acceptodds
Under review as a conference paper at ICLR 2027

Theta sweeps accelerate cognitive map learning by broadcasting information to unvisited regions

Abstract

To support flexible behavior, animals need to learn map-like representations of environments from limited experiences, but the underlying neural mechanisms remain unclear. Here, we propose that theta sweeps, a prominent feature in the activities of grid and place cells during active movement, can broadcast acquired sensory information to swept unvisited regions, accelerating cognitive-map learning. To test this idea, we built a network of direction, grid, and place-cell modules, in which theta modulation and firing-rate adaptation generate theta sweeps, while Hebbian plasticity enables place cells to bind environmental observations. Simulations and ablations showed that the dorsal–ventral grid organization produces multiscale, independent sweeps broadcasting acquired sensory information to a broad, fan-shaped unvisited area, enabling unvisited regions to build place-environment associations. The model further suggests that declining theta power as the environment becomes familiar acts as an annealing process, shifting from learning general information about the environment to making local refinements. The model makes three testable predictions: strong theta modulation accelerates early learning, environments with broader spatial correlations favor longer theta sweeps, and theta sweeps progressively shorten as environments become familiar. Together, our results support theta sweeps as a mechanism for accelerating cognitive-map learning from limited explorations, and it sheds light on developing brain-inspired algorithms for building world models.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Theta sweeps, cognitive maps, information broadcasting, sample-efficient learning, entorhinal--hippocampal circuits

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.50 over 4 reviews, above 99% of the others.
Reviewer_5BNVOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
3(+1)
Read review
Paper summary

The paper proposes that theta sweeps accelerate sensory-map learning by associating the animal’s current observation with internally represented nearby locations. It combines an established direction/grid-cell sweep mechanism with recurrent hippocampal populations and a plastic hippocampal-to-sensory readout. In simulated spatially smooth environments, the sweep-enabled model learns more accurate sensory predictions than its no-sweep counterpart and an adapted Vector-HaSH baseline, including at unvisited locations. Ablations examine multiple spatial scales and hippocampal recurrence. Parameter scans show that broader sensory correlations favor longer optimal sweeps, while a prescribed reduction in theta modulation illustrates a tradeoff between rapid early generalization and later accuracy. The principal contribution is a functional account of sensory broadcasting and its limitations, supported by simulations, but network-compatible mechanisms of theta strength annealing and sensory correlation dependent adaptation are not within the scope of this work.

Critical strengths

The paper provides a useful account of how internally generated spatial activity could support sensory generalization. Extending prior sweep-based learning of spatial connectivity (Marshall et al., 2025; Ji et al., 2026) to sensory-content associations exposes a consequential tradeoff: longer sweeps reaches more and locations farther away, but risk associating observations with locations whose sensory content differs. Further experiments connect the optimal sweep span to sensory correlation width, and illustrates a benefit from reducing broadcasting later in learning. These results are mostly well-validated and add a specific computational explanation beyond the general proposition that sweeps facilitate map formation.

Critical weaknesses

The attribution of improved learning to spatial-scale diversity is confounded by model capacity. Figure 4a,b compares 6,400 hippocampal cells in M1 with 32,000 cells across five populations (Appendix B.5), changing the number of trainable readout weights and the populations entering normalization along with scale and sweep diversity. The result therefore establishes an advantage of the full configuration, but does not isolate the contribution of diverse scales. This limits an important mechanistic interpretation of the architecture; a capacity-matched comparison would make that attribution more convincing.

The annealing result has limited validation relative to its familiarity-based interpretation. Figure 5 uses a prescribed elapsed-time schedule on one 4,000 s trajectory, with a shared learning coefficient rather than independently tuned fixed-modulation controls. Figure 7d also shows that sustained full modulation remains beneficial at persistently unvisited locations. The experiment supports a possible benefit from reducing broadcasting, but leaves unclear how robust that benefit is to exploration history and baseline tuning. This matters because familiarity-dependent sweep shortening is one of the paper’s three central predictions. A biological familiarity controller would be a valuable extension, but its implementation is not necessary to establish the narrower functional result.

Questions for authors

Major questions: please see criticial weaknesses.

Minor issues/clarifications:

  1. In figure 3(c) caption, “Under the same exploration trajectory” followed by “mean ± SD over three trajectories” is a little confusing, can the authors clarify?
  2. Spatial units vary: metres in 2c, “world units” in 2d, unspecified units in 2e, and normalized coordinates in 3d, is it feasible to conform the use of units?
  3. In line 21 (abstract) and line 466 (discussion), what does independent mean? The modules share a theta clock and movement inputs. “Module-specific sweeps with different directions and extents” may describe the demonstrated result more precisely.
  4. In section 4.3, lines 409–412, the manipulated theta parameter controls anchoring modulation strength, whereas the biological comparison concerns LFP theta power, can the authors make this clear?
  5. “sweep length” vs. “sweep span”, can the authors use a brief definition in the caption to avoid ambiguity in figure 6?
  6. “we add global inhibition to presynaptic input” — line 206. Equation 2 instead expresses inhibition through divisive firing-rate normalization, can the authors make it clearer in the text?
8 comments
AuthorsOpenReview ↗

Response to Reviewer 5BNv (Part I)

Thank you very much for your thoughtful, thorough review and for recognizing the core computational trade-off between generalizing to unexplored space and refining predictions along familiar paths. Your insightful suggestions—particularly regarding the capacity-matched control, trajectory-replicated and independently tuned annealing experiments, and terminological precision—have substantially strengthened the mechanistic clarity and rigor of our manuscript. We address each of your points in detail below.

W1. Does the multiscale advantage persist at matched capacity?

The attribution of improved learning to spatial-scale diversity is confounded by model capacity. Figure 4a,b compares 6,400 hippocampal cells in M1 with 32,000 cells across five populations (Appendix B.5), changing the number of trainable readout weights and the populations entering normalization along with scale and sweep diversity. The result therefore establishes an advantage of the full configuration, but does not isolate the contribution of diverse scales. This limits an important mechanistic interpretation of the architecture; a capacity-matched comparison would make that attribution more convincing.

Yes—the advantage of spatial-scale diversity persists with high statistical significance when network capacity is strictly matched.

