Tonic feedback motor commands predict visuomotor learning

Article activity feed

1. 

   eLife

   Jun 29, 2026

   eLife Assessment

   This valuable study rigorously examines how motor learning is influenced by the feedback response to a previous movement error. Using a series of well-conducted experiments, the authors provide solid evidence that the learning response following a cursor jump does not depend on the timing of the perturbation and is influenced by the tonic component of the feedback responses. Further work is needed to determine whether this generalizes to other perturbation paradigms and to more fully understand the relationship between learning and the tonic and phasic components of the feedback response.

   Read the original source
2. 

   eLife

   Jun 29, 2026

   Reviewer #1 (Public review):

   Summary:

   The authors investigate the relationship between feedback responses and trial-to-trial learning. In their paradigm, participants were constrained to a channel trial, and a cursor was visually perturbed. Using a channel-perturbation-channel structure, the authors obtain feedback responses to the perturbation and the learning response that ensues. In Experiment 1, the authors demonstrate that temporal dynamics of the learning response (LR) are poorly linked to temporal dynamics of the feedback response (FBR). The LR responses are yoked to the start of the movement, even in cases where the FBR is very delayed. Then, in Experiments 2 and 3, the authors dissect FBR and LR responses into two components: (1) a phasic component that has a peak point mid-movement and then declines, and (2) a tonic component …

   Reviewer #1 (Public review):

   Summary:

   The authors investigate the relationship between feedback responses and trial-to-trial learning. In their paradigm, participants were constrained to a channel trial, and a cursor was visually perturbed. Using a channel-perturbation-channel structure, the authors obtain feedback responses to the perturbation and the learning response that ensues. In Experiment 1, the authors demonstrate that temporal dynamics of the learning response (LR) are poorly linked to temporal dynamics of the feedback response (FBR). The LR responses are yoked to the start of the movement, even in cases where the FBR is very delayed. Then, in Experiments 2 and 3, the authors dissect FBR and LR responses into two components: (1) a phasic component that has a peak point mid-movement and then declines, and (2) a tonic component that grows over the movement time course and remains stable during the holding period. The authors provide evidence that LR responses are better predicted from the tonic component of the FBR than the phasic component. The idea that tonic FBR components drive learning over phasic components departs from prior models of error-based learning and provides a new theory to understand sensorimotor adaptation.

   Strengths:

   (1) The paper is well-written, and the contribution is important and timely. The authors provide clear experiments that change the way we conceptualize how trial-to-trial learning is driven by feedback responses to error.

   (2) The paper provides solid evidence to demonstrate that feedback (FBR) and learning (LR) responses are not linked by a fixed delay, in contrast to prior models.

   (3) The paper also introduces the concept that both tonic and phasic components of the FBR differentially influence the learning response. The paper provides solid evidence that the tonic forces maintained during holding still have an impact on the learning that proceeds on the next trial. This has implications for models of sensorimotor adaptation and our understanding of the physiology of learning.

   Weaknesses:

   While some conclusions are strong, I feel that the conclusions regarding FBR and LR relationships need additional analysis. All these concerns are elaborated below. Broadly speaking, there is a concern that some conclusions reached by the authors are linked to the particular phasic/tonic model they use to parse FBR and LR responses. Other models are not considered and could lead to differing results. Furthermore, it is assumed that LRs are scaled FBRs. This assumption excludes the possibility that LRs could be driven by FBRs and other mechanisms, which would alter the way the regression analyses are constructed. As described below, model-free analyses are warranted to corroborate the main findings. Further, the role that phasic-FBR plays in the adaptation process is understated in the Discussion despite evidence to the contrary in Figure 8. Much of the analysis is done on trial-averaged and participant-averaged responses, inflating R2 values. More analysis should be done at the trial level to better examine model performance and accuracy. And while valuable, the authors' experimental approach differs from standard force-field experiments that were initially used to test feedback error learning hypotheses. The paper could benefit from a Limitations section to discuss associated limitations.

