{"id":"22720880-e9ac-4a24-904c-c4e2f44bd54f","arxiv_id":"2507.15274","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"Temporal basis function models predict the spatiotemporal LFP response to optogenetic stimulation with test-set R2 around 0.46, beating linear state-space and LSTM baselines while training 30 to 100 times faster.","lead":"A new model, the temporal basis function model, predicts how an optogenetic brain stimulator's paired light pulses will change local field potentials in macaques, using just 20 milliseconds of prior brain activity. It trains in minutes on a desktop CPU and could make AI-driven closed-loop neural stimulation practical.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Closed-loop simulations use the recorded stimulated trajectory as ground truth for both 'stimulate' and 'no-stimulate' decisions (Sec. 4.9, App. 6.3), so the reported AUC measures forecast accuracy, not control. The closed-loop claim needs a counterfactual no-stimulation plant model.","rationale":"The forward prediction contribution is well-supported: the TBFM achieves held-out test R2 of 0.462 on a 164 ms horizon, clearly beats a state-agnostic baseline (0.006), and the efficiency claims (training time, latency, sample efficiency) are concrete and reproducible from the provided implementation. The closed-loop demonstration, however, is the second pillar of the paper and appears in the title, abstract, and Section 3.6. The simulations do not model the outcome of a 'no stimulation' decision; they assign the recorded stimulated trajectory as the ground truth for both branches of the controller's binary choice. This makes the reported AUC a measure of how well the model forecasts the fixed stimulation response, not whether the controller can shape neural activity. The reader's weakest assumption correctly identifies this conflation, and it is the most load-bearing concern because it directly undermines a central advertised contribution. The appropriate remedy is either to build a counterfactual no-stimulation plant model (e.g., using the TBFM itself with a zero stimulation descriptor, or matched resting-state trials) and re-evaluate the control performance, or to explicitly soften the claims to 'closed-loop triggering based on forecasted states' rather than 'closed-loop control.' A minor internal inconsistency also exists: the main text (Sec. 4.2) says the TBFM beats the AE-LSTM on all but 6 of 40 sessions (34), while Figure 14's caption says 32 of 40; this does not affect the core forward-model results but should be corrected. The conditional verdict is appropriate; the closed-loop claim needs substantial revision or re-analysis before the paper can be accepted as-is.","tokens_in":25734,"tokens_out":6259,"duration_ms":67243,"concrete_test":"Re-run both demonstrations with a proper plant model for the no-stimulation branch. For each test trial, use the TBFM itself (or a matched resting-state trial) to generate the trajectory that would occur if the controller chooses not to stimulate—e.g., by setting the stimulation descriptor to zero or using the model's prediction with no stimulation pulse. Then compute the ROC/AUC using the appropriate counterfactual for each decision: if stimulated, use the recorded stimulated trial; if not stimulated, use the generated no-stimulation trajectory. Additionally, compare the controller's decisions against a trivial policy (e.g., always stimulate or random) on the same outcomes. If the AUC remains substantially above 0.5 and beats the trivial policies, the closed-loop claim is supported; if not, the simulations must be reinterpreted as forecasting evaluations.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Demonstration 1 (Sec. 4.9.1), a 'false positive' is defined as the controller applying stimulation but the trajectory not crossing the target, and a 'false negative' as no stimulation but the trajectory crossing. In both cases 'the trajectory' is the recorded stimulated trial; the paper never generates or evaluates the no-stimulation outcome. For trials labeled 'do not stimulate' (App. 6.3.1), the target is chosen randomly, but the recorded trajectory still contains the stimulation response. Thus the ROC curve asks whether the model can predict which recorded stimulated trajectories will cross a target, not whether deciding not to stimulate would actually avoid the target. Demonstration 2 (App. 6.3.2) has the same flaw: reference trajectories are weighted averages of stimulated test trials, and the cost function compares the forecast (assuming stimulation) to the reference. The 'actual trajectory assuming stimulation' is used as the outcome; a no-stimulation trajectory is never simulated. Consequently, the AUC values (~0.70 and ~0.72) reflect the TBFM's forecasting skill under a fixed stimulation regime, not closed-loop control. The claim in the abstract and Sec. 3.6 that the model 'successfully allowed a closed-loop stimulator to control a neural circuit' is therefore unsupported. This is not a minor validation gap: the control problem and the prediction problem are conflated because the counterfactual branch of the controller's decision is missing.