Thank you for pointing out that our original comparison varied both spatial scale and the total number of hippocampal cells. Following your suggestion, we replaced the single-module baseline in Fig. 4a with a capacity-matched single-scale control consisting of five populations that all share the spatial scale of M1. Both the single-scale control and the multiscale model now have:

  • The exact same 32,000 hippocampal cells;
  • The exact same number of trainable sensory readout weights;
  • Five separate divisive normalization pools, while preserving the modules’ respective grid orientations, CA3 recurrent connectivity, and direction-cell adaptation parameters (detailed in revised Appendix B.5).

Across 10 matched trajectories, the multiscale model still achieved a substantially lower all-position prediction error at 400 s than the capacity-matched single-scale control (0.509 ± 0.008 vs. 0.653 ± 0.010, mean ± SD; two-sided paired t(9)=−49.78t(9) = -49.78, p=2.67×10−12p = 2.67 \times 10^{-12}). We have updated Fig. 4a with these capacity-matched learning curves and included Fig. 4b to illustrate the spatial coverage of M1 compared with the combined five-module population across ten trajectories (mean ± SD).

AuthorsOpenReview ↗

Response to Reviewer 5BNv (Part II)

W2. How robust is the benefit of theta annealing?

The annealing result has limited validation relative to its familiarity-based interpretation. Figure 5 uses a prescribed elapsed-time schedule on one 4,000 s trajectory, with a shared learning coefficient rather than independently tuned fixed-modulation controls. Figure 7d also shows that sustained full modulation remains beneficial at persistently unvisited locations. The experiment supports a possible benefit from reducing broadcasting, but leaves unclear how robust that benefit is to exploration history and baseline tuning. This matters because familiarity-dependent sweep shortening is one of the paper’s three central predictions. A biological familiarity controller would be a valuable extension, but its implementation is not necessary to establish the narrower functional result.

Thank you for this insightful critique. To rigorously evaluate the robustness of theta annealing against exploration history and baseline hyperparameter tuning, we repeated the 4,000-s experiment across 10 independent trajectories while using independently optimized learning coefficients for each fixed-modulation baseline.

  • Consistent advantage across trajectories: Across all 10 independent 4,000-s trajectories (where all three conditions followed the exact same path and received identical sensory observations within each replicate), annealing achieved the lowest final all-position prediction error on every single trajectory. The mean relative prediction error at 4,000 s was 0.244 ± 0.003 for annealing, compared with 0.271 ± 0.003 for fixed zero modulation and 0.376 ± 0.004 for fixed full modulation (mean ± SD across 10 trajectories).
  • Independent selection of learning coefficients: You are right that the original Fig. 5 used a shared learning coefficient (0.53840.5384). In our original parameter sweep (evaluating 15 candidate coefficients per condition to minimize mean 400-s error across three validation trajectories; Appendix C, Fig. 7a), the optimal learning coefficients were 0.1486 for fixed zero modulation and 0.5384 for fixed full modulation. In the revised Fig. 5, each fixed-modulation baseline uses its own dedicated optimal coefficient throughout the 4,000-s run. For the annealing schedule, each stage’s learning coefficient was tuned on a separate 4,000-s development trajectory (carrying learned weights and network states across stages) and fixed prior to testing on the 10 evaluation trajectories.
  • Trade-off at persistently unvisited vs. visited locations: We fully agree with your interpretation of Fig. 7d, and we have highlighted this mechanistic trade-off in the revised text. At persistently unvisited locations, sustained full modulation outperforms annealing (0.816 ± 0.008 vs. 0.884 ± 0.008) because long sweeps continue broadcasting sensory information from the path into unexplored regions. Conversely, at visited locations, repeated visits allow the animal to refine its sensory associations; shortening the sweeps localizes updates to the animal's true position, enabling annealing to achieve dramatically lower error than full modulation (0.068 ± 0.001 vs. 0.255 ± 0.003). Because visited locations cover 78.4% of the arena by 4,000 s, this substantial gain over familiar territory drives the lower overall error.
  • Candidate biological familiarity signal: We also appreciate your perspective on biological implementation. One possible plausible feedback signal might be sensory prediction error: in the hippocampal circuit model of Chen et al. (2024), CA3 predicts upcoming entorhinal inputs while CA1 computes the mismatch upon arrival. Large mismatches in novel environments could sustain strong theta modulation (and long sweeps), whereas declining prediction error with familiarity could down-regulate modulation and shorten sweeps. We have added this discussion to Section 5 as a concrete direction for future closed-loop modeling.

Q1. How are trajectories matched and averaged in Fig. 3c?

In figure 3(c) caption, “Under the same exploration trajectory” followed by “mean ± SD over three trajectories” is a little confusing, can the authors clarify?

Thank you for catching this ambiguity. Within each replicate, all models followed the exact same physical trajectory and received identical sensory observations; across the three replicates, three distinct trajectories were used. For each trajectory, we first averaged the relative prediction error across all spatial query locations at each time point, and Fig. 3c plots the mean ± 1 SD across the three resulting trajectory-level error curves. We have clarified this wording in the Fig. 3c caption (and noted in the Fig. 5a,c captions that the late-stage SD shading is narrower than the plotted line width).

AuthorsOpenReview ↗

Response to Reviewer 5BNv (Part III)

Q2. Can spatial units be standardized?

Spatial units vary: metres in 2c, “world units” in 2d, unspecified units in 2e, and normalized coordinates in 3d, is it feasible to conform the use of units?

Yes. We have standardized all spatial measurements across the manuscript and figures: positions, tuning widths, and sweep lengths are now consistently reported in metres (m\text{m}), and running speeds in metres per second (m/s\text{m/s}). We updated the axes and captions of Figs. 2c–e, 3d, and 6b–d, as well as all corresponding supplementary figures (with phase in radians and preferred direction in degrees).

Q3. What does “independent” mean for modules that share a theta clock?

In line 21 (abstract) and line 466 (discussion), what does independent mean? The modules share a theta clock and movement inputs. “Module-specific sweeps with different directions and extents” may describe the demonstrated result more precisely.

Thank you for suggesting this much more accurate phrasing. Because all modules share a common 10 Hz theta clock as well as common velocity and heading inputs, their sweeps are indeed coupled to the same global rhythm while diverging in trajectory due to each module’s own attractor state, adaptation dynamics, and spatial scale. We have replaced “independent” with your exact suggested phrasing—“module-specific sweeps with different directions and spatial extents”—in the Abstract and aligned the Discussion accordingly.

Q4. How does the model’s modulation parameter relate to LFP theta power?

