acceptodds
Under review as a conference paper at ICLR 2027

Mixture-of-Microzones for slow and fast sensorimotor adaptation

Abstract

Learning in naturalistic environments requires both rapid and long-term sensorimotor adaptation. How the brain and machines achieve such dual control is not clear. We introduce Mixture-of-Microzones (MoM), a cerebellar-inspired architecture that separates immediate error correction from slower learning. MoM uses error geometry to sparsely recruit specialized modules according to error direction and magnitude. Recruited modules perform local, state-conditioned coordinate transformations from sensory-error space to motor space, allowing their corrections to be composed and applied directly to the controller output. In parallel, the same gate determines which modules receive learning updates, progressively consolidating the response. In sensorimotor adaptation, this structured corrective pathway reduces adaptation time by approximately fourfold relative to ablated variants and substantially outperforms existing controller–adapter baselines. Ablations show that neither geometric gating nor a direct corrective pathway alone is sufficient: their combination is required to produce appropriately aligned motor corrections and rapid adaptation. We further isolate this principle in learned optimization, where directional gating generalizes to unseen coordinate transformations. Together, these results identify error-geometric coordinate transformation as a mechanism for correcting perturbations immediately while slower adaptive processes consolidate the response.

Primary area:
applications to neuroscience & cognitive science
Keywords:
biologically-plausible sensorimotor representations, cerebellum, sensorimotor adaptation, motor control, feedback control

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_GZAUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Mixture-of-Microzones (MoM), a cerebellum-inspired architecture that separates fast error correction from slower sensorimotor adaptation. It uses fixed directional gating to select modules that map sensory errors to motor corrections. In simulated reaching tasks, MoM achieves behavioral correction in substantially fewer trials than the baselines. The authors also include a small learned optimization experiment to illustrate how the directional gating works.

Critical strengths
  • Separation of fast correction and slow adaptation: I find this distinction interesting, and the overall architecture is fairly intuitive. Fixed directional gating is also a reasonable design choice, since it introduces some structure without having to learn a separate routing mechanism.
  • Strong behavioral correction results: The reduction from roughly 16–21 trials to 4 trials is quite noticeable. The ablations are helpful here, as they suggest that both the directional gating and the direct corrective pathway contribute to this result.
Critical weaknesses
  • Correction versus learning: My main concern is whether MoM actually learns faster or just corrects errors more quickly through the direct pathway. These are different things, and the experiments do not fully separate them. In fact, Table 2 shows that internalization is slightly slower for MoM than for the baselines (25 versus 21–24 trials). The cortical weights and transformation matrices are also frozen during adaptation, which makes me question how much of the improvement reflects lasting adaptation. I would like to see what happens if the direct corrective pathway is turned off after adaptation. If the improvement disappears, the claim about faster adaptation would need to be qualified.
  • Limited evaluation and missing baseline: Most of the evidence comes from a simulated 2D reaching task with predefined error directions. The learned optimization example is also restricted to 2D quadratic functions. I think the current experiments are enough to show that the architecture can work in these settings, but not enough to establish how broadly useful the microzone structure is. In particular, I would like to see a comparison with a simpler learned residual feedback controller that does not use directional experts. The current ablations remove individual components of MoM, but they do not answer whether multiple directional modules are necessary in the first place.
Questions for authors

What happens to MoM's performance if the direct corrective pathway is disabled after adaptation? Does the model retain any of the improvement?

2 comments

We thank the reviewer for the insightful critique and for pointing out the need to explicitly separate online correction from permanent weight-based learning, as well as the importance of comparing against a non-modular baseline.

Correction vs. Learning & Post-Adaptation Cutout: Our central claim is that MoM accelerates initial behavioral correction through its geometrically routed direct pathway, rather than accelerating weight-based internalization. As shown in Table 2 and Figure 4B, MoM achieves behavioral correction within 4 ± 2 trials, whereas internalization requires 25 ± 4 trials, compared with 21 ± 3 trials for Pemberton et al. (2024) and 24 ± 3 trials for Filipe et al. (2026). Thus, MoM's rapid behavioral compensation is accompanied by a slightly slower internalization timescale, more closely resembling the experimental adaptation dynamics shown in Figure 4B. One possible explanation is that rapid online correction reduces residual task errors and therefore the pressure for subsequent weight-based adaptation, whereas models without direct correction experience larger errors that may promote faster internalization.

