acceptodds
Under review as a conference paper at ICLR 2027

A Timescale-Adaptive RNN for Rule-Guided Reorganization of Sequence Working Memory

Abstract

Sequence working memory (SWM) is not merely a system for temporarily maintaining information, but an active computational workspace in which stored representations can be reorganized according to task demands. Recent recordings from primate frontal cortex revealed rank-content factorized representations and additional transient subspaces associated with sequence reordering, but the underlying circuit mechanisms remain unclear. Here, we show that a recurrent neural network with cue-triggered modulation of neuronal integration time constants can support flexible SWM manipulation while recapitulating key experimentally observed representational and dynamical signatures of SWM. After training on a rule-guided SWM manipulation task, the network developed a content-independent scaffold of ordinal-rank subspaces that served as reusable slots for different items. Before rule presentation, item-rank mappings were maintained; the forward rule preserved these mappings, whereas the backward rule reassigned items between rank slots. Item information was transiently routed into two additional subspaces during reordering, consistent with a buffering role. Comparisons across architectures and timescale parameters confirmed the importance of cue-triggered timescale modulation for learning this factorized memory organization. Three-item experiments further provided initial evidence that learned operations could be compositionally reused to support multi-step working-memory manipulation. These results suggest a concrete recurrent mechanism for flexible item–rank reorganization and provide mechanistic insight into how mental programming may operate through computations over factorized content and ordinal-rank representations.

Primary area:
applications to neuroscience & cognitive science
Keywords:
sequence working memory, recurrent neural network, order manipulation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_1eWBOpenReview ↗
Rating
2
Confidence
4
Informativeness
2(+1)
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper considers a sequential working memory task, where the information does not only need to be memorized throughout delay but also manipulated (reversing order depending on a rule cue). A model where certain units receive prescribed rule-cue-related time constant modulation was compared with other recurrent architectures. It is shown to outperform the other models at matching the experimental observations, especially the rule related change of representations.

Critical strengths

It is a right call to focus on the manipulation aspect of working memory.

Well made figures.

Critical weaknesses

The considered model received prescribed time resolved rule-cue-conditioned time constant modulation. It is not a comparison between architectures, but the effect of such prescribed time constant modulation.

Insufficient experiments on variants of time constant modulation.

The worse performance of LSTM and FS RNN is not convincing. It could simply be that they are not trained enough, while the adaptive timescale RNN already receives what they are supposed to be trained to do, thus being faster.

The paper offers little mechanistic insight into the link between timescale modulation and the advantage in the working memory task.

Insufficient analyses on the AT neurons. There is currently only the percentage of AT neurons in G1 G2 and O1 O2. Most of the analyses could be factorized by AT and FT populations. What about the representation in AT population? Does information in AT neurons persist?

Questions for authors

how does prescribing timescale modulation help solve this task? Do they get more receptive to new input which are from other memory subspaces?

what if you allow the time constant to be learned in an input dependent fashion (globally or in discrete groups)? What if you have fixed input independent modulations for each neuron?

3 comments
AuthorsOpenReview ↗

Response to reviewer 1eWB (Part 1/3)

We thank the reviewer for constructive comments. We have expanded the sensitivity analyses, introduced learned activity-dependent modulation, and added dynamical and causal analyses. We also correct an important error in our original description of Fig. 3. All compared architectures achieve 100% accuracy on the evaluated two-item task (Appendix C.8, Fig. S15A). Cross-condition decoding in Fig. 3 measures the consistency of item representations across conditions sharing the same ordinal rank. This correction is central to interpreting the comparisons below.

On Weaknesses

W1:

" The considered model received prescribed time resolved rule-cue-conditioned time constant modulation. It is not a comparison between architectures, but the effect of such prescribed time constant modulation."

We thank the reviewer for this careful distinction. We agree that the original comparison primarily examines how prescribed timescale modulation influences the solutions learned by recurrent networks. Our main finding is that cue-triggered timescale modulation can act as a dynamical inductive bias, facilitating the learning of experimentally observed rank-specific representations and their rule-guided reorganization. This provides a concrete computational account of flexible working-memory manipulation.

We further conducted two experiments to show that the prescribed timescale modulation strategy can emerge through learning and may have a plausible biological implementation.

1)We developed RNN-SMTS, in which a scalar modulation signal generated from recurrent population activity dynamically regulates the timescale shared across all recurrent neurons. Without a prescribed modulation schedule or timing supervision, the network learns when and how strongly to modulate its timescale through task training. The network spontaneously learns to shorten its time constant after the rule cue and recovers rank-specific organization, rule-dependent reassignment, and temporary item representations. It achieves 100% task accuracy and 0.97 ± 0.03 cross-condition decoding, closely reproducing the experimentally observed rank-specific representations and their rule-dependent reorganization. (Sec. 3.5, Fig. 6; Appendix C.9, Fig. S16).

2)Inspired by recent findings in Science [1] showing that dendritic signals encode task rules and modulate somatic computations, we further developed a dendrite–soma RNN in which rule-dependent dendritic activity modulates recurrent and feedforward inputs to the soma. Through learned rule-to-dendrite connections, the model develops similar cue-triggered timescale modulation that alters effective neuronal timescales and reproduces the key representational dynamics in both two- and three-item tasks (Appendix C.6, Figs. S12–S13). These results provide a biologically plausible implementation of timescale modulation through dendritic mechanisms.

W2:

" Insufficient experiments on variants of time constant modulation."

We thank the reviewer for this suggestion. We systematically varied the modulation magnitude, population size, onset, and duration to examine how these choices affect learning. Each condition was trained from initialization, with results summarized across five seeds.

  • Magnitude and population size. We varied the post-cue AT time constant from 0.05 to 0.95 and the AT proportion from 0% to 100%. All conditions achieved 100% accuracy on the evaluated two-item task. Rank-template agreement and cross-condition decoding remained high across a range of settings and declined with insufficient acceleration or too few AT neurons (Fig. 3B–C; Appendix A.4; Fig. S15B–C).
  • Modulation onset. We moved the onset across stimulus encoding, delay-1, cue presentation, delay-2, and retrieval. Behavioral accuracy remained 100%, while the experimentally observed rank-specific organization was best reproduced when modulation began around the cue-to-delay-2 transition, with weaker agreement at earlier or later onsets (Appendix C.8, Fig. S15D–F).
  • Modulation duration. We varied the acceleration window from 0 to 25 steps and included persistent modulation through retrieval. Accuracy remained 100%. A minimum modulation duration was required, beyond which rank-specific organization remained robust across a broad range of durations. (Fig. S15G–I).