In section 4.3, lines 409–412, the manipulated theta parameter controls anchoring modulation strength, whereas the biological comparison concerns LFP theta power, can the authors make this clear?

Thank you for pointing out the need to distinguish our model parameter from experimental LFP recordings. In our model, ϑ∈[0,1]\vartheta \in [0, 1] controls the modulation depth of the external anchoring inputs (position and heading cues) across each theta cycle, varying their gain between 1−ϑ1 - \vartheta and 11. When ϑ=0\vartheta = 0, anchoring remains constant; when ϑ=1\vartheta = 1, anchoring drops to zero at the trough of the theta cycle, freeing the activity bump to sweep outward before being pulled back.

By contrast, Penley et al. (2013) measured local field potential (LFP) theta power in CA1 and the dentate gyrus, reporting elevated theta power during novel-environment exploration. Our hypothesis links the two by proposing that stronger population theta oscillations (higher LFP theta power) reflect deeper rhythmic modulation of the inputs anchoring hippocampal activity, thereby enabling longer sweeps. Directly testing this quantitative link will require simultaneous recordings of LFP theta power, anchoring input strength, and sweep length.

Q5. How is sweep length defined?

“sweep length” vs. “sweep span”, can the authors use a brief definition in the caption to avoid ambiguity in figure 6?

We now use “sweep length” consistently throughout the paper. Specifically, within each complete 0.1-s theta cycle, we express the decoded position relative to the animal’s instantaneous position and compute the maximum Euclidean distance between any two points along that relative trajectory, averaging this distance across complete cycles. We have added this concise definition to the Fig. 6 caption and provided the full mathematical estimator in Appendix B.3.

Q6. How is global inhibition implemented in Eq. 2?

“we add global inhibition to presynaptic input” — line 206. Equation 2 instead expresses inhibition through divisive firing-rate normalization, can the authors make it clearer in the text?

Thank you for catching this discrepancy between the text and Eq. 2. You are completely right: rather than subtractive inhibition on presynaptic inputs, our circuit implements divisive firing-rate normalization within each place-cell population, where each cell’s squared, rectified input is divided by the summed squared, rectified inputs across all cells in that population. We have corrected Section 3.1 to state this explicitly and cited the continuous-attractor models from which this normalization formulation is adapted (Wu et al., 2008; Fung et al., 2008).

Thank you once again for your constructive guidance, your suggestion of the capacity-matched control, and your insightful analysis of the spatial trade-off in theta annealing. We hope that these new experiments, statistical analyses, and textual clarifications fully address all of your concerns. If so, we would be very grateful if you would consider raising your score.

Reviewer_E7aEOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper tests a possible functional benefit of theta sweep, that its alternating left and right path sweeps help associate relevant unvisited positional code to the same sensory (landmark) representations. The paper uses a more structured setting with landmark vector sensory input compared to previous papers such as VectorHash where sensory input is unstructured arbitrary vectors. In this setting, theta sweep greatly improves the performance. Even without thetasweep, the proposed model outperforms VectorHash, which the paper suggest due to the recurrence in CA3 and spatially structured grid-to-place projections.

Critical strengths

the paper proposes an interesting function for theta sweeps.

It is important to consider the statistics of sensory input during navigation, instead of assuming random sensory input.

Critical weaknesses

Novelty and significance of the broadcasting function of theta sweeps

It is an interesting proposal, but much of it is already captured in works such as Lin et al, 2026 ICLR (they call it smoothing). In comparison, Lin et al had more realistic sensory inputs, non-prescribed place fields, more complex environments and realistic navigation task instead of sensory binding. An explicit comparison with prior works with similar ideas would make the paper's contribution clearer and stronger.

why the left-right look-ahead broadcasting?

what about a gaussian broadcasting, past experience broadcasting, straight look-ahead broadcasting? Why and how is the proposed left-rightt look-ahead special?

Prescribed sensory code

I wonder why the it's needed to achieve the shown result. Intuitively as long as the sensory code is somewhat smooth, the broadcasting would provide a useful inductive bias. Is the specific sensory code (LVC) needed because the sensory code would allow spatial smoothing along specific angles, thus fitting the specific type of look-ahead broadcasting? For the same reason it feels also a bit unfair to compare to VectorHash which is designed for arbitrary sensory vectors, the paper needs to at least acknowledge that.

non-broadcasting case need a clearer positioning

the CA3 recurrence non-broadcasting model beat Vector-Hash. Is it a major achievement? The paper however did not provide enough mechanistic explanation. It also makes it hard to isolate the main contribution, is it the new model or the broadcasting mechanism? If it's the model, what about other prior works such as dense associative memory?

formatting issue

How are the line numbers generated? They don't seem to correspond to the lines at all.

Questions for authors

it's in the weakness section.

9 comments
AuthorsOpenReview ↗

Response to Reviewer E7aE (Part I)

Thank you for your interest in the function we propose for theta sweeps. We address all of your concerns point by point below.

W1. How does information broadcasting differ from the smoothing proposed by Lin et al. (2026)?

Novelty and significance of the broadcasting function of theta sweeps: It is an interesting proposal, but much of it is already captured in works such as Lin et al, 2026 ICLR (they call it smoothing). In comparison, Lin et al had more realistic sensory inputs, non-prescribed place fields, more complex environments and realistic navigation task instead of sensory binding. An explicit comparison with prior works with similar ideas would make the paper's contribution clearer and stronger.

Thank you for pointing out Lin et al. (ICLR 2026). While Lin et al.’s notion of “smoothing” might superficially appear related to our mechanism, the two works address fundamentally different functional hypotheses operating along orthogonal dimensions—temporal memory versus spatial generalization:

  • Lin et al. (2026) model smoothing across time (working memory): Hippocampal sequences act as a “temporal memory buffer” where “the CA3 module expands sparse DG codes into a temporally smoothed canonical basis set, providing long-horizon history.” Their mechanism retains past observations to help an agent choose actions when current sensory evidence is sparse.
  • Our model proposes broadcasting across space (spatial generalization): Our mechanism does not perform temporal smoothing; rather, it uses a single current observation at the animal's current location to generalize across space to unvisited nearby locations, accelerating cognitive map learning.
AuthorsOpenReview ↗

Response to Reviewer E7aE (Part II)

W2. Why use left–right alternating sweeps?

why the left-right look-ahead broadcasting? what about a gaussian broadcasting, past experience broadcasting, straight look-ahead broadcasting? Why and how is the proposed left-rightt look-ahead special?