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes temporal basis function models (TBFMs) for single-trial, spatiotemporal forward prediction of local field potential (LFP) responses to optogenetic stimulation, targeting sample efficiency, short training time, and low inference latency for closed-loop neural stimulation. The authors evaluate the model on 40 sessions of paired-pulse optogenetic stimulation data from two non-human primates, reporting a mean test-set R2 of 0.462 over a 164 ms horizon on time-domain data, with improvements over a linear state-space model and accuracy comparable to or better than a nonlinear AE-LSTM baseline. They also report statistical evidence for state-dependence of the stimulation response, benchmarks of training time and latency, and two simulated closed-loop controllers that use TBFM forecasts to decide whether to stimulate, with reported AUC values of roughly 0.65–0.72.","tokens_in":26090,"tokens_out":7133,"duration_ms":66415,"significance":"If the forward-model claims hold, the contribution is practically important: the model appears to offer a favorable trade-off among prediction accuracy, sample efficiency (about 5k trials), training time (2–5 minutes), and inference latency (about 0.2 ms on a CPU), directly addressing translational barriers for model-based neural co-processors. Strengths of the evaluation include a time-based train/test split that tests generalization across each session, a state-agnostic ablation showing the value of conditioning on the initial state, and a cross-validation control for non-stationarity and plasticity. The authors also release a PyTorch implementation and demo code. However, the closed-loop simulation results do not validly demonstrate closed-loop control, as detailed in the major comments; this limits the scope of the central claim but does not undermine the forward-modeling contribution itself.","major_comments":[{"comment":"The evaluation of Demonstration 1 conflates prediction with control. Because stimulation was always applied in the stimulation blocks (§3.1), no true no-stimulation trials exist in those blocks; nevertheless, Appendix 6.3.1 labels test-set trials as 'do not stimulate' trials and uses the recorded (stimulated) trajectory as the ground-truth outcome for both the stimulate and do-not-stimulate decisions. Consequently, the ROC analysis in §4.9.1 measures whether the model can forecast which already-stimulated recorded trials will cross the target range, not whether the controller's decision changes the outcome. The statement in §4.9.1 that 'These results indicate that the TBFM achieved overall controllability' does not follow from the reported analysis.","section":"§4.9.1, §6.3.1"},{"comment":"Demonstration 2 suffers from the same counterfactual missingness. The reference trajectories are weighted averages of stimulated test trials, and the cost function compares the TBFM forecast (which assumes stimulation) to the reference; the 'actual trajectory assuming stimulation' is the recorded stimulated trial, and no no-stimulation trajectory is ever simulated. The reported AUC values (0.721 raw time domain, 0.637 beta) therefore quantify the model's forecasting skill under a fixed always-stimulate regime, not its ability to control neural activity. The abstract and §3.6 claim that the model 'successfully allowed a closed-loop stimulator to control a neural circuit' is unsupported by the present evidence and should be withdrawn or substantially qualified.","section":"§4.9.2, §6.3.2"}],"minor_comments":[{"comment":"The text in §4.2 says the TBFM exceeded the AE-LSTM on 'all but 6 of the 40 sessions' (i.e., 34/40), while the Figure 14 caption says 'on 32 of 40 sessions'; these numbers should be reconciled.","section":"§4.2 vs. Fig. 14 caption"},{"comment":"The sentence reporting that test-set R2 was 'predictive of the AUC' gives only the p-value (p=1.01e−3); the correlation coefficient and the number of sessions should be reported.","section":"§4.9.2"},{"comment":"The phrase 'was loss-pass filtered' should read 'was low-pass filtered'.","section":"§6.1"},{"comment":"The column heading 'R2 mean-vs-mean' is not defined in the caption; please define it in the caption or in §4.2.","section":"Table 1"},{"comment":"The FSAM stopping rule is described as adding bases until validation-set accuracy no longer increases, but no quantitative convergence threshold is given; please