Together, these results show that the observed rank organization is robust across a broad range of modulation parameters, reducing concern that it depends on narrowly tuned settings. They also highlight the importance of cue-linked acceleration, a sufficiently large modulated population, and sufficiently fast post-cue dynamics for reliably learning the experimentally observed representations, despite equivalent task performance across conditions.

AuthorsOpenReview ↗

Response to reviewer 1eWB (Part 2/3)

W3:

"The worse performance of LSTM and FS RNN is not convincing. It could simply be that they are not trained enough, while the adaptive timescale RNN already receives what they are supposed to be trained to do, thus being faster."

We thank the reviewer for identifying this error. Our original text incorrectly described reduced cross-condition decoding as reduced task performance. We have corrected the wording and added explicit behavioral results showing that all comparison models, including LSTM and RNN-FS, achieve 100% complete-sequence accuracy on the evaluated two-item task (Appendix C.8, Fig. S15A). Within each training setting, the models use the same trial generator, targets, training objective, and optimizer (Appendix A.3).

The measures in Fig. 3 assess rank organization and the generalization of item representations across rank-equivalent conditions (Appendix A.4.2). The representational differences therefore persist at matched, ceiling-level task performance, supporting an effect of timescale modulation on the solution acquired through learning. RNN-SMTS further demonstrates that the modulation strategy itself can be learned while recovering the same key representational organization (Fig. 6; Fig. S16).

W4:

"The paper offers little mechanistic insight into the link between timescale modulation and the advantage in the working memory task."

We thank the reviewer for encouraging a deeper mechanistic analysis. Our model provides a computational account of working-memory manipulation and generates experimentally testable predictions beyond existing observations. We further conducted additional analyses to investigate why timescale modulation facilitates learning the experimentally observed representational organization.

First, our model links memory maintenance, rule-guided reorganization, and sequential retrieval within a concrete recurrent network. Items are maintained in reusable rank-specific subspaces, reassigned through temporary representations following the rule cue, and sequentially accessed through a shared readout population (Secs. 3.1 and 3.3, Figs. 2 and 4). Perturbing recurrent communication through temporary subspaces selectively impairs memory reversal, supporting their functional role in manipulation (Appendix C.12, Eq. 24, Fig. S20).

Second, our additional analyses suggest how timescale modulation shapes the learned network dynamics. Compared with Control RNNs, timescale-adaptive RNNs develop more stable, attractor-like memory dynamics rather than relying primarily on slow transient activity (Appendix C.10, Fig. S17). A possible explanation is that cue-triggered timescale shortening accelerates state evolution, potentially disrupting previously maintained representations and increasing their decay rates. To prevent memory loss, this may favor the learning of attractor-like dynamics that support both rapid representational reorganization and subsequent stable maintenance.

To examine this possibility, we compared networks with identical initializations and matched training sequences. Across learning epochs, the Jacobian eigenvalue distributions of timescale-adaptive RNNs progressively shifted leftward, away from 1, indicating more stable dynamics, while pronounced post-cue expansion emerged. Within trained networks, this resulted in a clear phase-dependent transition from post-cue expansion to late-delay contraction. In contrast, Control RNNs maintained eigenvalues clustered near 1 throughout training, with little phase-dependent modulation (Appendices C.10–C.11, Figs.S17–S19).

Third, our model generates new experimentally testable predictions. First, how can a fixed downstream readout sequentially access information stored in distinct rank-specific subspaces? Rather than independently gating different rank populations to the output, our model predicts that successive items are dynamically routed through a shared G1 population (Fig. 4F–H). This provides a potential solution to the fundamental problem of flexibly accessing distributed memory representations through fixed neural connections, and predicts specific population dynamics during sequential retrieval. Second, extending beyond the two-item experiments of Tian et al., our model predicts reusable rank-specific subspaces for three-item sequences (Fig. 5A–B), with successive rule cues driving sequential reorganization to support compositional memory manipulation (Fig. 5C–E). Both predictions can be tested through future neural population recordings and targeted perturbations.

Together, these findings suggest that timescale modulation acts as a dynamical inductive bias, favoring the learning of stable yet reconfigurable memory representations, while revealing how recurrent networks can coordinate rank-specific encoding, rule-guided manipulation, and shared readout to generate experimentally testable predictions.

AuthorsOpenReview ↗

Response to reviewer 1eWB (Part 3/3)

W5:

"Insufficient analyses on the AT neurons. There is currently only the percentage of AT neurons in G1 G2 and O1 O2. Most of the analyses could be factorized by AT and FT populations. What about the representation in AT population? Does information in AT neurons persist?"

We thank the reviewer for this insightful suggestion. We agree that examining AT and FT neurons separately could provide further mechanistic insights. Our current analyses primarily focus on population-level computations, examining how information is maintained in rank-specific subspaces, reorganized through temporary representations, and sequentially retrieved through a shared readout population (Fig. 4). We also provide initial evidence linking timescale modulation to functional organization, as G1 is enriched in AT neurons and AT neurons contribute strongly to readout during both output periods (Fig. 4I).

One challenge in interpreting AT and FT neurons is that similar rank-specific organization can emerge without a fixed fast/slow division, through different timescale modulation mechanisms. Networks with all neurons adaptive develop the characteristic rank organization without a fixed fast/slow division (Fig. 3B). Likewise, RNN-SMTS reproduces this organization through a shared, learned timescale (Fig. 6; Fig. S16), whereas the dendrite–soma RNN learns heterogeneous, rule-dependent timescale modulation, giving rise to four neuronal groups: neurons with consistently slow timescales, forward-specific shortening, backward-specific shortening, or shortening under both rules, while preserving the key representational geometry and transformations (Figs. S12–S13). These findings suggest that similar population-level computations do not necessarily imply identical underlying circuit mechanisms.

The reviewer's suggestion to analyze AT and FT populations offers a valuable direction for distinguishing different circuit implementations of the same computation. More broadly, this relates to the distinction between understanding memory through population-level attractor dynamics and identifying the specific neuronal and circuit interactions that implement them. We appreciate the reviewer for raising this important question and consider it a promising direction for future investigation.

