{"id":"c6cc0044-6053-41c9-a2bb-05b76e24f806","arxiv_id":"2608.04389","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Behavioral pretraining on macaque and robotic trajectories improves neural trajectory decoding and reduces calibration data needs across sessions, subjects, and tasks.","lead":"NeuroPB pretrains a movement encoder on large-scale behavioral trajectories, including robotic data, then aligns neural activity to that representation space with limited paired recordings, improving continuous motor decoding in macaques. The work suggests that abundant behavior data, not just scarce neural recordings, can be used to build calibration-efficient brain-computer interfaces.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Behavioral pretraining and freezing are confounded: the Scratch baseline trains the motor encoder, while all pretrained conditions freeze it, so reported gains may reflect a stable alignment target rather than transferable kinematic knowledge.","rationale":"The reader's verdict is CONDITIONAL, with the weakest assumption identified as the transferability of robotic trajectories to biological reaching and the undisclosed macaque pretraining source. My concern is different and more fundamental: the experimental design confounds pretraining with whether the motor encoder is frozen during neural alignment. This confound undermines the causal interpretation of the main results, including the robot-vs-macaque equivalence, because a frozen random encoder would also provide a stable target for contrastive alignment. The paper does not include such a control in any ablation (Table 2 ablates components of the neural encoder and removes contrastive, but never varies the frozen/trainable status independently of pretraining). This is not an external-consensus disagreement; it is an internal control flaw that can be tested directly. The proposed Frozen-Random and Shuffled-PT controls would settle whether the gains are attributable to behavioral knowledge. Until then, the central claim should be treated as unverified. Since the reader already recommended conditional acceptance pending further evidence, my analysis reinforces that condition rather than changing the verdict category; the specific missing control should be added to the acceptance criteria. I give credit for the thorough dataset evaluation and clear reporting of experimental settings, but the missing control is load-bearing because it targets the paper's main contribution.","tokens_in":13333,"tokens_out":3467,"duration_ms":34607,"concrete_test":"Add a 'Frozen-Random' control: initialize the motor encoder with random weights and freeze it for the entire alignment and decoding training, using the same architecture, data, and hyperparameters as the Macaque-PT condition, and evaluate on CO and RT. Also add a 'Shuffled-PT' control: pretrain the motor encoder on the same macaque or LIBERO trajectories but with temporal order randomly permuted within each trajectory segment (destroying kinematic structure while preserving statistics), then freeze during alignment. If either control matches or exceeds Macaque-PT/Robot-Matched-PT performance, the behavioral-pretraining benefit is not established; if both are substantially worse, the confound is mitigated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central causal claim—that behavioral pretraining, including robotic pretraining, improves neural decoding—rests on comparing pretrained motor encoders against a 'Scratch' baseline. But the Method section states: 'In Scratch, the motor encoder was randomly initialized and jointly optimized during neural alignment,' while for Macaque-PT, Robot-Matched-PT, and Robot-Large-PT the pretrained motor encoder is 'kept frozen during neural alignment.' Thus the comparison conflates two variables: the information content of the pretrained weights and the optimization dynamics (frozen vs. trainable target). A randomly initialized but frozen motor encoder, or one pretrained on non-informative/shuffled trajectories, would control for the freezing effect. Without such a control, the reported 11% and 8% R² improvements could arise not from behavioral structure but from the stabilizing effect of a fixed, nontrivial representation target during contrastive alignment. This concern applies directly to the central claim and to the robot-vs-macaque comparison: even if robotic and macaque trajectories yield comparable performance under matched scale, that comparability alone does not establish transferable kinematic structure if a frozen random encoder achieves similar results. The appendix's description of robotic trajectories (CoTracker point tracks) reinforces that the pretrained encoder may learn generic smoothness or temporal coherence, not behavior-specific semantics. The undisclosed macaque pretraining trajectory source is a secondary leakage risk, but the frozen-vs-trainable confound is more fundamental because it threatens interpretation of every pretraining comparison, including the headline numbers.