acceptodds
Under review as a conference paper at ICLR 2027

Geometry of Cognitive Maps in Minimal Representations

14 views

Abstract

An effective cognitive map organizes states into a compact coordinate system whose geometry—the relative arrangement of states—preserves task-relevant relations and makes them explicit for inference and planning. Yet determining how task structure constrains this geometry is difficult in neural population recordings and artificial neural networks, where representations are obscured by noise, finite sampling, learning dynamics, and architectural choices. Here, we model cognitive maps as regularized hidden-layer activations in linear neural networks trained to predict the next state. Building on previously derived theoretical representations, we prove that their dimensionality can be analytically computed as the number of connected components plus the number of globally compatible independent actions, minus one. Applying this theory across spatial and non-spatial tasks, our results reveal an organizing principle of cognitive maps not captured by standard state-transition models: how transitions are grouped into relational operations could shape cognitive-map geometry. The theory also provides several insights over neural representations, including how shared action encoding supports structural transfer and how increasing similarity among action representation (i.e. action merging) may serve as a mechanism for acquiring a low-dimensional cognitive map. Cognitive-map representations therefore reflect an agent's internal definitions of states and relational operations rather than the external environment alone. By translating candidate task models into quantitative predictions of neural population geometry, our theory may also open a principled route to inferring how an agent internally represents a cognitive map.

Primary area:
applications to neuroscience & cognitive science
Keywords:
cognitive maps, hippocampal-entorhinal representations, representation geometry, predictive linear neural networks

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_gd3aOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

Motivated by mammalian cognitive map representations, this paper gives a simple linear model to learn a map representation that reflects the task structure and shows some analyses on the resulting representation, with discussion relating these to biological facts. The task includes navigation through 2D or 3D space or a family tree.

Critical strengths

None

Critical weaknesses

The presented model is overly simplistic so that it does not give sufficient insight into the biological system. The model is essentially a linear model (though the authors emphasize ANN), where state transition is stipulated as linear addition of state and action vectors. Path integration is built in so that state indices are predefined. In such setting, representation learning and transfer are rather trivial and the presented results contain little surprise. Comparison with biology is only conceptual with no quantitative assessment.

Questions for authors

The actual mammalian hippocampal formation has various complex components like hippocampus and entorhinal cortex (grid cells). How do these compute and interact to give rise to cognitive maps and cross-domain generalizations?

What are quantitative relations between the obtained results and the experimental data? Sec 3.2 claims biological consistency, which is unclear.

I would not think that the family tree example works for larger scales. It is because nodes in deeper positions must be represented more densely so that displacements with the same type of relation depend on the depth. Is there any way to solve this problem in this framework?

What are relationships with previous studies: Sorscher et al (2023), Iyer et al (2025), Hosoya (2024), Gao et al (2021)?

== Minor comments ==

Sec 2.1 needs more informal explanation. What's the meaning of Definition and Theorem.

Sec 2.2. Why do you assign 1-dimensional coordinate to each state?

8 comments
AuthorsOpenReview ↗

Response (Part 1)

Thank you for your thoughtful suggestions.

We believe that some of the concerns may reflect a difference in how the scope of our work is interpreted. Our goal is not to explain how biological mechanisms generate specific phenomena such as grid-cell firing. Rather, we provide a normative theory of what the low-dimensional manifolds underlying cognitive maps are, how they arise from state–action structure, and, crucially, why different task structures produce different representational geometries. Our framework provides mathematically interpretable answers to these questions. We believe that identifying these organizing principles is important for understanding the nature of cognitive maps and can provide insights for experimentalists and theorists investigating their biological implementation.

We strongly encourage you to first read our clarification regarding the interpretation of the “action-consistent coordinate system” as a theoretical framework linking behavior, cognition, and neural population geometry, rather than as an account limited to neural population geometry alone. We sincerely hope that this clarification addresses some of your questions and offer more insights to the theory.

Below, we address your thoughtful questions.

Weakness. Model simplicity and biological insight.

