{"id":"a343daf8-51f5-4bc5-97e0-c349df68fcf3","arxiv_id":"2606.03481","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Short-term synaptic plasticity in a PFC-inspired reservoir model preserves goal-conditioned dynamics under noise, keeping multistep action selection robust where fixed-connectivity versions fail.","lead":"The paper tests whether short-term synaptic plasticity stabilizes goal representations in a reservoir model of prefrontal cortex during delayed multistep action planning. Under noise, the version with plasticity maintains high success rates while the version without drops sharply, suggesting a role for dynamic synaptic modulation.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.3","headline":"Grid search for reservoir/STP parameters plus specific noise model risks artifactual favor to STP condition","rationale":"The reader's weakest assumption directly names the load-bearing modeling choice. No other internal inconsistency (e.g., in the decoding or effective-connectivity analyses) is visible from the provided abstract and claim. The concern is therefore unchanged from the reader's assessment.","tokens_in":1818,"tokens_out":307,"duration_ms":12387,"concrete_test":"Re-run the identical grid search over reservoir parameters (no STP) to maximize success rate under the same state-noise condition used in the original paired comparison; if the optimized no-STP success rate rises above 80% while STP remains ~90%, the reported robustness advantage is likely due to asymmetric tuning rather than STP itself.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline robustness result (no-STP success drops 75.8%→49.5% under noise while STP stays ~90%) is the central empirical support. Both reservoir weights and the two STP time constants were selected by grid search, and the state-noise model is not independently validated against PFC data. This leaves open that the search identified a narrow regime in which STP's short-term modulation coincidentally compensates for the particular noise statistics, while the no-STP baseline was not re-optimized under the same noise. Gain-matched and perturbation controls address fixed scaling but do not test whether unequal optimization or noise-STP interaction produced the gap.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript claims that short-term synaptic plasticity (STP) stabilizes goal-conditioned dynamics in a PFC-inspired reservoir model for multistep goal-directed action planning. Paired simulations across 100 networks show goal identity remains decodable without STP, but under state noise success without STP drops from 75.8% to 49.5% while STP maintains performance (91.8% to 89.2%; dz=1.31). Time-resolved decoding, separability, action-value, and effective-connectivity analyses indicate STP preserves action-usable dynamics via history-dependent modulation rather than fixed scaling; a grid search identifies a facilitation-dominant STP regime.","tokens_in":1989,"tokens_out":621,"duration_ms":16162,"significance":"If the robustness result holds, the work supplies a concrete mechanistic account of how STP can convert recurrent activity into goal- and task-state-conditioned dynamics usable at delayed action opportunities, directly addressing a gap in PFC planning models. Credit is due for the paired-network design, multiple controls (gain-matched, STP-state perturbation), statistical reporting, and effective-connectivity analysis showing time-varying goal-specific patterning only with STP. These elements make the dynamic-modulation claim testable and stronger than a simple scaling account. The single-task, grid-searched parameter regime limits immediate generality but does not undermine the internal comparison.","major_comments":[{"comment":"Methods, model parameterization and grid search: Reservoir weights and the two STP time constants were selected by grid search to maximize task performance. The no-STP baseline was not reported as re-optimized under the identical state-noise model used for the key comparison. This leaves open the possibility that the reported gap (75.8%→49.5% vs. 91.8%→89.2%) partly reflects unequal optimization rather than an intrinsic STP effect. A direct test would be to re-optimize the no-STP weights under noise and repeat the paired evaluation.","section":"Methods (grid search and noise robustness)"},{"comment":"Results, noise-robustness paragraph and controls: The state-noise model is a fixed-amplitude perturbation whose statistics are not compared to empirical PFC variability. While gain-matched and perturbation controls address fixed scaling, they do not test whether the STP advantage persists under alternative noise regimes (e.g., multiplicative or input-dependent). This is load-bearing for the claim that STP confers general robustness.","section":"Results (noise robustness)"}],"minor_comments":[{"comment":"Figure legends and methods should explicitly state that all statistics are paired across the same 100 networks and report the exact grid ranges searched for STP time constants.","section":"Figures and Methods"},{"comment":"The effective-connectivity analysis would benefit from a quantitative comparison