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces NeuroPB, a three-stage framework for neural trajectory decoding. A transformer-based motor encoder is first pretrained on behavioral trajectories (macaque reaching trajectories, scale-matched robotic trajectories from LIBERO-Spatial, or large-scale robotic trajectories from LIBERO-100) using masked trajectory reconstruction. The pretrained motor encoder is frozen, and a Perceiver-based neural encoder is trained with a CLIP-style symmetric contrastive loss to align spike-activity representations with the frozen behavioral representation space, while a lightweight MLP decoder reconstructs positions from the aligned neural representation. The method is evaluated on macaque M1/PMd recordings from the Perich, Gallego, and Miller dataset, on center-out and random-target tasks. The authors report state-of-the-art single-session performance, an 11% R2 improvement over a randomly initialized and jointly optimized 'Scratch' motor encoder on center-out and 8% on random-target, comparable performance between Macaque-PT and Robot-Matched-PT, further gains with Robot-Large-PT, improved cross-session/cross-subject/cross-task generalization, and calibration efficiency requiring only about 10% of target data to match Scratch trained on 100% of the data.","tokens_in":13582,"tokens_out":3643,"duration_ms":38593,"significance":"If the reported effects are genuine, the paper identifies a potentially scalable alternative to neural-only pretraining: using large, readily available behavioral or robotic trajectory datasets to define a frozen representation space that shapes a neural encoder for motor decoding. This would be a useful contribution to the BCI literature. The paper has several strengths: it uses publicly available macaque data, provides ablations of key neural-encoder components (Table 2), includes a clear statement that the scale illustration in Fig. 4 is conceptual rather than measured, and compares two pretraining conditions (Macaque-PT and Robot-Matched-PT) at matched data scale. The central causal claim, however, is currently under-supported because the main pretrained-vs-scratch comparison conflates two variables, and the calibration-efficiency analysis uses asymmetric optimization protocols. The robot-vs-macaque source comparison is better controlled but still lacks a frozen-random or non-informative-pretraining baseline.","major_comments":[{"comment":"The central comparison between pretrained and Scratch motor encoders conflates pretraining with the freezing protocol. The paper states that in Scratch the motor encoder is randomly initialized and jointly optimized during neural alignment, whereas Macaque-PT, Robot-Matched-PT, and Robot-Large-PT are kept frozen during neural alignment. Thus the reported 11% R2 increase on CO and 8% on RT could reflect the stabilizing effect of a fixed, nontrivial representation target rather than transferable behavioral knowledge. To support the central claim, the authors should add a control condition with a randomly initialized but frozen motor encoder, and ideally a motor encoder pretrained on shuffled or non-informative trajectories and frozen. Without such controls, the 'behavioral pretraining is a critical component' conclusion in the Effect of Pretrained Motor Encoder section is not justified.","section":"Method, 'Motor Encoder'; Experiments, 'Effect of Pretrained Motor Encoder'"},{"comment":"The calibration-efficiency comparison is asymmetric. In the transfer experiments, pretrained models optimize only newly introduced unit and session embeddings on the calibration subset, while the Scratch baseline 'initialized the entire model randomly and optimized all components using each calibration subset.' This heavily disadvantages Scratch in the low-data regime and makes the claim that 'only 10% calibration needed to match training from scratch' an artifact of the comparison protocol rather than a demonstration of behavioral transfer. The authors should either train Scratch under the same frozen-encoder protocol (with a frozen random encoder) or optimize all components of the pretrained models during calibration, so the amount of learned plasticity is matched.","section":"Experiments, 'Calibration-Efficiency Validation'; Appendix, 'Calibration Details'"},{"comment":"Table 1 reports NEDS, NDT2, and POYO results obtained from other publications (Ryoo et al. 2026; Mao et al. 2026) rather than re-run under the same data-splitting and evaluation protocol. Because the authors explicitly state that the classical baselines were 'reproduced using the same protocol' but do not do so for these three neural-decoding baselines, the state-of-the-art claim on both CO and RT is not established under a controlled comparison. Different split conventions, session filtering, or window overlap could easily explain part of the reported gap, particularly the RT improvement from 0.7156 (POYO) to 0.8475 (NeuroPB). The authors should either re-run these baselines with their own protocol or explicitly restrict the SOTA claim to the reproduced baselines.","section":"Experiments, 'Comparison with Baselines'; Table 1"},{"comment":"The source of the macaque trajectories used for Macaque-PT