc manual alignment. PNBA maps neural activity and behavior into probability distributions in a shared latent space and aligns them with a probabilistic matching loss, while variational-autoencoder-style generative constraints—Poisson reconstruction of spikes and Gaussian reconstruction of behavior—prevent the encoders from collapsing to a constant code. A single network shared across subjects, with adaptive pooling to absorb varying neuron counts, produces latent codes that align to behavior on completely held-out subjects; the reported zero-shot cross-subject correlations are $R = 0.939 \\pm 0.033$ in monkey M1, $R = 0.856 \\pm 0.072$ in PMd, and $R = 0.892 \\pm 0.014$ in mouse V1. The authors interpret the V1 result, obtained with calcium imaging and passive viewing, as evidence that such preservation is a general property of cortical coding across species.","pith_inferences":["Editorial inference: the near-0.9 cross-subject correlations may be partly manufactured by the objective, because for identical behavior $y$, both subject encoders are trained to approach the same behavioral code $g_\\phi(y)$; a permutation null that shuffles which neural trials are paired with which behavioral trials would quantify how much of the similarity is enforced rather than discovered.","Testable extension: the same framework could be run with the behavioral encoder frozen to a random code or with behavior replaced by a nonsense variable; if cross-subject correlations remain high, preservation is driven by shared stimulus or behavior, and if they drop, the neural code itself contributes.","Neighbouring problem: PNBA's latent correlation can be used as a metric for representational preservation across developmental stages, learning, or perturbations, since it yields a distance between codes under fixed behavior without retraining.","Cross-species prediction: if the preservation is truly a general neural property, PNBA should report high cross-subject correlations even when the two subjects are different species performing the same behavioral task; that prediction is testable with existing primate-rodent behavioral datasets."],"forward_implications":["Calibration-free decoding: because held-out subjects can be mapped into the shared latent space without retraining, PNBA predicts that behavioral decoders trained on one set of subjects will work on new subjects, as shown for V1-guided running-speed decoding with a GRU decoder ($R^2 = 0.888$).","Hierarchical stability: preservation is graded by scale, with trial-level correlations highest ($R = 0.960$), session-level next ($R = 0.946$), and cross-subject still high ($R = 0.939$), so the paper predicts a nested structure of representational stability in cortex.","Generality across sensory and motor systems: the V1 result under passive viewing suggests preserved representations exist outside movement-related areas, so the same framework should transfer to other sensory cortices and species.","Implicit alignment replaces manual alignment: prior demonstrations of cross-subject preservation required post-hoc alignment, while PNBA makes the alignment intrinsic, so no per-subject optimization is needed for comparison or decoding.","A bridge to covert processes: using temporally delayed behavioral readouts, as in PMd, the authors propose that preparatory or cognitive states with no immediate behavioral correlate can still be anchored to behavior and tested for preservation."],"supporting_citations":[{"why":"Supplies the monkey M1/PMd center-out reaching dataset and the prior observation of preserved neural dynamics across subjects that PNBA extends with zero-shot validation.","marker":"(Safaie et al., 2023)"},{"why":"Supplies the mouse V1 calcium-imaging dataset with paired animals viewing identical visual stimuli, enabling the cross-subject V1 tests.","marker":"(Turishcheva et al., 2024)"},{"why":"Provides the probabilistic matching objective (sigmoid distributional matching) that PNBA adapts as its alignment loss.","marker":"(Chun, 2023)"},{"why":"Supplies the variational autoencoder and ELBO machinery used for the generative constraints that prevent representation collapse.","marker":"(Kingma & Welling, 2014)"},{"why":"Shows how to compose conditional Gaussian variational posteriors, underlies PNBA's joint posterior factorization.","marker":"(Johnson et al., 2016)"},{"why":"Connects the constrained optimization with Lagrangian/KKT multipliers, yielding the weighted total loss in Eq. (5).","marker":"(Higgins et al., 2017)"},{"why":"Establishes long-term stability of cortical population dynamics across sessions, the motivation for session-level preservation.","marker":"(Gallego et al., 2020)"},{"why":"Provides the low-dimensional population-dynamics framework and center-out reaching paradigm that motivate the latent-space approach in motor cortex.","marker":"(Churchland et al., 2012)"}],"fun_headline_variants":["Zero-shot neural alignment across brains via behavior","Neural codes persist across animals when aligned to behavior","Shared brain code found by aligning activity to behavior","Behavioral alignment reveals universal neural representations","PNBA: zero-shot brain mapping across species and cortices"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that high cross-subject similarity of the learned codes for matched behavior is evidence of biological preservation; the training objective itself pulls those codes together because both subjects' neural codes are aligned to the same behavioral code, so a null control is needed to show the similarity exceeds what the loss enforces.","fun_headline_variants_meta":{"raw":{"variants":["Zero-shot neural alignment across brains via behavior","Neural codes persist across animals when aligned to behavior","Shared brain code found by aligning activity to behavior","Behavioral alignment reveals universal neural representations","PNBA: zero-shot brain mapping across species and cortices"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000148,"raw_usage":{"total_tokens":1174,"prompt_tokens":912,"completion_tokens":262,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":528,"completion_tokens_details":{"reasoning_tokens":190}},"tokens_in":528,"tokens_out":262,"duration_ms":3063,"temperature":1.0,"reasoning_tokens":190,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:31:42.850550+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train PNBA on a dataset where the pairing between neural activity and behavior is permuted, replacing each trial's behavior with a randomly chosen behavior from another trial of the same subject, then measure cross-subject latent correlation under zero-shot testing; if the correlation stays near the intact-data values (e.g., $R \\approx 0.9$ in M1 or V1), the reported preservation is imposed by the objective rather than discovered in the brain.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the probabilistic matching objective (sigmoid distributional matching) that PNBA adapts as its alignment loss."}],"review_version":1}