acceptodds
Under review as a conference paper at ICLR 2027

One Circuit, Many Flow Fields: Mechanistic Models of Single-Trial Neural Dynamics

Abstract

A single neural circuit can exhibit qualitatively different dynamics across trials: one circuit, many flow fields. Standard models treat this trial-to-trial variability as noise around a fixed dynamical system or as discrete switches between regimes, yet neither captures how continuous internal-state variables, such as arousal or engagement, can gradually deform the circuit's flow field. We propose that single-trial fitting can be reframed as inferring the low-dimensional control parameters that reshape a shared circuit's flow field. We realize this with a low-rank recurrent network in which trial-specific static input biases act as bifurcation parameters: constant within a trial, they deform the flow field without directly driving activity over time. We validate the model in a teacher-student setting, where it recovers the ground-truth dynamics and bifurcation structure from activity alone. Applied to mouse and monkey motor cortex, the model infers control spaces with interpretable axes that track behavioral engagement in mouse and task structure in monkey. A generative extension reproduces the structure of trial-to-trial variability and partially transfers across sessions and subjects. Together, these results reframe a methodological problem of fitting single-trial activity as a scientific opportunity: reading off the control parameters of the underlying dynamics, and connecting data-driven inference of neural dynamics to mechanistic theories of how a single circuit reuses its dynamics for flexible behavior.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Computational neuroscience, Mechanistic models, Low-rank RNNs, Trial-to-trial variability, Dynamical systems, Bifurcations, Neural population dynamics

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_EQ58OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes that trial-to-trial variability in neural population activity reflects a single circuit whose flow field is continuously deformed by low-dimensional control parameters. The authors performed a synthetic validation, applied the model to mouse and monkey motor cortex recordings, and proposed a generative extension. They show that the model recovers ground-truth dynamics on synthetic data, identifies interpretable control axes in bias space on real data, and partially transfers across sessions and subjects.

Critical strengths

The paper is well motivated. The idea is clear and can be analyzed with existing low-rank theory. This makes the model more testable than methods that modulate weights.

The paper evaluates the model across different datasets and species. The transfer analysis includes a random control, and the reach-direction plane is tested on held-out angles rather than only on training trials.

Critical weaknesses
  1. “many flow fields” is only demonstrated on simulated data. On real data, there is no fixed-point or bifurcation analysis, so we can’t tell whether the disengaged regime is qualitatively different. A fixed-point analysis along the disengagement axis may help this.
  2. Although the paper compares reconstruction and decoding performance against other baselines, it does not compare alternative explanations of trial-to-trial variability in the mouse disengagement analysis, such as LFADS style models. Can these models reproduce the disengagement separation equally well?
  3. Table 2 lists only one session for the disengagement analysis. Replication across sessions would strengthen the biological conclusion.
  4. The cross-session transfer does not establish that the full latent dynamics are shared because part of the model is still trained. A stronger test may be needed.

Minor:

  1. In your figure 5 panel e, the text ‘remaining parameters’ on the figure has a typo.
  2. The cosine similarity is reported without a baseline.
Questions for authors

Please see weaknesses above.

3 comments
AuthorsOpenReview ↗

Response to Reviewer EQ58

We thank the reviewer for acknowledging the utility of our method and for the constructive suggestions. The first led to what we consider the most important addition to the paper.

“Many flow fields" is only demonstrated on simulated data.

We performed the fixed-point analysis the reviewer suggested (Sec. 3.3.1). It shows that the disengaged regime of the fitted network is qualitatively different: along the disengagement axis, the choice attractors disappear through saddle-node bifurcations and the network can no longer hold a choice. Please see the general response for details.

Alternative explanations of trial-to-trial variability.

To address alternative explanations of trial-to-trial variability, we added a comparison with models in which trials differ in their initial state or in stochastic inputs (App. D.9). It is described in the general response and in our response to Reviewer FvK9 (point 2). On the specific question of whether other models reproduce the disengagement separation: an expressive model such as LFADS may well do so, but this is not what distinguishes our model. What distinguishes it is its mechanistic account. Because its units are the recorded neurons and the bias is an input current to them, a control axis inferred by our model is a concrete input pattern over identified neurons. For the disengagement axis, this pattern is predicted to switch the circuit between engaged and disengaged regimes (Sec. 3.3.1), a prediction that can be tested in vivo with targeted photostimulation.

Replication across sessions.