To directly address whether the improvement persists after adaptation, we performed the suggested post-adaptation direct-pathway cutout (Appendix A.4.2, Table 6). After full adaptation, we disabled the direct corrective output (Δu = 0) without further learning. Task error remained essentially unchanged, from 0.05 ± 0.02 cm with the direct pathway active to 0.05 ± 0.01 cm after cutout. Similarly, take-off error was 0.5 ± 0.2° before cutout and 0.4 ± 0.1° afterward (mean ± SD across 10 seeds). These results demonstrate that the learned behavioral compensation persists without ongoing direct correction, supporting the conclusion that adaptation has been internalized into the learned controller–adapter dynamics rather than depending indefinitely on the direct corrective pathway.

Evaluation, Generalization, and Non-Modular Baseline: To assess whether directional microzones provide an advantage over a general-purpose nonlinear feedback controller, we implemented a non-modular baseline that retains the cortico-cerebellar architecture and mossy-fiber expansion but replaces the four directional microzones and geometric gating with a single global forward predictor and a nonlinear sensory-to-motor mapping. Specifically, a 2000 → 12 → 6 ReLU network predicts sensorimotor variables, while a separate 4 → 16 → 2 ReLU MLP maps delayed visual error and predicted position into joint-space corrections. This baseline contains 24,204 parameters, closely matching MoM's 24,028 active parameters. Both models receive the same sensory information and feedback delay and use the same training budget, with corrective weights frozen during adaptation (Appendix A.7).

Under visuomotor rotation, the global nonlinear controller required 13 ± 2 trials to achieve behavioral correction, compared with 4 ± 2 trials for MoM. Its corrective alignment was also lower (0.60 ± 0.15 vs. 0.91 ± 0.05; Table 1). To control for total model capacity, we additionally evaluated a wider global predictor (2000 → 24 → 6) with the same nonlinear direct pathway, yielding 48,288 parameters, closely matching MoM's 48,056 total parameters. Although this stronger baseline improved to 10 ± 3 correction trials and 0.65 ± 0.10 alignment, it remained slower and less accurately aligned than MoM (Appendix A.7, Table 8). All values are mean ± SD across 10 seeds.

These comparisons show that the benefit of MoM cannot be explained solely by the presence of a direct feedback pathway or by the number of model parameters. Instead, the results support the advantage of geometrically routing sensory errors to specialized local error-to-action transformations, producing more appropriately directed motor corrections and faster behavioral compensation. Beyond centre-out reaching, we also evaluate generalization to more complex MNIST drawing trajectories under a 30° visuomotor rotation (Appendix A.4.3, Figure 6), where MoM preserves trajectory structure immediately after perturbation and improves further with adaptation.

Thank you again for your review. As the discussion period comes to an end, we wanted to kindly ask whether our response and the revised manuscript have addressed your concerns. In brief:

  • Correction versus learning. As you suggested, we disabled the direct corrective pathway after adaptation without further learning: task error and take-off error remained unchanged (Appendix A.4.2, Table 6). Freezing plasticity instead preserves rapid correction but prevents internalization (Table 5). We therefore clarify that MoM accelerates behavioral correction, rather than weight-based learning.

  • Baseline without directional experts. We added a global, non-modular residual feedback controller matched for sensory inputs, feedback delay, training budget, and model capacity. MoM achieved faster and better-aligned corrections than this baseline (Table 1, Appendix A.7, Table 8).

  • Evaluation beyond 2D reaching. We additionally evaluated MoM as a corrective adapter for a frozen, pretrained SmolVLA policy on two LIBERO robotic manipulation tasks (Appendix A.4.4, Fig. 7), complementing our existing MNIST drawing experiments (Appendix A.4.3).

We note that other reviewers acknowledged the improvements and raised their assessments following the revisions. If these additions resolve your two main concerns, we would be grateful if you would consider reflecting this in your assessment. If anything remains unclear, we would be happy to clarify it before the deadline. Thank you again for your time and consideration.

