{"id":"1c331b6c-aa63-4880-b4cd-303db2c29b19","arxiv_id":"2607.23258","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A continuous-token autoregressive transformer pretrained on mouse calcium traces transfers across datasets, paradigms, and species with a frozen backbone, beating specialized baselines.","lead":"CAPT is a transformer pretrained on mouse calcium imaging that, with its backbone frozen, transfers to mouse, zebrafish, and C. elegans recordings for forecasting and behavior decoding, outperforming specialized baselines. The paper also shows CAPT embeddings carry cell-identity information in C. elegans via NeuroPAL labels.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Main 'consistently outperforms' claim lacks error bars/significance tests; several transfer margins are near noise (e.g., Brysch CAPT 0.1895 vs CalM 0.1872), so the central comparison is not yet pinned down.","rationale":"I read the paper as an empirical demonstration that a frozen continuous autoregressive backbone can transfer across datasets and species. For that demonstration to support the stated conclusion, the reported performance advantages must be real and robust, not artifacts of a single seed or small sample. The absence of error bars and significance testing is the most load-bearing weakness: even if the patch-length mismatch were resolved, the central 'consistently outperforms' claim would still be shaky without uncertainty quantification. The patch-length concern is worth testing, and the reader correctly identifies it as an assumption, but the numbers in Figures 2, 3, and 5—especially the near-tie in Brysch—make the statistical point more urgent. I therefore partially agree with the reader: their rationale already mentions missing error bars, but their 'weakest assumption' field focuses on temporal patch length. My stress-test puts the primary weight on the statistical robustness of the comparative claim. Since the reader's verdict is already CONDITIONAL, my concern does not change the recommended verdict; it reinforces the need for the same additional evidence (error bars, significance tests, code/data release).","tokens_in":16512,"tokens_out":5274,"duration_ms":50269,"concrete_test":"Rerun all transfer comparisons in Figures 2, 3, and 5 with at least 5 independent seeds under identical splits, and report mean ± std per dataset. For each dataset, compute per-session paired differences between CAPT and the best baseline and apply a paired bootstrap or Wilcoxon signed-rank test (with multiple-comparison correction across datasets). If any dataset—for example Brysch, where the margin is 0.0023—shows a non-significant difference, the claim should be downgraded from 'consistently outperforms' to 'generally competitive,' and the verdict should remain conditional.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that CAPT 'consistently outperforms' specialized and general-purpose baselines across eight transfer datasets. That claim is only as strong as the uncertainty attached to the reported point estimates. The main comparisons in Figures 2, 3, and 5 are presented as single correlation/R^2 values with no error bars, no number of seeds, and no significance tests, even though several margins are small. For example, the Brysch zebrafish forecasting correlation is 0.1895 for CAPT versus 0.1872 for CalM—a difference of 0.0023, which is almost certainly within run-to-run or session-level noise. Appendix E reports standard deviations only for a small subset (Table 11, POYO+ versus CAPT in C. elegans), not for the main transfer tables. The hierarchical aggregation described in Appendix A (per-neuron, per-trial, per-session, then across sessions) could be used to construct paired tests, but none are reported. Without knowing whether CAPT's advantage is stable across seeds and statistically significant after multiple-comparison correction, the 'consistently outperforms' assertion—the load-bearing empirical outcome of the paper—remains under-supported. The patch-length issue identified by the reader is a plausible architectural weak point, but the statistical uncertainty is more directly load-bearing because it applies to every comparison and to the paper's headline claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces CAPT, a continuous autoregressive transformer for calcium imaging population dynamics. CAPT tokenizes single-neuron calcium traces into continuous temporal patches, projects them with neuron/session embeddings, and trains a dual-axis transformer with an MSE next-patch objective. After pretraining on a large mouse dataset (Tseng et al.), the backbone is frozen and only dataset-specific embeddings or decoding heads are trained for transfer. The authors evaluate CAPT on neural population forecasting and behavior decoding across eight external mouse, zebrafish, and C. elegans datasets, comparing with POCO, POYO+, and CalM. They report that CAPT consistently outperforms the baselines, and additionally show that CAPT embeddings support NeuroPAL-based neuron identity classification across C. elegans datasets. The paper includes detailed dataset descriptions, hyperparameter tables, ablations, and