We have extended the disengagement analysis beyond a single session. In the new session-stitched model, all 72 sessions share one recurrent connectivity and one bias space (App. D.7), so disengagement axes from different recordings can be compared directly. An axis estimated from other sessions carries over to a session it was not estimated from (see the general response). The single-session analysis of Sec. 3.3 also gives consistent results in additional sessions, which we will include in the appendix of the final version.

Cross-session transfer.

The embedding U maps each session's recorded neurons onto the latent space. Because different sessions record different neurons, U must be learned separately for each one. We therefore transferred everything that can be shared: the latent dynamics and the distribution of biases, and compared the transfer against matched random models. We agree that, with U retrained, the transfer alone does not establish that the dynamics are shared. However, its success motivated a more demanding test, the session-stitched model (App. D.7). In that model, the latent dynamics and the bias space are shared by all 72 sessions from the start, so a single network must account for all of them. We show that it does, and a control axis estimated from some sessions carries over to a session that was left out (see the general response).

Minor points.

We thank the reviewer for catching these.

  1. The typo in Fig. 6e (Fig. 5e in the original submission) is fixed.
  2. We added a label-shuffled baseline for the cosine similarity between the disengagement axes (Sec. 3.3): 0.53 against 0.00 ± 0.24 (p = 0.01).

I thank all authors for answering my questions, I do not have more questions.

AuthorsOpenReview ↗

Thank you for the discussion

We thank the reviewer for the constructive discussion, and are glad that our responses addressed the reviewer's questions. These suggestions led to what we consider the most important additions to the paper: the bifurcation analysis of the ALM network and the cross-session validation. We would be grateful if the overall assessment could be updated accordingly.

Reviewer_FvK9OpenReview ↗
Rating
4
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Trial-to-trial variability in neural activity is often modeled by different initial conditions of the network, noise in the dynamics, and external inputs, while the flow field the underlying recurrent dynamics is fixed. Models that consider changes in the flow field usually assume a small number of regimes selected by a discrete variable. However, continuously-varying quantities, such as arousal or engagement, would be better represented by continuous, low-dimensional variables. This paper proposes to augment a low-rank recurrent network with trial-specific, low-dimensional input biases that operate within the same dynamically-relevant subspace. While this idea is not new (e.g. Shah et al. 2025), the authors use trial-specific biases in a novel way to capture continuously-varying quantities which modulate dynamics. They validate their model on synthetic data and use it to understand single-trial activity in two different motor cortex datasets. In a mouse ALM dataset, the trial-specific biases corresponded with animal engagement and identified a subspace in which they could interpolate between engaged and disengaged states in the model. On monkey motor and premotor cortex data, the identified subspace of biases corresponded to the target angle in a center-out task. Importantly, both of these results were obtained without supplying any labels of the captured variable.

Critical strengths

The approach is novel for the question being asked, and the results on real data suggest that the model can discover meaningful perturbations to the dynamics without explicit labels.

Critical weaknesses
  1. The motivation for the RNN being explicitly low-rank, and for the biases to be constrained within this subspace, needs to be better articulated.
  2. This approach accounts for trial-specific changes in state that are dynamically meaningful, effectively finding external inputs that operate on slower timescales than the task itself. However, it does not account for other quantities known to influence dynamics, such as trial-specific initial state and time-varying external inputs from unobserved populations that may influence dynamics on a faster timescale (e.g. in LFADS).
Questions for authors
  1. What advantages does assuming low-rank structure provide here, and would this structure not potentially emerge in an unconstrained task-trained model? Is there any situation where a bias that operates (partially) outside the recurrent subspace does something dynamically interesting, or would the outside component ultimately get attenuated?
  2. How would time-varying external inputs and trial-specific initial states alter the utility of this approach? Is there an identifiability concern here?
1 comment
AuthorsOpenReview ↗

Response to Reviewer FvK9

We thank the reviewer for the positive assessment and for recognizing the novelty of our approach.

Motivation for the low-rank constraint and for biases within the recurrent subspace.

The low-rank constraint is what makes the model mechanistically analyzable. The dynamics reduce to P latent variables, so fixed points and bifurcations can be computed directly. Our new bifurcation analysis of disengagement in ALM (Sec. 3.3.1) was obtained this way: a network of 3,940 neurons analyzed in 16 dimensions. An unconstrained network might also develop low-dimensional structure (as shown for example by Schuessler et al. Neurips 2024), but it would have to be extracted after fitting. We also find that a low rank is enough: increasing model dimensionality P beyond about 20 does not improve held-out performance (Sec. 3.2, Fig. 3b).