On Questions

Q1: Prescribed timescale modulation provides an effective inductive bias for learning stable, reusable rank representations and their rule-guided reorganization (Figs. S17–S19). It also remains essential for executing the learned solution, as disabling modulation after training reduces task accuracy from 100% to 1.30% (Fig. S14). Please see our response to W2 for the supporting analyses.

Regarding receptivity, shorter timescales enable neuronal states to respond more rapidly to recurrent inputs while accelerating the decay of previously stored information, potentially facilitating information transfer between rank-specific representations. This offers a valuable perspective for understanding the underlying circuit-level mechanisms. We thank the reviewer for this insightful comment.

Q2: Our additional experiments show that prescribed timescale modulation can emerge through learning. Both RNN-SMTS and the dendrite–soma RNN spontaneously develop cue-triggered timescale shortening and reproduce the key experimentally observed features of working-memory manipulation. This suggests that cue-triggered timescale modulation may be a natural computational strategy rather than an arbitrary design choice (see W1).

Interestingly, similar representational organization emerges under different modulation schemes. RNN-SMTS learns a shared timescale modulation across all neurons, whereas the dendrite–soma RNN develops four distinct groups with rule-dependent modulation profiles. This suggests that the same population-level computation can arise from different implementations of timescale modulation (see W5).

References

[1] Yuan Zhang et al., Science, 2026

Reviewer_WcrbOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper is best understood as a computational neuroscience study that uses a familiar form of recurrent control to investigate sequence working memory, modeling the macaque task of Tian et al. (2024). A continuous-time RNN receives a rule cue specifying an order permutation of input sequences, together with an externally prescribed, rule-independent schedule that reduces the integration time constant of 320 of its 400 units from 1 to 0.25. The schedule begins at rule-cue onset, so the faster regime marks the start of manipulation, which the network must complete during the subsequent delay, and it persists through recall. The paper reports rank-related representations, reassignment of items between rank subspaces, transient item representations, shared readout, successful state editing, and preliminary composition of individually trained operations. It argues that these signatures match the monkey recordings, and that standard recurrent architectures reach them less reliably, making timescale modulation critical.

The paper largely sits at the intersection of Computational Neuroscience and Machine Learning, and by my understanding, has different implications for both fields. As such, relevance has to be argued field by field. The mechanism is a prescribed, rule-agnostic update gate: a persistent phase signal that raises the integration rate after the cue. Conditioning recurrent computation on such signals is well established in machine learning, so "a control signal enables reorganization" is not a contribution in itself. The paper matters only insofar as this particular switch delivers something specific to specific audiences, be it neuroscience or machine learning.

Critical strengths

For neuroscience audiences:

  • The paper targets a question a recent dataset left open. Tian et al. described rank reassignment and temporary subspaces, but not the circuit dynamics behind them. A trained model that reproduces these signatures, analyzed with the same pipeline as the experiment, is a reasonable first step toward explaining them.

  • It turns a verbal idea into a concrete hypothesis. Tian et al. proposed that cue activity triggers sorting. The paper makes one version of that idea, a cue-locked change in integration speed.

For machine learning:

  • It is a clean, minimal testbed for compositional manipulation of memory. The two-cue test probes an untrained mapping from input order to output order, and the proposed model partly succeeds where the baselines fail.

  • It suggests a cheap inductive bias. A hand-set phase gate visibly changes which representation a network learns. If shown to match or beat learned gating, the framework may be a useful architectural insight.

Critical weaknesses
  1. Novelty of the gating/update formulation.

The update has the form xt+1=(1−αt) xt+αt F(xt,ut)x_{t+1} = (1-\alpha_t)\, x_t + \alpha_t\, F(x_t, u_t).

Using time- or input-dependent update rates as a switch is established in both neuroscience and ML:

  • Machine learning: gating derived from time warping (Tallec & Ollivier, 2018); clock-driven update rates (Clockwork RNN, Koutník et al., 2014; Phased LSTM, Neil et al., 2016); learned or input-dependent timescales (Heinrich et al., 2020; Khajehabdollahi et al., 2024); selective state-space models with an input-dependent step size (Gu & Dao, 2023); and structured neuromodulation of recurrent dynamics (Costacurta et al., 2024).

  • Neuroscience modeling: gain-controlled permutation circuits (Zuo et al., 2025; Bien et al., 2026). The paper's timescale modulation is mathematically a gain on the dynamics, so the main difference is learned weights under a prescribed schedule. Tian et al. themselves concluded that cue activity acts as a trigger for sorting; the schedule imposes that trigger by hand rather than explaining it. The paper overclaims here. It says its comparisons "confirmed the importance" of the mechanism, and claims "mechanistic insight" into mental programming. Experimental results do not prove mechanistic insights as claimed.

  1. The neuroscientific insights are further unclear. The model is presented as interacting fast and slow populations, yet it works just as well when every unit speeds up. That reduces the framework to a global post-cue speed-up, and the paper does not say what circuit could implement one or how it differs from gain control.

  2. The switching effect is not specific to timescale modulation. The experiments show that a cue-locked schedule suffices (proving that cues are sufficient, but not that they are necessary). At least two simpler explanations fit equally well:

  • Any explicit phase signal. The proposed model is persistently told that it is in the post-cue phase, and the baselines are not. If giving the baselines that signal as an ordinary input yields the same solution, the result reduces to "making task phase explicit helps."

  • Gain rather than timescale. The manipulation also amplifies the cue due to more frequent feedback. Threshold-based experiments can help isolate the effect of gain versus time.

  1. The paper offers no dynamical account of why a speed-up favors reuse. The "buffer" role of the transient subspaces is asserted from decodability rather than tested experimentally through perturbations or ablation studies. The paper also does not separate whether the switch matters for learning the solution or for executing it in a trained network.

  2. The mechanism is a hand-timed special case of known designs, and it is not compared with them. For an ML reader, the schedule is a content-blind change in update rate that must be told when the cue arrives. The obvious question is whether a model that learns input-dependent timescales recovers the same behavior and does as well or better. Without that comparison, the paper offers no architectural lesson beyond "a phase gate can help." Needing to know when manipulation begins also limits its use as a general architecture.