   Main Concern 1:

   The decomposition of FBR and LR into phasic/tonic components is based on a specific model (i.e., Equation (1)). The notion that tonic FBR predicts phasic/tonic LR is based on responses estimated from the model. Thus, it is unclear whether critical findings (e.g., LR responses are predicted by tonic FBR) are true of the "data" or true when the "data are analyzed in the context of their model". In other words, had the authors proposed a different model to decompose the LR/FBR into tonic/phasic components, would they obtain different results?

   There are many possible alternatives:

   (A) In Equation (1), the phasic and tonic components are assumed to add linearly at all times to obtain the force profile. But the phasic and tonic components could be applied at separate times. The tonic component could be invoked during holding, and the phasic component could be invoked during moving. This type of model will differ from the current version, especially in how the peak force during the moving period is assigned to the phasic/tonic components.

   (B) Another possibility is that the tonic and phasic components do indeed operate at the same time (like in Equation (1)), but they are separate, independent controllers. In the author's model, the tonic component is dependent on the phasic component.

   (C) Another possibility is that the tonic and phasic components are linked, but not by an integral.

   (D) Another possibility is that the phasic component is not a Gaussian function of time.

   Concern 1-1:

   While it is not possible to explore the entire model space described above, the authors should consider whether other phasic/tonic model classes could lead to qualitatively different results. The authors could also consider other phasic/tonic models if appropriate, and demonstrate that Equation (1) is superior based on an information criterion like AIC or BIC.

   Concern 1-2:

   I recommend that the authors pursue model-free, empirical analyses to support their findings. This would decrease the reliance on the "correctness" of a particular model. One logical choice would seem to be empirically estimating the phasic component as the peak force during the moving period and the tonic component as the average force during the holding period. In this model-free estimation of phasic and tonic commands, is it still the case that tonic FBR alone predicts LR components?

   Concern 1-3:

   Building on Concern 1-2, a clear case where the concern about using a model alone to estimate phasic and tonic components is in the across-subject variability analysis in Figure 7. Here, LR and FBR are compared to one another only in the context of the tonic-phasic model in Experiment 1. The result is that only the tonic FBR predicts the tonic LR. But investigating Figures 7b and 7c, it would appear that the peak force applied during the FBR during the moving period (which should reflect the phasic component in large part as in Figure 4a) would predict the peak (or average) force applied during the LR. Thus, the conclusion that tonic FBR only predicts tonic LR may be driven by how the model estimates tonic/phasic FBR/LR rather than a true property of the data. A model-free analysis, as suggested in Concern 1-2, would be helpful in addressing this concern.

   Main Concern 2:

   Analyses in Figures 4g, 4h, 6c, and 6d are based on relating LR and FBR components with no intercept: y = ax; the LR component is a scaled FBR component. It is unclear if the authors' conclusion would vary had a different model been used. For example, suppose that LR on trial n is partly determined by the FBR and also the sensory error (e) on trial n-1 (where c1 and c2 are constants):
   LR(n) = c1 FBR(n-1) + c2 e(n-1)

   Another model could suppose that the LR on trial n is due to the FBR on trial n-1, and also a non-specific adaptive component that is independent of both FBR and the sensory error:
   LR(n) = c1 FBR(n-1) + c2

   Concern 2-1:

   For these alternate models, y=ax (i.e., zero intercept) is not an appropriate relationship between LR and FBR components. Had the authors allowed a non-zero intercept in Figs. 4g, 4h, 6c, and 6d, will they still observe that only tonic FBR predicts LR components? In other words, would R2 improve for phasic FBR relationships with a non-zero intercept?

   Concern 2-2:

   Why was a non-zero intercept allowed for the between-subject analyses in Figure 7, but not for similar analyses in Figures 4 and 6?

