acceptodds
Under review as a conference paper at ICLR 2027

Cortico-Cerebellar Coordination Flexibly Recombines Reusable Representations for Zero-Shot Compositional Generalization

Abstract

Compositionality, the ability to flexibly recombine learned components to solve novel tasks, is a hallmark of biological intelligence and a central challenge for artificial systems. However, the circuit mechanisms and dynamical principles underlying compositional generalization remain poorly understood. Inspired by distributed cortico-cerebellar architecture, we develop a computational framework that couples recurrent cortical dynamics with adaptive cerebellar feedback. We evaluate this framework across cognitive and motor tasks requiring zero-shot execution of novel rule combinations and motor primitive sequences. While cortex-only models achieve accurate performance on learned compositions but fail to recombine previously acquired components in unseen contexts, cortico-cerebellar models exhibit robust zero-shot compositional generalization across both domains. Mechanistic analyses reveal that cortex-only models retain task-relevant latent representations but fail to flexibly deploy them under novel cognitive contexts and exhibit predecessor-dependent state carry-over during motor transitions. In contrast, cerebellar feedback stabilizes reusable cortical representations, steers successor dynamics, and adaptively regulates state transitions during novel compositional execution. Together, these mechanisms enable cortico-cerebellar models to achieve robust zero-shot compositional generalization across temporal and structural task variations. Our findings reveal a circuit-level mechanism for flexible compositional computation and establish a bio-inspired principle for developing artificial systems capable of zero-shot generalization.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Cortico-cerebellar coupling, Cognitive flexibility, Compositional recombination, Zero-shot generalization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_hF3mOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Summary of contributions:

  • The paper investigates mechanisms that could support compositional generalization in neural circuits: the ability to reuse learned components in previously unseen task combinations.
  • Inspired by distributed cortico-cerebellar circuitry, the authors propose a cortical RNN augmented with cerebellar-inspired feedback mechanism that supplies task-dependent content and regulates cortical state retention.
  • Across cognitive rule-combination and motor-sequencing tasks, the proposed model generalizes better than an approximately parameter-matched cortex-only baseline. The authors also report more robust performance as motor sequence length increases.
  • The authors interpret these findings as a candidate circuit mechanism for compositional computation and suggest that the underlying computational organization could inform the design of artificial systems.
Critical strengths
  • Clarity: The paper is clearly written and easy to follow. The architecture, experimental comparisons, and mechanistic analyses are presented in an accessible way, with no significant concerns about readability.
  • Originality of contributions: To my knowledge, using cortico-cerebellar-inspired feedback to investigate zero-shot compositional generalization is an interesting and potentially novel computational experiment. The paper offers a useful model for exploring how interactions between neural circuits could support the flexible recombination of learned computations.
  • Significance of results: The reported cognitive and motor results provide evidence that the proposed feedback architecture improves compositional generalization relative to the cortex-only baseline. Approximate parameter matching makes this comparison informative, and the longer-sequence results provide additional evidence of the proposed model’s flexibility.
Critical weaknesses
  • Weak baselines: Although the principal comparison approximately matches the total parameter counts of Ctx and CtxFb, the width-sensitivity experiments (in Appendix D) vary CtxFb alone. I would like to see comparisons at several parameter budgets, increasing the widths of both architectures while maintaining approximate parameter matching. This would establish whether the observed advantage persists as model capacity increases. A diminishing advantage at larger widths would still be informative, because it would clarify the conditions under which feedback provides a useful inductive bias.
  • Scope of claims: The claim that these findings inform the design of future AI systems is broader than the evidence presented. The experiments use relatively small recurrent models and structured cognitive and motor tasks. Modern AI systems differ substantially in architecture, training regime, and the settings in which compositional failures arise, so the relevance of this mechanism to those failures remains unclear. The results support a candidate computational mechanism in the studied setting; broader design claims should be qualified or supported by experiments in more representative AI settings.
Questions for authors

Please refer to my weaknesses above. Adding the comparison between Ctx and CtxFb models while sweeping their parameter budgets would be informative to address my questions about this paper's contribution.