We use left–right alternating sweeps in our model for both biological and computational reasons.

First, our aim is to explain the function of the sweep pattern actually observed in the brain. In freely moving rats, Vollan et al. (2025) found that the position represented by entorhinal grid cells and hippocampal place cells sweeps outward from the animal, alternating to the left and right of its heading across successive theta cycles. We therefore model this experimentally observed pattern to investigate how it facilitates spatial map learning along the animal's trajectory.

Second, from a computational perspective, alternating sweep directions samples nearby space more efficiently by minimizing spatial overlap between successive sweeps. Vollan et al. showed that choosing each sweep direction to minimize overlap with previously covered areas naturally produces stable left–right alternation, as sweeping to the opposite side covers more unexplored ground. In our sweep-generating network, firing-rate adaptation (slow negative feedback with a time constant about ten times that of neural activity) suppresses recently active neurons, making subsequent sweeps naturally avoid resampling the region just represented and thereby approximating this overlap-minimization strategy.

We compared left–right alternating sweeps with random local jitter and straight-ahead sweeps, matching their radial displacement scales.

For random local jitter, we drew an isotropic Gaussian displacement at each time step, matching the root-mean-square radial displacement of sweeps in each module and using the same learning rate. For straight-ahead sweeps, we directed each module’s sweep target along the animal’s instantaneous velocity while preserving its radial displacement. The table reports relative prediction error after 400 s as mean ± SD across the same three test trajectories.

Condition at 400 sAll locationsVisited locationsUnvisited locations
Ours, with sweeps0.506 ± 0.0120.244 ± 0.0040.618 ± 0.013
Ours, without sweeps0.611 ± 0.0140.211 ± 0.0060.783 ± 0.011
Random local jitter0.545 ± 0.0130.323 ± 0.0040.640 ± 0.012
Straight-ahead sweeps0.584 ± 0.0130.215 ± 0.0010.742 ± 0.010

Comparison with random local jitter: Alternating sweeps achieved lower error at both visited and unvisited locations. Independent Gaussian displacements can repeatedly sample the same nearby regions. Adaptation suppresses recently active neurons, directing successive sweeps toward complementary regions on either side of the animal’s heading. This organized coverage helps each new sensory observation reach nearby positions that have received less training.

Comparison with straight-ahead sweeps: Straight-ahead sweeps yielded lower error at visited locations, whereas alternating sweeps achieved lower error at unvisited locations and across the whole arena. Straight-ahead sweeps concentrate updates along the animal’s path. Left–right alternation extends learning laterally into the gaps between explored paths, improving coverage of unvisited space. The comparison with the no-sweep control also shows this trade-off: localized updates favor precision near the path, while sweeps improve prediction over the broader environment.

AuthorsOpenReview ↗

Response to Reviewer E7aE (Part III)

W3. Does the result depend on a particular sensory code, and how does the choice of code affect the comparison with Vector-HaSH?

Prescribed sensory code: I wonder why the it's needed to achieve the shown result. Intuitively as long as the sensory code is somewhat smooth, the broadcasting would provide a useful inductive bias. Is the specific sensory code (LVC) needed because the sensory code would allow spatial smoothing along specific angles, thus fitting the specific type of look-ahead broadcasting? For the same reason it feels also a bit unfair to compare to VectorHash which is designed for arbitrary sensory vectors, the paper needs to at least acknowledge that.

Our results do not depend on LVC or any single preconfigured sensory code. In the main text, we refer to the sensory representation as LEC activity, and we explicitly evaluated three biologically motivated encodings (SOVCs, LVCs, and BVCs) in Appendix C, Fig. 7:

  • Shared object-vector cells (SOVCs) encode displacement relative to an object and share their tuning across objects, with full visibility up to 3.5 m, a cosine decay from 3.5 to 3.6 m, and zero visibility beyond 3.6 m.
  • Landmark vector cells (LVCs) retain landmark identity.
  • Boundary vector cells (BVCs) encode the distance to a boundary along a preferred direction.

Across three matched 400-s trajectories, theta sweeps reduced mean relative prediction error over all locations by 17.3% for SOVCs, 18.5% for LVCs, and 18.4% for BVCs (Appendix C, Fig. 7i). Even when the surrounding walls determined the entire sensory pattern (BVCs), sweeps produced a learning benefit comparable to that obtained with object-based inputs. We agree with the reviewer that local spatial correlations allow a current observation to generalize across nearby locations; consistent with this interpretation, broader sensory tuning favored longer sweeps (Fig. 6b–d).

Furthermore, our model with theta sweeps outperformed Vector-HaSH under all three encodings. Using matched trajectories and sensory inputs, the mean relative prediction errors after 400 s were 0.530 ± 0.012, 0.506 ± 0.011, and 0.492 ± 0.010 for our model with SOVCs, LVCs, and BVCs, respectively, compared with 0.834 ± 0.066, 0.757 ± 0.023, and 0.706 ± 0.024 for Vector-HaSH (mean ± SD across three trajectories). To ensure a strictly fair comparison, we also disabled sweeps in both models and trained both sensory readouts using the exact same normalized delta rule on the main LEC prediction task; our model still achieved substantially lower error (0.611 vs. 0.930 for Vector-HaSH).

W4. Why does our model outperform Vector-HaSH even without broadcasting?

non-broadcasting case need a clearer positioning: the CA3 recurrence non-broadcasting model beat Vector-Hash. Is it a major achievement? The paper however did not provide enough mechanistic explanation. It also makes it hard to isolate the main contribution, is it the new model or the broadcasting mechanism? If it's the model, what about other prior works such as dense associative memory?

Our primary contribution is the spatial broadcasting function of theta sweeps, which consistently improves map learning over the non-broadcasting baseline across all sensory codes (e.g., reducing error from 0.611 to 0.506; Fig. 3). The fact that our non-broadcasting circuit also outperforms Vector-HaSH stems from two biologically grounded architectural differences that are already analyzed in the manuscript:

  1. Topographic vs. random projections: In Vector-HaSH, fixed random projections from grid cells to hippocampal units allow distant locations to activate overlapping sets of units. An update at the animal’s current location can therefore inadvertently interfere with sensory associations at distant locations where the sensory input differs. In our model, topographically organized grid-to-place projections bring spatially aligned inputs together, concentrating each update near the animal’s location.
  2. CA3 recurrent dynamics: Local recurrent excitation in CA3 reinforces the current activity peak and ensures it moves continuously as the animal navigates. Removing recurrence produces spatially separated activity peaks and increases error (Fig. 4c,d), whereas restoring recurrence under identical grid inputs eliminates erroneous jumps of the activity peak to distant locations (Appendix C, Fig. 8).

