acceptodds
Under review as a conference paper at ICLR 2027

Environment Diversity Drives Zero-Shot Generalisation of Predictive Spatial Representations

Abstract

The hippocampus rapidly establishes spatial responses in novel environments. Recurrent networks trained to predict their own sensory inputs develop similar spatial representations, with units tuned to place and head direction, but are typically trained and tested in a single environment. How such networks come to transfer to novel environments, and which parts of their representations carry over, is largely unexplored. Here we varied the diversity of training environments to address both questions. A recurrent network was trained only to predict its next view from egocentric views and self-motion, in 1, 5, 20, or 40 of 50 procedurally generated environments that differed in shape, size, texture, and landmarks. The weights were then fixed and the network tested in the remaining, unseen environments. Spatially and directionally tuned units emerged in every network, but generalisation depended on training diversity. The increase in prediction error from seen to unseen environments fell from 89% after training in one environment to 3% after 40 and, with the total amount of training data matched, was 5% after training in 40 environments against 50% in five. In nearly every network, a ring of directionally tuned units rotated together between environments, as head direction cells in the brain do. Their preferred directions were bunched after training in one or five environments but tiled the circle more evenly after training in 20 or 40, and this evenness predicted generalisation across training conditions. Place units tended to rotate with the ring between environments, more closely than with environment geometry. An evenly spread head direction signal, conserved across species and present early in development, may therefore be a spatial primitive that allows maps to form rapidly in novel environments. We predict that, between differently shaped environments, part of the arrangement of place fields is preserved, rotated with head direction cells.

Primary area:
applications to neuroscience & cognitive science
Keywords:
self-supervised learning, recurrent neural networks, zero-shot generalization, spatial representation, NeuroAI, training distribution diversity, cognitive maps

Reviews

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

The authors trains a recurrent neural network to predict upcoming views from self-motion and current view. The training is done on either 1, 5, 20, or 40 environments generated with RatInABox (George et al., 2024) and RatVision (Abrate, 2026). The main findings are:

  1. The prediction loss improve on seen and unseen environments when increasing the number of training environments
  2. A selection of 25 ring units are more uniformly distributed in the models trained on more environments
  3. Place units tends to rotate with the ring units
Critical strengths

The question of how spatial representation remaps across environments is an important question.

Critical weaknesses
  1. Positioning relative to prior work and novelty

Prior work is only briefly covered in the introduction section, which makes it hard to asses the contributions. This work appears closely related to Abrate et al. (2026). As I understand, it uses the same data generation tools RatInABox (George et al., 2024) and RatVision (Abrate, 2026) to train the same architecture, with the same unit classification pipeline. The paper acknowledges this connection, for example by describing the model as “similar to that of Abrate et al. (2026)”. However, it would help to explicitly state how much of the pipeline is inherited and to highlight the contributions that arises from building upon it.

As I understand it, the main difference is the training data generated. Abrate et al. (2026) generated different training groups based on different locomotor development phases in a single environment, whereas this work replaced that with a fix set of environments explored with a single trajectory type. Another difference is that the model is trained to autoregressively predict the next five visual observations, which I assume follows Levenstein et al. (2024) whose setup is also similar. However, because this work does not properly situated itself relative to prior work, the motivations, differences and novelty are not clear.

  1. The main generalization result has previously been discussed

One of the main contributions is that training on more environments improve performance on held-out environments generated by the same procedure (i.e. from the same distribution space), compared with model that were trained on a smaller subset of that space. This finding is consistent with a substantial body of prior work, for example [1-6]. The paper would benefit from discussing these results and clarifying what this finding adds to previous work.

[1] Baxter, J. (2000). A model of inductive bias learning. Journal of artificial intelligence research, 12, 149-198.

[2] Maurer, A., Pontil, M., & Romera-Paredes, B. (2016). The benefit of multitask representation learning. Journal of Machine Learning Research, 17(81), 1-32.

[3] Cobbe, K., Klimov, O., Hesse, C., Kim, T., & Schulman, J. (2019, May). Quantifying generalization in reinforcement learning. In International conference on machine learning (pp. 1282-1289). PMLR.

[4] Zhang, C., Vinyals, O., Munos, R., & Bengio, S. (2018). A study on overfitting in deep reinforcement learning. arXiv preprint arXiv:1804.06893.

[5] Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., & Sutskever, I. (2019). Language models are unsupervised multitask learners. OpenAI blog, 1(8), 9.

[6] Raventós, A., Paul, M., Chen, F., & Ganguli, S. (2023). Pretraining task diversity and the emergence of non-bayesian in-context learning for regression. Advances in neural information processing systems, 36, 14228-14246.