   Main Concern 3:

   The main results in Figures 4g, 4h, 6c, and 6d are based on an R2 value that is calculated on a linear fit to the mean response averaged across participants and trials. This raises the concern that the R2 value is being inflated, and it also misses the rich trial-to-trial variation and subject-to-subject variation that could be used to examine the model's accuracy. A couple of concerns here:

   Concern 3-1:

   As can be seen from the horizontal and vertical error bars in Figures 4g and 4h, there is considerable variability across participants. While not shown, it is almost certainly the case that there is considerable variability across trials within a participant (as alluded to in the Fig. 8 analyses). The authors should evaluate their model performance and report goodness-of-fit (or error) at the single-trial level. For example, the model could be fit to individual trial data, and the R2 values from the trial fits could be used for comparing the various relationships in Figures 4 and 6. Another idea would be to keep the alpha, beta, T and sigma estimates obtained from the average data, and then apply these parameters to individual trial responses and report the model error. Do phasic FBR commands similarly predict LR components at the trial level, or do trial-level analyses corroborate the current conclusions on tonic FBR superiority?

   Concern 3-2:

   The authors report on Line 200 that the R2 values of 0.635 and 0.698 have modest predictive power. It would be helpful for the authors to statistically compare the R2 values between Figures 4g and 4h. One idea would be to obtain an R2 value for each individual participant. Then the distribution of R2 values across participants could be compared between the different relationships in Figure 4g/4h (e.g., via a t-test). This would help to better support the idea that Figure 4h shows better model fits than Figure 4g. These analyses could also be conducted for the relevant parts of Figure 6 (Experiment 3). The authors should consider allow a y-intercept in this process as they do in Figure 7.

   Main Concern 4:

   The authors compare tonic and phasic FBR predictive power in Figure 4. There are other places where the analyses in Figures 4g and 4h should be repeated:

   Concern 4-1:

   Tonic and phases FBR responses appear to vary in Experiment 1 (Figure 2c), but the authors do not test whether they predict the LR component magnitudes in Figure 2d. Analyses in Figures 4e,4f, 4g, and 4h should be added to the Experiment 1 analysis.

   Concern 4-2:

   While I understand the rationale behind computing differences in Figure 6 to isolate the second-shift effect on FBR/LR, the authors should still perform the primary investigation in Figures 4e, 4f, 4g, and 4h on the FBR and LR responses in Figures 5b-g (without subtracting the "Maintained" component). In other words, before analyzing the contributions of the second shift in Figure 6, the authors should repeat their analysis in Figure 4 applied to the FBR and LR responses in Figure 5 (without subtracting off the maintained response). How well does Equation (1) and y=ax capture the FBR and LR responses in Figures 5b-g?

   Main Concern 5:

   Given current practices in human sensorimotor adaptation, the current n=10 (or n=12) group sizes appear limited in size, raising concerns on statistical power.

   Concern 5-1:

   The authors should consider a power analysis or provide some other justification to support their chosen sample sizes.

   Concern 5-2:

   It is unclear why cross-correlation analyses in Figure 2e, 3d, and 5h have error bars, but no other FBR or LR time courses have error bars. Error bars should be provided in Figures 2b, 2c, 2d, 3b, 3c, 5b, 5c, 5d, 5e, 5f, 5g, 6a, and 6b.

   Concern 5-3:

   The subject counts are reported as n=10 for Experiment 1, n=12 for Experiment 2, and n=12 for Experiment 13, but the subject-to-subject analysis in Figure 7 says n=33.

   Main Concern 6:

   I agree that the author's model suggests that LR responses are most strongly predicted by the tonic FBR component. But I feel the narrative and Discussion surrounding this point are too strong. They paint the picture that only tonic FBR is important in learning. To do this, the role that phasic FBR plays is discounted, and mixed results concerning tonic FBR are overlooked. I feel that the Discussion should be broadened to acknowledge that the authors find evidence that both tonic and phasic FBR appear to influence the learning response, with tonic FBR making the stronger contribution in this task. Here are key areas that require attention:

   Concern 6-1:

   Importantly, the authors downplay their result in Fig. 8h, that the phasic FBR predicts phasic LR in their Results on Line 350. This argues against the idea that only tonic FBR influence LR parameters. On Line 485, the authors state that "trial-by-trial variability in LR amplitude was explained by the tonic component of the FBR, but not by the phasic component (Fig. 8)." This is not correct. Both the tonic and phasic components of the FBR altered LR components in Figure 8.