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

This paper proposes a Mixture-of-Microzones (MoM) model, a cortico-cerebellar recurrent controller with fixed error-direction top-2 gating. In this model, selected microzones integrate cortical states with delayed sensory feedback via a predictive pathway, while a direct pathway maps weighted task-space errors and predicted states into corrective adjustments in joint space. A unified gating mechanism governs both the selection of modules contributing to the output and the determination of which modules are eligible for learning. In a planar reaching task involving a rotation of the controller's output, the full model requires only 4±2 trials to return to baseline error levels; in contrast, ablation experiments in which the direct pathway is removed or the gating mechanism is replaced with a learnable alternative—require more trials to achieve the same performance. Furthermore, the direct corrective adjustments produced by this model exhibit a higher degree of alignment with the task errors. In a separate benchmark test based on gradient-based optimization under rotation, the model also demonstrates the ability to interpolate across previously unseen rotation angles within a two-dimensional environment.

Critical strengths

This study integrates both immediate correction and gradual adjustment into a single, concrete system framework. In this model, a direct pathway enables the rapid reduction of task errors following a perturbation, while predictive microzones facilitate continuous learning across multiple trials; a shared gating mechanism provides a unified organizational structure for these processes. This offers a useful framework for exploring how performance can be maintained while the controller itself undergoes adjustments. Through factorial ablation experiments, the study further suggests that this combination is indeed effective within the proposed architecture.

Critical weaknesses

The conceptual contributions of this study require more precise definition, and more compelling points of comparison should be introduced. For instance, the feedback-error learning model proposes using feedback-generated teaching signals to train an inverse model, while the MOSAIC framework employs a modular responsibility-assignment approach to combine and train pairs of forward and inverse models. More directly, research by Feulner et al. 2025 demonstrates that a single delayed error signal can simultaneously support online correction and drive trial-by-trial plasticity; furthermore, the architecture proposed by Pemberton et al. 2024 serves as a foundation for the present study. A specific contribution of the MoM model is its use of a fixed routing mechanism—based on the geometry of the error—to map local errors to actions, utilizing a single gating mechanism to facilitate both online correction and learning-based updates. However, Table 4 indicates that in the No geometry variant, this fixed, direction-based gating is replaced by a learnable gating mechanism. Table 1 therefore establishes an advantage over that learned-gate implementation, but does not isolate the benefit of microzone geometry against a general feedback controller matched for error/state information, sensory delay, capacity, and training budget. A global state-conditioned error-to-action mapping baseline under those matched conditions would clarify whether the proposed decomposition improves zero-learning first-trial correction, sample efficiency, or held-out rotation generalization.

Questions for authors
  1. Distinguishing correction from learning. In experiments involving perturbations, when are the Adam updates applied,at each time step, after the entire sequence, or between trials? Please report performance for two scenarios: first, when all plastic parameters are held fixed (i.e., updates are disabled) while the feedback pathway remains intact; and second, after adaptation has occurred, with only the direct correction mechanism disabled. These interventions allow for the differentiation between immediate online correction and newly acquired compensation. A decline in the contribution of the direct pathway, as shown in Figure 2D, does not in itself constitute evidence of internalization, because, the error-dependent component may decline as task error shrinks.

  2. Equation (13) applies a rotation to W_out, which alters hand velocity in Cartesian coordinates while leaving the arm's kinematic properties unchanged. How does this operation correspond to the rotation of visual feedback implemented in the experiments? Furthermore, if a general-purpose feedback controller—configured with the same error and state information, latencies, parameter-scale constraints, and training regimen—were employed, could it achieve similarly rapid corrections under such a perturbation?

  3. What are the functional advantages of sharing a directional gate between immediate correction and learning? Please compare the full model against variants where the geometric routing mechanism is applied exclusively to either direct output correction or learning allocation. Furthermore, examine the performance in the contexts of early-stage correction and late-stage internalization. Given that the preference directions are fixed, directional tuning alone is insufficient to establish the superiority of this shared architecture. Does this gating mechanism, therefore, provide benefits that extend beyond those of a single-path approach?

  4. Auditability of Key Data. Please verify the data for the “4±1” and “4±2” experimental groups in the “full-model,” as well as the “16±3” and “21±3” groups in the “learned-gate” model. Indicate whether there were any differences in the protocols or random seeds used for these experimental groups, and provide the value corresponding to each random seed. Additionally, provide a detailed definition of the metrics used to construct Figure 4C—including the model population, reference benchmarks, specific measurement criteria, normalization procedures, averaging methods, and the scaling applied relative to the experimental data. The answers to these questions will determine the extent to which I can rely on the key numerical comparisons and the comparisons of neural activity presented in the paper.

4 comments

We thank the reviewer for the rigorous critique. We agree that isolating the benefits of microzone geometry against a fully matched, non-modular global error-to-action baseline is essential. Below, we address each point and detail how the revised manuscript incorporates these clarifications and validations.