specify the criterion used in Figure 10.","section":"§3.5/§4.8"},{"comment":"The code availability statement says the implementation is available under an open-source license but does not name the license; please specify it.","section":"§5.3"}],"recommendation":"major_revision","confidential_remarks":"The forward-model portion of the paper is sound and publishable in principle. The closed-loop claims, however, are overreaching and need to be reframed or removed before acceptance; I would ask the authors to either supply a no-stimulation plant model or clearly relabel the simulations as forecast-based triggering analyses. I do not see this as a reject, because the forward-model contribution is independent and substantial."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the forward model is a solid, practical contribution; the closed-loop simulation is not. The TBFM architecture and the 40-session evaluation are worth engaging with seriously. But the claim to have demonstrated closed-loop control is unsupported by the simulations as designed.\n\nWhat the paper does well: TBFM is a genuinely simple and effective idea — an MLP generates temporal bases from stimulation descriptors, an affine map from the pre-stimulation runway produces per-channel weights, and the output is a weighted sum. The engineering metrics are the real story: 5k trials (about 17 minutes) of data, 2–5 minutes of training on a CPU, 0.115 ms compiled inference, and test R² of 0.462 over 164 ms, beating an AE-LSTM on 34 of 40 sessions and LSSM on all sessions. The sample-efficiency and latency analysis is careful and directly addresses the translational barrier. The state-dependence analysis in Sections 4.1 and 4.3 is also solid: the KSG-MI and HSIC tests are reasonable, and the state-agnostic ablation is a fair control.\n\nThe soft spot is Section 4.9 and Appendix 6.3. The simulations evaluate the controller against the recorded stimulated trajectory in every trial, whether the controller chooses to stimulate or not. For a \"do not stimulate\" decision, the correct outcome is the counterfactual no-stimulation trajectory, which is never generated. The AUC of about 0.70 therefore measures whether the model can predict which already-stimulated recordings will cross a target — a forecasting test, not a control test. The abstract's phrase \"successfully allowed a closed-loop stimulator to control a neural circuit\" overstates what the evidence shows. This is a fixable flaw: replace the ground truth with a TBFM-generated no-stimulation prediction, or collect data where stimulation is withheld in some trials, or reword the claim to \"simulated assessment of forecast-triggered stimulation.\" The rest of the paper does not depend on this section.\n\nBottom line: this deserves a serious referee. The forward model and efficiency comparisons are a real contribution to the neuromodulation and BCI literature. I would ask for a major revision on the closed-loop claims before acceptance.","headline":"Solid forward model with real translational metrics, but the closed-loop simulation claim conflates forecasting with control because no no-stimulation counterfactual is ever generated.","tokens_in":26574,"tokens_out":2140,"would_cite":true,"duration_ms":23792,"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":"A temporal basis function model predicts single-trial, spatiotemporal responses to optogenetic stimulation from a 20 ms runway of multichannel LFP, with held-out $R^2$ of 0.462 over a 164 ms horizon; the paper argues this makes…","keywords":["temporal basis function model","closed-loop neural stimulation","optogenetics","local field potentials","state-dependent stimulation response","non-human primate","brain co-processor","sample-efficient machine learning"],"falsifier":"Run the trained TBFM in a real-time closed-loop experiment on the same non-human primate preparation, with the controller instructed not to stimulate on a random subset of trials and the true unstimulated LFP recorded under each initial state; if the decision performance computed against those genuine no-stimulation outcomes does not track the reported simulated AUCs, the closed-loop claim is not supported.","tokens_in":25549,"feed_emoji":"🧠","tokens_out":10215,"duration_ms":107739,"temperature":0.7,"pith_summary":"This paper argues that a simple temporal basis function model can learn, from about 5,000 trials collected early in a session, a single-trial forward prediction of how paired-pulse optogenetic stimulation will reshape local field potentials across an electrode array. On held-out trials the model achieves a mean test-set $R^2$ of 0.462 over a 164 ms forecast horizon, and it beats a more complex recurrent model on most sessions and a linear state-space model on every session while