  1. Ring unit analyses would benefit from further support

Ring units are defined as the 25 directional units with the highest co-rotation score. However, Figure 4b shows a continuous spectrum of co-rotation score with not clear separation at this cutoff. To better support that the results follow from this property and not this particular cutofff, all ring-based results should be repeated for different set sizes, e.g. top 10%, 10-20%, etc or by range of co-rotation score (1-0.9, 0.9-0.8, etc.). As of now, it is unknown if the results are robust beyond trends shown for only 25 units. Moreover, Ring units are also drawn from directional units, defined with a less stringent criterion than the head direction units also used in the paper. It would be informative to see whether the results hold when ring units are selected using that standard criterion. Furthermore, the common population rotation is estimated by the mode. Showing the distribution of changes in preferred direction across directional units would confirm that it is unimodal and that the mode is an appropriate summary.

Questions for authors

Section 2 states that the training environment of each arm where "drawn at random from the pool of 50 for each seed, and the remaining environments were unseen and used only for testing.". Does that mean the test set differs across arms? Is my understanding correct that model trained on 1 environment are tested the other 49 environments, and those trained on 40 environment tested on 10? Could the arms be compared on a fixed common held-out set, exact same environments excluded for every arm? This would ensure that the comparison is not influenced by a difference in test environments.

Appendix D mentions that “fields were grown with a threshold rising from 0.1 in steps of 0.05”, but it does not precise how those increasing threshold are applied. Could the authors describe the procedure in more detail?

3 comments

Response to Reviewer oeAu

We thank the reviewer for concrete suggestions, all of which we have taken up.

"explicitly state how much of the pipeline is inherited"

A new table (Appendix B) does this. Our network belongs to a well-established class of predictive models of hippocampal function (Recanatesi et al., 2021; Uria et al., 2020; Levenstein et al., 2024). Inherited from Abrate et al. (2026): its single-layer form, the RatInABox trajectories and the classification thresholds. Adapted: the five-step rollout (following Levenstein et al., 2024, as the reviewer surmises), a self-motion embedding, a carried hidden state, and the RatVision renderer (Abrate, 2026), which we substantially extended to generate procedural environments. New: the procedurally generated environments, training arms that vary the number of environments, the matched-data control, and the ring, clamping and field-frame analyses. The question also differs: what transfers to unseen environments, rather than how locomotor development shapes coding.

"This finding is consistent with a substantial body of prior work [1-6]"

Agreed, and thank you for the references, now discussed in Section 1. We present the setting, not the effect, as new. The paper adds that the effect reflects diversity rather than the amount of data (the increase in prediction error from seen to unseen environments was 5% after training in 40 environments, against 50% in five, with the total amount of training data matched), and shows which parts of the representation carry over: a coherent directional ring forms with any training, but its even tiling needs diversity. It also makes a testable prediction for place cells, a fundamental component of the hippocampal cognitive map, now supported by a causal test: clamping the ring weakens place maps' alignment with it (Fig. A8).

"all ring-based results should be repeated for different set sizes [...] standard criterion [...] distribution of changes in preferred direction"

All three are done (Appendix G):

  • Ring size (Fig. A6): co-rotation (1 if all units rotate by the same angle) falls smoothly from 0.90 for the top 10 units to 0.80 for the top 50 in the 40-environment networks, against chance of 0.28–0.13. From 25 to 50 units, every network trained in 1 or 5 environments has a more bunched ring than every network trained in 20 or 40; this weakens at 10–15 units, where tiling is noisy.
  • Standard criterion (Fig. A6): a ring drawn only from head direction and conjunctive units gives the same results, with tiling tracking the gap more strongly (Spearman correlation across networks ρ = 0.81, against 0.66 for the paper's ring).
  • Distribution: changes in preferred direction had no second peak even half as high as the main one in 88% of environment pairs, against 0–4% after shuffling unit identities, so the mode is a fair summary. Ring co-rotation and tiling do not use it.

Repeating the clamping and place-field analyses for each ring definition needs new runs, so these are not yet done; we would be glad to add them for the camera-ready if the reviewers find them useful.

"Could the arms be compared on a fixed common held-out set"

Your reading is correct: the held-out sets differ (49, 45, 30 and 10 environments for networks trained in 1, 5, 20 and 40), though all come from the same generator and the gap is relative to each network's own seen environments. We are generating a fresh set that no network has seen, as part of the out-of-distribution tests, and will report every network on it during the discussion.

"Appendix D [...] does not precise how those increasing threshold are applied"

We have rewritten it step by step (now Appendix E). Each rate map is min–max normalised; bins above 0.1 form candidate regions (at least 10 bins, peak at least 0.5); each is grown by raising the threshold in steps of 0.05, following regions that split; the field is the lowest threshold that covers at most half the map with split-half stability of at least 0.25.

We hope these revisions address the reviewer's concerns, and we would be glad to answer further questions.