   Concern 6-2:

   Again, it is stated on Line 502, that the phasic FBR component "had only a modest effect on the LR". This again seems to underplay the result. The authors should amend their Results and Discussion to better acknowledge that their data support a role for both tonic and phasic FBR contributions to LR, but the tonic component appears to make a larger contribution in their model.

   Concern 6-3:

   While the role of phasic FBR in determining LR amplitude appears to be understated, the role of tonic FBR is, on occasion, overstated. The Discussion should mention that there is mixed evidence for the role of tonic FBR in LR parameters. For example, in their between-subjects analysis in Figure 7f, the authors do not find that phasic LR can be predicted by tonic FBR. Thus, across subjects, no component of the FBR appears to predict phasic LR.

   Concern 6-4:

   To better investigate the role that both phasic FBR and tonic FBR may play in adaptation, it would be advisable for the authors to consider this hypothesis. As it stands, tonic LR or phasic LR is regressed only onto tonic FBR or phasic FBR individually. In Figures 1 (Experiment 1), 3 (Experiment 2), and 5 (Experiment 3), the authors could regress tonic LR and phasic LR onto both phasic FBR and tonic FBR simultaneously. Models where LR = c1 phasic-FBR + c2 tonic-FBR could be considered and compared against univariate models, LR = c phasic-FBR and LR = c tonic-FBR using AIC or BIC to determine whether a mixed model that predicts LR with both phasic and tonic FBR is warranted.

   Irrespective of the result, the authors should be careful (Concerns 6-1 and 6-2) to state that when levels of tonic-FBR were controlled in Figure 8 (which is likely the cleanest way to look at the role phasic FBR plays in learning), phasic-FBR showed a clear influence on LR.

   Major Concern 7:

   On Line 577, it states the "hand was automatically returned to the starting position". Does this mean that the robot moved the hand back to the start location? If so, was the hand ever released from a force channel in between the perturbation trial and the following channel trial? A concern is that the holding forces from the perturbation trial could "bleed over" into the forces applied during the subsequent channel trial if the subject always remains in a channel trial in between the trials. Suppose we label the 3-trial structure as Channel 1 (C1) - Perturbation (P) - Channel 2 (C2). The authors should confirm that the holding forces on P are not correlated with baseline force (i.e., the channel force prior to movement onset) in C2. I do not expect there to be a strong correlation given that the learning responses in Figs. 2d, 3c, and 5e-g appear near-zero at t=-400ms, but this should still be verified.

   Major Concern 8:

   In Supplementary Figure 1, there appears to be an error in the "Amplitude of phasic LR (N)". In Supplementary Figure 1f, the phasic LR magnitudes appear in line with Supplementary Figure 1d, but there is a mismatch in the magnitudes for the phasic LR in Supplementary Figures 1e & 1d (the phasic LR magnitudes appear to be too low in Supplementary Figure 1e, peaking at around 0.1N when they should peak at around 0.15N).

   Major Concern 9:

   The authors should provide a Limitations section, highlighting unanswered concerns listed above, mixed results, and differences from prior work. These are touched upon in the Discussion section (particularly in Perspectives for future studies) but should be expanded further. At a minimum, the authors should consider including a discussion of the following points:

   Differences from prior work:

   9-1: There are methodological differences between this work and past studies highlighted by the authors. It could be that there are multiple error-based learning mechanisms that drive the FBR. Here, the authors find that visually-driven FBR responses do not drive LRs at a "common temporal shift". Instead, LRs are broadly expressed at the start of the movement (regardless of when the FBR was timed). However, tasks that have other components (e.g., a proprioceptive error) might invoke different learning mechanisms. For example, proprioceptive-driven FBRs might invoke LRs that have different temporal properties than visually-driven FRBs.

