{"id":"b82eb3eb-215f-4527-9bd3-a1c8ae2f8540","arxiv_id":"2607.14086","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Adding masked-autoencoder self-supervision to POYO-style spike-token decoders improves decoding accuracy and label efficiency across species and neural modalities.","lead":"MOJO adds a masked-autoencoder self-supervised loss to spike-tokenizing neural decoders, so they can train on unlabelled recordings alongside labelled ones. Across monkey, mouse, and human ECoG datasets it beats purely supervised versions, especially when only a few labelled trials are available.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Appendix D.6 contradicts the paper's headline: adding unlabelled data during pretraining does not improve finetuning transfer, undercutting the central 'leverage unlabelled data' claim.","rationale":"The reader's verdict was CONDITIONAL, and their rationale already listed Appendix D.6 as issue (1). However, their formally stated 'weakest_assumption' was evaluation fairness / preprocessing changes, not this contradiction. My stress-test identifies the D.6 contradiction as the single most load-bearing concern because it directly undermines the paper's title and headline claim about using unlabelled data. If unlabelled pretraining data does not improve finetuning transfer, the main novelty is substantially weakened — the method may still improve decoding through joint SSL-SL on labelled data, and the few-shot finetuning result (Figure 2a) shows a benefit from unlabelled data during finetuning, but that result is single-session with no error bars. The interpretability analyses are independent and credible, but they do not rescue the unlabelled-pretraining claim. The paper should be accepted only if the authors either reconcile D.6 with the abstract/title or explicitly narrow the claim. Since the reader's verdict already required conditions and this concern reinforces that conditionality, the verdict remains CONDITIONAL; no change to the label is needed, but the condition should explicitly include resolving this internal contradiction. This is an internal inconsistency, not an external disagreement or ad hominem, and it is verifiable by reproducing the D.6 experiment.","tokens_in":29852,"tokens_out":6563,"duration_ms":59941,"concrete_test":"Reproduce the Figure 4a / D.6 experiment: pretrain MOJO-POYO (or MOJO-POGRU) on the Perich et al. dataset with 10%, 20%, and 50% labelled data, with and without the remaining unlabelled data, following the paper's stated protocol. Then evaluate both UI and FT finetuning R2 on all held-out sessions (C-CO, T-CO, T-RT) over at least 3 seeds. If adding unlabelled data yields no significant positive mean R2 difference (or a negative difference) in either finetuning scheme, the 'leverage unlabelled data during pretraining' and '90% unlabelled downstream improvement' statements must be retracted or restricted to the few-shot finetuning setting.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central value proposition is that MOJO leverages unlabelled data to improve decoding — stated in the title, abstract, and §3.1 ('MOJO can effectively leverage unlabelled data during pretraining', Figure 2b). However, Appendix D.6 (Figure 4a) reports the direct test of this claim: adding unlabelled data during pretraining 'does not lead to better finetuning performance, especially in UI.' Since finetuning to new sessions is the operational objective of pretraining and the source of the strongest gains claimed (Table 1, Figure 2a), this is not a minor caveat — it removes the evidence for the '90% unlabelled' downstream-improvement claim and reduces the unlabelled-data contribution to a single-session few-shot finetuning effect (Figure 2a) plus interpretability analyses. The paper does not reconcile these; the Discussion repeats 'significant downstream improvement even if up to 90% of the pretraining data is unlabelled,' directly contradicting D.6. A reader cannot tell whether the method's benefit is joint SSL-SL on labelled data (a valuable but different claim) or actual use of unannotated data.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces MOJO, a joint self-supervised (masked autoencoder) and supervised training objective for spike-tokenizing neural decoders of the POYO family. MOJO is evaluated on monkey reaching, mouse vision and decision-making, and human ECoG speech, with comparisons against supervised spike-tokenizing models and binned baselines (NDT-2/3, NEDS, Du-IN, EEGNet). The authors report consistent gains from MOJO when finetuning to held-out sessions, especially in few-shot regimes, as well as more interpretable unit embeddings. However, Appendix D.6 reports that adding unlabelled data during pretraining does not improve finetuning transfer, which directly conflicts with the title, abstract, and Discussion claim of leveraging unlabelled data for generalizable decoding.","tokens_in":30123,"tokens_out":5948,"duration_ms":53718,"significance":"The