is never disclosed. If these trajectories come from the same dataset or, more seriously, from the same recording sessions used for neural-decoding evaluation, the Macaque-PT results could benefit from information leakage, and the robot-vs-macaque comparison would not be clean. The authors should state exactly how macaque trajectories were obtained, how many trajectories/sessions they comprise, and confirm that all pretraining data are disjoint from the evaluation sessions at the session level.","section":"Method, 'Scaled Pretrained Motor Encoder'; Experiments, 'Datasets'"}],"minor_comments":[{"comment":"The manuscript contains several grammatical errors that should be corrected in a language pass. Examples include 'For adapt to different numbers of spikes varies across windows' in the Neural Encoder section, 'The output of the final self-attention block as the neural representation' (missing verb), and 'multi-layer perception' in the Appendix.","section":"Various sections"},{"comment":"The left panel of Fig. 4 is labeled as a conceptual illustration of relative pretraining-data scales and is explicitly not derived from measured scores. This is an honest disclosure, but a quantitative table listing the number of trajectories, total time steps, and data diversity metrics for Macaque-PT, Robot-Matched-PT, and Robot-Large-PT would be more informative than a conceptual schematic.","section":"Figure 4 and Appendix, 'Details of robotic pretraining datasets'"},{"comment":"Table 1 reports R2 values with standard deviations but does not include the number of sessions or a significance test across methods. Given the small number of datasets/conditions, the authors should clarify how many sessions each mean covers and whether the differences between NeuroPB and the best baseline are statistically significant.","section":"Table 1"},{"comment":"The heading 'Generation Result' appears to be a typo for 'Generalization Result.' Please correct it.","section":"Appendix, 'Additional Results'"},{"comment":"The PCA visualization in Fig. 6B is qualitative. Since the paper claims that alignment 'suppresses session-dependent nuisance variation while preserving behaviorally relevant movement information,' a quantitative metric (e.g., silhouette score computed separately for direction labels and session labels) would strengthen this claim.","section":"Visualization Analysis"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely and potentially impactful idea, and the authors are transparent about some limitations (e.g., the conceptual nature of Fig. 4). However, the evidence for the headline claim is currently confounded by the frozen-vs-trainable motor encoder comparison and by the asymmetric calibration protocol. The borrowed baseline numbers are a further obstacle to the SOTA claim. These are fixable with additional experiments and reruns, so I recommend major revision rather than rejection. I would also ask the editor to check whether the appended 'Here, I will explain...' passage in the Appendix is an artifact that slipped into the manuscript text, as it is not in the formal style of the rest of the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth reading: it is the first to my knowledge to pretrain a motor encoder on large-scale robotic trajectories for neural decoding, and it combines masked trajectory pretraining with CLIP-style contrastive alignment in a sensible way. The framing—use behavior data instead of scarce neural data—is genuinely new in this subfield. The ablations are careful, and the comparisons among frozen pretrained encoders (Macaque-PT vs Robot-Matched-PT vs Robot-Large-PT) are clean and informative, since those conditions all freeze the motor encoder. The scaling result (larger, more diverse robotic data helps) is the strongest evidence in the paper, and it is not affected by the confound I am about to raise.\n\nThe soft spot is the Scratch baseline. Scratch randomly initializes the motor encoder and jointly optimizes it during alignment, whereas every pretrained condition freezes the encoder. So the reported 11% and 8% R2 gains conflate two variables: the information in the pretrained weights and the stability of a fixed alignment target. The paper does not include a control with a randomly initialized but frozen encoder, or one pretrained on shuffled trajectories. Without that, the headline claim—that behavioral pretraining, including robotic pretraining, improves decoding—is not yet established. A frozen random target might stabilize contrastive learning just as well. This is a load-bearing issue for the main claim, and the authors should run that control before I would trust the conclusion. The robot-vs-macaque comparison is less affected, since both are frozen, but even there, comparable performance from a frozen random encoder would undercut the 'transferable kinematic structure' interpretation.\n\nOther issues are real but secondary: the Macaque-PT pretraining trajectory source is undisclosed, which is a leakage risk; several baseline numbers in Table 1 are borrowed from other papers without re-running under the same protocol; and the 10% calibration claim lacks uncertainty quantification. No code or data are provided, which is increasingly expected for this kind of empirical work.