We thank the reviewer for raising this concern. We agree that our model is not intended to reproduce the biophysical details of neural circuits; rather, it is a deliberately simplified normative model. The purpose of using a linear network is not to claim that biological computation is linear, but to isolate and characterize exactly the component of representational geometry that is necessarily imposed by task structure, independently of architectural choices, initialization, and optimization dynamics. Importantly, the model is provided only with categorical state and action identities: no spatial coordinates, distances, or path histories are supplied. Thus, although state identities are indexed, their representational geometry is not predefined and must be learned through next-state prediction.

We also agree that the additive state–action form constitutes a compositional inductive bias. However, this assumption does not make the resulting geometry or dimensionality trivial. A shared action displacement must be globally consistent across every occurrence of that action and across all alternative paths and cycles in the transition graph. These constraints can make nominally distinct actions dependent, collapse task dimensions, or prevent a globally consistent coordinate from existing. Our main theoretical contribution is to characterize these nonlocal constraints and to derive exact predictions for the learned dimensionality and principal-component subspace. For example, the theory predicts when shared action codes permit structural transfer and when merging relations reduces representational dimensionality, predictions that are not specified merely by assigning state indices.

We therefore view the linear solution as an analytically tractable and falsifiable baseline, rather than as a complete mechanistic model of the biological system. It identifies which population-level geometric signatures follow from a proposed state–action organization; deviations from these signatures in nonlinear networks or neural data can then reveal contributions beyond this minimal account. Our nonlinear-network analyses further show that the predicted geometric signatures are not restricted to an exactly linear implementation. Nevertheless, we agree that the present comparison with biology is primarily at the level of representational geometry rather than a direct quantitative fit to neural recordings, and we will try our best to revise the manuscript to state this scope and limitation more explicitly.

Q1. How might hippocampal and entorhinal components interact to support cognitive maps and cross-domain generalization?

Our current framework is formulated at the computational rather than circuit-mechanistic level and therefore does not explicitly model interactions between the hippocampus and entorhinal cortex. One possible interpretation is that the state-specific component represents hippocampal-like coding of particular entities or locations, whereas shared action increments provide an entorhinal-like relational scaffold that can be reused across environments or cognitive domains. Their combination would allow distinct contents to occupy a common relational coordinate system. We emphasize, however, that this is a hypothesis suggested by the model, not an established mapping between network variables and brain regions.

AuthorsOpenReview ↗

Response (Part 2)

Q2. What is the quantitative relationship between the model and experimental data?

We agree that the wording in Sec. 3.2 was too strong. The present analysis does not directly fit neural recordings. Quantitatively, the model predicts that using shared rather than environment-specific actions reduces the task-determined dimensionality from five to three: two relational dimensions are aligned across environments, while one dimension distinguishes environment identity. This is qualitatively consistent with the reported preservation of environment identity in hippocampal representations and stronger cross-environment alignment in prefrontal representations, but it is not yet a quantitative account of those data.

Q3. Can the family-relation example scale to deeper and more densely branching trees?

We respectfully disagree that increasing tree depth requires relation displacements to become depth-dependent. This concern applies to drawing a growing tree within a bounded two-dimensional region, whereas our coordinates are neither bounded nor required to uniquely encode every individual. A parent–child increment can remain constant as the generation coordinate extends with depth. Importantly, this result is non-trivial: the network receives only categorical state and action identities, without predefined family coordinates. Global consistency across alternative relational paths reduces five nominal inverse-relation pairs to three independent dimensions, from which generation- and sex-related structure emerges. Thus, the geometry is derived from the action-labeled graph rather than imposed by the visualization.

Q4. How does the framework relate to Sorscher et al., Iyer et al., Hosoya, and Gao et al.?

These studies and ours address complementary questions. They primarily investigate the neural mechanisms or learning processes through which grid-cell-like and relational representations emerge. In contrast, we ask how a given state–action task structure constrains the geometry and dimensionality of its minimal representation. Thus, their work concerns how cognitive maps may be learned or implemented, whereas ours characterizes what representation geometry is implied by the task structure.