Following up on the common held-out set: all 20 networks were scored on one fresh set that no network had seen. The gaps were 79%, 40%, 16% and 2% after 1, 5, 20 and 40 environments (paper: 89%, 50%, 22%, 3%), and 3% for the matched-data networks, so the ordering is unchanged. Details are in our follow-up general comment and the revised PDF (Appendix G, Fig. A9).

I appreciate that the revision improves the positioning of the paper relative to prior work. Beyond that, however, the responses to my concerns were unfortunately too incomplete to resolve them. The paper’s central result, that training on a more complete subset of data from a generator improves performance on held-out data from the same generator, is expected from prior work. My concerns about the ring unit analyses were only partially addressed. For example, the new results vary the ring size from the top 10-50 units instead of examining disjoint bands across the full score spectrum, which would show that the results follow from co-rotation and not from an arbitrary cut-off of a few top units. Moreover, these new results show that the difference in tiling between training arms weakens for the top 10-15 units, which increases this concern. Therefore, I will unfortunately maintain my score.

Reviewer_sst4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how environmental diversity influences the generalization of predictive spatial representations. The authors train a single-layer recurrent neural network (RNN) with 500 sigmoid units to predict the next egocentric visual observation from the current observation and self-motion signals. The model is trained in 1, 5, 20, or 40 procedurally generated environments and evaluated on held-out environments without further weight updates.

The authors report that increasing the number of training environments substantially reduces the generalization gap between seen and unseen environments. Specifically, the reported gap decreases from approximately 89% with one training environment to approximately 3% with 40 environments. A matched-data experiment further suggests that environmental diversity is more beneficial than simply increasing the amount of training data. The authors also analyze the learned representations, reporting that directionally tuned units form a ring-like organization, that the uniformity of directional tuning correlates with generalization performance, and that place-field bearings are associated with the rotation of the direction ring across environments.

The paper addresses an interesting intersection of self-supervised predictive learning, spatial representation learning, and computational neuroscience. The empirical observation that environmental diversity improves generalization is potentially valuable. However, I have concerns about the strength of the mechanistic claims, the identification of causal factors underlying the observed improvements, the scope of the generalization demonstrated, and the extent to which the results support the proposed neuroscientific interpretation.

Critical strengths
  1. A relevant and interesting research question. Understanding how environmental diversity shapes reusable spatial representations is important for both representation learning and computational neuroscience.
  2. A useful matched-data comparison. Comparing training regimes with different numbers of environments while approximately controlling the amount of training data is a positive aspect of the experimental design. It provides evidence that diversity may contribute beyond data quantity alone.
  3. Potentially valuable scientific hypothesis. The observation that place-field bearings tend to follow the directional population's rotation suggests a testable hypothesis about the role of directional representations in aligning spatial maps across environments.
Critical weaknesses
  1. The causal mechanism is not established, and the representational evidence is limited. The central mechanistic claim—that training diversity produces a more evenly distributed directional ring, which in turn drives generalization—is not causally supported. The correlation between ring tiling and generalization gap (Spearman ρ = 0.66 across 19 networks) is driven primarily by differences between training conditions, while the association vanishes within conditions (ρ = 0.09, p = 0.73). This strongly suggests that ring uniformity is merely a byproduct of training diversity rather than an independent explanatory mechanism. Furthermore, the clamping intervention demonstrates functional importance but does not directly manipulate uniformity to prove causation. Similarly, the place-field analysis is observational, relying on 1D bearings from the environment center, which cannot establish the preservation of a complete 2D spatial map nor prove that the ring causally determines place-field organization. The strong neuroscientific narrative (i.e., the ring as a "general spatial primitive") therefore remains speculative and insufficiently validated.
  2. Statistical reliability is limited, and the scope of generalization is narrow. Several core analyses depend on only 3–5 independent network seeds, which is insufficient for robust statistical inference, particularly when many observations share the same trained network. The lack of clear reporting on statistical units, confidence intervals, and multiple comparison corrections further undermines confidence. In addition, the "zero-shot" claim is overstated: evaluation environments are held out from the same procedural generator family. This is merely an in-distribution or mildly shifted test, not broad out-of-distribution (OOD) generalization. The conclusions should be strictly confined to the current environment generator and model setting.
  3. The simplicity of the architecture limits the breadth of the conclusions, and the contribution remains primarily descriptive. The study relies on a single-layer RNN with 500 sigmoid units and 32×16 grayscale inputs. While controlled simplicity is useful for isolating recurrent dynamics, the concern is that the breadth of the conclusions exceeds the range of settings validated. It is unclear whether the reported relationship between environmental diversity, ring uniformity, and generalization persists with richer visual encoders (e.g., CNNs) or alternative architectures, which might exploit visual cues differently. Consequently, the contribution remains largely descriptive. Environmental diversity improving generalization is an intuitive prior; the paper's novel claim (ring uniformity as a mechanism) lacks independent explanatory power. Moreover, the paper does not demonstrate how this insight can inform a better training strategy or directly test its role in rapid online adaptation since weights are frozen. Without establishing a rigorously validated mechanism or a generalizable design principle, the conceptual and practical significance of the work is limited.
Questions for authors
  1. Causal Mechanism. Given the vanishing within-condition correlation (ρ=0.09), does ring tiling independently explain generalization? Can you directly manipulate ring uniformity (beyond global clamping) to establish a causal link with place-field alignment?
  2. Evidence and OOD Scope. In the place-field analysis, can you provide metrics of full 2D map similarity (e.g., spatial correlation of rate maps after rigid rotation), rather than relying solely on 1D bearings from the environment center? Furthermore, can the frozen network exhibit true out-of-distribution generalization when tested on systematically shifted visual statistics (e.g., unseen textures) or a substantially different procedural generator?
  3. Architecture and Shared Frame. Is this phenomenon (ring organization and improved generalization) dependent on the minimal single-layer RNN architecture? Does this mechanism persist if a deeper visual encoder (e.g., a CNN) or alternative recurrent architecture is introduced? Additionally, can a decoder trained in one environment recover position in an unseen environment without refitting?
  4. Adaptation and Reproducibility. How does a frozen directional ring facilitate "rapid mapping"? Have you tested if brief fine-tuning accelerates adaptation? Will the complete code and generator be provided in an accessible anonymous repository?
4 comments