1 comment
  • Weakness 1

As suggested, we have extended the width-sensitivity analysis to seven approximately parameter-matched budgets, spanning more than a 600-fold increase in parameter count. Specifically, the cortical widths of Ctx-cb are set to 24, 48, 96, 192, 384, 768, and 1536, while those of Ctx-only are adjusted to match the total parameter counts. All configurations are evaluated across 10 random seeds.

Hidden units (Ctx-cb)Hidden units (Ctx-only)Parameters (Ctx-only/Ctx-cb)Ctx-only Acc(%)Ctx-cb Acc(%)Ctx-only NMSECtx-cb NMSE
24524167 / 410453.14±3.9453.14 \pm 3.9499.05±0.21\boldsymbol{99.05 \pm 0.21}0.24401±0.063920.24401 \pm 0.063920.00042±0.00005\boldsymbol{0.00042 \pm 0.00005}
48798460 / 842459.42±2.8359.42 \pm 2.8399.23±0.18\boldsymbol{99.23 \pm 0.18}0.22187±0.041210.22187 \pm 0.041210.00018±0.00003\boldsymbol{0.00018 \pm 0.00003}
9613020547 / 2052062.95±1.5662.95 \pm 1.5699.36±0.15\boldsymbol{99.36 \pm 0.15}0.21223±0.040870.21223 \pm 0.040870.00015±0.00002\boldsymbol{0.00015 \pm 0.00002}
19222858375 / 5853664.69±1.6364.69 \pm 1.6399.52±0.12\boldsymbol{99.52 \pm 0.12}0.19355±0.019630.19355 \pm 0.019630.00011±0.00003\boldsymbol{0.00011 \pm 0.00003}
384422189907 / 18986466.12±1.5166.12 \pm 1.5199.66±0.08\boldsymbol{99.66 \pm 0.08}0.18114±0.018520.18114 \pm 0.018520.00009±0.00002\boldsymbol{0.00009 \pm 0.00002}
768807673852 / 67370468.09±1.4868.09 \pm 1.4899.78±0.01\boldsymbol{99.78 \pm 0.01}0.16274±0.0174120.16274 \pm 0.0174120.00007±0.00002\boldsymbol{0.00007 \pm 0.00002}
153615752524732 / 252612069.85±1.2669.85 \pm 1.2699.91±0.01\boldsymbol{99.91 \pm 0.01}0.14164±0.016230.14164 \pm 0.016230.00003±0.00001\boldsymbol{0.00003 \pm 0.00001}

Ctx-cb consistently outperforms Ctx-only across all tested parameter budgets in both tasks, indicating that its advantage is not limited to small networks. Notably, Ctx-cb performance approaches saturation, whereas Ctx-only continues to improve with increasing model size. Even the smallest Ctx-cb model (4104 parameters) outperforms the largest Ctx-only model (2524732 parameters) in both tasks, despite the latter having approximately 615 times as many parameters. These results demonstrate the substantial parameter efficiency and computational advantage of Ctx-cb within the tested range. We will include these results and the corresponding discussion in the revised manuscript.

  • Weakness 2

We agree that the current experiments, conducted using relatively small recurrent networks and structured cognitive and motor tasks, may not establish the general applicability of the proposed mechanism to modern AI systems. We will revise the manuscript accordingly to clarify the scope and limitations of our conclusions.

  • Question

Please refer to our responses to the Weaknesses above.

Reviewer_4q37OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a recurrent network with structured feedback and adaptive gating, inspired by cortico-cerebellar interactions, and evaluates compositional generalization in cognitive and motor tasks. The results are promising, but the comparisons do not sufficiently establish what the proposed mechanism adds beyond established gating and feedback approaches.

Critical strengths

The model achieves substantial improvements on held-out combinations across two task domains, with a range of different experiments.

The feedback and gate ablations, targeted perturbations, and analyses of internal activity provide useful evidence about how the trained model operates.

The distinction between representing task information and successfully using it is nice.

Critical weaknesses

Stronger comparisons are needed. A vanilla recurrent network is a limited baseline for an architecture whose central components are gating and feedback. Comparisons with GRUs/LSTMs, with comparable parameters and importantly comparing with the optimum of GRU/LSTMs would help establish the contribution, and would immediately increase my score. Forcing/feedback like algorithms (generalized teacher forcing and error forcing) are also relevant because they directly influence recurrent activity using the output or error, which is kind of similar to using predicted state to gate the internal dynamics. Their relationship to the proposed mechanism can be discussed.