Q5. What is the informal meaning of the definition and theorem in Sec. 2.1?

Informally, Definition 2.1 selects, among all networks realizing the required next-state predictions, those that use the smallest combined hidden-representation and readout norm. This provides a precise notion of a minimal predictive representation. The following theorem states that, although different solutions may rotate this representation across hidden units, they all preserve the same intrinsic distances and PCA geometry. Thus, the individual neurons are not uniquely determined, but the population-level geometry relevant to our analysis is.

Q6. Why is a one-dimensional coordinate assigned to each state in Sec. 2.2?

A scalar function f(s)f(s) represents one candidate axis of the cognitive map, not the complete representation of a state. For example, horizontal and vertical position are two separate scalar coordinates. The space F\mathcal F contains all scalar coordinates whose changes are consistent with the action labels, and a basis of its nonconstant part supplies the multiple axes of the full representation. Starting with scalar coordinates therefore allows us to count the number of independent representational dimensions without assuming that the cognitive map itself is one-dimensional.

Reviewer_gd3aOpenReview ↗

Reply to the authors

I thank the authors for clarification. I carefully read through the comments. However, although the authors stress that the goal of the study is representational insight rather than biological relevance, the gained insight seems still simple and it is not clear how much these are significant in the presence of the raised list of prior studies.

Regarding the family relation example, even though the model manages to learn a linear representation for such relations, it is not clear whether such representation gives us meaningful insight. I imagine that the nodes deep in the tree can be overlapping to each other, and it may not show any discernible structure. Perhaps, this part needs much more thorough investigation.

Together, I could not find enough materials to change my initial assessment.

Reviewer_xdh9OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

Cognitive map is a widely used concept for understanding the structure of the brain’s internal representations that support behavioral tasks. So far, it remains unclear how the structure of the tasks may constrain the geometry of the cognitive maps in trained the brain (as well as in artificial neural networks). The authors molded cognitive maps in linear neural networks using a next-state prediction task. They developed some theoretical results on how dimensionality of cognitive maps are determined by the task structure, and applied this theory to several spatial and non-spatial tasks. The key insight is that the structure of cognitive maps reflect the states and relational operations (some of which may not be captured by the external environment alone).

Overall, this is an interesting study with substantial merit. The logic of the paper is well presented and clear. I believe these results will help the community to better understand the structure of cognitive maps. While the writing is generally clear, the paper may benefit from clarifying several important technical points to (i) better convey the intuition; (ii) clearly state how the theory can be falsified.

Critical strengths
  • The combination of analytical results and numerical results in the context of several applications is impressive. The authors first presented theoretical results, based on adapting prior results on the minimum-representation-norm solutions presented in Braun et al (2025) to the domain of cognitive maps. They numerically tested the theoretical predictions in a number of tasks, including both spatial and non-spatial tasks and their combinations. They also showed that the results generalize to non-linear neural networks.

  • The findings that (i) action merging affects the structure of cognitive maps (section 3.3) and (ii) factorization of task variables to form shared representations (section 3.4) are interesting insights, and may have implications in understanding the neural data collected in a variety of behavioral tasks.

Critical weaknesses
  • The theoretical results in Section 2 can be better organized to emphasize the intuition behind Theorem 2.5. It would also help to clearly state what was already known in Braun et al (2025), and what was new. Furthermore, it would also be useful to explain in what ways the theoretical predictions can be falsified. The last point also limited the ability to assess how strong the empirical results are.
Questions for authors
  1. What is the precise theoretical prediction in terms of the dimensionality of neural activity (obtains in the PCA analysis) in the trained linear neural network? For example, Fig. 1c stated that d_{task} =2. The plot shows two equally important PCs as the top two PCs, but there are also many more PCs that are needed to fully account for the variance of neural response. How well does the theoretical prediction match the empirical results in the simulations? Similar questions apply to the PCA analysis for other tasks. This seems to be a critical point that need to be clarified.