Response to Reviewer sst4 (1/2)

We thank the reviewer for a thorough review, which has led to narrower claims and new analyses.

"The causal mechanism is not established [...] the association vanishes within conditions"

Agreed. Tiling tracks generalisation across conditions but not within them, so it does not show that evenness causes generalisation. We now say "tracks", state the within-condition result (contributions, Section 3.4, Discussion), and frame the "spatial primitive" as a hypothesis. What holds: the diversity effect on tiling survives other ring sizes and rings of standard head direction units (Fig. A6), and clamping the ring raises prediction error 2.1 times as much as clamping matched controls in seen environments, and 2.4 times as much in unseen ones (five 40-environment networks), so the ring is used.

"the place-field analysis is observational, relying on 1D bearings"

Full maps (Section 3.4, Fig. A7). For each pair of held-out environments, we rotated one environment's activity through a full turn in 5° steps, rebinned it as rate maps, and correlated each place unit's map with its own map in the other environment, over the bins the two environments share. Each correlation is compared with those of other place units at the same rotation, so at each rotation the same bins are compared. We then compared the ring's rotation with the geometric rotation. Because the geometric rotation maximises the area the two environments share, which favours it, we controlled for overlap by regression across pairs and by matching pairs whose overlaps differ by at most 5 bins.

  • Directional tuning sorts the place units: on matched pairs, 95% of the most direction-tuned quarter were better aligned by the ring's rotation, against 11% of the least tuned, whose maps followed geometry. Over all place units the advantage is modest: 56% (95% CI over networks 52–61%), against 72% with bearings.
  • Overlap matters. Uncorrected, the two rotations look tied, because the geometric rotation shares about 9 more bins on average and each shared bin adds about 0.006 to the correlation (about 0.06 in total). At equal overlap, the ring's rotation is ahead by a similar amount, +0.058 (95% CI 0.038–0.079).
  • Controls: 94% of ring units were better aligned by the ring's rotation, and occupancy maps followed geometry.

Causal test (Section 3.4, Fig. A8). Using the clamping runs of Fig. 4d, we repeated this analysis with the 25 ring units, or each of 20 matched control sets, held at their mean activation (five 40-environment networks). Place maps' extra correlation at the ring's rotation fell from 0.046 without clamping to 0.033 with the ring clamped, against 0.038 with controls clamped (ring against controls −0.006, 95% CI −0.011 to −0.001, paired t test over networks p = 0.035). The share of place units better aligned by the ring's rotation at equal overlap fell by 7 percentage points, to 50%, with the ring clamped, against 3 points with controls clamped (difference −4.1 points, 95% CI −5.6 to −2.5, p = 0.002). The effect is modest, as might be expected: the clamp fixes only the 25 ring units, while the other directional units, which also carry heading, are left intact. It nonetheless shows that the ring contributes causally to place-field organisation.