of time-invariance (e.g., a statistical test on the time-dependence of goal-specific edges).","section":"Results (effective connectivity)"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We are grateful to the referee for their detailed and insightful comments, which have helped us improve the manuscript. We address each major comment below and outline the revisions we will make.","responses":[{"response":"We thank the referee for this suggestion. The original grid search was performed without noise to identify parameters that maximize baseline performance. To address this concern, in the revised manuscript we will re-optimize the no-STP reservoir weights under the state-noise condition using the same grid-search procedure and re-evaluate the paired comparison. This will clarify whether the robustness advantage persists when both models are optimized under identical conditions.","revision_made":"yes","referee_comment":"Methods, model parameterization and grid search: Reservoir weights and the two STP time constants were selected by grid search to maximize task performance. The no-STP baseline was not reported as re-optimized under the identical state-noise model used for the key comparison. This leaves open the possibility that the reported gap (75.8%→49.5% vs. 91.8%→89.2%) partly reflects unequal optimization rather than an intrinsic STP effect. A direct test would be to re-optimize the no-STP weights under noise and repeat the paired evaluation."},{"response":"We agree that the noise model is a specific choice and its statistics were not matched to empirical data. The manuscript's claim is specifically that STP stabilizes performance under this additive state-noise perturbation, as evidenced by the controls showing dynamic modulation rather than scaling. We do not claim generality to all possible noise regimes. In revision, we will add a paragraph in the Discussion explicitly stating the assumptions of the noise model and noting that testing multiplicative or input-dependent noise would be a valuable extension. This addresses the scope of the robustness claim.","revision_made":"partial","referee_comment":"Results, noise-robustness paragraph and controls: The state-noise model is a fixed-amplitude perturbation whose statistics are not compared to empirical PFC variability. While gain-matched and perturbation controls address fixed scaling, they do not test whether the STP advantage persists under alternative noise regimes (e.g., multiplicative or input-dependent). This is load-bearing for the claim that STP confers general robustness."}],"tokens_in":1603,"tokens_out":478,"duration_ms":17970,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The central result is that adding short-term plasticity to the reservoir keeps task success near 90% even when state noise is injected during the delay, whereas the matched model without STP falls from 76% to 50%. The paper shows this across 100 networks with paired tests and an effect size of 1.31.\n\nWhat is new is the combination of a multistep goal-directed task, explicit noise during the delay, and the follow-up analyses of time-resolved decoding, state-space separability, and effective connectivity. The connectivity measure is particularly useful: with STP it develops goal-specific patterning that strengthens toward the action window, while without STP it stays flat. The gain-matched and perturbation controls are straightforward and rule out a simple fixed-gain account.\n\nThe controls are solid enough for a modeling paper. The authors also report that goal identity remains decodable without STP, so the claim is not that STP is required for representation but that it stabilizes the dynamics needed for later readout.\n\nThe soft spot is the parameter regime. Reservoir weights and the two STP time constants were chosen by grid search, and the noise amplitude is a free parameter. The no-STP baseline was not re-optimized under the same noise, so part of the performance gap could reflect that difference rather than a general property of STP. The noise model itself is not checked against PFC data, which keeps the result inside the simulation.\n\nThis is a paper for people who build or analyze reservoir-style models of prefrontal function and want to see how STP interacts with goal maintenance under perturbation. A reader who values careful controls on a single task will get something from it.