- 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
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.
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
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:
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:
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.
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.