\n\nWho is this for? Researchers in neural decoding and BCI who care about calibration-efficient methods and transfer across embodiments. The paper deserves peer review—the idea is timely and the architecture is plausible—but the review should explicitly ask for the frozen-random/shuffled control and the macaque-PT data source. I would not cite it in my own work until that control is in place.","headline":"The core idea is fresh, but the headline result currently conflates pretraining with freezing; the scale/source comparisons are cleaner and deserve a control to fully land.","tokens_in":14163,"tokens_out":2011,"would_cite":false,"duration_ms":22654,"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":"NeuroPB shows that a motor encoder pretrained on robotic trajectories, frozen, and used as an alignment target for macaque spike activity improves continuous trajectory decoding and scales with behavioral data.","keywords":["brain-computer interfaces","neural decoding","motor trajectory decoding","behavioral pretraining","contrastive representation alignment","robotic trajectories","macaque motor cortex","transfer learning"],"falsifier":"Pretrain the motor encoder on robotic trajectories whose temporal order has been shuffled, or on random-walk point trajectories with the same marginal speed and target statistics; if decoding $R^2$ does not fall below Robot-Large-PT, then the benefit is not due to transferable kinematic structure. A complementary check is to rerun Macaque-PT with any trajectory from a session that appears in the evaluation split removed; if the performance gain disappears, leakage is the explanation.","tokens_in":13129,"feed_emoji":"🧠","tokens_out":6509,"duration_ms":61670,"temperature":0.7,"pith_summary":"NeuroPB is a framework for decoding continuous arm movements from neural spikes that treats behavior, not neural data, as the scalable resource. The paper's central claim is that a motor encoder pretrained on large quantities of two-dimensional trajectories, whether macaque or robotic, defines a behavioral representation space that can serve as an alignment target for neural activity, and that this transfer improves trajectory decoding. On macaque M1/PMd recordings, aligning neural activity to a behavioral space pretrained on LIBERO-100 robotics data raises $R^2$ by 11 percentage points on center-out reaching and by 8 percentage points on random-target reaching over training the motor encoder from scratch. At matched pretraining scale, robotic trajectories perform as well as macaque trajectories, and increasing robotic data scale further improves decoding, cross-session/subject/task generalization, and calibration efficiency. The paper concludes that kinematic structure transfers across biological and artificial embodiments, making behavioral pretraining a practical route to high-performance, calibration-efficient brain-computer interfaces under limited neural data.","feed_headline":"Robot trajectories improve brain-signal decoding in macaques","feed_subtitle":"Aligning spikes to a frozen behavioral representation beats training from scratch and needs 10% calibration data.","key_machinery":"The load-bearing object is the pretrained motor encoder, an eight-layer transformer trained by masked trajectory reconstruction on two-dimensional point trajectories and then frozen. It produces a normalized behavioral representation by pooling contextualized tokens, and this representation is the fixed target space for the neural side. The neural encoder is a Perceiver-style architecture that turns spike events into a fixed set of latent tokens with rotary positional embeddings, pools them, and projects them into the same space; a CLIP-style symmetric contrastive loss aligns temporally paired neural and behavioral representations while pushing mismatched pairs apart. A lightweight MLP decoder then reconstructs the trajectory from the aligned neural representation, and at inference the motor encoder is discarded. This three-stage design, pretrain a behavioral space, align neural activity to it, decode from the aligned representation, is what carries the argument.","core_discovery":"The central discovery is that a frozen behavioral representation space, pretrained without any neural data, can act as an anchor that organizes neural population activity for motor decoding. The paper shows that pretraining a transformer motor encoder on masked trajectory reconstruction, using either macaque reaches or robotic manipulation point tracks, produces a space into which spike activity can be aligned with a CLIP-style contrastive loss, and that this alignment is what drives the decoding gains. Removing the contrastive alignment to the pretrained space lowers center-out $R^2$ from 94.78% to 82.96%, the largest drop in the ablations. Equally central is the scaling result: a motor encoder pretrained on LIBERO-100, containing 5,000 robotic trajectories, outperforms both a scale-matched robotic pretraining with 500 