Finally, regarding dense associative memory (Krotov and Hopfield, 2016), that formulation does not specify an online learning rule for acquiring a spatial sensory map from successive observations during exploration, and thus does not provide a direct baseline for online map learning performance.

Reviewer_LgSNOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper posits that theta sweeps (an already well-documented phenomenon in neuroscience) allow for the acceleration of cognitive map learning in models. On the topic of theta sweeps, the authors also discuss how declining theta power leads to a shift from learning general information about the environment to making local refinements. The authors build a model that incorporates the MEC and HPC along with Hebbian plasticity, and build a theta sweep generator for the model. They also show that multiple spatial scales and CA3 recurrence support sensory learning, and that theta annealing balances learning speed and prediction accuracy. Finally, the authors made predictions that they (have not yet) posit can be tested directly in biology to provide useful information about the brain: a) stronger theta modulation should accelerate learning during early exploration, b) theta-sweep length that best supports learning should depend on the spatial correlation length of the sensory representation, and c) theta sweeps should progressively shorten as the environment becomes familiar and theta modulation declines.

Critical strengths

The novelty of the paper is high, and the work is well-motivated and interesting. Additionally, the 'biological' model that the authors use seems useful, and the improvements that they demonstrate theta sweeps cause in cognitive learning are definitely interesting.

Similarly, the predictions that the authors make about the nature of theta sweeps would be directly relevant to furthering our understanding of the way the brain processes information, if proven in biology.

I appreciate the rigorous hyperparameter search, and the breadth of experiments run to understand the role of theta sweeps in their particular model.

Critical weaknesses

The authors do not mention the p-values or establish the statistical significance of their results anywhere in the paper. Additionally, the authors only use 3 seeds for the experiments and, for some results, only use one seed (ie. e., Figs. 4, 5, 6a, and 7d). This would make it hard to disprove that the improvements or results the authors find are not a result of statistical noise or variance.

For the main results in figure 3, the authors only use one control (the Vector-HaSH), and do not use any other control to validate their MEC-LEC model. Also, the Vector-HaSH model is justified as a control, but it isn't enough in my eyes to completely validate the MEC-LEC model.

Questions for authors

Could you release the code (anonymized, of course) for me to view your exact implementations and rerun the experiments, and validate the results? What were the p-values for figs 3c and 3e? Did you average across seeds or pick the best ones for the figures? Could you run up to 10 seeds for each condition and report if the results hold, along with the p-values? Why does the Vector-HaSH control show an increase in L2 error at the start of training, but the "biological" models don't?

4 comments
AuthorsOpenReview ↗

Response to Reviewer lgSN (Part I)

Thank you very much for your encouraging and positive assessment of our work’s novelty and biological motivation, as well as for your insightful, constructive suggestions. Your comments directly helped us strengthen the statistical rigor, baseline comparisons, and reproducibility of the paper. We address each of your points in detail below.

W1 & Q1: Statistical significance and multi-seed replication (W1 & Q1)

The authors do not mention the p-values or establish the statistical significance of their results anywhere in the paper. Additionally, the authors only use 3 seeds for the experiments and, for some results, only use one seed (ie. e., Figs. 4, 5, 6a, and 7d). This would make it hard to disprove that the improvements or results the authors find are not a result of statistical noise or variance.

What were the p-values for figs 3c and 3e? Did you average across seeds or pick the best ones for the figures? Could you run up to 10 seeds for each condition and report if the results hold, along with the p-values?

Thank you for highlighting the importance of statistical rigor. Following your suggestion, we expanded the primary 400 s comparisons below to 10 independent seeds, with one paired endpoint per trajectory.

Figs. 3c and 3e retain the mean ± SD across all three original test trajectories (we did not cherry-pick best seeds); the table below reports the expanded ten-seed comparison. Within each seed, every model experienced the same trajectory and sensory observations.

Across all 10 seeds, every effect remains statistically significant. For each condition, we paired its final prediction error with that of Ours, with sweeps on the same trajectory and applied a two-sided paired t-test across the 10 seeds. The table reports mean ± SD of relative prediction error over all locations at 400 s, together with the p-value from this test and the p-value after Holm–Bonferroni correction across the seven planned comparisons. All displayed comparisons remain significant after correction. For the unvisited comparison in Fig. 3e, the ten-trajectory paired test gives p = 1.44 × 10⁻¹⁴ (Appendix C.2).

Condition at 400 s (10 seeds)Mean ± SDTwo-sided paired t-test p-valueHolm-adjusted p-value
Ours, with sweeps0.509±0.0080.509 \pm 0.008——
Ours, without sweeps0.617±0.0090.617 \pm 0.0091.14×10−131.14 \times 10^{-13}3.41×10−133.41 \times 10^{-13}
Single scale0.653±0.0100.653 \pm 0.0102.67×10−122.67 \times 10^{-12}5.34×10−125.34 \times 10^{-12}
Without CA3 recurrence0.639±0.0110.639 \pm 0.0114.81×10−114.81 \times 10^{-11}4.81×10−114.81 \times 10^{-11}
Fixed theta modulation: 0.250.660±0.0080.660 \pm 0.0086.57×10−166.57 \times 10^{-16}4.60×10−154.60 \times 10^{-15}
Fixed theta modulation: 0.500.658±0.0080.658 \pm 0.0087.24×10−167.24 \times 10^{-16}4.60×10−154.60 \times 10^{-15}
Fixed theta modulation: 0.750.643±0.0080.643 \pm 0.0082.38×10−152.38 \times 10^{-15}1.19×10−141.19 \times 10^{-14}

For the long-duration annealing experiments, we completed the simulations across 10 independent seeds. Each trajectory spans 4,000 s, ten times the duration of the 400 s comparisons. Annealing achieved the lowest final all-position prediction error on every trajectory. The table reports mean ± SD and two-sided paired t-test p-values against annealing, with Holm–Bonferroni correction across the three pairwise comparisons among these conditions.