  3. Scale and generality are limited. The task uses a handful of items, sequences of at most three, and four fixed rules. There are no long sequences, larger vocabularies or other manipulations. That is reasonable for a neuroscience model, but it limits what an ML reader can conclude about compositional memory. I would refer the authors to linear memories like Gated DeltaNets or Kimi Delta Attention modules for examples of more general tasks that their framework could benefit from to prove scalability and generality.

Questions for authors

Do the baseline recurrent networks (leaky RNN, fixed fast/slow RNN, GRU, LSTM) reach the same solution if they are simply given the post-cue phase signal as an ordinary input (my intuition is yes)? If so, what does the timescale form add?

Does a model that can learn when to change its timescale discover a cue-locked speed-up on its own, and how does it compare?

For more evidence on the "mechanistic" claim, what circuit could implement a global post-cue speed-up, and how would it differ from gain modulation? Additionally, why does the speed-up lead to swapping items within existing subspaces, while the control network moves them into new ones? A dynamical account of this contrast could be a nice mechanistic contribution.

In general, I like the direction of the work, but I think some claims are unsupported, and the experimental evidence does not show generality/scalability or account for other explanations of the observed results. I am curious to see responses to my questions and weaknesses, and will raise my score if these points are addressed sufficiently with both intuitions and experimental results.

3 comments
AuthorsOpenReview ↗

Response to reviewer Wcrb (Part 1/3)

We thank the reviewer for the detailed comments. To address the concerns, we added explicit phase-input controls, learned timescale modulation, a dendrite–soma implementation, training and execution interventions, local dynamical analyses, and a causal test of temporary-subspace involvement.

On Weaknesses

W1:

" time- or input-dependent update rates are established ... Machine learning ... "

We thank the reviewer for highlighting these connections. We agree that gating and adaptive update rates are well-established principles in machine learning. Our main finding is that changing neuronal timescales at the transition from memory maintenance to manipulation helps a task-trained network learn reusable representations of sequence positions and move remembered items between them, closely matching the organization observed experimentally. All comparison models (including GRUs and LSTMs) can perform the task successfully (Appendix C.8, Fig. S15A), yet they develop different internal representations (Fig. 3).

Neuroscience modeling: ... its comparisons "confirmed the importance" of the mechanism, and claims "mechanistic insight" into mental programming....

We thank the reviewer for highlighting the connection to gain-controlled permutation circuits. We agree that timescale modulation can be expressed as a gain on the dynamical vector field, and we have moderated our discussions accordingly (see the revised version). Our model shows how timescale modulation promotes the emergence of rank-specific representations and memory manipulation through task learning.

We now further show that modulation itself can be learned and examine the resulting dynamics, as detailed below.

First, cue-linked modulation can emerge through learning and may have a plausible biological implementation.

1)We developed RNN-SMTS, in which a scalar modulation signal generated from recurrent population activity dynamically regulates the timescale shared across all recurrent neurons. Without a prescribed modulation schedule or timing supervision, the network learns when and how strongly to modulate its timescale through task training. The network spontaneously learns to shorten its time constant after the rule cue and recovers rank-specific organization, rule-dependent reassignment, and temporary item representations. It achieves 100% task accuracy and 0.97 ± 0.03 cross-condition decoding, closely reproducing the experimentally observed rank-specific representations and their rule-dependent reorganization (Sec. 3.5, Fig. 6; Appendix C.9, Fig. S16).

2)Inspired by recent findings in Science [1] showing that dendritic signals encode task rules and modulate somatic computations, we further developed a dendrite–soma RNN in which rule-dependent dendritic activity modulates recurrent and feedforward inputs to the soma. Through learned rule-to-dendrite connections, the model develops similar cue-triggered timescale modulation that alters effective neuronal timescales and reproduces the key representational dynamics in both two- and three-item tasks (Appendix C.6, Figs. S12–S13). These results provide a biologically plausible implementation of timescale modulation through dendritic mechanisms.

Second, the added analyses clarify how timescale modulation shapes the learning of attractor-like dynamics.

Our analyses suggest that timescale-adaptive RNNs tend to develop attractor-like dynamics supporting stable memory representations, whereas control RNNs rely more on slow transient dynamics (Fig. S17). A possible explanation is that cue-triggered timescale shortening accelerates state evolution, potentially disrupting previously maintained representations and increasing their decay rates. To prevent memory loss, this may favor the learning of attractor-like dynamics that support both rapid representational reorganization and subsequent stable maintenance.

To examine this possibility, we compared networks with identical initializations and matched training sequences. Across learning epochs, the Jacobian eigenvalue distributions of timescale-adaptive RNNs progressively shifted leftward, away from 1, indicating more stable dynamics, while pronounced post-cue expansion emerged. Within trained networks, this resulted in a clear phase-dependent transition from post-cue expansion to late-delay contraction. In contrast, Control RNNs maintained eigenvalues clustered near 1 throughout training, with little phase-dependent modulation (Appendices C.10–C.11, Figs.S17–S19).

Together, these findings suggest that timescale modulation acts not only as a control mechanism during task execution, but also as an inductive bias that shapes the dynamical solutions learned by the network, favoring attractor-like dynamics over slow transient solutions.

AuthorsOpenReview ↗

Response to reviewer Wcrb (Part 2/3)

W2: We thank the reviewer for raising this important point. Our results suggest that cue-locked timescale modulation, rather than a fixed division between fast and slow neurons, is the key computational ingredient. Our primary focus is on population-level computations, including rank-specific encoding, rule-guided reorganization, and sequential readout (Fig. 4). Although G1 is enriched in adaptive-timescale neurons (Fig. 4I), the all-adaptive condition also develops the characteristic rank-specific organization (Fig. 3B), indicating that a fixed fast/slow partition is not necessary.

Our additional experiments reveal that similar population-level computations can emerge through different timescale modulation strategies. Our RNN-SMTS learns a shared, cue-triggered timescale modulation through recurrent feedback (Fig. 6; Fig. S16). In contrast, our biologically motivated dendrite–soma RNN learns heterogeneous, rule-dependent gain modulation, giving rise to four neuronal groups: consistently slow, forward-specific, backward-specific, and fast under both rules (Figs. S12–S13). In this model, dendritic gain modulates somatic inputs and thereby influences effective network timescales, providing a possible biological link between gain control and timescale modulation.