The question about biases outside the recurrent subspace is a good one. Such a component can be dynamically meaningful, and it would not simply be attenuated. In a linear network it would have no effect on the latent dynamics. In our network it passes through the nonlinearity, so the change it produces in the firing rates has a component within the recurrent subspace. It therefore acts on the latent dynamics and can reshape the flow field, as an in-subspace bias does. Our model excludes such components by design, so that each trial is described by only P parameters. We find that this is sufficient to reproduce the single-trial structure of the recorded activity (e.g., Sec. 3.3, Fig. 4b). Whether out-of-subspace components add explanatory power is an interesting open question.

Initial states, time-varying inputs, and identifiability.

The reviewer is right that our model includes neither trial-specific initial states nor fast time-varying inputs. Our aim is to capture internal-state variables that change slowly relative to a trial, such as engagement or arousal, and our hypothesis is that such variables act by changing the flow field. The new bifurcation analysis (Sec. 3.3.1) shows the kind of explanation our model can give: along the disengagement axis, the fitted network moves through bifurcations into a regime in which it can no longer hold a choice.

To directly compare our model with these other forms of variability, we tested each of them as an alternative to the bias in the ALM data (App. D.9). Both fit held-out data worse in most sessions. More importantly, unlike the alternatives, our model generalized consistently to a task condition it was not trained on, which supports the mechanism it proposes (see the general response).

There is indeed an identifiability concern in general: the same recorded activity can often be explained in more than one way, but each explanation is a different hypothesis about the circuit. An initial state changes where activity starts within a shared flow field, while a bias changes the flow field itself, so they make different predictions outside the fitted condition, which is what the new generalization test examines. A related ambiguity, between changes in inputs and changes in recurrent weights, is discussed in our response to Reviewer v4VA. Adding initial states or time-varying inputs to the model could make it more expressive but could add to this ambiguity, and we leave such extensions to future work.

Reviewer_v4VAOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper focuses on modeling the trial-by-trial variability of neural data. The authors approach the problem through the inclusion of a "control signal" that biases the dynamics towards different corners of the flow field. This approach is nice since it bridges between a completely uninterpretable version, i.e., LFADS, that compresses time sequences into initial states that serve a similar purpose, and pure dynamical systems where the bias is learned and fixed for all trials.

While I appreciate the approach, there are serious drawbacks to the current manuscript. For a modeling paper, there's effectively no serious comparison with other models. There are plenty of models cited (and a few not) that would be viable alternative hypotheses and are not compared to. The one exception is the Neural Latents Benchmark dataset, where the values for the presented model seem OK at best. Similarly, the model presents a very strict hypothesis and it would be very good to understand when the hypothesis is violated what the model returns. There's really three critical issues in this regard (see below) that I feel prevent me from being more positive about the work.

Critical strengths
  • The modeling is well justified and addresses a specific challenge in neural data analysis. Moreover the analysis is nice in that there are both quantitative and explanatory elements.
  • The emphasis on interpretability is appreciated.
  • There are multiple datasets analyzed, which helps know that the results are not dataset-specific
Critical weaknesses
  • There are some newer methods not discussed that are relevant, specifically in terms of "changing netwroks. Specifically, CLDS, dLDS, and the behavior-oriented variant thereof (behavior-dLDS) would fall into the alternative hypothesis that variability comes from modulatory effects on the weights rather than different inputs. More critically, probabilistic dLDS (p-dLDS), the probabilistic version of dLDS also had a slowly changing (albeit not completely constant) bias that was inferred that is very similar to the model presented. If anything, that model was more flexible in terms of not being locked into trial-structured data.

  • If changes were due to circuit changes, what would the control-based approach fit? A simulation in this direction would help build confidence in the approach and that the bias would not hallucinate generic differences.

  • There's no real model comparison. The one real exception is Table 4 in the supplement, however there the model doesn't seem to do that well. I tend to agree with the authors that those metrics lack the ability to really suss out mechanistic interpretability, but a comparison is still needed somewhere and somehow. If the goal is to test the control signal hypothesis, generic null models don't really count as much as testing against the competing hypotheses of, e.g., modulatory changes to the network itself (as discussed in the background section), or other possibilities.