\n\nSend it for review. The simulation comparison and the connectivity analysis are concrete enough to justify referee time, even though the parameter and noise choices will need scrutiny.","headline":"STP keeps success rates stable under state noise in this reservoir model while the no-STP version drops sharply, but the grid-searched parameters and single noise model are the main limits on how far the result travels.","tokens_in":2474,"tokens_out":452,"would_cite":false,"duration_ms":17489,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Short-term synaptic plasticity stabilizes goal-conditioned dynamics against state noise in a PFC-inspired reservoir model.","keywords":["short-term synaptic plasticity","reservoir computing","prefrontal cortex","goal-directed action planning","multistep tasks","state noise","effective connectivity","temporal difference learning"],"falsifier":"Re-running the simulations with STP disabled after the delay period begins and finding that success rates under noise remain as high as when STP stays active throughout would falsify the claim that ongoing dynamic modulation is required.","tokens_in":2713,"feed_emoji":"🧠","tokens_out":727,"duration_ms":20672,"temperature":0.7,"pith_summary":"The paper asks whether short-term synaptic plasticity can keep goal information in a form that remains directly usable for later actions inside recurrent circuits modeled on prefrontal cortex. It builds paired reservoir networks, one with and one without STP, that learn action values through temporal-difference rules and tests them on a multistep goal-directed selection task with delays. Goal identity stays readable in both versions, yet only the STP version keeps high success rates when state noise is added; performance without STP falls sharply while the STP version stays near 90 percent. Time-resolved decoding, separability measures, and effective-connectivity analysis show that STP creates goal-specific, time-varying recurrent patterns that align with upcoming action opportunities. Grid search further locates a facilitation-dominant range of STP time constants that supports this robustness.","feed_headline":"Plasticity keeps goal signals usable in noisy PFC models","feed_subtitle":"Success stays near 89 percent with STP under noise while dropping to 50 percent without it, via dynamic goal-specific connectivity.","key_machinery":"Short-term synaptic plasticity as an online modulator of synaptic strengths that dynamically reshapes effective recurrent connectivity in a goal- and task-state-dependent manner.","core_discovery":"Incorporating short-term synaptic plasticity into the PFC-inspired reservoir model preserves high task success under state noise by maintaining goal-conditioned dynamics that remain available for action selection at later times. This occurs through online, history-dependent modulation that produces goal-specific patterning in effective recurrent connectivity, which grows stronger toward the end of the delay period; without STP the connectivity stays time-invariant and performance collapses under noise. Gain-matched and STP-state perturbation controls indicate the benefit is not explained by simple fixed scaling.","pith_inferences":["The same STP mechanism could allow biological PFC to hold multiple goals across behavioral timescales without requiring persistent firing in every neuron.","Analogous short-term plasticity rules might stabilize delayed decisions in other recurrent circuits that must operate under internal noise.","Optogenetic or pharmacological disruption of STP during noisy multistep tasks in behaving animals would provide a direct test of whether the modeled robustness appears in vivo."],"forward_implications":["STP is unnecessary for forming a linearly readable goal representation but essential for keeping that representation in an action-usable dynamical form under noise.","Effective connectivity becomes goal-specific and increases in strength toward later trial epochs only when STP is present.","Facilitation-dominant STP time constants identified by grid search reliably produce the highest success rates.","Perturbation of STP states during the delay disrupts the preserved action-value differences, confirming history dependence rather than static gain."],"fun_headline_variants":["STP stabilizes goal dynamics under noise in PFC models","Plasticity preserves goal signals in noisy reservoir networks","STP maintains goal-conditioned dynamics despite state noise","Short-term plasticity supports goal-specific connectivity patterns"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The reservoir and STP parameters chosen by grid search together with the specific state-noise model accurately represent biological prefrontal dynamics and do not create an artificial advantage for the STP condition.","fun_headline_variants_meta":{"raw":{"variants":["STP stabilizes goal dynamics under noise in PFC models","Plasticity preserves goal signals in noisy reservoir networks","STP maintains goal-conditioned dynamics despite state noise","Short-term plasticity supports goal-specific connectivity patterns"]},"model":"grok-4.3","cost_usd":0.007315,"raw_usage":{"total_tokens":3425,"prompt_tokens":783,"num_sources_used":0,"completion_tokens":55,"cost_in_usd_ticks":73149500,"prompt_tokens_details":{"text_tokens":783,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2587,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":783,"tokens_out":55,"duration_ms":18971,"temperature":1.0,"reasoning_tokens":2587,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-28T07:42:27.586820+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Re-running the simulations with STP disabled after the delay period begins and finding that success rates under noise remain as high as when STP stays active throughout would falsify the claim that ongoing dynamic modulation is required.","supporting_citations":[],"review_version":1}