Critical Weakness: Matched Non-Modular Baseline To isolate whether MoM's performance gains arise from geometric microzone routing rather than simply adding a direct corrective pathway, we implemented a non-modular nonlinear feedback controller that preserves the cortico-cerebellar architecture, sensory inputs, feedback delay, and training budget while replacing the geometrically routed microzones with a global forward predictor and nonlinear error-to-action mapping.

Architecture and capacity matching: The global baseline uses a 2000 → 12 → 6 ReLU forward predictor and a separate 4 → 16 → 2 ReLU MLP that converts delayed visual error and predicted position into joint-space corrections. With 24,204 parameters, this closely matches MoM's 24,028 active parameters. We additionally evaluated a wider 2000 → 24 → 6 global predictor with the same nonlinear direct pathway, giving 48,288 parameters, closely matching MoM's 48,056 total parameters (Appendix A.7, Table 8).

Performance comparison: Under visuomotor rotation, MoM achieves behavioral correction in 4 ± 2 trials, compared with 13 ± 2 trials for the active-capacity-matched global nonlinear controller. Corrective alignment is also higher for MoM (0.91 ± 0.05) than for the global controller (0.60 ± 0.15; Table 1). Increasing the global controller's total capacity improves its performance to 10 ± 3 correction trials and 0.65 ± 0.10 alignment, but does not close the gap with MoM (Appendix A.7, Table 8).

Why geometric decomposition helps: These results indicate that the advantage of MoM cannot be explained solely by the presence of a direct corrective pathway or by total model capacity. Geometrically routing errors to specialized local transformations provides a useful inductive bias for generating appropriately directed motor corrections. The results support the importance of directional routing.

Q1 optimization updates: Cerebellar parameters are optimized with Adam, with updates applied at each simulation time step using delayed sensory feedback, following Pemberton et al. (2024). During adaptation, cortical RNN weights and the direct transformation matrices are frozen, while the cerebellar parameters remain plastic unless explicitly disabled.

Take-off error versus task error: Behavioral correction time measures how quickly task error returns to its pre-perturbation level. In contrast, internalization is evaluated using take-off error, defined as the angular deviation between the target direction and the initial movement direction estimated from the first 10 trajectory points after movement onset (Appendix A.5). This measures the evolving initial movement plan before later feedback-dependent corrections substantially alter the trajectory.

MoM achieves behavioral correction in 4 ± 2 trials, while internalization requires 25 ± 4 trials. By comparison, internalization requires 21 ± 3 trials for Pemberton et al. (2024) and 24 ± 3 trials for Filipe et al. (2026) (Table 2). Thus, MoM accelerates behavioral correction without accelerating weight-based internalization.

Frozen-plasticity intervention (Appendix A.4.2, Table 5): We additionally froze cerebellar parameters during adaptation while retaining the direct corrective pathway. Behavioral correction remained rapid, but take-off error did not return to baseline within 500 adaptation trials. This demonstrates that immediate feedback-driven behavioral correction can occur without weight updates, whereas recovery of the initial movement plan depends on plasticity in the experimental conditions tested.

Post-adaptation direct-pathway cutout (Appendix A.4.2, Table 6): After full adaptation, we disabled only the direct corrective output (Δu = 0) without further learning. Task error remained essentially unchanged, from 0.05 ± 0.02 cm with the pathway active to 0.05 ± 0.01 cm after removal. Take-off error was similarly low (0.5 ± 0.2° versus 0.4 ± 0.1°). These results show that acquired compensation persists without ongoing direct corrective output, supporting internalization within the learned controller–adapter dynamics rather than indefinite dependence on the fast feedback pathway.

Q2

Visuomotor Perturbation Protocol: We thank the reviewer for identifying an error in the original manuscript. We incorrectly described the perturbation as a rotation of the cortical output matrix WoutW_{\mathrm{out}}. This was a mistake in the methodological description. The visuomotor rotation is applied to visual cursor feedback, while the physical arm kinematics remain unchanged. We have corrected and clarified the protocol in Section 3.2.1 of the revised manuscript and apologize for the confusion.

Q3

We performed pathway-specific routing ablations to determine whether geometric gating primarily benefits direct motor correction, cerebellar learning updates, or their combination (Appendix A.4.1, Table 4). All conditions use the same architecture, training budget, and random seeds.