References:

  • Geadah, et al. Modeling neural activity with conditionally linear dynamical systems. NeurIPS. 2026.
  • Mudrik, et al. Decomposed linear dynamical systems (dlds) for learning the latent components of neural dynamics. JMLR. 2024.
  • Yezerets et al. Behavior-dLDS: A decomposed linear dynamical systems model for neural activity partially constrained by behavior. TMLR. 2026.
  • Chen et al. Probabilistic decomposed linear dynamical systems for robust discovery of latent neural dynamics. NeurIPS. 2024.
Questions for authors

To recap from critical weaknesses

  • Where are the model comparisons?
  • What is the difference from slow drift baises, such as in p-dLDS (or for that matter data-driven biases as in rSLDS, which is only really mentioned in the supplement)
  • How does the model behave when presented with data from the "wrong" hypothesis?
4 comments
AuthorsOpenReview ↗

Response to Reviewer v4VA (part 1/2)

We thank the reviewer for the thoughtful and helpful feedback.

Relation to CLDS, dLDS and p-dLDS.

We thank the reviewer for pointing us to these methods, which we now discuss in the related work. The difference is how changing dynamics are generated. CLDS, dLDS, behavior-dLDS and p-dLDS describe variability through time- or condition-dependent linear operators, which gives tractable inference and clear local interpretations. We take a different approach and instead keep the recurrent circuit fixed and nonlinear, following low-rank network theory (Mastrogiuseppe and Ostojic, 2018; Dubreuil et al., 2022), because this is what lets us analyze attractors and bifurcations directly in the fitted model.

Crucially, holding our control input fixed leaves a nonlinear system that can support multiple coexisting attractors. Changing that input shifts neuronal operating points and gains, modifying effective interactions and potentially eliminating attractors while recurrent weights remain fixed: "one circuit, many flow fields." An affine system with a fixed dynamics matrix cannot change stability through an additive bias alone. In p-dLDS, for example, holding the dynamics coefficients fixed leaves the slow offset as a moving origin for the local dynamics; it does not encode the same relationship between input, neuronal gain, and recurrent stability.

Our new fixed-point and bifurcation analysis on ALM data, prompted by the comment of Reviewer EQ58, makes this mechanism explicit (Sec. 3.3.1, new Fig. 5). In models fitted to ALM recordings, moving along the inferred disengagement direction induces saddle-node bifurcations that eliminate the choice-selective attractors, leaving only the pre-stimulus attractor. A constant input thus qualitatively reorganizes the fitted circuit's dynamics, beyond shifting activity to another region of state space. Locally linear models might capture such phenomena through suitable regime combinations; our framework generates these transitions directly through the nonlinear response of a fixed recurrent circuit.

Finally, we agree that the current implementation assumes trial-structured data. Slowly varying latent biases provide a natural extension to continuous recordings (as in p-dLDS) while retaining the fixed nonlinear circuit, although their inference would require appropriate temporal constraints. While this is a nice extension, we leave it as future work.

What the model recovers when the circuit itself changes.

This is a valid concern. We assume that biases change across trials, while in reality the recurrent interactions might change. One reason for our choice is that modulatory inputs seem to be a plausible source of variability across repeated trials of the same task, without requiring synaptic plasticity. However, we agree that synaptic modulation can also change effective recurrent interactions without plasticity.

Disentangling changes in recurrent interactions from changes in feedforward drive in nonlinear networks is not that simple and often leads to an identifiability problem. We therefore do not necessarily expect our model to distinguish whether the inputs or recurrent interactions changed. A version with trial-dependent recurrent interactions might also explain the ALM dynamics and recover the bifurcation. Indeed, a saddle-node bifurcation, such as the one we find in ALM, might be driven by changes in either recurrent or input parameters. Distinguishing these mechanisms can therefore be difficult, regardless of the model used.

To test this more rigorously, we added a teacher–student experiment with a deliberate mismatch (App. C.5): across trials, the teacher's recurrent interactions vary, while the student tries to reproduce the teacher's activity through bias changes alone, with fixed recurrent connectivity. As expected, the student succeeds, confirming an identifiability issue. Two results nonetheless hold. First, the student recovers the teacher's saddle-node bifurcation rather than hallucinating a different transition. Second, the control axis identified by the student can still be used to bifurcate the teacher: applying the corresponding input perturbation to the teacher drives it through a bifurcation, even though its original variability came from recurrent changes. The student therefore identifies populations through which we can control the original system, and our method can identify a correct control axis that we can then test experimentally.