additional results in appendices.","tokens_in":16896,"tokens_out":4726,"duration_ms":44596,"significance":"If the reported transfer results are statistically robust, the paper would be a valuable step toward general-purpose calcium-imaging foundation models: the frozen-backbone transfer protocol is clean, the benchmark spans three species and multiple laboratories, and the NeuroPAL embedding analysis is a useful extension beyond predictive accuracy. The authors provide extensive implementation detail, consistent preprocessing, and ablate key design choices. However, the central comparative claim currently rests on point estimates without uncertainty quantification, and the fixed patch-length tokenization raises a cross-species architectural concern. These issues are fixable within the scope of the paper and do not invalidate the approach, but they must be addressed before the headline claim can be accepted.","major_comments":[{"comment":"The central claim that CAPT 'consistently outperforms' baselines is supported only by point estimates. The main transfer comparisons in Figs. 3 and 5, and the embedding classification in Fig. 4, lack error bars, number of seeds, and significance tests. Several margins are very small; for example, Brysch forecasting correlation is 0.1895 for CAPT versus 0.1872 for CalM. Appendix A describes a hierarchical aggregation (per neuron, per trial, per session) that could support paired tests, and Appendix E reports standard deviations for only one comparison (Table 11), but not for the main tables. The limitation paragraph in §6 also acknowledges that transfer gains are generally smaller, making the missing uncertainty quantification more consequential. Please add confidence intervals or error bars, report the number of seeds, and provide paired significance tests with multiple-comparison correc","section":"§5.1–5.4, Figs. 2–5"},{"comment":"The patch length L=8 is used for all datasets, but it is selected on 6 Hz mouse pretraining data (Appendix D). At target sampling rates, the same L corresponds to very different real-time windows: about 0.5 s at 10 Hz (Sun) and about 4.8 s at 1.67 Hz (C. elegans). The frozen temporal attention is therefore asked to operate over substantially different temporal receptive fields. The ablation in Table 9 only varies L on the 6 Hz source subset and does not test whether L=8 is appropriate for 1.67–15.6 Hz target data. This is load-bearing for the cross-species transfer claim because the continuous patch tokenization is the main mechanism that makes the traces commensurable. Please either adapt L to each target dataset's sampling rate, add a per-dataset patch-length sensitivity analysis, or explicitly show that the real-time span of a patch does not affect the transfer conclusions.","section":"§3.1, Appendix D, Table 1"},{"comment":"Baseline comparability is not fully established. POCO is restricted to fixed context/horizon of 40/24 steps (Appendix B), and for the Lavian whole-brain dataset it is reported as 'unavailable' because the downstream setup is incompatible (Appendix E). Thus the claim that CAPT outperforms specialized baselines on all transfer datasets is partly a consequence of protocol constraints rather than model quality alone. Similarly, POYO+ fine-tuning only adjusts sequence length and latent step based on real time; Table 11 shows that in-domain POYO+ performs much better than cross-species fine-tuning, which supports the authors' distribution-shift interpretation but does not fully establish that the fine-tuning protocol is equally favorable to all methods. Please either evaluate all baselines under the same transfer protocol where possible, or clearly qualify the comparisons for which a baseline","section":"§4.3, Appendix B, Appendix E, Fig. 5"}],"minor_comments":[{"comment":"Typo: the text says 'we freeze the POYO backbone' but should say 'POCO backbone'.","section":"Appendix B, POCO paragraph"},{"comment":"The decoding R2 axis is shown from 0 to 1, but the text reports negative values such as -0.0124 for POYO+ on Atanas head-angle decoding. The negative bars are likely clipped or invisible; please adjust the axis or annotate the values.","section":"Fig. 3E/F"},{"comment":"The embedding classification results would be easier to interpret with details on SVM hyperparameters, class sizes, and whether the 5-fold CV is stratified; also, the 'chance' levels should state whether they are based on label permutation or majority-class proportion.","section":"§5.4, Fig. 4B"},{"comment":"The ablation table reports mean±std but not the number of sessions or seeds; adding the number of runs would help assess the stability of the patch-length comparison.","section":"Table 9"},{"comment":"The related work section introduces 'POCO' for forecasting and 'POYO+' for decoding, but the names are close; a brief clarifying sentence distinguishing them would improve readability.","section":"§2.