Condition at 4,000 s (10 seeds)Mean ± SDTwo-sided paired t-test p-valueHolm-adjusted p-value
Theta annealing0.244±0.0030.244 \pm 0.003——
Ours, without sweeps0.271±0.0030.271 \pm 0.0031.75×10−131.75 \times 10^{-13}1.75×10−131.75 \times 10^{-13}
Ours, with sweeps0.376±0.0040.376 \pm 0.0041.01×10−151.01 \times 10^{-15}3.03×10−153.03 \times 10^{-15}
AuthorsOpenReview ↗

Response to Reviewer lgSN (Part II)

2. An independent control beyond Vector-HaSH (W2)

For the main results in figure 3, the authors only use one control (the Vector-HaSH), and do not use any other control to validate their MEC-LEC model. Also, the Vector-HaSH model is justified as a control, but it isn't enough in my eyes to completely validate the MEC-LEC model.

Thank you for this great suggestion. To provide an independent, strong baseline beyond Vector-HaSH, we implemented and evaluated ALPaCA (Harrison et al., 2018), a meta-learning framework that learns spatial basis features and a prior distribution over readout weights across environments, enabling rapid online Bayesian regression from sparse observations in novel environments.

Notably, this experimental setup gives ALPaCA a substantial pretraining advantage: prior to testing, ALPaCA was pretrained via backpropagation on 1,024 distinct landmark layouts, entering each test arena with prior structural knowledge of the task family. In contrast, our MEC–LEC model received zero pretraining and learned the spatial sensory map strictly online from scratch along a single trajectory. At test time, both models received identical sensory observations along the exact same three test trajectories.

Even against this pretrained meta-learning baseline, our model with theta sweeps achieved substantially lower error after 400 s:

  • All locations: Our model with sweeps achieved 0.506 ± 0.012, significantly outperforming pretrained ALPaCA (0.628 ± 0.004, mean ± SD across three trajectories).
  • Unvisited locations: Without sweeps, our model had higher error than ALPaCA (0.783 vs. 0.687), reflecting the benefit of ALPaCA’s cross-environment meta-learned prior. However, adding theta sweeps reduced our model’s unvisited-location error to 0.618, surpassing ALPaCA without requiring any prior training across environments.

We have added the all-position learning curves to Fig.4e and discuss this comparison in the revised main text.

3. Initial increase in Vector-HaSH prediction error (Q2)

Why does the Vector-HaSH control show an increase in L2 error at the start of training, but the "biological" models don't?

Thank you for pointing out this interesting phenomenon—it highlights a key mechanistic difference between the two architectures.

As shown in Appendix C, Fig. 7b–c, Vector-HaSH’s prediction error rises at unvisited locations during early training. Because unvisited locations constitute the vast majority of the arena early on, this error increase dominates the all-position average in Fig. 3c. Here, the visited region is defined by the complete 400-s trajectory, so its early error also includes locations that the animal will reach later.

Mechanistically, this occurs because Vector-HaSH relies on fixed random projections from grid cells to hippocampal units. Consequently, distant locations in the arena share overlapping subsets of active hippocampal units. Updating readout weights to fit the sensory observation at the animal's current location inadvertently perturbs predictions at distant, unvisited locations that have completely different sensory content. In contrast, our model combines topographically organized grid-to-place projections with local recurrent excitation in CA3, which concentrates learning locally around the represented position and prevents destructive interference at distant locations (Fig. 4d and Appendix C, Fig. 8). We have added this mechanistic explanation to the Results section.

4. Access to anonymized reproduction code (Q1)

Could you release the code (anonymized, of course) for me to view your exact implementations and rerun the experiments, and validate the results?

  • We originally planned to release the code publicly after acceptance. We appreciate your generous assessment of the work and the specific suggestions you made for testing it more rigorously. To support your review, we are happy to share an anonymized subset of our code that reproduces the paper’s core results, together with the configurations and commands needed to run the experiments. The relevant code has been provided in the supplementary material.

Once again, thank you for your thoughtful and constructive review. Your suggestions have substantially strengthened the statistical validation and empirical grounding of our work. We hope that these new multi-seed statistical tests, the pretrained ALPaCA comparison, our mechanistic explanation, and the anonymized codebase fully address all of your questions. If you find that all of your concerns have been resolved, we would be deeply grateful if you would consider raising your score to reflect this strengthened version of the paper.

Reviewer_LgSNOpenReview ↗

Concerns resolved

My concerns have all been satisfactorily addressed. I am raising my score to a 4 in order to reflect this. Good work!

AuthorsOpenReview ↗

Thank You for Your Encouraging Feedback and Increased Score

Thank you for confirming that your concerns have been addressed and for raising your score. Your suggestions prompted stronger statistical validation and broader baseline comparisons, which have improved the paper. We greatly appreciate your encouragement and constructive engagement throughout the discussion.

Reviewer_uMFfOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a computational model with proposed function for theta sweeps, which is to allow near-distance sensory observations from locations the animal hasn’t visited to become bound to spatial representations for the current location. Because sensory content is spatially correlated, each observation then improves the map at many locations, so a cognitive map can be learned from sparse exploration.

The authors test their idea in an entorhinal–hippocampal circuit model in which theta modulation and firing-rate adaptation generate left–right alternating sweeps across grid modules, as seen in experiments.

The model’s architecture builds on but also elaborates the Vector-Hash model: the “building on” portion takes the idea that grids are invariant and the plastic connections are heteroassociatively learned place-sensory synapses that are trained online. The elaboration includes CA3 recurrence, topography in the grid-to-hippocampal projection to respect differences in spatial scale, and a model of sweep generation.

With sweeps, the model learns a landmark-vector map faster than without sweeps and faster than a Vector-Hash control, because it covers unvisited positions. The paper identifies a speed–accuracy tradeoff: that broad sweeps speed-up early learning but limit final precision, and the authors propose an annealing schedule. Annealing improves perofrmance and can be related to the decline of theta power with familiarity.

Critical strengths
  1. The paper provides amechanistic and falsifiable function for theta sweeps.It assembles published components into one circuit, allowing circuit function/performance to be tied to measurable variables such as theta modulation depth, adaptation strength and sweep span. This is useful and important.

  2. The annealing result links a known decline of hippocampal theta power with familiarity to a functional speed–accuracy tradeoff in map learning. It will be an interesting prediction to test in experiments if the modulation can be experimentally manipulated. Also, the appendix appears sufficiently detailed for model reproduction.