These findings suggest that similar rank-specific representations and memory operations can arise from different underlying modulation circuit mechanisms. Distinguishing how these mechanisms distribute memory information and coordinate transitions between rank subspaces is an interesting circuit-level question. We appreciate the reviewer's suggestion, which motivates further investigation of population-specific dynamics and circuit interactions.

W3: We thank the reviewer for raising these alternative explanations. We addressed them through two complementary controls.

First, stronger and sustained post-cue input was insufficient to recover the observed rank organization. All models already receive rule-cue and fixation inputs (Fig. S1). We additionally supplied Control RNN, GRU, and LSTM with a phase-input channel active throughout delay-2, varied its amplitude from 1 to 5, and retrained each condition from initialization. All conditions achieved 100% task accuracy, while rank-template agreement remained at 0.13–0.39 and cross-condition decoding at 0.12–0.28 (Appendix C.8, Fig. S15J–L). This experiment directly tests both explicit phase information and increased post-cue drive. Across the tested amplitudes, neither was sufficient to reproduce the characteristic rank-subspace switching.

Second, the timing of timescale modulation systematically shaped the learned representations. We varied modulation onset across encoding, delay-1, cue presentation, delay-2, and retrieval. Task accuracy remained 100% throughout, while rank organization was strongest around the cue-to-delay-2 transition and substantially weaker with earlier or later modulation (Fig. S15D–F). With trained weights held fixed, disabling modulation reduced AT-RNN accuracy from 100% to 1.30%, establishing its necessity for executing the learned solution (Appendix C.7, Fig. S14).

Together, these results support the need for appropriately timed timescale modulation to reliably learn the observed rank-subspace organization.

W4: We thank the reviewer for this suggestion. As discussed above, the dynamical and training analyses support a role for modulation in learning reusable memory representations (W1, Figs. S17-S19).

We directly tested the functional role of temporary subspaces by perturbing their contribution to recurrent information transmission. From cue onset until retrieval, we projected the joint Temp-1/Temp-2 activity component out of the recurrent drive while keeping learned weights, external inputs, the timescale schedule, and the readout unchanged. Forward trials preserve the existing item–rank assignments and show little recruitment of these temporary representations. Consistent with this, the intervention left forward performance unaffected. Backward trials recruit the temporary subspaces during item–rank reassignment, and the same intervention disrupted 68.33 ± 2.36% of backward trials across five seeds (Appendix C.12, Eq. 24, Fig. S20). This selective impairment provides causal evidence that recurrent communication through the temporary subspaces contributes to sequence reversal.

Finally, learning and execution are tightly coupled to yield distinct, non-interchangeable solutions. With weights fixed at test time, disabling modulation in our timescale-adaptive RNN reduced accuracy from 100% to 1.30%, whereas imposing it post hoc on the Control RNN reduced accuracy from 100% to 5.30% (Appendix C.7, Fig. S14).

AuthorsOpenReview ↗

Response to reviewer Wcrb (Part 3/3)

W5: We agree that a general model should be able to learn its modulation strategy. To address this, we developed RNN-SMTS, in which a learned feedback readout of recurrent activity dynamically controls the neuronal time constant. Connectivity and feedback weights are jointly optimized through task training, allowing the network to learn when and how strongly to modulate its timescale (Sec. 3.5; Appendix C.9.1).

The network spontaneously learns cue-triggered timescale shortening without prescribed modulation schedules or explicit timing supervision (Fig. 6B). Across five seeds, it achieves 100% task accuracy and 0.97 ± 0.03 cross-condition decoding accuracy. Importantly, the learned modulation reproduces key representational dynamics, including rank-specific organization, rule-dependent reassignment, output-aligned state convergence, and temporary item representations (Fig. 6C; Appendix C.9.2, Fig. S16).

These results demonstrate that the modulation strategy can emerge through task learning. Together with the dendrite–soma implementation (Appendix C.6, Figs. S12–S13), they provide a learnable computational mechanism and a candidate biological realization of timescale control.

W6: We thank the reviewer for this valuable suggestion. We agree that larger vocabularies, longer sequences, and more diverse manipulation tasks are important directions for evaluating scalability and generality, as already acknowledged in the Discussion. Our additional results with RNN-SMTS and dendrite–soma RNN further suggest that the proposed modulation mechanism is not restricted to a prescribed schedule. These mechanisms may potentially be integrated into more scalable architectures, such as Gated DeltaNets and Kimi Delta Attention, to explore their applicability to more general sequence-memory tasks.

On Questions

Q1. No. Providing the post-cue phase signal as an ordinary input does not reproduce the characteristic rank-subspace organization learned by our RNNs, despite all baseline models achieving 100% task accuracy. Please see our detailed response to W3 and the additional experiments in the revised manuscript (Appendix C.8, Fig. S15J–L).

Q2. Yes. Our newly developed RNN-SMTS learns to modulate its timescales without a prescribed schedule, spontaneously exhibiting cue-locked timescale shortening that closely resembles our original design. Importantly, it reproduces the experimentally observed rank-specific representational organization and dynamics. See our response to W2 and (Sec. 3.5, Fig. 6; Appendix C.9, Fig. S16) for details.

Q3. We thank the reviewer for these insightful questions. We address them from two complementary perspectives.

First, a biologically plausible implementation. Our dendrite–soma RNN demonstrates how learned dendritic gain modulation can regulate effective neuronal timescales and reproduce key representational dynamics (Figs. S12–S13). Thus, we view gain and timescale modulation as closely related mechanisms, with dendritic gain providing one possible biological implementation.

Second, a learning inductive bias for subspace reuse. Our additional analyses suggest that cue-triggered timescale shortening acts not merely as an execution-time control signal, but as an inductive bias that shapes the dynamical solutions learned by the network. During training, timescale-adaptive RNNs progressively develop attractor-like dynamics with phase-dependent expansion and contraction, whereas control RNNs exhibit dynamics more consistent with slow transient solutions (Figs. S17–S19). We propose that this learning bias favors stable, reusable rank-specific representations while allowing rapid transitions between them, rather than relying on distinct rule-dependent trajectories. Although the precise causal link remains to be established, these findings provide a possible dynamical explanation for the different representational strategies.