   9-2: As noted by the authors, Reference [10] studied FBR-driven learning in muscle commands, as opposed to forces. Muscle responses may have differing temporal and/or magnitude (for phasic/tonic) components that qualitatively differ from the force-based conclusions made here. Thus, the learning mechanisms at the muscle level may differ from those observed at the force level.

   9-3: While the tonic FBR is a strong predictor of the learning response in this experiment, most of the experimental conditions are done where the cursor remains deviated from the target throughout the trajectory and into the holding period. This differs from past work on feedback error learning, where feedback was veridical, and the cursor (and hand) ended on the target. This persistent displacement from the target during the prolonged holding period may influence the learning process and could enhance the tonic-FBR contribution to learning.

   9-4: The authors state in the present study that subjects were told not to use "explicit strategies" and move as straight as possible to the target. For past work, participants were able to use explicit strategies during feedback and learning responses. It could be that the lack of (or reduction in) explicit responses alters single-trial learning mechanisms relative to past work.

   Alternate models:

   9-5: No alternate models are considered here for the tonic-phasic relationship. Other models could relate these two processes differently, which could lead to different conclusions.

   9-6: It is assumed that both the tonic and phasic controllers are active at the same moment in time and sum linearly to generate the overall force output. Other models could have applied each "controller" to different phases of the reach in a differential manner (e.g., two separate controllers, a moving controller and a holding controller operating at different moments in time).

   9-7: It is assumed here that the LR should be a scaled FBR: y = ax. Conclusions made here could change if the LR is due to multiple processes, FBR-driven learning only being one of them. Other models where the LR is driven by both FBR and the sensory error were not considered here.

   Mixed results:

   9-8: While tonic FBR was a good predictor of phasic LR at the group-level (e.g., 4g), it did not predict phasic LR between subjects (Fig. 7f) and in fact tended toward a negative relationship.

   9-9: Phasic FBR predicts Phasic LR at the trial-level (Figure 8h) but not as well at the subject-level (Figure 7d).

   9-10: Overall, with the exception of Figure 8, most analyses look at the relationship between LR and tonic FBR or phasic FBR separately. In Figures 4c, 4d, 6c, 6d, and 7d-g, the authors look at the marginal effect of tonic or phasic FBR on learning, but do not control for variations in the other FBR component (e.g., they look at phasic FBR on tonic LR, but do not control for tonic FR). The only analysis that controls for the other component is in Figure 8, suggesting that both tonic and phasic FBR contribute to LR.

   Minor concerns

   (10) I'm not sure I follow the cross-correlation analysis in Figure 3. Overall, to me, both the FBR in Figure 3b and the LR in Figure 3c look quite similar in their temporal profiles, irrespective of the shift magnitude. The authors state on Line 158 that their cross-correlation analysis "...revealed that the overall shape of the cross-correlation function changed systematically with error magnitude". However, to me, in Figure 3d, the shape of the many curves looks similar.

   What is confusing to me here is including a phasic movement period and a tonic holding period inside the cross-correlation. The tonic "static" component during the holding period will likely greatly influence how well the cross-correlation is able to match the phasic peaks during the LR/FBR moving periods. In other words, the reach consists of a "movement" and a "holding" period. But the cross-correlation is blending the two together, and thus, I am not sure how reliable this measure will be for truly estimating the temporal shift between conditions. For example, if you look at the shaded gray area in Figure 3b, the "Movement period" looks almost identical in temporal properties. The "peaks" and "troughs" happen at nearly the same moment in time across all conditions. The onset of the FBR at approximately 200 ms is also identical across shift magnitudes. Thus, to me, the temporal properties of the FBR seem very similar during the moving period (where the FBR is responding to the error). But including the holding force (the tonic force after the 600ms period) seems to be causing the cross-correlation function to estimate differences at very high lags. If these differences are being driven solely by the holding forces, I am not sure this is meaningful.

   It seems that the authors might want to repeat this analysis, excluding the holding force period from the calculation of the cross-correlation coefficients.