Response to Reviewer sst4 (2/2)

"statistical units, confidence intervals, and multiple comparison corrections"

Every headline test now treats each trained network as one independent sample (new Appendix D; Table A3), with network-level intervals, e.g. the Spearman correlation between ring tiling and the generalisation gap across 19 networks, ρ = 0.66 (bootstrap 95% CI 0.20–0.91, permutation p = 0.004). The Mann–Whitney test in Section 3.3 (ring tiling after 1 or 5 against 20 or 40 training environments) is now two-sided (p = 7×10⁻⁴), and the only family of tests (Fig. 3c) is Holm-corrected. More seeds would help, and we would be glad to add them for the camera-ready if the reviewers find them useful.

"the 'zero-shot' claim is overstated" / Q2 out-of-distribution

Agreed. Zero-shot now means no weight updates in held-out environments from the same generator, and our conclusions are confined to it. We are testing every frozen network on unseen textures and landmarks, more walls and larger arenas; rendering and evaluating the new environments takes some time, so we will report the results later in the discussion.

"richer visual encoders (e.g., CNNs) [...] largely descriptive"

The minimal model is deliberate: place and head direction units emerge in it and every unit can be analysed, and a 128-unit network generalised similarly (generalisation gap 2%, against 3% for 500 units). Other architectures are untested, as the Limitations now say, and we see testing them as important future work. The paper does offer a design principle: at a fixed data budget, diversity matters more than data per environment (the increase in prediction error from seen to unseen environments was 5% after training in 40 environments, against 50% in five, with the total amount of training data matched).

Q1, Q3, Q4

  • Q1: tiling does not independently explain generalisation. Manipulating uniformity directly needs new clamping runs; we would be glad to add it for the camera-ready if the reviewers find it useful.
  • Q3: not yet tested; we would expect a decoder to transfer partly once rotated by the ring's offset, and would be glad to add this for the camera-ready if the reviewers find it useful.
  • Q4: "rapid" describes the brain. With frozen weights, spatial tuning in unseen environments forms through the recurrent dynamics alone: a linear position decoder fitted in each environment had an error of 6.8 cm in unseen environments and 5.4 cm in seen ones, against 37 cm by chance. Whether brief fine-tuning accelerates adaptation needs new training runs, and we see it as important future work. Code was provided with the submission as an anonymised zip; the updated code and the environment generator will be made public at the anonymised link in the paper for the camera-ready.

We hope these revisions address the reviewer's concerns. If so, we would be grateful if the reviewer would consider raising their score. We would be glad to answer further questions.

Following up on W2/Q2: our follow-up general comment and the revised PDF (Appendix G, Fig. A9) report the out-of-distribution tests. The ordering by training diversity held in all five new sets (p = 7×10⁻⁴ in each), and the cost of the larger shifts was smaller after diverse training. Our main claims remain confined to the generator; these tests show how far the effect extends beyond it.

Thank you for the revisions. I acknowledge the improvements in statistical reporting (network-level units, CIs, Holm correction), the full 2D place-field analysis with overlap control, and the new clamping-based causal test for place-field alignment. These are substantive improvements.

I have also noted the other reviewer’s positive assessment that the multi-environment RNN setting is a novel contribution. I agree that this setting has value. However, from methodological standpoint, novelty of the setting does not resolve my core concerns regarding causal mechanism and experimental completeness.

I am therefore maintaining my score of 2.

Key concerns:

  1. Causal mechanism. The within-condition correlation still vanishes (ρ = 0.09), and a direct manipulation of ring uniformity—beyond global clamping—has not been performed. Framing the “spatial primitive” as a hypothesis is honest, but it effectively withdraws the paper’s original mechanistic claim. Another reviewer also notes that ring uniformity lacks independent explanatory power.

  2. Novelty. The authors state they present “the setting, not the effect, as new.” I accept that the setting is useful, but the matched-data control mainly validates a known principle, and the distinction from closely related prior work remains largely engineering-based. The concern that the contribution is primarily descriptive still stands after revision.

  3. Experimental completeness. Cross-environment decoder transfer, broader OOD tests (beyond the same generator), alternative architectures, and repeating the clamping/place-field analyses under each ring definition are all deferred to camera-ready. The current evidence chain is therefore incomplete.

Reviewer_y8hPOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper described results based on training recurrent neural networks to predict future sensory inputs in diverse environment. They found that the network generalizes poorly when trained on a single environment, but can generalize well when trained on many environments. They also reported that, in networks trained on diverse environments, a fraction of the units that are tuned to head direction tend to rotate together between environments. Units identified as place cells also tend to rotate with them.

I found some of the results reported to be interesting. Recurrent networks trained on many environments in the spatial setting has not been studied systematically in previous work. So there is novel contribution in this paper. The results on on head direction units and place cells are interesting has some empirical support from the rodent literature, although I also noted several substantial discrepancy to rodent data.

Critical strengths
  1. The authors trained recurrent neural networks (RNNs) to predict upcoming views in multiple spatial environments and test how the generalization performance of RNNs depend on the task. It was found that the training in many environments enables the RNNs to generalize to new environments.