References

[1] Yuan Zhang et al., Science, 2026

Reviewer_ytSAOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

The authors train recurrent neural networks to solve a cognitive control task, where subjects first have to remember a sequence, and then reproduce it in either forward or reverse order. This is inspired by Tian et al. (2024), who found sequence representations in dlPFC of monkeys solving the task. Critically, the sequence representation was ‘reversed’ through the engagement of two temporary ‘swap’ subspaces on the reversal trials. The authors find that a similar solution is learned by RNNs trained with adaptive timescales, and this is not the case for other baseline architectures.

Critical strengths

This paper addresses an interesting question of how neural networks can implement ‘cognitive control’, in this case through sequence manipulations. The authors do a comprehensive job comparing the trained RNNs to the biological data from Tian et al. The paper is generally well-written and the figures are mostly clear.

Critical weaknesses

My main concerns are centered around the model comparisons. Specifically, the authors are trying to argue that the AT-RNN is (i) superior to the other models, and (ii) more like the biological data.

However, the AT-RNN model feels somewhat ad-hoc. What is the reason that the timescales are changed specifically during the cue presentation? And what is the motivation for changing the time constant from 1 to 4 instead of some other number? These feel like very specific choices for this very specific task, and they are not particularly well motivated. If the authors want to argue that the AT-RNN is a generally useful model, it would be important to show that there is some general way to make such architectural choices. A natural choice might be to learn the timescales jointly with the connectivity, or to make the time constants an output of the RNN itself (not unlike how GRUs work).

The fact that the AT-RNN feels over-engineered is a potential concern in the context of the model comparisons in the paper. For example, the networks in Figure 3B-C fail to learn the task when they do not have enough AT units. However, this is somewhat surprising given that it is a fairly simple task, and it makes me worry that the authors simply did not try very hard to optimise their baselines compared to their AT-RNN model. This is particularly worrying since the representational results in 3B-C exactly mirror the performance results – it is not very surprising that RNNs do not learn a useful representation when they fail to solve the task! It is unclear whether the GRU and LSTM networks learn the task, and how that relates to the representational results for those models.

Finally, it feels like the authors have not taken full advantage of their ability to interrogate the trained RNNs. It is not obvious that we learn much from training an RNN and redoing analyses that have already been done in the biological data, since that does not tell us anything new about the brain. However, understanding why these representations emerge in the artificial system could provide new hypotheses for how the same computations are implemented in biological brains. The paper would be greatly strengthened if the authors could elucidate the mechanisms driving the observed representations.

Questions for authors
  1. How sensitive is the AT-RNN to different choices of the timescale interpolation time, duration, magnitude, ...?

  2. Is there a way to learn these choices instead of having to manually craft an appropriate AT-RNN?

  3. Why do baseline models fail to learn this fairly simple task?

  4. Do the baselines exhibit less 'biology-like' representations because they fail to learn the task?

  5. How does the AT-RNN solve the task -- i.e. what is the circuit mechanism?

3 comments
AuthorsOpenReview ↗

Response to reviewer ytSA (Part 1/3)

We thank the reviewer for the constructive comments. To address reviewer's concerns, we have added an activity-dependent timescale model, a dendrite–soma implementation, sensitivity analyses, and dynamical and causal tests. We also correct an error in our original description of Fig. 3B–C. All architectures and parameter conditions achieve 100% accuracy on the evaluated two-item task (Appendix C.8, Fig. S15A–C). The measures in Fig. 3 quantify representational organization. Below, we address each concern in turn and provide concise answers to the questions.

On Weaknesses

W1:

“The authors are trying to argue that the AT-RNN is (i) superior to the other models, and (ii) more like the biological data.”

Our comparisons demonstrate specific advantages in representational organization and generalization, rather than overall task performance. All five architectures (our timescale-adaptive RNN, Control RNN, RNN-FS, GRU, and LSTM) achieve near-perfect accuracy on the two-item task (Fig. S15A). However, our model more consistently reproduces the experimentally observed rank-specific organization and rule-dependent reassignment (Figs. 2–3), while showing better generalization in the tested memory-editing (Fig. 4E), compositional manipulation (Fig. 5D), and extended-delay conditions (Fig. S17B). These results support the specific computational advantages of timescale modulation in learning biologically relevant representations and applying them flexibly across task conditions, without implying that the model is universally superior to other recurrent architectures.

W2:

“The AT-RNN model feels somewhat ad-hoc. What is the reason that the timescales are changed specifically during the cue presentation?”

We thank the reviewer for this important comment. We conducted additional experiments examining four aspects of cue-triggered timescale modulation:

  • The specificity of cue-triggered timescale modulation: We varied modulation onset and retrained each network. Despite 100% task accuracy across conditions, the experimentally observed rank-specific organization was best reproduced when modulation began around rule-cue onset, with weaker agreement at earlier or later onsets (Appendix C.8, Fig. S15D–F).

  • Whether it can emerge through learning: We developed RNN-SMTS, in which a learned feedback signal generated from recurrent population activity dynamically modulates the effective timescale shared across all neurons. Without a prescribed modulation schedule or timing supervision, the network learns when and how strongly to modulate its timescale, spontaneously developing cue-triggered acceleration. It also reproduces the experimentally observed rank-specific organization and temporary representations(Sec. 3.5, Fig. 6; Appendix C.9, Fig. S16; details below).

  • Whether it can be biologically implemented: Inspired by [1], who showed that dendritic signals encode task rules and modulate somatic computations, we developed a dendrite–soma RNN in which rule-dependent dendritic inputs modulate recurrent and feedforward inputs to the soma. The network learns the strength and neuronal specificity of gain modulation through training, rather than relying on prescribed modulation. The model exhibits cue-triggered timescale shortening and reproduces the key rank-specific organization, rule-dependent reassignment, and temporary item representations, extending to three-item sequences (Appendix C.6, Eqs. 13–18, Figs. S12–S13).

  • How cue-locked timescale modulation influences the dynamics: Our analyses suggest that timescale-adaptive RNNs tend to develop attractor-like dynamics supporting stable memory representations, whereas control RNNs rely more on slow transient dynamics (Fig. S17). A possible explanation is that cue-triggered timescale shortening accelerates state evolution, potentially disrupting previously maintained representations and increasing their decay rates. To prevent memory loss, this may favor the learning of attractor-like dynamics that support both rapid representational reorganization and subsequent stable maintenance.