   (11) It would appear that the authors have a significant main effect of their ANOVA (p=0.028) in Fig. 3f, but no post-hoc tests are reported to indicate which group means differ.

   (12) When plotting FBR, a [0,600]ms period is shaded as the movement period. On Line 580, it says that feedback was provided on peak movement speed. Was any feedback provided as to the movement duration? If not, did participants complete the movement within the 600 ms window labeled as movement speed? Were movements during perturbation trials longer than non-perturbed trials?

   (13) Over what time period is Equation (1) fit to the data? Is it the [-200,700]ms window shown in Figure 4a? A concern is that including too much of the "holding period" in the model fit will cause the model to be biased toward fitting the holding period well and not the moving period. This, in turn, might lead to better estimates for the beta parameter than the alpha parameter. In addition to clarifying the fitting process, the authors should also include R2 values for the moving and holding periods separately.

   (14) The procedure is clear from Figure 1e, but it would be helpful on Line 91 to explain that "collapsing" FBR and LR across rightward and leftward means that the FBR and LR were negated for one of the directions (prior to collapsing).

   (15) Are the "Amplitude of tonic LR (N)" supposed to be negative in Figures 6c and 6d?

   (16) Overall, the parameter distributions in Figures 4e and 4f are similar to those in Supplementary Figures 1c and 1d. The FBR amplitudes look nearly identical. Only the Phasic LR amplitudes in Supplementary Figure 1d appear to be larger than the Phasic LR amplitudes in Figure 4f. Can the authors provide an intuition for why the phasic LR contributions increase when T and sigma parameters are allowed to vary between participants?

   (17) There are two points where the authors should consider softening their language:

   17-1: The authors state at multiple points (e.g., Line 154) that "...the waveforms of LRs remained largely similar across conditions, while their amplitudes showed only modest modulation with cursor shift magnitude". However, in Figure 3c, the LR amplitude for the 0.4 cm shift is approximately 0.2 N, and the LR amplitude for the 3 cm shift is approximately 0.3 N - a 50% increase. The authors should consider softening the language here to appreciate the variations in LR amplitude.

   17-2: On Line 258, it is stated that the FBR during holding "diverged only slightly" for the 16 cm condition in Fig. 5b. This seems too strong a statement. The "Maintained" FBR holding force is about 0.2 N, and the reverse is about 0.1 N. Thus, the "Maintained" condition is doubled. While I agree that the LR diverges more than the FBR (i.e., 5b vs. 5e), I think the language choice here should be more careful.

   Read the original source
3. 

   eLife

   Jun 29, 2026

   Reviewer #2 (Public review):

   Summary:

   The authors find a strong trial-level relationship between tonic feedback responses and tonic learned responses.

   Strengths:

   The authors have performed several well-conducted experiments and thoughtful analyses to test the relationship between feedback responses and subsequent learned responses. The strength of the paper is the experimental control to probe this relationship and, eventually, oppugn the feedback error learning hypothesis.

   Weaknesses:

   In general, the processes studied in this manuscript and the past work have not explained the underlying mechanisms for the observed phenomena. Without knowing the mechanisms, the results are largely observational/correlational when linking feedback responses to learned responses, and there are no strong alternative hypotheses to explain the results. Most …

   Reviewer #2 (Public review):

   Summary:

   The authors find a strong trial-level relationship between tonic feedback responses and tonic learned responses.

   Strengths:

   The authors have performed several well-conducted experiments and thoughtful analyses to test the relationship between feedback responses and subsequent learned responses. The strength of the paper is the experimental control to probe this relationship and, eventually, oppugn the feedback error learning hypothesis.

   Weaknesses:

   In general, the processes studied in this manuscript and the past work have not explained the underlying mechanisms for the observed phenomena. Without knowing the mechanisms, the results are largely observational/correlational when linking feedback responses to learned responses, and there are no strong alternative hypotheses to explain the results. Most of the larger comments below stem from this theme, including:
   (i) what causes the phasic and tonic portions of the feedback response,
   (ii) justifying the phasic learned response,
   (iii) what are some alternative hypotheses that can explain the current results and past literature?