empirical scope is substantial: multiple species, modalities, tasks, and backbones, with ablations of mask ratio, pathway integration, and alternative SSL objectives, plus a running-time comparison. The interpretability analyses (region classification, spike-statistic prediction, probe-distance geometry) are thoughtful and go beyond standard decoding metrics. If the central claims were fully supported, MOJO would be a useful step toward using unannotated neural recordings in foundation-model pretraining. However, the manuscript's own D.6 undermines the main unlabelled-data claim, and the lack of released code weakens reproducibility of the baseline comparisons. The results still suggest that joint SSL-SL training improves over purely supervised training on labelled data, but the title and discussion overstate the evidence for leveraging unlabelled data.","major_comments":[{"comment":"The paper's central claim that MOJO leverages unlabelled data for generalizable decoding is contradicted by its own Appendix D.6 (Fig. 4a): 'having additional unlabelled data during pretraining does not lead to better finetuning performance, especially in UI.' Since finetuning to new sessions is the operational objective of pretraining and the source of the strongest reported gains, the §5 statement that 'We see significant downstream improvement even if up to 90% of the pretraining data is unlabelled' is not supported. Fig. 2b shows only pretraining R2, not downstream finetuning. The authors must reconcile this, e.g., by reporting finetuning results for the unlabelled-pretraining sweep or by reframing the unlabelled-data contribution to the few-shot finetuning setting (Fig. 2a) and interpretability analyses.","section":"§3.1 'Unlabelled Data During Pretraining', Appendix D.6, §5"},{"comment":"The reported superiority over external baselines rests on modifications to the baselines and data splits, but no code is released. The paper acknowledges IBL split and re-sorting changes and reruns NEDS, but for NDT-2/3, Du-IN, and EEGNet it states only that original code was used 'with some modifications' (C.1, C.2, C.4). Given the causal evaluation and non-trial-aligned preprocessing choices, the gains could in part be an artifact of evaluation setup. I recommend releasing code and exact preprocessing scripts, and providing a sensitivity analysis (e.g., evaluating NDT-3 under the same causal protocol) to verify that the comparisons are apples-to-apples.","section":"§3.3, §3.4, Appendix F"},{"comment":"The few-shot finetuning experiment is the main surviving evidence for unlabelled-data benefit, but the protocol is underspecified. The text says 'incorporating up to 32 trials of unlabelled data' without explaining how the SSL loss is applied during finetuning, which parameters are updated, how unlabelled trials are selected, or how they are balanced with labelled trials. Without these details, the reader cannot determine whether the gain comes from the SSL objective on unlabelled data or from increased training signal/regularization. Please provide the full finetuning protocol and, ideally, an ablation that replaces the unlabelled trials with additional labelled trials or discards them.","section":"§3.1 'Few-shot Finetuning', Figure 2a"}],"minor_comments":[{"comment":"The notation 'MOJO-POYO(J)' and 'MOJO-POYO-L(J)' is used in the main table but only defined in the text; similarly, 'MOJO-POYO(A)' appears in Table 21. Please add definitions to the table captions for readability.","section":"Table 2 and Table 21"},{"comment":"The phrase '47% improvement for classifying syllables' should report absolute accuracies (e.g., from X% to Y%) to avoid ambiguity between relative and absolute improvement.","section":"§3.4"},{"comment":"There are typos: 'certain extend' should be 'certain extent' in D.5, and 'underling neural dynamics' should be 'underlying neural dynamics' in §5.","section":"Appendix D.5 and §5"},{"comment":"The 'NEDS+bugfix' entry is not defined in the main text; please state what the bug fix is and why it is reported.","section":"Appendix F/Table 24"},{"comment":"The preprocessing changes relative to Zhang et al. [16] (removing trial alignment and firing-rate exclusion, different whisker normalization) are important for interpreting Table 3; they should be stated prominently before the results rather than only in Appendix A.3.","section":"§3.3"}],"recommendation":"major_revision","confidential_remarks":"The D.6 result is a genuine internal contradiction: the paper's title and abstract promise 'leveraging unlabelled data' for generalizable decoding, but the authors' own downstream finetuning sweep shows no benefit from additional unlabelled pretraining data. The paper is fixable by narrowing the claims to the few-shot finetuning setting and to unit-embedding interpretability, while presenting joint SSL-SL on labelled data as the core contribution. However, as written the overclaim is substantial. The absence of released code also makes it difficult to audit the baseline modifications, which is a practical barrier for a methods paper in this area."