training in minutes rather than hours. The point is translational: if a forward model is this cheap to build and this fast to evaluate, closed-loop stimulation can be shaped to the patient's current brain state within a realistic loop latency. The authors also present statistical evidence that the stimulation response depends on the pre-stimulus brain state, and they show that removing that state dependence from the model collapses its accuracy.","feed_headline":"Simple model beats deep networks at predicting brain responses","feed_subtitle":"Trained in minutes, it forecasts state-dependent LFP responses in under 0.2 ms.","key_machinery":"The central object is the temporal basis function model (TBFM), defined by $\\hat{y}_c = x_{c,r}\\mathbf{1} + \\sum_{i=1}^b W(X)_{c,i} B_{i,*}$: the forecast for each channel is the last measured LFP value plus a weighted sum of $b$ temporal basis functions. The basis functions are produced by a small multilayer perceptron from a stimulation descriptor (here, the timing of the two light pulses), and the weight matrix $W$ is an affine function of the Z-scored \"runway\"—the final 20 ms of multichannel LFP before the forecast begins. This decomposition carries the argument because the nonlinear dependence on stimulation parameters lives in the basis generator, while the trial-specific dependence on brain state lives in the linear weight estimator, and the whole horizon is produced in one forward pass instead of being unrolled through time. After training, the nonlinear generator can be compiled away for a discrete set of stimulation parameters, leaving a linear map that runs in about 0.115 ms on a desktop CPU.","core_discovery":"The paper's central claim is that the forward model for optogenetic stimulation does not need to be a deep recurrent network: a weighted sum of a small set of learned temporal basis functions, with weights read off affinely from the recent multichannel LFP, is enough to predict the spatiotemporal response on individual trials. The authors report a mean test-set $R^2$ of 0.462 (stdev 0.207) over the full 164 ms forecast, 0.787 over a 40 ms forecast, and a state-dependent $R^2$ of 0.878 when predictions are binned by initial state. When the model is given a sham runway, test $R^2$ falls to 0.006 (stdev 0.012), which the authors take as evidence that the model is conditioning on the pre-stimulus brain state rather than emitting a mean evoked response. They interpret the overall result as showing that the practical barriers to model-based closed-loop stimulation—sample efficiency, training time, and loop latency—can be addressed without giving up accuracy.","pith_inferences":["One extension the paper leaves implicit is that the same architecture should generalize to continuously varying pulse parameters such as width, amplitude, or site, since the stimulation descriptor can be extended; the paper demonstrates only timing variation, so this remains untested. ","Because the basis generator can be compiled into a fixed linear map once the stimulation set is discrete, the paper's latency figures suggest a version of the model could run on low-power embedded hardware for an implantable controller, though the authors do not test this. ","The decisive test of the closed-loop claim would be a real-time experiment with genuine no-stimulation trials, comparing trajectories after a no-stimulate decision with unstimulated baselines matched on initial state; the reported AUCs are computed against recorded trials in which stimulation was always delivered. "],"forward_implications":["A session can be modeled from about 5,000 early trials (under 20 minutes of collection), trained in 2 to 5 minutes, and then used for stimulation decisions on the rest of the session. ","A clinically deployed stimulation model should feed the controller the ongoing LFP rather than only the stimulation schedule, because state-agnostic models in this dataset achieve near-zero $R^2$. ","The measured forecast latency of about 0.2 ms on a desktop CPU leaves ample room within a 20 ms control loop for sensing, filtering, and stimulation hardware. ","If the comparison with the autoencoder-LSTM baseline generalizes, complex recurrent dynamics are not necessary for this forecasting task, and simpler, easier-to-verify models can be used in safety-critical settings. ","Even a model with test $R^2$ around 0.46 can, in the paper's simulations, time stimulation to predicted target states and toward reference trajectories, suggesting that useful closed-loop control may not require near-perfect prediction. "],"supporting_citations":[{"why":"Supplies the 40-session paired-pulse optogenetic micro-ECoG dataset from two rhesus macaques on which the TBFM is trained and evaluated.","marker":"[27]"},{"why":"Provides the general