The cognitive experiment does not isolate the benefit of recurrent feedback. Explicitly combining the predicted decision with the response rule is fine, however, the paper should compare feeding this combined information back into the recurrent network with using the same information directly in the output. Removing feedback demonstrates that the trained model relies on that pathway, but does not distinguish this alternative. Equivalent auxiliary loss objectives across models should also be documented.

The motor task permits simpler solutions that deserve comparison. Both movement identities, the active sequence position, movement phase, and amplitude are supplied. Therefore the required current velocity seems like does not require remembering the preceding movement. Controls that reset the recurrent state at movement boundaries, or use the same movement-selection setup with a feedforward velocity predictor, would clarify the benefit beyond selection and removal of irrelevant history.

The neuroscience claims exceed the demonstrated specificity. The biological motivation is valuable, but the experiments establish benefits of a neuroscience-inspired architecture rather than identifying a specifically cortico-cerebellar mechanism. Claims such as “reveal a circuit-level mechanism” should be moderated.

Other minor comments:

“Predecessor-dependent state carry-over” is not understandable when reading the abstract.

Figure 3C largely repeats the train–test comparisons in Figures 3A–B. Removing it or moving it to the appendix could improve focus.

The AI-use statement already provides useful disclosure. For additional transparency, please consider elaborating on this in more detail.

Questions for authors

see weaknesses

2 comments
  • Weakness 1

As suggested, we have added two sets of comparisons: (1) parameter-matched GRU (Cho et al., 2014) and LSTM (Hochreiter & Schmidhuber, 1997) baselines, and (2) generalized teacher forcing (GTF; Hess et al., 2023) and error forcing (EF; Sağtekin et al., 2025) algorithms. For each architecture and task, we search 16 hyperparameter configurations and select the best based on validation accuracy or NMSE. All configurations are evaluated across 10 random seeds.

ModelHidden units / parametersCognitive Acc(%) ↑\uparrowMotor NMSE ↓\downarrow
Ctx-only (RNN)79 / 846059.42±2.8359.42 \pm 2.830.22187±0.041210.22187 \pm 0.04121
GRU42 / 836570.99±5.6370.99 \pm 5.630.20434±0.030130.20434 \pm 0.03013
LSTM36 / 861167.73±6.5467.73 \pm 6.540.37841±0.035160.37841 \pm 0.03516
RNN + GTF79 / 846046.90±1.1746.90 \pm 1.170.26373±0.059930.26373 \pm 0.05993
RNN + EF79 / 846051.10±3.5451.10 \pm 3.540.25118±0.060920.25118 \pm 0.06092
Ctx-cb48 / 842499.23±0.18\boldsymbol{99.23 \pm 0.18}0.00018±0.00003\boldsymbol{0.00018 \pm 0.00003}

Ctx-cb outperforms GRU, LSTM, GTF, and EF in both cognitive and motor tasks, suggesting that its advantage extends beyond generic recurrent gating or training-time forcing. We will include these results in the revised manuscript.

Ctx-cb selectively regulates feedback during inference, binding predicted decisions to response rules in cognition and using sequence position to select active movement primitives in motor control. Its learned gate also controls the retention of prior cortical activity (See Figs. 4A, 5, and 8 in the manuscript). Although GRUs and LSTMs can learn these relationships, their gates do not explicitly encode such task-specific operations, nor does training-time forcing provide them at test. These findings support the role of learned feedback during execution. We will include this discussion in our final manuscript.

Ref:

[1] Cho, K., et al. (2014). Learning phrase representations using RNN encoder–decoder for statistical machine translation. EMNLP, 1724–1734.

[2] Hochreiter, S., et al. (1997). Long short-term memory. Neural Computation, 9(8), 1735–1780.

[3] Hess, F., et al. (2023). Generalized teacher forcing for learning chaotic dynamics. ICML, 202, 13017–13049.

[4] Sağtekin, A. E., et al. (2025). Error forcing in recurrent neural networks. NeurIPS, 38.

  • Weakness 2

As suggested, we have conducted a direct-output control experiment that uses the same combined information but routes it directly to the output. All experiments were evaluated across 10 random seeds.