Critical weaknesses
  1. The paper doesn’t separate the sweep mechanism from general spatial smoothing in forming sensory associations. During a sweep, the current observation is bound to place representations within a distance set by the sweep span. The paper should test a control in which the association is broadened without sweeps: for example, a no-sweep model whose place-cell activity is convolved during learning with a Gaussian of width matched to the measured sweep span, or whose readout is trained on position-jittered targets. Without this control I can’t tell whether the specific structure of theta sweeps (left–right alternation, the fan shape, scale dependence, etc.) contributes anything specific beyond the association kernel width, in which case the central claim would reduces to the statement that smoothing improves generalization when sensory inputs are spatially correlated. This and the next comment are my main concern for soundness score and the rating.

  2. What drives the main functional differences between modols? There are a few differences between vector-Hash and the sweep model: the readout rule is different (batch pseudoinv without ridge regularization, versus an online normalized delta rule), the grid-to-place projection (random versus spatially structured) is different; and CA3 recurrence versus not is different. Given that no-sweep model (Section 4.1) has an advantage over Vector-HaSH, and the authors attribute it to latter two differences, it means sweeps alone are driving model performance differences.

  3. Though the authors mention Vector-Hash and TEM in their related work and introductory text, the connections between their model and vector-Hash and TEM should be made explicit when describing their model in Section 3. For instance, the fixed grid scaffold, fixed grid-to-place projections, and a single plastic place-to-sensory association uses the Vector-Hash design and principle that grid scaffolds are invariant and memory is formed through heteroassociative plasticity at the sensory interface. Then, it’s clear that the present model innovates by elaborating on/adding to this basic vector-Hash architecture with a sweep generator, and CA3 recurrence topographic grid-hippocampal projections. A model presented as an elaboration of an established architecture, with additional mechanisms, is a larger and more important contribution than a one-off circuit, because it represents forward and building-up progress for the field and users can transfer existing model results onto the current model, so it is more powerful. So, in Section 3 please state which components are inherited from Vector-HaSH and TEM and which are new. The relationship suggests another experiment: run vector-Hash with and without the present model’s sweep generator, with the readout rule fixed.

  4. One of the predictions (sweeps shorten with familiarity) conflicts w/ the literature (Feng, Silva and Foster, J. Neurosci. 2015: theta sequences are absent on the first lap and emerge over laps). The paper should discuss and address this difference.

Questions for authors
  1. Suggestion: please add the smoothing kernel control described above.

  2. Clarification+suggestion: sweeps have 0.25 to 1 m scale (Figure 6d), the LVC tuning scale is 0.6 m, and "unvisited" means more than 0.15 m from the trajectory. How does unvisited-position error depend on distance from the trajectory, in units of the LVC width? I expect the benefit to vanish beyond one correlation length; please show where it vanishes.

  3. Clarification: the abstract and Section 1 call the plasticity Hebbian, but Equation 4 is an error-correcting delta rule that requires the prediction error to be available in LEC. Please clarify the terminology and make consistent. Also, please discuss how the error is computed in/rerpesented in LEC?

  4. Clarification: how does the model reconcile prediction 3 with Feng et al. 2015 and with the dependence of sweep length on running speed?

  5. Suggestion: Show: what happens if you run the Vector-HaSH-plus-sweeps comparison described above?

  6. Suggestion: Show: does the sweep advantage persists if the LEC target is corrupted with noise and when the same LEC pattern occurs at several locations (aliasing)?

  7. Suggestion: please add a single diagram of the complete model that marks the band-cell anchor, the sweep generator, the fixed connections and the one plastic connection.

  8. Suggestion: the no-sweep condition sets theta modulation to zero but retains adaptation, so the grid-cell activity bump may still move away from the anchored position. Please report the measured sweep span in this condition.

8 comments
AuthorsOpenReview ↗

Response to Reviewer uMFf(Part I)

Thank you for your thoughtful review and for highlighting the mechanistic, falsifiable account of theta sweeps and its link to theta annealing. Below, we address your questions regarding spatial generalization versus smoothing, circuit component contributions, architectural connections to Vector-HaSH and TEM, and biological dynamics.

W1 & Q1: How do bump width and sweep extent affect learning, and why use alternating sweeps rather than general spatial smoothing or jitter?

R: Static bump width and dynamic theta sweeps address the trade-off between local positional precision and spatial generalization in different ways. A narrow place-cell activity bump localizes synaptic updates to the animal's immediate coordinates, supporting high precision at visited locations but providing little generalization to unvisited regions. Conversely, broadening a static bump spreads updates across a wider area at the cost of blurring positional selectivity. Theta sweeps address this trade-off by keeping the instantaneous activity bump narrow while moving it sequentially across nearby coordinates as sensory input varies smoothly along the physical path. In this way, a single sensory observation updates multiple distinct place representations while preserving narrow instantaneous place fields. As exploration progresses and the environment is more densely sampled, reducing the sweep range shifts learning from broad spatial generalization to local refinement.

The regional learning curves in revised Appendix C, Figure 7b,c show the transition from early generalization to later local refinement. The visited region includes query locations within 0.15 m0.15\text{ m} of the complete trajectory, including positions reached later in the run. Sweeps initially improve predictions around those future path segments. As exploration proceeds, the condition without sweeps gives greater precision at visited locations, while sweeps retain their advantage in unvisited regions.

We tested a bump broadened to sweep extent on the same three trajectories. At each learning step, we spatially convolve each module's HPC firing-rate map with a Gaussian kernel before applying the original readout feature transformation. The kernel width is calibrated to that module's RMS sweep displacement. This spreads the current sensory association across neighboring place cells simultaneously. Evaluation uses the original stationary query representation.

We also added a random local jitter control. At each step, an isotropic Gaussian displacement is shared across modules and scaled to match each module's radial RMS sweep displacement. Jitter and sweeps use the same learning coefficient. The table compares both controls with sweeps and the default no-sweep reference:

Condition (400 s, n=3)All locationsVisited locationsUnvisited locations
Ours, with sweeps0.506 ± 0.0120.244 ± 0.0040.618 ± 0.013
Bump broadened to sweep extent0.603 ± 0.0140.237 ± 0.0060.760 ± 0.012
Ours, random local jitter0.545 ± 0.0130.323 ± 0.0040.640 ± 0.012
Ours, without sweeps0.611 ± 0.0140.211 ± 0.0060.783 ± 0.011

Broadening the bump to sweep extent lowers all-position and unvisited error relative to the narrow-bump no-sweep reference, while increasing visited error. This illustrates the trade-off between wider generalization and local precision. Alternating sweeps achieve lower all-position and unvisited errors than the broadened bump, while the broadened bump has a slightly lower mean visited error. Sequentially moving a localized bump therefore provides more effective generalization in this task than simultaneously spreading each update across nearby cells.