  2. G is an important quantity that represents the number of connected components of the state-transition graph. It is difficult for the readers to understand how this is computed in the current version. I think adding a figure to SI to explain how G is computed in one or two simple examples would help.

  3. Can you provide more intuitions of Theorem 2.5?

  4. The theorems appeared to rely on L_2 norm. Would training RNNs using L_1 regularization change the results?

5 comments
AuthorsOpenReview ↗

Response (Part 1)

Thank you for recognizing the value of our work and for your thoughtful suggestions.

We first clarify how the “action-consistent coordinate system” provides a theoretical framework connecting behavior, cognition, and neural population geometry in a message to all reviewers. We hope this clarification helps explain the intended scope of the theory and addresses some of your questions.

In response to your comments, we also conducted additional experiments and expanded the methodological explanations and visualizations. These results are available in the “Reviewer xdh9” folder of the newly submitted supplementary material.

We address each of your questions below.

Weakness. Innovativeness concern.

It would also help to clearly state what was already known in Braun et al (2025), and what was new.

We wish to clarify our theoretical contribution. Braun et al. (2025) undoubtedly provide powerful mathematical tools. However, our central concept—the action-consistent coordinate system, namely the vector space FF introduced in Section 2.2—is independent of their MRNS framework. Within the network implementation, our contribution beyond Braun et al. (2025) is to identify the importance of balanced visitation, a factor that has received little attention. We were surprised that such an elegant result had not, to our knowledge, been reported previously, given the simplicity of MRNS and the widespread use of MDP to study cognitive-map representations. Moreover, the well-structured representations induced by BV enable us to mathematically characterize how representational geometry changes when a perfect representation is perturbed (see Appendix C).

More broadly, as noted in our clarification, our theory can be understood as a framework linking behavior, cognition, and neural population geometry. Establishing this link is itself a non-trivial contribution, as illustrated by several examples in our paper.

AuthorsOpenReview ↗

Response (Part 2)

Q1. Can the task-determined subspace and its dimensionality be characterized precisely?

What is the precise theoretical prediction in terms of the dimensionality of neural activity obtained in the PCA analysis of the trained linear neural network? How well does the theoretical prediction match the empirical results in the simulations?

We first clarify the scope of our theoretical prediction. The action-consistent coordinate space FF is defined at the level of the task: given a set of states, action identities, and action-labeled transitions, it contains all state coordinates for which each action produces a state-independent displacement. Its dimensionality is therefore determined by the specified task structure and, in particular, by which transitions are treated as instances of the same action. After removing the globally constant component, the theory predicts dtask=dim⁡F−1=G−1+dincd_{\mathrm{task}}=\dim F-1=G-1+d_{\mathrm{inc}} independent nonconstant action-consistent coordinates.

This task-level result should be distinguished from our prediction about neural population geometry. The latter applies to the specific linear predictive architecture and minimum-representation-norm solutions analyzed in our paper. Under balanced visitation and the other assumptions of the theorem, the nonconstant subspace of FF determines the dominant dtaskd_{\mathrm{task}}-dimensional PCA subspace of the representation. Thus, the theory predicts the dimensionality and orientation of a task-determined leading subspace, rather than the total dimensionality or full PCA spectrum of an arbitrary neural representation. It does not imply that every trained network capable of solving the task, particularly a nonlinear network or one that has not reached the relevant solution, must have exactly dtaskd_{\mathrm{task}} dimensions.