ModelTcorrT_{\mathrm{corr}} (trials)TintT_{\mathrm{int}} (trials)Zero-shot error (cm)
Full MoM (Shared Gate)4±24 \pm 225±425 \pm 41.5±0.31.5 \pm 0.3
Correction-Only Gate6±16 \pm 126±326 \pm 31.8±0.41.8 \pm 0.4
Learning-Only Gate19±419 \pm 422±422 \pm 43.2±0.43.2 \pm 0.4

Correction-Only Gate: Geometric routing is retained for the direct corrective pathway, while learning updates are ungated and all microzones receive updates. This condition preserves most of MoM's rapid behavioral correction (6±16 \pm 1 versus 4±24 \pm 2 trials) and produces a similar internalization timescale (26±326 \pm 3 versus 25±425 \pm 4 trials).

Learning-Only Gate: Learning updates remain geometrically routed, but the direct pathway combines contributions from all microzones with uniform weights. Behavioral correction becomes substantially slower (19±419 \pm 4 trials) and zero-shot error increases to 3.2±0.43.2 \pm 0.4 cm, despite internalization occurring within 22±422 \pm 4 trials.

Conclusion: These results identify geometric routing of the direct corrective pathway as the principal contributor to rapid behavioral compensation. Routing learning updates alone is insufficient to achieve the same speed. Sharing the directional gate between correction and learning provides a smaller additional benefit, but the main advantage arises from selecting appropriately directed local transformations for online motor correction.

Q4 Great catch—thanks for pointing this out. The difference between 16 and 21 arises because the non-directional case in Table 1 corresponds to a random gate rather than a learnable gate (which was inadvertently mislabeled in earlier drafts due to a compilation error, which we have now corrected).

Numerical consistency and baseline definitions: We have revised the manuscript to distinguish the non-modular global nonlinear baseline from the random-gate ablation and to ensure consistent numerical values throughout the main text and appendix.

Specifically, Table 1 reports:

  • Full MoM: Tcorr=4±2T_{\mathrm{corr}} = 4 \pm 2 trials; corrective alignment A=0.91±0.05A = 0.91 \pm 0.05.
  • No Geometry (Fixed Random Gate): Tcorr=16±3T_{\mathrm{corr}} = 16 \pm 3 trials; corrective alignment A=0.24±0.15A = 0.24 \pm 0.15.
  • Forward Model + Direct Pathway (Global Nonlinear Controller): Tcorr=13±2T_{\mathrm{corr}} = 13 \pm 2 trials; corrective alignment A=0.60±0.15A = 0.60 \pm 0.15.

The random-gate condition selects two experts independently of sensory-error direction, whereas the non-modular baseline uses a global ReLU predictor and a separate nonlinear direct controller. The latter is further evaluated under both active- and total-capacity matching in Appendix A.7, Table 8. All reported comparisons use mean ±\pm SD across 10 random seeds.

Figure 4C methodology: Following Feulner et al. (2025), we quantify changes in cortical RNN activity during early adaptation relative to pre-perturbation activity. We calculate the absolute activity difference for each cortical unit, average across units and trials, and align the resulting activity profiles to the go cue. Activity is baseline-subtracted using the pre-movement interval [−200,0][-200, 0] ms and subsequently max-normalized. For visualization, the model activity-change curves are scaled by a factor of 30 to facilitate comparison with experimental primary motor cortex (M1) recordings from Perich et al. (2018), as analyzed by Feulner et al. (2025). This scaling is for visualization and is not used to calculate behavioral adaptation metrics.

Thank you for the additional experiments and for correcting the description in the original manuscript. However, Although I can see some improvements, my remaining concern is the broader contribution. It is still unclear what new insight MoM offers into the neural mechanisms of control. On the engineering side, the results do not yet show whether its advantage extends to other control tasks. I therefore maintain my rating as weak rejection.

We thank the reviewer for the constructive feedback. We agree that the broader contribution needed to be clearer, and we have revised the Discussion accordingly.

Neural-mechanistic contribution. MoM proposes a computational link between two findings: rapid motor adjustment is associated with alternating bands of Purkinje-cell activity encoding the sign and magnitude of sensorimotor mismatch (Nguyen et al., 2026), and cerebellar error signals are organized by movement direction and motor output (Herzfeld et al., 2018; Doostmohammadi et al., 2026). We hypothesize that error direction may determine which microzones are recruited for rapid correction and how they contribute to motor output, alongside slower synaptic learning. MoM implements this idea through geometrically routed, learned local error-to-action transformations, building on the cortico-cerebellar framework of Pemberton et al. (2024). It thereby offers one way in which immediate behavioral correction and slower internalization could coexist in the same circuit.