   Suggestions to improve the paper are below.

   (1) As mentioned above, it appears that there is limited mechanistic understanding of the underlying processes. For the feedback response, there is clearly a phasic and tonic component. It is not until one gets to the discussion that a potential mechanism is proposed, where presumably the phasic response may be velocity dependent, and the tonic response may be position dependent. On a somewhat related tangent, these responses somewhat mirror muscle spindles, which are known to have velocity and position-dependent responses, leading to the phasic and tonic firing during muscle stretch experiments.
   a) Can the authors provide more discussion on the work that they currently cite, which studied position and velocity dependent responses?
   b) Relatedly, did the authors put any thought into developing a model, using error inputs from the experimental trials, that can capture the feedback responses? For example, dF/dt * tau = a*pe + b*ve - cF + e, where F = force response, tau is a time constant to generate the force, a is a gain on position error (pe), b is a gain on velocity error (ve), c relates to the leak, and e is Gaussian noise. The leak would be needed to explain the equilibrium / steady state at the end of the trial. It could be very insightful if this, or some other similar flavour of model, could explain the phasic and tonic components of the feedback response. The advantage of a model in this form is that there are experimental inputs and the process evolves over time, rather than fitting static curves to the data.

   (2) Aligned with past literature, the authors have characterized the early and late phases of both the feedback responses and learned responses as phasic and tonic. It is clear from the data that the feedback response data are composed of a phasic and tonic phase. However, it is less clear from the data in many of the figures that there is an actual phasic response in the learned response. Further, from a modelling perspective, it is conceivable that the fitting algorithm would partition the variance between the two components of equation 1, even though there may only be one true underlying process. This may also explain why there was no correlation between tonic feedback responses and phasic learned responses in Figure 7F.
   a) Can the authors provide more rationale on why the learned response would also have a phasic response? Is the assumption here that since the feedback response had a phasic response, the learned response should as well?
   b) Can the authors fit the learned response with only the tonic portion of the equation? Then, perform model comparison between the phasic+tonic learned response model and the tonic only learned response model using AIC/BIC, to justify whether or not a phasic portion of the model is needed to explain the data.
   c) Can the authors comment on the possibility that the learned response may just rise and then decay over time, without being the outcome of two distinct processes?

   (3) The nicely controlled experiments do well to provide evidence against the feedback error learning hypothesis, which alone is a valuable contribution to the literature. However, the authors do not provide a strong alternative hypothesis. There is a proposal of alternative hypotheses. For example, on lines 494-498, referring to state estimation, which the authors then state could not explain all the results in the preceding paragraph. It would be beneficial to further bolster the possible explanations. Perhaps further discussion details on what the mechanisms are for the feedback responses (e.g., position or velocity dependence), and what states (position error, velocity error, motor commands, etc.) transfer into the learned response. Are they stored? Are they the outcome of a continuous process? This may be difficult given the current state of understanding in the literature, but it could substantially improve the paper.

   Read the original source
4. 

   eLife

   Jun 29, 2026

   Reviewer #3 (Public review):

   I believe that the paper is excellent and very well executed. I have several reservations about the meaning of the tonic component of the feedback responses and about the more general interpretation from a computational standpoint. These aspects may not require extensive adjustments, but some key points could be discussed or better justified:

   (1) It is true that most papers view adaptation as a trial-by-trial update and that several models summarise motor errors by a scalar quantity for a model fit. The importance of feedback control in visuomotor control has also been overlooked, as several studies explicitly instructed not to correct. I also agree about the fact that the temporal aspects of sensory encoding and control are often neglected in motor adaptation studies. However, there have been some …

   Reviewer #3 (Public review):

   I believe that the paper is excellent and very well executed. I have several reservations about the meaning of the tonic component of the feedback responses and about the more general interpretation from a computational standpoint. These aspects may not require extensive adjustments, but some key points could be discussed or better justified:

   (1) It is true that most papers view adaptation as a trial-by-trial update and that several models summarise motor errors by a scalar quantity for a model fit. The importance of feedback control in visuomotor control has also been overlooked, as several studies explicitly instructed not to correct. I also agree about the fact that the temporal aspects of sensory encoding and control are often neglected in motor adaptation studies. However, there have been some developments about adaptive control in the context of force field learning to express the error signal and learning rule based on continuously evolving state variables as those formulated in online control models (Crevecoeur et al., 2020, eNeuro 7(1); Kalidindi and Crevecoeur, 2023, Curr Opin Neurobiol, 83, 102810). Could the authors consider discussing whether this framework could or not be consistent with the current dataset?

   (2) The choice of a cursor jump may require more in-depth justification. From an experimental standpoint, it is clear from the authors' data that a cursor jump does evoke an aftereffect and hence the developments are clearly validated empirically. The nature of the adaptive response is less clear: indeed, cursor jumps can be represented as an external perturbation to a variable that may be independent of the hand (e.g. Kasuga et al., 2022, J Neurophysiol, 127 (2), 354-372). In contrast, a visuomotor rotation requires a change in state space representation parameters (it is not clear which ones) that is more closely related to the update of an internal model. Could the authors explain why they believe that a learning response to a cursor jump is consistent with adaptation in general?

   (3) The relationship between the tonic component of the feedback response and the learning response is very clear from an experimental perspective again. However, I would suggest being very cautious about the interpretation of this effect. My concern is that it is not clear that this tonic response is irrelevant from a behavioural standpoint, and I am left wondering what the correlation with the learning response truly means. Indeed, in real-life conditions, there should be no net force produced in the end during a static phase, as the force during stabilisation is by definition zero; only the net force produced against constant external loads is required. There can be co-contraction but not net resultant force, unless external forces are applied. So if the tonic response vanishes in real conditions, should there be no learning response? This aspect is also relevant if one attempts to generalise the findings to force field learning: since velocity-dependent force fields vanish during stabilisation, how can there be a tonic component?

   Read the original source
5. 

   eLife

   Jun 29, 2026

   Author response:

   We thank the reviewers for their thoughtful and constructive comments, and we plan to implement many of their suggestions to improve the paper. We agree that the manuscript would benefit from a clearer and more evidence-based presentation of how feedback responses relate to subsequent learning responses. To address this point, we will perform additional analyses and modeling, including model-free analyses of the phasic and tonic components. These analyses will allow us to test whether the tonic component remains the dominant predictor of the learning response without relying on the specific assumptions of the tonic/phasic decomposition model.

   We also agree that the manuscript would benefit from a more detailed discussion of the mechanisms that may shape the temporal evolution of feedback responses and their relationship …

   Author response:

   We thank the reviewers for their thoughtful and constructive comments, and we plan to implement many of their suggestions to improve the paper. We agree that the manuscript would benefit from a clearer and more evidence-based presentation of how feedback responses relate to subsequent learning responses. To address this point, we will perform additional analyses and modeling, including model-free analyses of the phasic and tonic components. These analyses will allow us to test whether the tonic component remains the dominant predictor of the learning response without relying on the specific assumptions of the tonic/phasic decomposition model.

   We also agree that the manuscript would benefit from a more detailed discussion of the mechanisms that may shape the temporal evolution of feedback responses and their relationship to subsequent learning. We will therefore expand the discussion of this issue and relate our findings to adaptive feedback control and continuous-time models of motor adaptation, which may provide useful frameworks for interpreting the relationship between feedback responses and learning responses.

   Finally, we agree that the scope and limitations of the current experimental paradigm should be discussed more explicitly when considering the generality of our findings. We will therefore discuss whether and how the present results may generalize to broader forms of sensorimotor learning and adaptation. We will also

   Read the original source
6. 

   Version published to 10.64898/2026.01.30.702978 on bioRxiv

   Feb 2, 2026