  • To examine this possibility, we compared networks with identical initializations and matched training sequences. Across learning epochs, the Jacobian eigenvalue distributions of timescale-adaptive RNNs progressively shifted leftward, away from 1, indicating more stable dynamics, while pronounced post-cue expansion emerged. Within trained networks, this resulted in a clear phase-dependent transition from post-cue expansion to late-delay contraction. In contrast, control RNNs maintained eigenvalues clustered near 1 throughout training, with little phase-dependent modulation (Appendices C.10–C.11, Figs.S17–S19).

Together, these findings suggest that cue-triggered timescale modulation is not simply an arbitrary design choice, but a learnable, biologically plausible strategy that may favor the emergence of stable yet reconfigurable memory representations.

AuthorsOpenReview ↗

Response to reviewer ytSA (Part 2/3)

W3:

“What is the motivation for changing the time constant from 1 to 4 instead of some other number?”

“These feel like very specific choices for this very specific task.”

We evaluated the following sensitivity analyses to clarify this concern.

  • Magnitude and population size. We varied the post-cue time constant from 0.05 to 0.95 and the AT fraction from 0% to 100%. All conditions reached 100% task accuracy(Fig. S15B–C). Rank-template agreement and cross-condition decoding were high across a range of settings and declined with insufficient acceleration or too few AT neurons (Fig. 3B–C).
  • Onset. We moved modulation onset from stimulus presentation through memory maintenance, cue presentation, and retrieval. All conditions reached 100% accuracy. Representational scores were highest around the cue-to-delay-2 transition and lower with substantially earlier or later modulation (Appendix C.8, Fig. S15D–F).
  • Duration. We varied the acceleration window from 0 to 25 steps and included persistent modulation through retrieval. Accuracy remained 100%. Increasing the window from 0 to 15 steps raised mean cross-condition decoding from 0.14 to 0.94, with persistent modulation reaching 0.98 (Appendix C.8, Fig. S15G–I).

Together, these results show that the experimentally observed rank-subspace organization does not depend on narrowly tuned parameters, remaining robust across a broad range of modulation strengths, adaptive population sizes, and durations. However, learning this organization requires a sufficiently large adaptive population, adequate modulation duration, and modulation aligned with the rule cue. Furthermore, our additional RNN-SMTS demonstrates that the modulation schedule can emerge through task learning rather than being prescribed (Sec. 3.5, Fig. 6; Appendix C.9, Fig. S16).

W4:

“A natural choice might be to learn the timescales jointly with the connectivity, or to make the time constants an output of the RNN itself.”

Yes. We thank the reviewer for this insightful suggestion. To address this, we developed RNN-SMTS (Fig. 6A), in which a scalar modulation signal is generated from recurrent population activity, m(t)=Wmr(t)+bmm(t)=\mathbf{W}_m\mathbf{r}(t)+b_m ,and used to dynamically regulate the effective timescale shared across all recurrent neurons, τeff(t)=σ(m(t))τ\tau_{\mathrm{eff}}(t)=\sigma(m(t))\tau. Task training jointly optimizes the feedback and recurrent network weights, without auxiliary supervision of the modulation signal (Sec. 3.5; Appendix C.9.1, Eqs. 19–22). Thus, the network learns when and how strongly to modulate its timescale without a prescribed modulation schedule.

The model spontaneously learned cue-triggered timescale shortening, closely resembling the prescribed modulation in our original timescale-adaptive RNN. It achieved 100% task accuracy and 0.97 ± 0.03 cross-condition decoding, comparable to our original model (0.98 ± 0.02), while reproducing rank-specific organization, temporary item representations, and population remapping (Fig. 6; Appendix C.9, Fig. S16). These results demonstrate that the timescale modulation strategy can emerge through task learning without a prescribed schedule.

W5:

“It makes me worry that the authors simply did not try very hard to optimise their baselines compared to their AT-RNN model.”

“It is unclear whether the GRU and LSTM networks learn the task, and how that relates to the representational results for those models.”

Thank you for raising this concern. All five architectures—our RNN, RNN-FS, Control RNN, GRU, and LSTM—achieve 100% accuracy on the evaluated two-item task (Appendix C.8, Fig. S15A). The same holds across the AT-fraction and time-constant sweeps (Fig. S15B–C). Our original text incorrectly described reduced cross-condition decoding as deteriorating task performance. We have corrected this interpretation and explicitly reported behavioral accuracy.

AuthorsOpenReview ↗

Response to reviewer ytSA (Part 3/3)

W6:

“It is not obvious that we learn much from training an RNN and redoing analyses that have already been done in the biological data, since that does not tell us anything new about the brain.”

Our work provides three insights beyond reproducing the experimental observations of Tian et al.:

1. A computational mechanism for working-memory manipulation. Our model links rank-specific encoding (Fig. 2A–I), rule-guided reorganization through temporary representations (Fig. 4A–E), and sequential retrieval (Fig. 4F–I) within a single recurrent network. Importantly, cue-triggered timescale modulation facilitates learning these representational dynamics, suggesting a possible mechanism for flexible memory manipulation.

2. A testable prediction for sequential memory readout. How can a fixed downstream readout sequentially access information stored in distinct rank-specific subspaces? Two possible mechanisms are to selectively gate different rank populations to the output or to dynamically route their contents through a shared readout population. Our model supports the latter: during retrieval, item information is transferred from G2 to G1, allowing successive items to be accessed through the same fixed output connections (Fig. 4F–H). This predicts a shared neuronal population that sequentially represents different memory items during retrieval, providing a testable circuit-level hypothesis beyond the existing experimental observations.

3. Testable predictions for three-item memory manipulation. Extending beyond the two-item experiments of Tian et al., our model predicts reusable rank-specific subspaces for three-item sequences (Fig. 5A–B), with successive rule cues driving sequential reorganization of item representations to support compositional manipulation (Fig. 5C–E).

Together, these findings go beyond reproducing neural data by proposing computational mechanisms and experimentally testable predictions for how recurrent circuits organize, manipulate, and retrieve working-memory representations.

On Questions

Q1: Sensitivity analyses of modulation onset, duration, magnitude, and adaptive-neuron fraction show that rank-specific organization is robust across a broad parameter range but deteriorates with insufficient or poorly timed modulation, despite 100% task accuracy (Further details are provided in our response to W3; Fig. 3B–C; Appendix C.8, Fig. S15A–I).