Our ablations support this interpretation. Routing only the corrective pathway largely preserves the rapid correction of the full model, whereas routing only the learning updates does not accelerate it (Appendix A.4.1, Table 4). Freezing plasticity preserves rapid correction but prevents internalization (Table 5).

Testable prediction. If this mechanism contributes to biological control, selectively perturbing microzones associated with one error direction should preferentially impair rapid corrections in that direction. We present this as a hypothesis, not an established mechanism.

Broader applicability. Our original submission already included MNIST drawing trajectories under visuomotor rotation (Appendix A.4.3) and generalization to held-out rotations in learned optimization (Fig. 5). To address the reviewer’s concern about other control tasks, we added experiments with a frozen, pretrained SmolVLA policy controlling a 7-DOF robotic arm on two LIBERO manipulation tasks (Appendix A.4.4, Fig. 7). MoM achieved higher task success than the unmodified VLA and a global corrective adapter under two perturbations introduced only at test time, while preserving performance under clean conditions. Together, these results provide preliminary evidence that the benefits of geometric routing extend to other tasks and to a different controller. We have clarified the scope and limitations of these findings in the revised Discussion.

We hope these revisions clarify the contribution.

Reviewer_h6p5OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces Mixture-of-Microzones (MoM), a cerebellar-inspired error-correction adapter that influences motor controls through a modulatory feedback to the continuous-time recurrent cortical controller (an RNN) supporting slower adaptation and an immediate kinematic error correction (directly summed to the output of the RNN). Each microzone is associated with a frozen preferred direction distributed around the circle and the top 2 "experts" (microzones) are retained at every step. The model is tested on two experiments:

  • A synthetic, instructed-delay reaching task inspired by monkey experiments (Perich et al., 2018) where trials are 3 seconds (300 steps) long, and online error correction and adaptation are evaluated by perturbating output projections with rotations.
  • A learned optimisation task, where the optimiser must recover appropriate updates when observed gradients are rotated relative to optimisation coordinates.

The model shows faster behavioural adaptation with respect to the baseline (another recurrent controller-adapter model; Pemberton et al., 2024) and to its geometry- or direct-feedback- ablated versions. The authors show some experimental evidence - a comparison with a single biological neuron (Herzfeld et al., 2018) - that MoM is a biologically-plausible architecture, while model responses comparison with experimental data (Perich et al., 2018) throughout adaptation, number of trials, and time is more thorough. The authors conclude by claiming a contribution to neuroscience, which is in my opinion partially supported (see below); and to the ML/control field, without supportive evidence.

Critical strengths

The paper introduces a simple and interpretable architecture of cortico-cerebellar interactions. The model is supported by a relatively broad set of previous experimental results, but this validation could be enhanced (see weaknesses below).

The methods and results are clearly presented, both in the main text and in figures.

Ablations of components from MoM show substantive evidence that both geometric gating and direct feedback pathways are important mechanisms for effective and rapit adaptation. In particular, results show that direct correction enables rapid learning after a perturbation, while slower recurrent adaptation develops afterwards, in agreement with the proposed timescale separation.

The directional tuning of a single expert shows promising alignment with an experimentally recorded biological neuron (taken from Herzfeld et al., 2018). Furthermore, the introduced MoM gating is shown to be consistently preferable over alternatives: hard (top-1), soft (keep all experts), or learned gating mechanism.

Direct feedback and learned correction show distinct adaptation scales, resembling responses reported in experimental data (Perich et al., 2018).

Critical weaknesses

Figure 3a compares the directional tuning of a single microzone expert with that of one experimentally recorded neuron (Herzfeld et al., 2018). It is not clear how representative this comparison is, and if a similar correspondence could be generalised across a population of recorded neurons. Quantifying this agreement across multiple biological neurons could substantially strengthen this contribution.

Table 2 presents results that are quite central to the claims of the paper, in particular with respect to gating strategies. However, these results receive little attention or interpretation in the main text. A more thorough coverage would help clarify significance.