  2. The paper reported that a group of head direction tuned neurons seem to generalize and maintained a fixed relationship across different environments. Furthermore, the rotation of place fields is better explained by the rotations of the head direction cells compared to the change of the geometry of the environment. This is an interesting observation.

Critical weaknesses
  1. The observation that generalization improves substantially with more training environments, while interesting, isn’t too surprising given the environments generated and tested shared similar statistics and structures. It would be useful if they authors could perform an out-of-distribution test to assess generalization to other types of environments. This limits the strength of the main conclusion in the paper.

  2. While the analysis of the tuning properties in the RNNs are interesting, there are several discrepancies to data from rodent spatial navigation circuits. For example, as mentioned in the paper, head direction cells exist before eye opening, it seems that the model described in the paper (training RNNs in vision prediction tasks in many environments) isn’t adequate in explaining the development of head direction circuit. Also, most of the units classified as head direction cells in Fig. 2c still have substantial spatial selectivity. Meanwhile, the units classified as place cells seem to encode the boundary rather than the interior of the environment. These seem to inconsistent with rodent hippocampus data.

Questions for authors

Fig. 2a, what does “prediction vs. input” mean?

Better quantification is need for Fig. 2b. Some further summary statistics would be useful. It seems that even for the top 25 units (labeled as ring neurons), the residual remains substantial from the heatmaps.

Fig. 2d. What are the units on the y-axis for each panel?

Various Figures show the prediction from the null. How were these computed?

When describing the Model, it was stated that the model is similar to a model used in Abrate et al (2026). How is this model different from the RNNs used in earlier papers that used RNNs to study navigation?

Many studies of rodent hippocampus also reported global remapping when changing environments. It seems that the types of changes in the place fields from the model (e.g., rotation of the place fields) are different from global remapping, which may involve the formation of place fields in seemingly random new locations. Any thoughts on this?

3 comments

Response to Reviewer y8hP

We thank the reviewer for the positive assessment and helpful questions.

"It would be useful if the authors could perform an out-of-distribution test"

We are now testing every trained network on out-of-distribution environments, with unseen textures and landmarks, more walls and larger arenas, and will report results during the discussion. We also agree that the in-distribution result is expected: "zero-shot" now means held-out environments from the same generator, and our conclusions are confined to it (Section 1). Beyond the expected benefit of diversity, the matched-data control shows that diversity, not the amount of data, drives it: the increase in prediction error from seen to unseen environments was 5% after training in 40 environments, against 50% in five, with the total amount of training data matched.

"head direction cells exist before eye opening"

Agreed: the model does not explain how head direction coherence first arises, which we now state (Sections 1 and 4); its visual input stands in for multisensory cues. Ring units rotated together after any amount of training, while even tiling needed diverse experience.

"head direction cells [...] still have substantial spatial selectivity [...] place cells seem to encode the boundary"

We apologise for not making this clear. The network is a model of the hippocampus and its cortical inputs, including parietal and entorhinal cortex, not of the hippocampus alone, so its units show more mixed tuning than neurons in any one region. This is a deliberate simplification, now stated in Section 2 and the Limitations. The tuning classes follow Abrate et al. (2026) and are thresholds, not pure types, and conjunctive place and head direction tuning is found in hippocampal cells of rats, bats and mice (Acharya et al., 2016; Rubin et al., 2014; Jercog et al., 2019). Boundary-related firing is also common in rodents (O'Keefe & Burgess, 1996; Solstad et al., 2008; Lever et al., 2009), and in our new 2D analysis the maps of weakly direction-tuned place units followed arena geometry, while those of strongly tuned ones followed the ring (Fig. A7). A ring drawn only from standard head direction units gives the same results (co-rotation 0.84, against 0.86 for the original ring; correlation of ring tiling with the generalisation gap ρ = 0.81, against 0.66; Fig. A6).

Figs. 2a, "2b", "2d" and the nulls

  • "Prediction vs. input" is the distance between the prediction and the current view. It is close to "label vs. input" (true next view against current view), so the network predicts change rather than copying; its one-step L1 prediction error in unseen environments is 0.08, against 0.120 for copying the input (40-environment network).
  • "2b" and "2d" are, we believe, Figs. 4b and 4d. In 4b, ring units are within 15° of the common rotation of all directional units in 39% of pairs of environments, against 21% for other directional units and 8% by chance (40-environment networks; 43% relative to the ring's own rotation, Section 3.3): the claim is about the population. In 4d, the y-axis is the effect of clamping divided by the mean effect of 20 matched control sets.
  • Each null recomputes the statistic after shuffling unit identities (co-rotation, bearings), or uses random directions (tiling) or matched control sets (clamping).

"How is this model different from the RNNs used in earlier papers?"