Held-out performanceFeedbackDirect output
Cognitive Acc(%) ↑\uparrow99.23±0.18\boldsymbol{99.23 \pm 0.18}60.88±5.1960.88 \pm 5.19
Motor NMSE ↓\downarrow0.00018±0.00003\boldsymbol{0.00018 \pm 0.00003}0.01138±0.001120.01138 \pm 0.00112

Our new results demonstrate that feedback substantially outperforms direct output in both cognitive and motor tasks, suggesting that the advantage of Ctx-cb extends beyond task-relevant information integration to recurrent regulation of cortical dynamics. We will include these results and the corresponding discussion in the revised manuscript.

Regarding auxiliary loss objectives, all cognitive models use the same response cross-entropy, computed from logits averaged over three response steps, together with an auxiliary loss (2×2 \times mean binary cross-entropy) supervising two selected sensory bits and their logical decision at the final pre-response recurrent state. The binding uses only the model’s predicted decision, while ground truth entered only the loss. No model received an additional binding or gate loss. Both motor routes likewise used identical objectives: velocity MSE, first-difference velocity MSE (weight 0.05), and selector entropy regularization (weight 0.02), with no auxiliary motor labels. We will revise the manuscript to explicitly clarify these equivalent training objectives.

  • Weakness 3

We appreciate the reviewer’s suggestion. As suggested, we have added two experiments: (1) resetting recurrent states at movement boundaries; (2) using the same movement-selection setup with a feedforward velocity predictor. All experiments are evaluated across 10 random seeds.

ModelHeld-out Motor NMSE ↓\downarrow
Ctx-only0.22187±0.041210.22187 \pm 0.04121
Boundary reset0.14242±0.031010.14242 \pm 0.03101
Feedforward predictor0.16734±0.039230.16734 \pm 0.03923
Ctx-cb0.00018±0.00003\boldsymbol{0.00018 \pm 0.00003}

Ctx-cb substantially outperforms both boundary-reset and feedforward controls in motor tasks, indicating that its advantage cannot be explained solely by movement selection or history removal. These results further support the contribution of structured cerebellar feedback to motor compositional generalization. We will include these results and the corresponding discussion in the revised manuscript.

  • Weakness 4

We agree that our results demonstrate the computational benefits of a neuroscience-inspired architecture rather than directly establishing a biologically realistic cortico-cerebellar mechanism. Nevertheless, our model is biologically motivated and incorporates key features of cortico-cerebellar circuitry, although some biological details are simplified or omitted.

We have replaced statements such as “reveal a circuit-level mechanism” with more precise expressions such as “suggest a candidate computational mechanism.” We will also clarify that the biological implementation of the proposed mechanism remains to be experimentally established.

  • Comment 1

We agree that “predecessor-dependent state carry-over” requires clarification. It refers to how information from preceding movements, retained in the recurrent cortical state, influences the execution of the next movement. We will therefore replace this term with “the influence of prior movement history on subsequent movement execution” in the abstract and manuscript for clarity.

  • Comment 2

We will move Figure 3C to the Supplementary Information to reduce redundancy and improve the focus of the main text.

  • Comment 3

We have expanded the AI-use statement to clarify its role in language editing and simulation-code debugging during the original submission, as well as data analysis and figure preparation.

  • Question

Please refer to our responses to the Weaknesses above.

Reviewer_PoT7OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2(+1)
Read review
Paper summary

The paper proposes an architecture mimicking cortex and cerebellum, and tests its performance on two compositional tasks. The paper then analyses the effect of different ablations.

Critical strengths

very interesting topic of compositionality and cortico-cerebellar interaction.

Critical weaknesses

Computational validity of the model

With quite convoluted components in the model, does the proposed cortico-cerebellar vs. cortical designs essentially reflect a simple difference between a leaky RNN and a GRU? It would be more informative to start with minimal addition of gating instead of going into a few cell types that do not necessarily add computational capacity.

It seems that the Cθ(τ)C_\theta^{(\tau)} and Aθ(τ)A_\theta^{(\tau)} in their cerebellar deisgn already contain privileged information about which task and context it is, making the comparison not about architecture but this privileged information.