Thus, although the source of variability, recurrent or feedforward, remains ambiguous, the student still recovers the relevant dynamics and identifies a useful intervention direction. This gives us greater confidence in the dynamical interpretation of the ALM disengagement axis, while leaving open whether its underlying source is input modulation, recurrent modulation, or both. We added the teacher–student mismatch experiment and now discuss this issue in the limitations section.

AuthorsOpenReview ↗

Response to Reviewer v4VA (part 2/2)

Model comparisons.

We agree that comparisons against competing explanations of trial-to-trial variability are needed, and we have added them (see below). As the reviewer notes, Table 4 evaluates our model on the Neural Latents Benchmark (NLB), an established dataset and evaluation protocol against which existing methods have been extensively optimized. Our model achieves reasonable predictive performance, although modestly below that of the strongest deep models (NDT, AutoLFADS), while providing mechanistic interpretability. As the reviewer acknowledges, these generalization metrics do not directly measure how well a model captures the underlying mechanism.

Our framework represents one hypothesis about trial-to-trial variability. The most direct way to test it is to compare alternative explanations within the same modeling framework. We therefore added controlled comparisons within the same nonlinear recurrent architecture, varying the proposed source of variability. Specifically, we compare our bias model against alternatives based on time-dependent stochastic drive, inspired by Pals et al. 2024, or trial-specific initial conditions, inspired by LFADS-like models (see also our response to Reviewer FvK9). We evaluate these alternatives using co-bps and R2R^2, with our bias model performing better in most sessions. As discussed above, our teacher–student experiment also addresses recurrent modulation and the difficulty of distinguishing it from input modulation.

More importantly, following the reviewer's comment, we added a stronger test of the inferred mechanism: extrapolation to held-out catch trials in the ALM data, in which the go cue was delivered 600 ms later than usual. If the model has identified the attractors underlying delay-period activity, it should maintain that activity when the expected go cue is withheld. This tests whether the learned dynamics sustain the state beyond the usual delay duration, rather than simply reproducing trajectories under familiar task timing. This is what we consistently observe in our bias model, but not in the initial-condition or stochastic-drive alternatives. We therefore view this extrapolation as a more direct test of the proposed mechanism than standard held-out prediction alone. These comparisons are now reported in the new appendix section (App. D.9, see general comment).

Reviewer_v4VAOpenReview ↗

Appreciated

Thank you for the detailed response.

I appreciate the additional experminets that were run and feel that these are in the right direction.

One residual point is: if both inputs and changing weights capture the same qualitative behavior, and have similar data representation, where is the main benefit of the proposed method? There's still limited comparison to the current literature, which would potentially address this more directly. Right now the added comparisons are pretty limited as baselines and not actually checking if something is to be gained with the proposed model. Without this the novelty and impact is hard to understand.

As an aside, there's a chance I misunderstood some of the edits. The text is still a little hard to follow at times, with some very awkward phrases that don't quite make sense.

Given this, I feel that this work still falls in the "weak reject" category.

AuthorsOpenReview ↗

Response to remaining concerns

The main benefit and novelty of our method:

We thank the reviewer for clarifying the remaining concern, and address it directly below. Our goal is not only to fit single-trial activity, which many models do well, but to capture how internal states, such as engagement, reshape a circuit's dynamics from trial to trial. Our method does this by inferring control axes of a fixed circuit that are at once bifurcation parameters, whose effect on the dynamics can be analyzed directly, and input patterns over the recorded neurons, which can be delivered by photostimulation. To our knowledge, no existing method infers from single-trial activity how internal states may reshape a circuit, in a form that can be both analyzed as bifurcation parameters and tested by stimulation. In ALM, engagement became such an axis: along it, the circuit loses its choice attractors through saddle-node bifurcations (as our new analysis shows), and the axis specifies the stimulation pattern predicted to switch the circuit between engaged and disengaged regimes. More broadly, our framework aims to offer a general way to turn trial-to-trial variability into mechanistic explanations of a circuit, and these explanations into causal experiments.

Our novelty therefore lies both in the framework and in what it reveals: in networks fitted to ALM recordings, disengagement corresponds to a bifurcation that eliminates the choice attractors.