The model is compared with experimental data from Herzfeld et al. (2018) in Figure 3 and Perich et al. (2018) in Figure 4. The two datasets are never clearly introduced or described in the main text, and the reason for using two different experimental datasets is not sufficiently explained. Confusingly, Section 3.2.1 describes the synthetic reaching task citing Perich et al. (2018). Clarifying the distinction between synthetic tasks, experimental tasks and recordings, and the two sources would help clarify this.

Figure 4 is a central contribution of the paper, as it gives evidence for the proposed separation of fast and slow timescales. It is not clear to me how the model-derived signals presented in Figures 4b and 4c are derived. Additional methodological details about these two metrics would be useful.

Although I appreciate the comparison with experimental neural recordings, the claimed contributions to ML/control is less convincing. Evaluations are mainly carried out in synthetic settings, leaving doubts on the broader applicability of the proposed architecture. Proposing to test the model on a more representative benchmark and identifying a concrete suitable dataset - even as future work - would greatly help.

I would be happy to increase my score if my main concerns are addressed.

Questions for authors

For major, see weaknesses. Minor, but could be useful for my understanding:

  • Why is Section 4.1 with ablation studies introduced before Section 4.2 where the main comparison with the baseline is made?
  • I am not entirely sure about what Figure 3e is showing, are "opposing experts" the same as "directional gating"? There might be a labelling mismatch going on.
  • Will the code be made public?
3 comments

We thank the reviewer for the thoughtful feedback and constructive comments. Below, we address each point and detail how the revised manuscript incorporates these clarifications and updates.

  • Fig 3A & Population-Level Correspondence: We thank the reviewer for this observation. Importantly, Fig. 1B of Herzfeld et al. (2018) demonstrates population-level directional tuning across 67 Purkinje cells aligned to each neuron's preferred error direction (CS-on). Similarly, MoM microzones share the same geometric tuning function, differing only in their preferred directions. Thus, when aligned to their respective preferred directions, the correspondence shown in Fig. 3A extends across the entire microzone population rather than representing a single isolated expert. We chose to highlight a single expert example to make the color-matching visualization clear, and we have expanded the figure caption and Section 4.3 to clarify this population-level interpretation.

  • Table 2 and Gating Strategies: We have substantially expanded the discussion of Table 2 in Section 4.3 and Section 4.4 to clarify its significance. Table 2 quantifies the behavioral results in Fig. 2 and evaluates alternative gating strategies. The rationale for top-2 gating stems from the directional tuning of Purkinje cells (Herzfeld et al., 2018): each cell responds preferentially to errors along its CS-on direction, with weaker responses to neighboring directions and minimal responses to opposite errors (180º). Similarly, MoM’s top-2 gate recruits two neighboring experts for each error direction, enabling the composition of intermediate directions (e.g., 45º) while excluding oppositely tuned experts. Hard top-1 gating prevents this composition (producing step-like responses), whereas soft gating activates all experts, diminishing the distinction between neighboring and opposing responses. Top-2 thus provides an optimal balance between behavioral performance and biologically motivated directional tuning.

  • Regarding Fig. 3E, “adjacent” and “opposing” refer specifically to experts whose preferred directions differ by 360º/N (for N experts) and 180º, respectively, not to different gating strategies. We have updated the text and figure captions to explicitly eliminate any labelling ambiguity.

  • Datasets: We have clarified the distinct roles of the two experimental datasets in Section 3.2.1 and Section 3.2.3. Herzfeld et al. (2018) provide Purkinje-cell recordings used to compare cerebellar directional tuning with MoM's gating mechanism (Fig. 3). Perich et al. (2018) provide monkey center-out reaching data, including hand trajectories and primary motor cortex (M1) recordings, used to validate the model's behavioral output and cortical RNN activity dynamics (Fig. 4). Our synthetic reaching task replicates this experimental center-out paradigm (reaching from the center toward one of eight equally spaced targets); we cite Perich et al. (2018) as the source of this behavioral paradigm rather than as the raw data source for our synthetic environment.

  • Details for Figures 4B and 4C: We have expanded the methodological descriptions in Section 4.5 and Appendix A.5:

Fig. 4B (Take-off error): Measured as the angular difference between the target direction and the initial movement direction, estimated from the first 10 trajectory points after movement onset, capturing motor intention prior to feedback correction.