evidence that brain and behavioral responses to stimulation depend on the current brain state, motivating the runway-conditioned design.","marker":"[17]"},{"why":"Documents state-dependent effects of cortical stimulation in non-human primates, supporting the paper's state-dependence hypothesis and analysis.","marker":"[20]"},{"why":"Defines the electrode-rejection criteria and the plasticity and non-stationarity results used for the early-session to late-session split.","marker":"[38]"},{"why":"Supplies the linear state-space modeling approach that the authors adapt and then exceed in both accuracy and speed.","marker":"[33]"},{"why":"Provides the earlier state-space optimal feedback control formulation for optogenetically driven activity used as the simpler baseline family.","marker":"[11]"},{"why":"Underlies the AE-LSTM baseline's recurrent dynamics and training procedure, the more expressive model the TBFM is compared against.","marker":"[44]"},{"why":"Informs the tripartite loss used to train the AE-LSTM latent dynamics model.","marker":"[45]"},{"why":"Supplies the phase-locked stimulation context and the few-millisecond loop latency figures that motivate the controller timing demonstrations.","marker":"[21]"}],"fun_headline_variants":["Simple model predicts brain responses in 0.2 ms","Basis functions rival complex models for brain control","Fast-learning model shapes brain stimulation in real time","Low-latency brain model can tailor stimulation per patient","Minutes of data train brain stimulation forecaster"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that a trial recorded with stimulation always delivered is the correct ground-truth outcome even when the controller decides not to stimulate; if true no-stimulation trajectories differ, the simulated closed-loop results do not establish that the controller can shape brain activity.","fun_headline_variants_meta":{"raw":{"variants":["Simple model predicts brain responses in 0.2 ms","Basis functions rival complex models for brain control","Fast-learning model shapes brain stimulation in real time","Low-latency brain model can tailor stimulation per patient","Minutes of data train brain stimulation forecaster"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000283,"raw_usage":{"total_tokens":1729,"prompt_tokens":1063,"completion_tokens":666,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":679,"completion_tokens_details":{"reasoning_tokens":592}},"tokens_in":679,"tokens_out":666,"duration_ms":8017,"temperature":1.0,"reasoning_tokens":592,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:36:32.725812+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained TBFM in a real-time closed-loop experiment on the same non-human primate preparation, with the controller instructed not to stimulate on a random subset of trials and the true unstimulated LFP recorded under each initial state; if the decision performance computed against those genuine no-stimulation outcomes does not track the reported simulated AUCs, the closed-loop claim is not supported.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the 40-session paired-pulse optogenetic micro-ECoG dataset from two rhesus macaques on which the TBFM is trained and evaluated."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the general evidence that brain and behavioral responses to stimulation depend on the current brain state, motivating the runway-conditioned design."},{"cited_title":"IEEE Engineering in Medicine and Biology Society","cited_arxiv_id":null,"evidence_quote":"Documents state-dependent effects of cortical stimulation in non-human primates, supporting the paper's state-dependence hypothesis and analysis."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the electrode-rejection criteria and the plasticity and non-stationarity results used for the early-session to late-session split."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the linear state-space modeling approach that the authors adapt and then exceed in both accuracy and speed."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the earlier state-space optimal feedback control formulation for optogenetically driven activity used as the simpler baseline family."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Underlies the AE-LSTM baseline's recurrent dynamics and training procedure, the more expressive model the TBFM is compared against."},{"cited_title":"Learning to Correspond Dynamical Systems","cited_arxiv_id":"1912.03015","evidence_quote":"Informs the tripartite loss used to train the AE-LSTM latent dynamics model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the phase-locked stimulation context and the few-millisecond loop latency figures that motivate the controller timing demonstrations."}],"review_version":1}