Earlier navigation RNNs (e.g. Banino et al., 2018; Cueva & Wei, 2018) are trained to report position or heading. Ours sees only egocentric views and self-motion and is trained only to predict its next view. Appendix B lists what is inherited, adapted and new.

"global remapping"

Remapping is often treated as random. Our results suggest it has a partly predictable component: part of the arrangement of place fields rotates with the head direction ring. This is not the whole story, as much of each map still changes. In a fixed frame, as remapping is usually measured, a place unit's rate map correlated with its own map in another environment at r = 0.33 (Pearson correlation over spatial bins), against 0.18 with other place units' maps (40-environment networks): partial rather than complete remapping, consistent with evidence that remapping is less random than often assumed (Whittington et al., 2020; Fenton, 2024). At equal arena overlap, maps were better aligned by the ring's rotation than by geometry, and clamping the ring weakened this (Figs. A7 and A8). The model predicts that realigning place maps by the rotation of co-recorded head direction cells should reveal the preserved part.

We hope these revisions address the reviewer's concerns. We would be glad to answer further questions.

Following up on your out-of-distribution question: results are in our follow-up general comment and in the revised PDF (Appendix G, Fig. A9). With unseen textures, unseen landmarks, larger arenas and more walls, the benefit of diverse training held in every set, and networks trained in 40 environments predicted as well as in their seen environments.

Thank you for your detailed response and explaining the additional analysis. The revision clarified certain aspects of critiques from the reviews. I consider that there are some interesting results reported in this paper, so I would like to maintain my positive rating. But I also agree with some of the weaknesses raised by other reviewers (such as the results are mainly discriptive even with the clamping manipulation), and could see why they viewed it less favorably.

Comments on the paper

3 comments

General response to all reviewers

We thank all three reviewers for careful and constructive reviews. We have uploaded a revised PDF; all changes and additions are highlighted in blue. In brief:

Text changes

  • Narrower claims: ring tiling tracks generalisation across training conditions, not within them; "zero-shot" means held-out environments from the same generator; the "spatial primitive" is a hypothesis.
  • Statistics: networks are the unit of every headline test (new Appendix D), and the Mann–Whitney test in Section 3.3 (ring tiling after 1 or 5 against 20 or 40 training environments) is now two-sided (p = 7×10⁻⁴).
  • Positioning: a table of what is inherited, adapted and new relative to Abrate et al. (2026) (Appendix B), and a related-work paragraph on diversity and generalisation (Section 1).
  • Clarifications: the network is a model of the hippocampus and its cortical inputs, including parietal and entorhinal cortex; the tuning classes are thresholds, not pure types; eye opening; field detection.

New analyses (Appendix G of the revised PDF)

  • Ring robustness: rings of 10–50 units, and rings drawn from standard head direction units (Fig. A6).
  • Full rate maps compared at equal arena overlap (Fig. A7).
  • A causal test: clamping the ring weakens place maps' alignment with it (Fig. A8).
  • A table of network-level statistics (Table A3).

In progress: out-of-distribution tests of every trained network (unseen textures and landmarks, more walls, larger arenas), with a fresh in-distribution control set.

Beyond this revision: more seeds need new training runs, which we would be glad to add for the camera-ready if the reviewers find them useful; we see other architectures and fine-tuning as important future work. Code was provided with the submission as an anonymised zip; the updated code and environments will be made public at the anonymised link in the paper for the camera-ready version.

Follow-up: out-of-distribution results

As promised in our general response, we tested every trained network, with weights frozen, on five new sets of 10 environments (2,000 s each): a fresh control from the same generator with a new seed, which no network has seen (also the common held-out set Reviewer oeAu asked for); new textures, with every wall and floor texture replaced by a procedural one not in the training pools, matched in mean and contrast, in the same arenas and along the same paths as the fresh set; new landmarks, ten new shapes on the same walls; larger arenas, 1.5–2.0 m against 0.8–1.2 m in training; and more walls, 11–12 against 4–9. We report the generalisation gap as in the paper: the increase in one-step prediction error relative to each network's seen environments.

Check. On the fresh control, the error of using the current view as the prediction was 0.120, as on the original unseen environments, and the gaps after training in 1, 5, 20 and 40 environments were 79%, 40%, 16% and 2%, against 89%, 50%, 22% and 3% in the paper.

Results.

  1. The benefit of diversity held in every new set: every network trained in 1 or 5 environments had a larger gap than every network trained in 20 or 40 (exact two-sided Mann–Whitney p = 7×10⁻⁴ in each set). Across the four shifted sets, gaps were 67–92% after one environment, 29–47% after five, 6–14% after 20, and −6% to −1% after 40.
  2. The 40-environment networks predicted as well in every shifted set as in their seen environments. Because frames in the shifted sets change less from step to step (copying error 0.101–0.113, against 0.120), we also compared prediction error relative to that copying error, against the fresh set. Larger arenas and new textures cost every arm something, but less after diverse training: larger arenas +25% after one environment, +23% after five and +11% after 40; new textures +16%, +13% and +4%. New landmarks and more walls cost every arm 3% or less.
  3. Networks trained in 40 environments with the five-environment data budget had a smaller gap than five-environment networks in every set (Welch p ≤ 0.012).