Neuroscience validity of the model

The entire cortex is modeled as a leaky RNN, whereas in reality neither experimental evidence or architectural prior exclude the cortex from doing what their cerebellum exclusively does in their model, namely context processing and gating. None of their design contrasts between cortex and cerebellum really reflects any known physiological difference between cortex and cerebellum (such as the input expansion, time dependence, inhibition-excitation, neuronal density, firing rate and learning mechanism).

Validity of the tasks

Despite resembling real experimental stimuli in the illustrations in Figure 1, the tasks are not more than one-hot encoded inputs. Trial stage, context, motor effects, etc are also hard coded. The unnecessary complexity hides the simple essence where it could have been somewhat informative to the reader - what architectural component lead to learning a represention of two one-hot input vectors orthogonally vs. conjunctively, and what would lead to a better learning when stimuli are presented sequentially (which is well known, gating).

Very difficult notations for simple tasks.

Unclear what Figure 2B wants to show.

Confusing use of "successor states"

It really seemed like successor representation (Dayan, 1993) in a world model / RL setting, while actually the corresponding paragraphs are about how memory in the hidden state influence the subsequent input which are by design independent and are contextualy cued.

Questions for authors

see weaknesses

5 comments
  • Weakness 1

We thank the reviewer for these comments and regret that our original description did not make the computational role of the cerebellar module sufficiently clear. Our response has three main points:

  • Architecture: Unlike a GRU, Ctx-cb uses a separate cerebellar-inspired module to generate structured feedback that enters cortical recurrent dynamics additively.
  • Performance: A parameter-matched GRU baseline outperforms the Ctx-only RNN but remains substantially below Ctx-cb, suggesting that generic recurrent gating alone cannot explain the observed advantage of cerebellar feedback.
  • Computational contribution of cerebellar components: Component ablations and feedback perturbations (Fig. 4A, C in original manuscript) suggest that the information carried by cerebellar feedback contributes to performance beyond gating alone.

(1) Architectural distinction from a GRU

A GRU uses separate update and reset gates to regulate recurrent computation, with the update gate controlling interpolation between the previous and candidate states:

ht+1=(1−zt)⊙ht+zt⊙h~t,h_{t+1} = (1 - z_t) \odot h_t + z_t \odot \tilde{h}_t,

In contrast, Ctx-cb uses a separate cerebellar-inspired module to gate and transform cortical activity into structured feedback signal ctc_t, which enters the cortical recurrent drive additively:

ht+1=(1−α)ht+αϕ(Wrecht+Winxt+ct+bh).h_{t+1} = (1-\alpha)h_t + \alpha\phi(W_{\text{rec}}h_t + W_{\text{in}}x_t + \boldsymbol{c}_t + b_h).

(2) Comparison with parameter-matched GRU

We introduce a parameter-matched GRU baseline. Across 10 random seeds, the GRU outperforms the Ctx-only RNN but remains substantially below Ctx-cb in the following Table. These results suggest that generic recurrent gating alone does not account for the performance advantage of Ctx-cb.

ModelHidden units / parametersCognitive Acc (%) ↑\uparrowMotor NMSE ↓\downarrow
Ctx-only (RNN)79 / 846059.42±2.8359.42 \pm 2.830.22187±0.041210.22187 \pm 0.04121
GRU42 / 836570.99±5.6370.99 \pm 5.630.20434±0.030130.20434 \pm 0.03013
Ctx-cb48 / 842499.23±0.18\boldsymbol{99.23 \pm 0.18}0.00018±0.00003\boldsymbol{0.00018 \pm 0.00003}

(3) Computational contribution of cerebellar components

The GRU baseline, together with component ablations and targeted feedback perturbations (Fig. 4A, C in original manuscript), show that the Ctx-cb advantage extends beyond gating, with feedback-content removal causing greater performance impairment than gating removal alone. These findings suggest that structured information processing mediated by distinct cerebellar cell types contributes to the computational advantage.

(4) No privileged task or context information

We clarify that the gating signals in Ctx-cb are learned from cortical states and shared task inputs, without access to privileged task or context labels. In the cognitive task, Ctx-cb learns to associate its predicted decision with the instructed response rule, without receiving the ground-truth decision as input.

We will include these results and discussion in the revised manuscript.