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core finding here is real: adding a masked-autoencoder SSL loss to spike-tokenizing POYO/POSSM models improves decoding over purely supervised training across monkey reaching, mouse vision, mouse decision, and human ECoG speech. The gains are consistent, the ablations are informative, and the unit-embedding analyses are genuinely thorough. Parameter overhead is small. This is a solid empirical systems paper.\n\nWhat is actually new is the application of masking-based SSL to spike-token models — prior SSL work in this family used calcium traces or binned spikes — and the joint SSL/SL formulation with shared backbone. The authors also deserve credit for reporting Appendix D.6 at all, since it directly undermines their own headline.\n\nThe problem is exactly what the stress-test note says. The title and abstract promise that MOJO leverages unlabelled data to improve pretraining and downstream decoding. But D.6 (Figure 4a) shows that adding unlabelled data during pretraining does not improve finetuning performance, especially in UI. The few-shot finetuning result in Figure 2a is the main positive evidence for unlabelled data, and it is a single session with no error bars. The discussion repeats the \"90% unlabelled\" claim without reconciling this contradiction. So the strongest form of the paper's contribution — that unlabelled data helps — is not supported. What is supported is that joint SSL/SL on labelled data helps, which is still valuable but a different claim.\n\nOther soft spots are minor in comparison but worth noting. No code is released, so the baseline re-runs (NDT, NEDS, Du-IN) and preprocessing choices cannot be audited. The handling of unlabelled batches in the joint loss is unspecified. The IBL and ECoG modifications are acknowledged but could disadvantage baselines. These are fixable.\n\nThe paper deserves a serious referee. It is not a desk reject. But the referee should send it back for major revision: reframe the title and abstract around the supported claim, add error bars to the few-shot result, release code, and either reconcile D.6 or drop the unlabelled-data framing. I would probably cite the joint-training result, and I'd bring it to a reading group as a good case study in overstated claims with honest appendices.","headline":"MOJO's joint SSL–SL objective consistently beats supervised baselines, but the paper's headline claim about leveraging unlabelled data is contradicted by its own Appendix D.6.","tokens_in":30680,"tokens_out":2248,"would_cite":true,"duration_ms":24090,"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":"MOJO shows that adding a masked-autoencoder objective to spike-tokenizing neural decoders improves decoding, especially when labeled data are scarce.","keywords":["neural decoding","self-supervised learning","masked autoencoder","spike tokenization","few-shot finetuning","unit embeddings","brain-computer interface","neuro-foundation models"],"falsifier":"Run a pre-registered, code-released head-to-head on several held-out sessions from new animals, finetuning each model with only 2–4 labeled trials (and unlabeled data for MOJO); if MOJO's few-shot advantage over its purely supervised counterpart disappears or reverses across sessions, the central claim collapses.","tokens_in":29707,"feed_emoji":"🧠","tokens_out":4067,"duration_ms":43977,"temperature":0.7,"pith_summary":"This paper sets out to show that spike-tokenizing neural decoders, which normally require paired behavioral labels, can be pretrained with unlabeled neural data by adding a masked-autoencoder reconstruction objective alongside supervised decoding. The proposed MOJO framework shares almost all parameters between the two pathways, adding only a small overhead (about 231K parameters on a 9.88M-parameter model). Across monkey reaching, mouse vision, mouse decision, and human speech electrocorticography, MOJO variants outperform their purely supervised counterparts under the same finetuning strategy, with the largest gains in few-shot finetuning. The learned per-neuron embeddings also become more interpretable: linear probes predict brain region and single-neuron spike statistics better than supervised-only embeddings. The paper argues this opens a path to using vast unlabeled neural recordings to train scalable neuro-foundation models.","feed_headline":"Unlabeled spikes boost neural decoding when labels are scarce","feed_subtitle":"Joint masked-autoencoding and supervised decoding beats supervised-only models on monkey, mouse, and human speech","key_machinery":"The central object is the MOJO training objective L = α_SSL L_SSL + α_SL L_SL, a weighted sum of a Poisson negative log-likelihood for reconstructing spike rates from masked latents and a