Fig. 4C (Cortical population activity change): Following the methodology of Feulner et al. (2025, Fig. 3C), we compute the absolute difference in each cortical RNN unit's activity between adaptation epochs and average across units and trials, aligned to the go cue. This reveals temporally distinct early learning-related and later feedback-related activity changes, which match experimental M1 recordings from Perich et al. (2018) as analyzed by Feulner et al. (2025, Fig. 3D).

  • Broader ML/Control Applications: We agree that evaluating MoM beyond synthetic reaching tasks is a crucial direction. A direct and concrete application for future work is robotic manipulation using vision-language-action models evaluated on benchmarks such as LIBERO. Integrating MoM as a cerebellum-inspired adaptation module could enable rapid online correction when executed actions deviate from intended outcomes (e.g., when a robot encounters unmodeled dynamics, actuator lag, or a grasping failure). Rather than continuing an unsuccessful action sequence, the module uses the resulting error signal to adjust subsequent motor commands immediately.

Minor Responses:

Q1 (Section 4.1 Ordering): We presented the ablation studies first to motivate the structural necessity of both components before comparing the full model against baseline architectures. We have added text in Section 4.1 explaining this structural progression.

Q2 (Figure 3E): Addressed above

Q3 (Code): Yes, the code will be made fully public upon publication to support reproducibility and future extensions.

I thank the authors for the thorough rebuttal. All of my concerns were addressed. Moreover, I appreciate the authors' replies to the other reviewers and the clarity in presenting additional results and updated passages in the article. I believe the manuscript now surpasses the acceptance threshold. Hence, I will raise my score to weak accept.

We sincerely thank the reviewer for taking the time to carefully review our response and the revised manuscript. Thank you again for your valuable feedback and for raising your score.

Comments on the paper

1 comment

We sincerely thank all reviewers for their thoughtful and constructive feedback. We appreciate their positive observations regarding the simplicity and interpretability of the architecture, the clarity of the presentation, the separation between rapid correction and slower adaptation, and the experimental comparisons and ablation studies. Their comments have helped us substantially strengthen the manuscript.

We have uploaded a revised version, with all additions and substantial modifications highlighted in blue. The main revisions are as follows:

  • Clarified the biological motivation and novelty (Introduction and Related Work). We expanded the discussion of cerebellar error-direction tuning, synaptic plasticity, and descending motor pathways. We also clarified MoM's distinction from existing cerebellar-inspired models, emphasizing its use of error geometry to organize direct sensory-to-motor corrections.

  • Strengthened the architectural ablations (Section 4.1, Table 1, Appendix A.7). We introduced global forward-model baselines matched for capacity, feedback delay, and training budget, incorporating a direct corrective pathway without geometric routing. These comparisons demonstrate that rapid behavioral correction is enabled not simply by adding a direct pathway, but by geometrically partitioning the error space into directionally specialized modules. This structure selectively recruits local error-to-action transformations relevant to the observed error, enabling substantially faster correction than a global mapping that must accommodate errors across all directions.

  • Expanded the analysis of directional gating (Section 4.3, Fig. 3, Table 2). We clarified the population-level comparison with Purkinje-cell recordings and expanded the comparison of top-1, top-2, soft, and learned gating strategies.

  • Distinguished fast correction from slow internalization (Section 4.4, Tables 2 and 4–6, Appendix A.4). We explicitly separated behavioral correction from weight-based internalization using distinct metrics. We further introduced pathway-specific gating ablations, a frozen-plasticity experiment, and a post-adaptation direct-pathway cutout. These experiments demonstrate that MoM accelerates immediate behavioral correction without accelerating weight-based learning, while the learned compensation is retained even after the direct corrective pathway is disabled.

  • Corrected and clarified the experimental methods (Section 3.2.1, Table 1, Appendix A.7). We corrected the description of the visuomotor perturbation, clarifying that the rotation is applied to visual feedback while physical arm kinematics remain unchanged. We also clarified the random fixed-gate ablation and ensured numerical consistency across the reported comparisons.

Overall, the revised manuscript provides a more precise and experimentally supported account of MoM's central contribution: geometrically routing local sensory-error-to-action transformations enables rapid behavioral correction, while slower plasticity supports subsequent internalization. The additional baselines and interventions help isolate these mechanisms and distinguish online correction from longer-term learning.

We hope these revisions address the reviewers' concerns and substantially improve the clarity, rigor, and scope of the manuscript. We thank the reviewers again for their valuable feedback and the opportunity to strengthen our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.