In an ideal scenario with known action labels, the dimensionality can be directly determined from the null-space formulation, exactly as dtask=G−1+dincd_{\mathrm{task}}=G-1+d_{\mathrm{inc}}. But for an arbitrary network representation, exact eigenvalue multiplicity is unlikely to hold in practice. Theorem C.1 in Appendix C instead provides a sufficient criterion for assessing whether its leading subspace remains close to a candidate task subspace. Specifically, we can construct candidate models with different action assignments and derive a predicted subspace U0U_0 and dimensionality for each. We then compare each candidate with the observed Gram matrix G1G_1 of the network. Specifically, relative to U0U_0, we define A1=U0⊤G1U0A_1=U_0^\top G_1U_0, B1=U0⊤G1U0,⊥B_1=U_0^\top G_1U_{0,\perp}, and D1=U0,⊥⊤G1U0,⊥D_1=U_{0,\perp}^\top G_1U_{0,\perp}, together with δ1=λmin⁡(A1)−λmax⁡(D1)\delta_1=\lambda_{\min}(A_1)-\lambda_{\max}(D_1) and b1=∥B1∥2b_1=\lVert B_1\rVert_2. If δ1>2b1\delta_1>2b_1, Theorem C.1 guarantees a positive eigengap at the predicted dimensionality and bounds the rotation of the observed leading subspace away from the candidate task subspace. Thus, the null-space formulation determines dimensionality from a specified task model, while the eigengap criterion assesses the robustness of its predicted subspace in an observed representation. This supports evaluating candidate internal action assignments, although it does not by itself establish that a candidate is uniquely identifiable.

AuthorsOpenReview ↗

Response (Part 3)

Q2 & Q3. Schematic and intuition for Theorem 2.5

G is an important quantity that represents the number of connected components of the state-transition graph. It is difficult for the readers to understand how this is computed in the current version. I think adding a figure to SI to explain how G is computed in one or two simple examples would help. Can you provide more intuitions of Theorem 2.5?

We thank the reviewer for this helpful suggestion. We have added a new schematic in the supplementary material and hope it will help to clarify the intuition behind Theorem 2.5 together with the explanation below.

The key problem Theorem 2.5 addresses is how many coordinate, or dimensions, are needed to represent a cognitive map. This is not an obvious question, because a task structure can be highly complex, with arbitrarily many defined actions and a complicated graph structure. We prove that the required dimensionality can be decomposed into two parts: G−1G-1 and dincd_{\mathrm{inc}}.

Here, GG is the number of connected components. In the proof, GG equals the dimension of the null space of (Y⊤−S⊤)(Y^\top-S^\top). The key insight is that when a state is connected to a graph of other states, its coordinates in the subspace are constrained by those states. If snews_{\mathrm{new}} is connected to solds_{\mathrm{old}} through action ato-newa_{\mathrm{to\text{-}new}}, then the coordinates of snews_{\mathrm{new}} must equal those of solds_{\mathrm{old}} plus the increment associated with ato-newa_{\mathrm{to\text{-}new}}. In contrast, if two graphs are disconnected—for example, if they represent New York and Paris—the coordinates within each graph can be translated independently without affecting the other. Intuitively, the G−1G-1 dimensions are used to distinguish different maps. Even if two cognitive maps have the same latent structure, their identities must still be distinguished.

The term dincd_{\mathrm{inc}} represents the number of latent variables within a cognitive map. To understand its value, consider a task with one Stay action and pp pairs of opposite actions, such as Left and Right. In principle, each action could contribute one dimension. However, Stay contributes no dimension because its increment is always zero. For inverse action pairs, once the increment of one action in an opposite pair is determined, the increment of the other must be its negative. Thus, these action pairs can contribute at most pp dimensions. Moreover, the actual dimensionality can be reduced further by path relations. For example, in a family graph, a father may reach his son either directly or through the mother. The increment from father to son must therefore equal the increment from father to mother plus the increment from mother to son. This equality means that these three actions can contribute at most two dimensions.

Q4. Theorem application to the L1L_1 norm and RNNs.

The theorems appeared to rely on L2L_2 norm. Would training RNNs using L1L_1 regularization change the results?

We thank the reviewer for this suggestion. We trained ANNs and RNNs with different initializations on the 4×54\times5 grid-navigation task (see the supplementary material for details). Across random seeds and initializations, networks exhibited diverse and inconsistent representational geometries (That's why we want to study a linear theory to make things cleaner). Neither architecture consistently recovered the ideal geometry, with deviations being particularly pronounced for RNNs.