These are 10 environments per set, within one simulator and renderer. We have uploaded a revised PDF that adds them to Appendix G (Fig. A9) and says in Sections 1 and 4 how far the claims extend; all changes are in blue.

Dear Area Chair,

Thank you for handling our submission, and the three reviewers for their engagement. Below we briefly summarise the discussion, the changes we made, and the points that remain open.

Reviewer positions. Reviewer y8hP keeps a positive score and finds the results interesting. Reviewer sst4 describes the network-level statistics, the overlap-controlled 2D place-field analysis and the new causal clamping test as "substantive improvements" and agrees the multi-environment setting has value, but keeps their score over mechanism, novelty and completeness. Reviewer oeAu acknowledges the improved positioning but keeps their score, mainly over the ring analyses and the expectedness of part of the result.

Changes during the discussion. We focused on analyses we could do rigorously with the 20 networks already trained:

  1. Statistics: every headline test uses networks as the unit, with intervals (Appendix D, Table A3).
  2. Place fields: full 2D rate maps compared at equal arena overlap (Fig. A7), and a causal test: clamping the ring weakens place maps' alignment with it more than clamping matched controls (paired t test over networks, p = 0.002; Fig. A8).
  3. Ring robustness: rings of 10–50 units, rings from standard head direction units, and the unimodality of changes in preferred direction (Fig. A6).
  4. Out-of-distribution tests of all 20 frozen networks on five new sets: a fresh held-out set no network had seen, unseen textures, unseen landmarks, larger arenas and more walls (Appendix G, Fig. A9).
  5. Positioning: what is inherited, adapted and new (Table A1), related work on diversity, and narrower claims throughout.

Remaining concerns.

  • Mechanism (sst4, y8hP). We agree that ring tiling is correlational within training conditions, and we now say so. The paper does include three interventions: on the training distribution (the matched-data control separates diversity from quantity), on the ring (clamping it raises prediction error 2.1–2.4 times as much as matched controls), and on place maps (clamping the ring weakens their alignment with it). A direct manipulation of ring uniformity needs a new intervention design and its own controls; we could not do this properly within the discussion period, and offered it for the camera-ready.
  • Broader OOD tests (sst4). These were completed after sst4's reply. New textures and landmarks outside the training pools, arenas up to 2.0 m, and 11–12 walls all lie outside the training generator's range. In every set, every network trained in 1 or 5 environments had a larger gap than every network trained in 20 or 40 (exact two-sided Mann–Whitney p = 6.7×10⁻⁴ in each), and the 40-environment networks predicted as well as in their seen environments. Tests outside this simulator remain open.
  • Disjoint score bands (oeAu). We computed these during the discussion but reported only the cumulative top-N rings in the revision and in our reply; we should have included them. With the co-rotation ranking split into disjoint deciles, co-rotation in the 40-environment networks falls from 0.85 in the first decile to 0.56 in the fifth (its null 0.16). The diversity effect on tiling is complete in the top two deciles (every 1- or 5-environment network more bunched than every 20- or 40-environment network, p = 6.7×10⁻⁴) and fades in deciles 3–5. So the tiling difference is carried by the most strongly co-rotating units and weakens as co-rotation weakens; it does not depend on a cut at 25 units. At 10–15 units, tiling is measured on very few directions and is noisy (the null mean is 0.42 at 10 units, against 0.15 at 25).
  • Novelty (sst4, oeAu). We agree that diversity improving in-distribution generalisation is expected, and the revised introduction says so. Beyond this, the setting shows that the benefit reflects diversity rather than quantity (5% against 50% with data matched), which parts of the representation carry over (a coherent directional ring forms with any training, but even tiling needs diversity), and gives a testable prediction for place cells; the new tests also show that the ordering by training diversity holds outside the training distribution.

Deferred analyses. More seeds, alternative architectures and fine-tuning require retraining networks; each takes multiple hours on our single GPU, so one new architecture across all arms is several days of compute before any analysis. Cross-environment decoder transfer, repeating the clamping and place-field analyses for each ring definition, and a direct manipulation of ring uniformity need new runs and controls. Instead of reporting rushed versions, we prioritised the concerns raised as central (statistical units, the place-field analysis, causality and OOD generalisation).

We believe the revision addresses what was feasible within the discussion period, and hope this summary is useful when weighing the reviews.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.