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is built on the same group's CalM/DAT framework, and the evaluation protocol closely follows the earlier CalM paper. The novelty is the continuous patch tokenization and the broader transfer study. This is a legitimate increment, but the authors should be asked to provide the statistical machinery that would make the 'consistently outperforms' claim credible across the reported datasets. I would also encourage the editor to ask for code/data release, since none is mentioned and the paper is highly empirical."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe genuinely new thing here is the systematic frozen-backbone cross-species transfer test for calcium imaging: pretrain on one large mouse dataset, then adapt only neuron/session embeddings or task heads on eight independent datasets across mice, zebrafish, and C. elegans. That is a useful benchmark and a sensible protocol. The continuous patch tokenization is a modest change over CalM's discrete tokens, but it removes a separate tokenizer and keeps the forecasting objective in the raw trace space, which is a reasonable design choice. The NeuroPAL embedding analysis (pooled and cross-dataset identity classification) is a nice bonus that goes beyond predictive metrics.\n\nI agree with the reader's conditional verdict, and the stress-test note lands. The central claim—\"consistently outperforms\"—is not pinned down because the main comparisons in Figs 2, 3, and 5 are point estimates without error bars, seeds, or significance tests. Some margins are tiny (Brysch: 0.1895 vs 0.1872). The hierarchical aggregation in Appendix A could support paired tests, but none are reported. Without that, a few of those wins could be noise.\n\nThe patch-length concern is plausible but secondary. L=8 was chosen on 6 Hz mouse data and then applied everywhere, so the real duration per patch ranges from about 0.5 s to 4.8 s across datasets. That could matter, but the model still transfers reasonably, so it is a possible architectural weakness, not a demonstrated one.\n\nI don't see circularity. Transfer is tested on held-out animals and external labs. The self-citation to CalM for the backbone is legitimate, not a hidden reuse of the test set.\n\nWhere the paper is weaker: baselines are sometimes compared under protocol restrictions (POCO needs fixed context/horizon; POYO+ degrades under cross-species distribution shift, and the appendix does a good job showing that's not just hyperparameters). No code or data release yet, which limits reproducibility. That is a real soft spot for a foundation-model claim.\n\nWho is this for: anyone working on neural foundation models for calcium imaging, and neuroscientists wanting a cross-species transfer baseline. It deserves a serious referee, but the revision needs uncertainty quantification and ideally code release before I'd trust the headline.\n\nRecommendation: send to peer review, with the expectation of major revision focused on statistical rigor.","headline":"Useful cross-species transfer benchmark, but the 'consistently outperforms' claim needs error bars and significance tests before I'd trust the headline.","tokens_in":17322,"tokens_out":3061,"would_cite":true,"duration_ms":23880,"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 continuous autoregressive transformer pretrained on mouse calcium imaging, with its backbone frozen, generalizes to new mouse, larval zebrafish, and C. elegans datasets for neural forecasting and behavior decoding.","keywords":["calcium imaging","foundation model","autoregressive transformer","cross-species transfer","neural population forecasting","behavior decoding","continuous tokenization","NeuroPAL cell identity"],"falsifier":"Retrain CAPT with per-dataset patch lengths chosen to equalize real time per patch, and compare with the fixed L=8 result on larval zebrafish and C. elegans forecasting. If per-dataset patch alignment substantially improves or fixed L=8 fails on a new species at a different sampling rate, the cross-species transfer claim rests on a coincidental token size rather than a general mechanism.","tokens_in":16426,"feed_emoji":"🧠","tokens_out":4150,"duration_ms":38220,"temperature":0.7,"pith_summary":"The paper proposes CAPT, a transformer that models continuous calcium-imaging traces directly, without first turning them into discrete tokens. Pretrained on a large mouse dataset, the model's backbone is frozen and only small per-dataset neuron/session embeddings or task-specific decoders are updated. Across eight independent datasets spanning mice, larval zebrafish, and C. elegans, this frozen backbone forecasts neural activity and decodes behavior better than specialized and general-purpose baselines. In C. elegans, the learned neuron embeddings also support cross-dataset identification of cell identity using NeuroPAL labels. The authors argue this is evidence that a simple continuous autoregressive objective can produce a reusable neural foundation model for calcium imaging.","feed_headline":"Mouse-pretrained neural model transfers to fish and worm recordings","feed_subtitle":"With the backbone frozen, one continuous autoregressive transformer beats task-specific baselines on calcium imaging.","key_machinery":"Continuous patch tokenization: each neuron trace is cut into non-overlapping patches of length L=8 samples and projected directly into model space with a