We would also like to clarify the derivation and scope of our theorem. An action-consistent coordinate system is a network-independent concept, but its emergence in learned network representations follows from the MRNS and balanced-visitation condition. We are not sure that other regularizers or RNNs could not yield the same result in general. Characterizing such conditions especially for RNNs is an interesting direction for future work.

Reviewer_HG1LOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper models a cognitive map as the hidden layer representation of a two layer linear network trained to predict the next state from one hot (state, action) inputs, and analyzes the minimum-representation-norm solutions (MRNS) of Braun et al. (2025), whose PCA geometry is unique up to rotation and therefore determined by task structure alone. Under balanced next state visitation (Assumption 1) and a shared stay action (Assumption 2), the authors prove that the dominant PCA subspace of any MRNS is spanned by the nonconstant action-consistent task coordinates, that is, real valued functions on states whose change along every observed transition depends only on the action label (Theorem 2.4). They further show that the task dimensionality decomposes as dtask=G−1+dincd_{\mathrm{task}} = G - 1 + d_{\mathrm{inc}}, where GG is the number of connected components of the transition graph and dincd_{\mathrm{inc}} counts globally path consistent independent action increments (Theorem 2.5). The theory is illustrated on a series of constructed tasks: 2D and 3D grid navigation and a family relation pedigree (predicted dtaskd_{\mathrm{task}} of 2, 3, 3, matching numerical spectra); two disconnected mazes with shared versus map specific action labels (3 versus 5 dimensions), motivated by Tang et al. (2023); the 12 object graph of Garvert et al. (2017) with edge specific versus direction merged action labels (11 versus 2 dimensions, participation ratio 11.00 versus 7.70), where a width constrained linear network trained by gradient descent spontaneously makes same direction action embeddings similar; and a composed navigation plus family task showing collapse under state superposition and coexistence under Cartesian product states. Nonlinear network simulations (Tables S1 and S2) and neuron subsampling (Figure S4) probe generalizability. The authors conclude that cognitive map geometry reflects the agent's internal definition of states and relational operations, not the external environment alone, and that shared action encoding and action merging may support structural transfer and the acquisition of low dimensional maps.

Critical strengths
  1. An exact, training independent link between task structure and representational geometry. By working with the analytically characterized MRNS family, the paper separates task determined geometry from initialization and optimization artifacts, which is a genuine methodological advance for normative theories of neural representation. The resulting dimensionality formula dtask=G−1+dincd_{\mathrm{task}} = G - 1 + d_{\mathrm{inc}} is interpretable and nonobvious in its consequences: in the family pedigree, five nominally independent inverse action pairs collapse to three globally consistent directions because of path constraints, a prediction one could not have made by counting named relations. This matters most because it converts competing hypotheses about an agent's internal state and action definitions into quantitative, falsifiable predictions of neural population geometry, something transition matrix based models such as the successor representation cannot express. The analytic results are exactly reproducible without any network training, and I verified representative cases independently.

  2. A conceptually consequential reframing with well chosen worked examples. The identification of edge identity as a determinant of cognitive map geometry redirects attention from state transition structure to the agent's relational ontology, and it generates concrete, experimentally accessible hypotheses: transfer under shared action codes, compression under action merging, and coexistence of multiple relational structures under factorized state representations. The example set (grids, pedigree, disconnected mazes, the Garvert graph, composition) is thoughtfully constructed to make each consequence intuitive, and the framework could be genuinely useful for designing experiments that aim to infer how a subject internally organizes a task.