supervised behavioral loss. The self-supervised pathway applies temporal masking to latent tokens and reconstructs spike rates for sampled units; the supervised pathway uses output cross-attention queries to predict behavior. The key design is pathway integration: the two pathways share the input cross-attention output and backbone parameters, so the self-supervised pathway adds only one output cross-attention module. This lets unlabeled neural data improve representations without doubling the model.","core_discovery":"The central claim is that augmenting spike-tokenizing models with the MOJO joint self-supervised-supervised objective improves decoding performance over purely supervised training, especially in label-impoverished few-shot finetuning, and yields more interpretable unit embeddings. Operationally, MOJO trains a masked autoencoder on latent tokens produced by input cross-attention while simultaneously training a supervised behavioral decoder, sharing the input cross-attention and backbone parameters. Evaluated on monkey reaching, mouse vision, mouse decision, and human ECoG speech, MOJO consistently outperforms purely supervised baselines under identical finetuning strategies, achieves the larg","pith_inferences":["If this result generalizes, brain-computer interface calibration could shift from collecting hours of labeled data per session to a short labeled calibration plus passive unlabeled recording, which would make clinical deployment faster and less burdensome.","The finding that self-supervision makes unit embeddings more interpretable suggests masked autoencoding could serve as a generic pretraining scheme for neural recordings where behavior labels are absent or noisy, such as sleep or freely moving behavior.","A testable extension would add spatial masking (masking neurons or brain regions) in addition to temporal masking; the paper notes this is absent and could further improve embeddings for unseen neurons or regions.","The joint loss coefficients were set to 1 for both pathways across experiments; on more heterogeneous data or with imbalanced label availability, adaptive weighting might be needed, and this could be tested directly."],"forward_implications":["MOJO-trained decoders can be finetuned with very few labeled trials from a new session or animal, and can use unlabeled trials during finetuning to recover 60–75% of fully supervised performance with only 2–4 labeled calibration trials.","Unlabeled data can be used during pretraining: even when up to 90% of pretraining data are unlabeled, MOJO maintains strong decoding and unit-embedding quality.","Unit embeddings learned by MOJO support linear-probe classification of brain regions and regression of spike statistics with substantially higher accuracy than purely supervised embeddings.","Joint pretraining across monkey reaching and mouse vision transfers positively and scales with model size and data, improving convergence and finetuning on mouse tasks while preserving monkey performance.","MOJO generalizes beyond spiking data to human ECoG speech decoding, outperforming supervised POYO and matching a specialized continuous-signal foundation model."],"fun_headline_variants":["Masked autoencoding sharpens neural decoders with scant labels","MOJO pairs self-supervision with supervision to decode brains","Few-shot neural decoding gets a boost from unlabeled spike data","Unlabeled spikes improve brain decoders across species and tasks","Self-supervised pretraining elevates spike-tokenizing models"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the public datasets, with the paper's preprocessing and re-splitting, fairly measure cross-session and cross-animal generalization—in particular, that the modified IBL data preparation and the re-run baselines do not systematically favor MOJO.","fun_headline_variants_meta":{"raw":{"variants":["Masked autoencoding sharpens neural decoders with scant labels","MOJO pairs self-supervision with supervision to decode brains","Few-shot neural decoding gets a boost from unlabeled spike data","Unlabeled spikes improve brain decoders across species and tasks","Self-supervised pretraining elevates spike-tokenizing models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000123,"raw_usage":{"total_tokens":970,"prompt_tokens":812,"completion_tokens":158,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":83}},"tokens_in":556,"tokens_out":158,"duration_ms":2987,"temperature":1.0,"reasoning_tokens":83,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T02:48:06.882462+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a pre-registered, code-released head-to-head on several held-out sessions from new animals, finetuning each model with only 2–4 labeled trials (and unlabeled data for MOJO); if MOJO's few-shot advantage over its purely supervised counterpart disappears or reverses across sessions, the central claim collapses.","supporting_citations":[],"review_version":1}