  • Weakness 2

We regret that the original manuscript did not sufficiently clarify the architectural and physiological rationale for distinguishing cortical and cerebellar computations. We address these concerns from three perspectives: (1) model architectures, (2) physiological distinctions between cortex and cerebellum, and (3) the rationale for coupling across multiple brain regions.

(1) Comparison of Different Model Architectures

We agree that context processing and gating are not exclusive to the cerebellum. We introduce two cortex-only alternatives by replacing the cerebellar-inspired controller with a second RNN. All three coupled models process task context and use learned gating to regulate cortical-state retention.

  • Generic RNN–RNN, which learns directly from task cues;
  • Structured RNN–RNN, which retains the same decision-rule binding and movement selection mechanisms as Ctx-cb.

Ctx-cb outperforms the other models at comparable parameter counts, although Structured RNN–RNN achieves comparable cognitive accuracy (99.82% vs. 99.91%). These findings suggest that structured cerebellar feedback provides benefits beyond context processing and gating, particularly in continuous motor tasks.

ModelparametersCognitive Acc (%) ↑\uparrowMotor NMSE ↓\downarrow
Single RNN (Ctx-only)846059.42±2.8359.42 \pm 2.830.22187±0.041210.22187 \pm 0.04121
RNN+RNN (Generic)843272.44±5.0772.44 \pm 5.070.19514±0.074560.19514 \pm 0.07456
RNN+RNN (Structured)845498.82±0.2698.82 \pm 0.260.03442±0.012640.03442 \pm 0.01264
RNN+Feedforward (Ctx-cb)842499.23±0.18\boldsymbol{99.23 \pm 0.18}0.00018±0.00003\boldsymbol{0.00018 \pm 0.00003}
  • Weakness 2

(2) Physiological distinctions between cortex and cerebellum

We acknowledge that our model does not explicitly capture all physiological details and biological mechanisms. Instead, it focuses on functional cortico-cerebellar interactions at an abstract computational level while incorporating several physiologically inspired distinctions. Specifically,

  • Temporal dependence arises primarily from cortical recurrent dynamics and task-related inputs, while the cerebellar module remains feedforward, reflecting a fundamental architectural distinction between cortical and cerebellar processing.

  • Circuit-level processing is captured by effective population-level transformations inspired by the established cerebellar anatomical organization and signal-processing roles of MF, GrC, PC, and DCN pathways.

  • Input expansion and nonlinear transformation are captured by granule-cell-like nonlinear processing of cortical states.

(3) Rationale for coupling across multiple brain regions

Multi-region brain coupling plays an important role in cognitive and motor functions, as supported by growing experimental evidence. Consistent with recent computational studies (Pemberton et al., 2024; Zheng et al., 2024; Clark & Beiran, 2025; Perich et al., 2026), we use simplified recurrent networks to investigate how such interactions may support compositional generalization, a possibility supported by physiological evidence, although the underlying computational mechanisms remain poorly understood.

In cortico-cerebellar interactions, experimental findings on task-dependent recruitment of cortical inputs to the cerebellum and context-dependent modulation of granule-cell activity (Shahshahani et al., 2024; Garcia-Garcia et al., 2026) motivate our cerebellar-inspired feedback design. These mechanisms represent a biologically inspired computational hypothesis rather than functions exclusive to the cerebellum.

Ref:

[1] Pemberton, J., et al. (2024). Cerebellar-driven cortical dynamics can enable task acquisition, switching and consolidation. Nat. Commun., 15, 10913.

[2] Zheng, W.-L., et al. (2024). Rapid context inference in a thalamocortical model using recurrent neural networks. Nat. Commun., 15, 8275.

[3] Clark, D. G., & Beiran, M. (2025). Structure of activity in multiregion recurrent neural networks. PNAS, 122(10), e2404039122.

[4] Perich, M. G., et al. (2026). Inferring brain-wide interactions using data-constrained recurrent neural network models. Neuron, published online August 7, 2026.

[5] Shahshahani, L., et al. (2024). Selective recruitment of the cerebellum evidenced by task-dependent gating of inputs. eLife, 13, RP96386.

[6] Garcia-Garcia, M. G., et al. (2026). Granule cells reorient cortical trajectories to separate contexts. Nature, 658, 977–986.