linear layer plus learnable neuron and session embeddings; the Dual-axis Transformer processes them with causal temporal attention across time and neuron-axis attention across the population, and predicts the next continuous patch in raw trace space under an MSE loss. Two corruptions (pseudo scheduled sampling and temporal neighborhood replacement) make training robust to imperfect histories. The same continuous tokens are used for pretraining, forecasting, and as input to a nonlinear low-rank per-neuron readout for behavior decoding; keepi","core_discovery":"CAPT's central claim is that a single continuous autoregressive transformer, pretrained with mean-squared-error next-patch prediction on one mouse calcium-imaging dataset, produces representations reusable across datasets, paradigms, and species. The authors show that freezing the pretrained backbone and updating only trainable neuron/session embeddings (for forecasting) or a task head (for behavior decoding) consistently outperforms specialized forecasting and decoding models, including on larval zebrafish and C. elegans, whose recording rates, neuron counts, and anatomy differ substantially from the mouse source. Additionally, linear classifiers on CAPT neuron embeddings recover NeuroPAL-d","pith_inferences":["A testable extension: fix patch length per dataset so each patch covers the same real time rather than a constant L=8; if performance survives, the mechanism is scale-invariant calcium dynamics, and if not, the fixed-token assumption is load-bearing.","If the shared embedding space holds, one could predict NeuroPAL identities for unannotated neurons by nearest-neighbor transfer, effectively using CAPT as a functional cell-identity mapper.","Because the decoder readout is low-rank per neuron, behavior decoding performance likely depends on global neuron ordering and alignment; comparing with permutation-invariant readouts would isolate whether population-level structure or per-neuron alignment drives transfer.","CAPT embeddings could provide a common latent space for cross-species registration of neural populations, a step the paper gestures at but does not implement."],"forward_implications":["If true, a single frozen backbone can serve as a general feature extractor for calcium imaging, so new datasets require only small per-session or per-neuron embeddings rather than full retraining.","Cross-species transfer would mean shared low-level dynamics across mouse, fish, and worm calcium signals are captured by the same temporal attention weights, opening the door to multi-species pretraining.","Because embeddings line up with anatomical cell identity across datasets, the same functional coordinates could link calcium activity to connectomic and transcriptomic atlases in C. elegans and beyond.","The simple MSE objective removes the need for a separate vector-quantization tokenizer, reducing hyperparameters and keeping forecasting in the trace domain.","The authors note transfer gains are smaller than in-domain gains, implying the source distribution still constrains the backbone and that wider pretraining data is the next step."],"fun_headline_variants":["One transformer, many species: mouse-trained model nails fish and worm brains","Frozen backbone, broad reach: calcium model transfers across species","Pretrained on mice, works on worms: CAPT transfers across species","Single transformer for neural dynamics transfers across species","Calcium imaging model: pretrain once, transfer to any species"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The fixed patch length of 8 samples is assumed to give temporally commensurate tokens across datasets, even though it spans roughly 0.5 seconds of mouse data, 0.8–5 seconds of zebrafish data, and about 4.8 seconds of C. elegans data; if normalized calcium dynamics are not scale-invariant across those durations, the frozen temporal attention weights lose their meaning.","fun_headline_variants_meta":{"raw":{"variants":["One transformer, many species: mouse-trained model nails fish and worm brains","Frozen backbone, broad reach: calcium model transfers across species","Pretrained on mice, works on worms: CAPT transfers across species","Single transformer for neural dynamics transfers across species","Calcium imaging model: pretrain once, transfer to any species"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000698,"raw_usage":{"total_tokens":3020,"prompt_tokens":804,"completion_tokens":2216,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":2129}},"tokens_in":548,"tokens_out":2216,"duration_ms":13203,"temperature":1.0,"reasoning_tokens":2129,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T23:54:44.576049+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain CAPT with per-dataset patch lengths chosen to equalize real time per patch, and compare with the fixed L=8 result on larval zebrafish and C. elegans forecasting. If per-dataset patch alignment substantially improves or fixed L=8 fails on a new species at a different sampling rate, the cross-species transfer claim rests on a coincidental token size rather than a general mechanism.","supporting_citations":[],"review_version":1}