trajectories and macaque pretraining, even though the robotic data come from a different embodiment and task distribution. The authors take this as evidence for transferable kinematic structure shared across biological and artificial movement systems, and as a demonstration that behavioral data can substitute for scarce neural recordings in pretraining.","pith_inferences":["If the scaling trend extends beyond 5,000 trajectories, behavioral pretraining on much larger embodied datasets could push neural decoding toward plug-and-play calibration for new sessions and subjects, a regime the paper does not test.","Because the target space is purely kinematic, the same alignment recipe could be applied to non-invasive recordings such as EEG or fMRI, provided those signals carry enough movement-related information; the paper only evaluates invasive macaque data.","A direct extension would be to ablate the content of the behavioral representation, for example pretraining on speed-only or angle-only trajectories, to identify which kinematic features the neural encoder actually latches onto; the current study compares sources and scales but not feature content.","The paper does not disclose the session provenance of the macaque trajectories used for Macaque-PT; verifying that they do not overlap with evaluation sessions is an open check, and the robotic-pretraining results are independent of that provenance."],"forward_implications":["Behavioral pretraining is a scalable axis: with neural data held fixed, decoding accuracy improves as robotic pretraining data grow from 500 to 5,000 trajectories, so adding more behavior data should continue to help.","Calibration requirements drop sharply: with only 10% of target-session data, the pretrained model matches the scratch model trained on 100% of the data across session, subject, and task transfer settings.","The pretrained behavioral space suppresses session-specific nuisance variation: after alignment, neural representations cluster by reach direction rather than by recording session, supporting better cross-session generalization.","The motor encoder is not needed at deployment, so the added pretraining cost is paid once offline; the deployed model is just the neural encoder plus a light decoder."],"supporting_citations":[{"why":"Supplies the POYO Perceiver-based neural encoder architecture that NeuroPB adapts to spike-event inputs.","marker":"Azabou et al. 2023"},{"why":"Supplies the CLIP-style symmetric contrastive objective used for neural-behavioral alignment.","marker":"Radford et al. 2021"},{"why":"Supplies the macaque M1/PMd center-out and random-target reaching dataset with paired spikes and trajectories.","marker":"Perich, Gallego, and Miller 2018"},{"why":"Supplies the LIBERO robotic manipulation datasets whose trajectories are converted into point tracks for motor-encoder pretraining.","marker":"Liu et al. 2023"},{"why":"Supplies the masked trajectory reconstruction objective used to pretrain the motor encoder.","marker":"Wen et al. 2023"},{"why":"Supplies rotary positional embeddings used to encode spike timing in the neural encoder's attention.","marker":"Su et al. 2024"}],"fun_headline_variants":["Robot motion pretraining boosts brain-signal decoding","Macaque brain decoding improved by robot movement data","Pretraining on robot reaches sharpens neural decoding","Robot trajectories help decode macaque brain signals","Moving robots teach decoders to read macaque brains"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim rests on the assumption that a motor encoder trained only on two-dimensional point trajectories from robots captures kinematic structure that is compatible with macaque motor cortex activity, so that freezing that space and aligning spikes to it helps rather than distorts decoding.","fun_headline_variants_meta":{"raw":{"variants":["Robot motion pretraining boosts brain-signal decoding","Macaque brain decoding improved by robot movement data","Pretraining on robot reaches sharpens neural decoding","Robot trajectories help decode macaque brain signals","Moving robots teach decoders to read macaque brains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000204,"raw_usage":{"total_tokens":1434,"prompt_tokens":1036,"completion_tokens":398,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":652,"completion_tokens_details":{"reasoning_tokens":326}},"tokens_in":652,"tokens_out":398,"duration_ms":4555,"temperature":1.0,"reasoning_tokens":326,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T18:26:05.967190+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Pretrain the motor encoder on robotic trajectories whose temporal order has been shuffled, or on random-walk point trajectories with the same marginal speed and target statistics; if decoding $R^2$ does not fall below Robot-Large-PT, then the benefit is not due to transferable kinematic structure. A complementary check is to rerun Macaque-PT with any trajectory from a session that appears in the evaluation split removed; if the performance gain disappears, leakage is the explanation.","supporting_citations":[],"review_version":1}