  • Weakness 3

We regret that the original manuscript did not clearly explain the rationale for input setting and learning mechanisms. We provide following clarifications:

(1) Regarding input encoding, we adopt one-hot encoding for categorical task rules, consistent with the standard setting used in established computational neuroscience models (Yang et al., 2019; Ito et al., 2022; Driscoll et al., 2024). We will explicitly highlight this, retaining this setting to maintain consistency with prior work in the revised manuscript.

(2) Regarding the contributions of different architectural components, the content pathway combines task-relevant information, while the learned gate regulates cortical state retention. Neither mechanism explicitly imposes orthogonality on learned cortical representations. Instead, our analysis in Fig. 6 shows that Ctx-cb exhibits stronger cross-context representational alignment than Ctx-only (mean cosine similarity: 0.995 vs. 0.761), supporting more consistent representation reuse across contexts rather than enforced orthogonality.

Ref:

[1] Yang, G. R., et al. (2019). Task representations in neural networks trained to perform many cognitive tasks. Nat. Neurosci., 22, 297–306.

[2] Ito, T., et al. (2022). Compositional generalization through abstract representations in human and artificial neural networks. NeurIPS, 35, 32225–32239.

[3] Driscoll, L. N., et al. (2024). Flexible multitask computation in recurrent networks utilizes shared dynamical motifs. Nat. Neurosci., 27, 1349–1363.

(3) Regarding sequential processing and gating, we acknowledge that adaptive gating is an established computational mechanism for maintaining information across temporally separated stimuli. Our contribution is not to claim gating itself as a novel mechanism, but to investigate its interaction with structured cerebellar content feedback. The GRU baseline shows that conventional recurrent gating alone cannot fully explain the generalization advantage of the proposed Ctx-cb model. The component ablations in Fig. 4 reveal the complementary contributions of adaptive gating and structured content feedback.

  • Weakness 4

(1) Notations for tasks

We regret that our task notation was not presented clearly enough in the original manuscript. Our task designs and mathematical formulations follow established neurocomputational paradigms adopted in previous studies (Ito et al., 2017; Tan et al., 2026; Tian et al., 2026). We will revise Section 3.1 and Figure 1 to clarify the task notation before presenting the formal definitions. Specifically, we will make three changes:

  • Cognitive task (Appendix A.1). Simplifying the response mapping into three steps: selecting the relevant stimulus features, applying the logical rule, and mapping the resulting decision to a response. For example, applying “BOTH” to features 1 and 0 yields false, corresponding to response 1 when response 0 represents true. This example clarifies the mapping without requiring readers to interpret the permutation.

  • Motor task (Appendix A.2). Simplifying the target velocity into a single expression during movement. For example, a rightward-to-upward movement sequence follows the corresponding velocity profiles in order. The exact timing and velocity definitions are retained in Appendix A.2 for reproducibility.

  • Task cues versus the learned gate (Section 3.3, Eq. 2). Reserving qtq_t for the learned gate, with 1−qt1 - q_t denoting the fraction of the preceding cortical state retained, and distinguishing externally supplied cues as “trial phase” in cognition and “sequence position” in motor control. This clarifies the distinction between predefined inputs and learned computations.

Refs:

[1] Tan, J. B., et al. (2026). Compositional recombination is facilitated by a distributed cortico-cerebellar network. Cell Reports, 45(7), 117572.

[2] Tian, L. Y., et al. (2026). Neural representation of action symbols in primate frontal cortex. Nature, 654, 152–162.

[3] Ito, T., et al. (2017). Cognitive task information is transferred between brain regions via resting-state network topology. Nature Communications, 8, 1027.

(2) Terminology of “successor states”

We regret that the term “successor states” was not clearly explained in the original manuscript. We clarify the distinction as follows:

  • In our motor task, “successor” refers to the next movement in a sequence, whose associated neural dynamics are influenced by preceding movements. Figs. 7–8 show that Ctx-cb reduces this predecessor-dependent variability compared with Ctx-only, enabling more consistent execution across different movement histories.

  • In Dayan (1993), successor representations encode the expected discounted occupancy of future states under a given policy.

To avoid confusion, we will replace “successor states” with the clearer expression “cortical activity during the next movement” in the revised manuscript.

Ref:

[1] Dayan, P. (1993). Improving generalization for temporal difference learning: The successor representation. Neural Computation, 5(4), 613–624.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.