Q2: Yes. RNN-SMTS spontaneously learns cue-triggered timescale acceleration and reproduces the experimentally observed rank-specific organization, without a manually prescribed modulation schedule (see W4; Sec. 3.5, Fig. 6; Appendix C.9, Fig. S16).

Q3 - Q4: All baseline architectures reach 100% accuracy on the evaluated two-item task. The original wording conflated task accuracy with cross-condition decoding, and we have corrected it and added explicit behavioral results (see W5; Appendix C.8, Fig. S15A–C).

Q5: Our model provides a computational account of how working-memory representations are maintained, manipulated, and retrieved within a recurrent network. Items are encoded in reusable rank-specific subspaces and reassigned across these subspaces through temporary representations following the rule cue (Fig. 4A–C). Disrupting recurrent communication through temporary subspaces selectively impairs reversal (Fig.S20), supporting their functional role in memory manipulation. During retrieval, successive items are dynamically routed to a shared G1 population, enabling sequential access through a fixed downstream readout (Fig. 4F–H). Together, these results reveal how rank-specific encoding, rule-guided reorganization, and shared readout can be coordinated within a learned recurrent network.

However, how specific recurrent connections implement information transfer between rank-specific subspaces remains unresolved, particularly given the mixed selectivity of neurons in the trained RNN.

References

[1] Yuan Zhang et al., Science, 2026

Comments on the paper

5 comments

Dear Reviewer,

Thank you for your constructive feedback. We have carefully addressed your concerns through extensive additional experiments and analyses in our revised submission.

We noticed that your latest comment may have been intended for another submission. We would greatly appreciate it if you could review our responses and reconsider your rating before the discussion period closes.

Thank you for your time and consideration.

Sorry for the prior confusion.

I appreciate the additional experiments, which make the role of timescale modulation as an inductive bias more convincing.

However, I remain somewhat unconvinced by the framing of the computational contribution. Zuo et al. (2025) and Bien et al. (2026) have already proposed explicit gain-controlled circuits implementing rule-dependent information transfer between working-memory and temporary storage populations, as also discussed in the rebuttal to another reviewer.

The present study offers an interesting alternative in which the recurrent connectivity is learned rather than explicitly designed. To me, its most interesting finding is that a relatively simple form of gating biases learning toward the experimentally observed representational organization, despite other architectures solving the same task equally well.

While the additional dynamical analyses provide some insight into this effect, they do not fully explain why timescale modulation favors this particular solution, or how the learned recurrent connectivity implements rule-dependent information transfer. The alternative solutions learned by the baseline architectures also remain largely uncharacterized.

Overall, I find the empirical evidence substantially more convincing after the rebuttal, but the emphasis on timescale modulation as a mechanism for working-memory manipulation somewhat obscures the paper's more specific contribution. I think the work would benefit from a substantial reframing around this contribution. In its current form, I remain unconvinced that the paper fully realizes its scientific potential.

We thank the reviewer for acknowledging the substantial improvements in our revised manuscript. We would like to clarify several important points regarding our computational contributions.

  1. 1. Distinct computational contributions: learning and population-level mechanisms. Zuo et al. (2025) and Bien et al. (2026) adopt highly simplified, hand-crafted circuit settings, in which memory representations, functional modules, and information-transfer operations are explicitly designed to demonstrate how working-memory manipulation could be implemented. These models establish the feasibility of particular circuit implementations, but do not explain how the rich population representations and their rule-dependent transformations observed experimentally emerge through learning. This is precisely what our work addresses.

In our learning-based framework, recurrent connectivity and internal representations emerge through task training, reproducing the experimentally observed population geometry and its temporal evolution without prescribing these structures. Beyond systematically reproducing experimental observations, we uncover the population-level computational mechanisms underlying working-memory manipulation: reusable rank-specific subspaces support maintenance, temporary representations mediate rule-dependent reorganization, and a shared readout population enables sequential retrieval (Fig. 4). Causal perturbations establish the functional importance of these computations (Fig. S20), while learning-dynamics and Jacobian analyses explain how timescale modulation biases learning toward stable yet rapidly reconfigurable dynamics (Figs. S17–S19). Our framework further extends to three-item sequences and compositional manipulations (Fig. 5).

Our contribution is therefore fundamentally more than learning the connectivity of a hand-designed circuit: we reveal how experimentally observed population computations emerge through learning, identify their underlying computational mechanisms, and explain how timescale modulation shapes their emergence.

  1. Timescale modulation offers a more suitable dynamical perspective for understanding how neural networks balance stable memory maintenance with rapid, rule-dependent transformations. This can be realized through diverse circuit mechanisms, including gain modulation. Our dendrite–soma RNN illustrates this connection by showing how learned dendritic modulation alters effective timescales and reproduces experimental representations (Figs. S12–S13). We have discussed these connections in the manuscript and our response to Reviewer 2, who subsequently acknowledged our clarifications and raised their rating.

  2. The appropriate level of mechanistic explanation. We emphasize that population-level computation is an appropriate level of mechanistic understanding for our study. As demonstrated by our analyses, similar rank-specific computations can emerge through different circuit implementations, including networks with widely varying proportions of adaptive-timescale neurons. This suggests that population-level organization captures computational principles that are not tied to a particular circuit implementation. A finer-grained dissection of recurrent connectivity and alternative circuit solutions is beyond the scope of the present study and has already been acknowledged in our Limitations as an important direction for future work. The absence of a complete circuit-level explanation should not overshadow the substantial mechanistic insights and computational contributions established at the population and dynamical levels.

  3. Proposed reframing of the manuscript. We appreciate the reviewer's suggestion to sharpen the framing of our computational contribution and are willing to reorganize the manuscript accordingly. Specifically, we plan to introduce the learning-based timescale modulation model as our primary framework from the outset, emphasizing how recurrent connectivity, timescale modulation, and internal representations emerge through task training. We will then reposition the prescribed-modulation model as a simplified control framework for systematically isolating the effects of cue timing, modulation magnitude, duration, and other timescale-modulation variants. Since the learned-modulation model reproduces the key experimental representations, this reframing can preserve the main findings while placing the learning-based emergence of population computations and the dynamical inductive bias of timescale modulation at the center of the paper.

We believe this reframing directly addresses the reviewer's constructive suggestion and more clearly communicates the scientific contributions supported by our results.

We sincerely thank the reviewer for their constructive suggestions and consideration.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.