Recurrent Vs. feedforward:

As the reviewer notes, activity alone generally cannot reveal whether such changes come from inputs or from recurrent weights, and this limitation is not specific to our model. The main output of our method is an axis along which the circuit can be moved between regimes, and such axes can remain valid even when inputs are not the source of variability. Our teacher-student test checks exactly this. The teacher's variability came only from changes in its recurrent weights. Yet the axis inferred by our model, applied to the same teacher as an input with its weights unchanged, drove it through the same bifurcation. This makes our model of ALM directly testable, even though the source of the saddle-node bifurcations cannot be identified from activity: changes to synaptic weights cannot currently be imposed in a comparably targeted way, whereas inputs along the inferred control axis can be delivered by targeted photostimulation, and are predicted to move ALM dynamics through the bifurcations.

Comparisons with other models:

Our new comparisons were designed to test alternative explanations of trial-to-trial variability: how do predictions differ when variability is attributed to initial conditions, to stochastic inputs, or to static inputs that reshape the circuit's flow field? These hypotheses were raised in the reviews. We compared them within the same recurrent architecture, changing only the source of variability, so that differences reflect the hypothesis itself. Comparisons across published methods also differ in architecture, parameterization and inference, and reconstruction scores alone cannot isolate which assumption accounts for a difference in performance.

This follows the standard practice in the field: new methods are typically compared within a shared model class, on what they can capture that the other models cannot (e.g., the p-dLDS paper compares it with other linear dynamical methods), or across model classes on standard benchmarks (as we did with the NLB). Our comparisons therefore do both. In the extrapolation to trials with a delayed go cue, only our model consistently predicted that choice activity persists, consistent with the choice attractors reported in ALM. If the reviewer has other specific comparisons in mind, we would be glad to discuss them.

Finally, the reviewer's original assessment rated the presentation as clear and easy to follow, so we understand that the unclear passages are likely in the newly added sections. We will revise these carefully, and would welcome specific examples so that we can address them directly.

In light of this response, the new results in the revised manuscript and the detailed answers to the other reviewers, we would appreciate the reviewer's reassessment of the novelty and scientific impact of our work.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers

We thank the reviewers for their constructive comments. The revised manuscript contains four new analyses that we believe substantially strengthen the paper and answer the reviewers’ concerns. All changes in the manuscript are marked in blue.

The central claim, "one circuit, many flow fields", is now demonstrated in a network fitted to cortical recordings (new section in the results). Following Reviewer EQ58's suggestion, we performed an extensive fixed-point analysis of a network fitted jointly to 72 ALM sessions from 24 mice (Sec. 3.3.1). With the recurrent weights fixed, moving a constant input along the disengagement axis induces saddle-node bifurcations that eliminate the choice attractors and leave only the pre-stimulus attractor: in this regime the fitted network can no longer hold a choice. This explains the mechanism behind our original finding that shifting the bias makes trajectories ignore-like. The bifurcations occur within the range of the recorded trials, with ignore trials lying on average near the point where the last choice attractor disappears.

Disengagement corresponds to a common control axis across sessions. The bifurcation analysis above was conducted in a new session-stitched model, in which all 72 sessions share one recurrent connectivity and one bias space (App. D.7), so the same direction can be compared across recordings. An axis estimated only from other sessions still distinguishes ignore from responded trials in a held-out session, with most of the separation achieved by that session's own axis. Disengagement is therefore not merely a session-specific feature of the fit, but a shared direction along which the fitted circuit changes regime.

The model predicts a condition it never saw, where alternative models fail. All three reviewers asked how our model compares with other explanations of trial-to-trial variability. We tested models in which all trials share one flow field and differ only in their initial state or in stochastic inputs (App. D.9). Our model fit held-out data better in most sessions. More importantly, it was the only one that consistently predicted the persistent choice activity observed when the go cue is delayed, a condition excluded from training which we now add to the analysis. Our model thus generalizes beyond the condition it was fitted to, which the single-flow-field alternatives did not do consistently. This extrapolation supports the mechanism our model proposes, a per-trial change in the flow field, over the single-flow-field alternatives.

The inferred control axis remains valid even when the true source of variability is recurrent. Following Reviewer v4VA, we fit our model to a teacher whose bifurcation is driven by changes in its recurrent weights (App. C.5). The student cannot tell where the change came from, which we now discuss as a limitation. However, it recovers the correct bifurcation, and the input axis it infers drives the teacher itself through that bifurcation. In this test, the model identified a direction along which the true system can be controlled, which is the kind of prediction we propose to test experimentally.

Detailed responses follow under each review.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.