Mechanistically, memoryless isotropic jitter repeatedly resamples the immediate neighborhood of the already-traversed path, whereas left–right alternating sweeps direct excursions laterally into surrounding unvisited territory. This agrees with the experimental and normative coverage findings of Vollan et al. (2025, Nature), who showed that entorhinal–hippocampal theta sweeps alternate left and right into unvisited space and that minimizing overlap with prior sweep footprints naturally generates left–right alternation. Combined with the dorsoventral gradient of sweep angles reported by Ji et al. (2026, Science Advances) and our finding that broader sensory correlations favor longer sweeps in revised Figure 6b–d, structured multiscale sweeps sample complementary sectors around the animal more effectively than memoryless random local jitter.

AuthorsOpenReview ↗

Response to Reviewer uMFf(Part II)

W2: What are the respective contributions of the MEC–HPC projection, CA3 recurrence, and theta sweeps?

R: The circuit components contribute in three complementary ways:

  1. Topographic MEC$\to$HPC projections (Solstad et al., 2006) align periodic grid-cell inputs whose phases match a given location in space. Because individual grid modules are periodic, however, feedforward alignment alone still leaves secondary input peaks at other locations, and random MEC$\to$HPC projections distribute shared activity more broadly across distant coordinates. During associative readout learning, shared activity across distant locations causes cross-location interference: updating the weights for the current observation perturbs predictions at unrelated positions.
  2. CA3 local recurrent excitation: The recurrent CA3 circuitry and pattern completion studied by Guzman et al. (2016) provide the biological motivation. We combine local excitation with divisive inhibition following the continuous-attractor model of Wu, Hamaguchi & Amari (2008). In our circuit, this suppresses competing secondary peaks and stabilizes continuous bump movement. Adding recurrence eliminates within-cycle nonlocal jumps in the two-module diagnostic in Appendix C.3, Figure 8. Removing it increases prediction error in the five-module learning task in Figure 4c; the activity snapshots in Figure 4d show how recurrence suppresses secondary peaks. The replication results and paired tests are reported in our response to lgSN W1 & Q1.
  3. Theta sweeps extend each sensory observation to surrounding unvisited positions. Figure 3c and Appendix C, Figure 7b,c show reduced error in both our circuit and Vector-HaSH. The latter retains its random MEC$\to$HPC projection, zero CA3 recurrence and pseudoinverse readout. Structured projections and recurrence provide a localized spatial code, while sweeps expand the spatial reach of learning. The Vector-HaSH comparison is detailed in W3 & Q5 below.
AuthorsOpenReview ↗

Response to Reviewer uMFf(Part III)

W3 & Q5: How does the model relate to Vector-HaSH and TEM in Section 3, and can Vector-HaSH itself benefit from theta sweeps when the readout rule is held fixed?

R: Section 3.1 now explicitly explains the relationship to TEM (Whittington et al., 2020, Cell) and Vector-HaSH (Chandra et al., 2025, Nature). We build on their shared principle of separating invariant spatial structure in MEC from environment-specific sensory representations in LEC and binding them through associative plasticity. To study how internal dynamics accelerate map learning from sparse exploration, we integrate the continuous-attractor sweep mechanisms developed in prior work—adaptation-driven bump sweeps (Chu et al., 2024, eLife), the direction–conjunctive–grid circuit for left–right alternation (Ji et al., 2025, Current Biology), the upstream band-cell path integrator (Chu et al., 2025, NeurIPS), and the dorsoventral gradient in sweep angles (Ji et al., 2026, Science Advances)—with topographic MEC$\to$HPC projections, CA3 recurrence, and online HPC$\to$LEC sensory readout (Sections 3.1–3.2).

Our original Vector-HaSH baseline was designed to align with the published model, retaining its random projection architecture and pseudoinverse-based sensory association. We compute the readout with a batch pseudoinverse, jointly fitting the accumulated observations. Following your suggestion, we tested whether Vector-HaSH itself benefits from our theta-sweep generator when its random projection architecture and batch pseudoinverse readout are held fixed (revised Figure 3c and Appendix C, Figure 7b,c). We also compared batch pseudoinverse against sequential normalized delta updates on the identical Vector-HaSH random representation at ϑ=0\vartheta=0 using the best delta coefficient from the search on independent validation trajectories:

Condition (400 s, n=3)All locationsVisited locationsUnvisited locations
VH, with sweeps (batch pseudoinverse)0.552 ± 0.0090.120 ± 0.0040.738 ± 0.006
VH, without sweeps (batch pseudoinverse)0.769 ± 0.0220.195 ± 0.0021.015 ± 0.018
VH, without sweeps (normalized delta rule)0.930 ± 0.0050.832 ± 0.0050.972 ± 0.006

With the pseudoinverse readout held fixed, theta sweeps reduce Vector-HaSH's all-position error, showing that their learning benefit transfers to this architecture. The batch pseudoinverse jointly accounts for overlapping representations across the accumulated observations, reducing the interference that arises when a new association changes the readout at other locations. Sequential delta updates act on one observation at a time, so these changes can overwrite earlier associations. Even with its validation-selected learning coefficient, sequential delta learning gives higher mean all-position and visited-location errors in the three-seed comparison. In our circuit, topographic projections and CA3 recurrence localize place activity, allowing online delta updates to support accurate sensory predictions.

Comments on the paper

2 comments

Dear reviewers,

The authors have now responded to your comments/critiques. As we are nearing the end of the discussion period, please read through the authors' rebuttal and respond if there are points that remain unclear (or to let them know they have addressed your concerns).

Sincerely,

AC

AuthorsOpenReview ↗

To All Reviewers

Dear Reviewers,

Thank you for your careful reading and thoughtful engagement throughout the discussion. Your questions prompted additional experiments and closer examination of our results, which have made the paper stronger. Your feedback also helped us explain the underlying mechanisms more clearly and identify questions worth pursuing further.

We are grateful for the opportunity to discuss our work with you and for the care you have taken in evaluating it. We will continue to build on these discussions as we prepare the final version and pursue the next steps in this research.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.