Critical weaknesses
  1. The central quantity is fragile and describes only a thin slice of the representation. dtaskd_{\mathrm{task}} is the exact multiplicity of a theoretically maximal eigenvalue, which requires exactly balanced visitation and deterministic transitions. In my own reconstruction of the 2D grid (following Appendix C), duplicating only a few transitions already breaks the exact tie and changes the value of the criterion, even though the leading components continue to encode position well (R2≈0.8R^2 \approx 0.8 in my check), so the geometry appears more robust than the defined measure, a point the paper neither shows nor discusses. Moreover, the "dominant" subspace often carries little variance: about 24% in the 2D grid with the next eigenvalue at roughly 0.81 of the maximum, and only about 8% in the 3D grid where the fourth component is close behind (visible in Figure 1c). In the disconnected maze task the participation ratio changes only from 26.62 to 25.95. Claims of "low dimensional" or "compressed" maps are therefore not supported as stated. Finally, for periodic structures such as rings or tori the theory yields dtask=0d_{\mathrm{task}} = 0 because no nonconstant action-consistent coordinate exists, yet in my ring control the top principal components of the MRNS encode the circle perfectly (R2=1.0R^2 = 1.0). This blind spot, inherited from modeling relations purely as additive translations, is consequential because periodic codes are central to head direction and grid cell representations, and the paper does not mention it.

  2. The broader insights are largely established by construction, and the one emergent result is unexplained and arguably misattributed. Shared action labels producing shared coordinates, and product states producing summed dimensions, follow directly from the definitions; the difficult scientific question of how an agent comes to identify transitions across environments as the same operation is assumed away in the input encoding, and no held out transition test supports the "structural generalization" claim. The superposition collapse in the composition experiment depends on an arbitrary person to location assignment. The bottleneck action merging result lies outside the theory, is reported only qualitatively, and its robustness to initialization scale, optimizer, and visitation statistics is unknown; in a linear additive predictor the converged action embedding must equal the displacement between endpoint state embeddings, so merging may be a readout of the learned state layout rather than a mechanism that produces the map, the reverse of the causal direction the paper suggests. The nonlinear evidence is weak in places (3D grid scores of 0.07 to 0.18 in Table S1) with no chance levels reported for the similarity metric. Technical novelty over Braun et al. (2025) is modest, resting on a Rayleigh quotient argument and rank nullity counting. Finally, the paper misses connections that would substantially situate the contribution: relational embeddings as translations in knowledge graphs (TransE) and their rotational extensions (RotatE), Hinton's classic family tree network, actionable representations (Dorrell et al., 2023), group theoretic accounts of grid and head direction codes, and a direct comparison with the action conditioned transition structure of the Tolman Eichenbaum Machine.

Questions for authors

Clarifications whose answers could materially change my assessment:

  1. Can the task determined subspace and its dimensionality be characterized robustly, for example through the null space formulation in Appendix B or an eigengap criterion, rather than through exact eigenvalue multiplicity? Please report, for every task, the fraction of variance in the dominant subspace and the gap to the next eigenvalue, and show how the subspace behaves under non uniform visitation and stochastic transitions.
  2. For periodic environments the theory yields dtask=0d_{\mathrm{task}} = 0 even though the representation clearly encodes the cyclic structure in its leading components. How should the theory be read in this case, and would multiplicative or rotational action representations be required to cover head direction and grid cell codes?
  3. Regarding action merging in the bottleneck network: what selects the merged solution, given that the objective may admit many equally optimal low rank solutions? Does merging survive larger initialization scales, other optimizers, and unbalanced visitation? Is the learned state code at Dh=2D_h = 2 effectively one dimensional? In what quantitative sense does the learned representation in Figure 3d resemble the successor representation? And since at convergence each action embedding equals the displacement between its endpoint state embeddings in this model class, how should the claimed causal direction (merging produces the map) be understood?
  4. What is the chance level of the subspace similarity metric for each task in Tables S1 and S2, and how do the learned networks compare against stronger baselines such as the successor representation or a code trained to predict the next state identity directly?
  5. In the family relation task, what is the third independent action-consistent coordinate beyond generation and sex?

Suggestions for future work (not required for my assessment):

  1. Test transfer directly by holding out transitions in the second map and comparing shared versus map specific action encodings.
  2. Apply the framework to an existing behavioral or neural dataset, for example Garvert et al. (2017) or Tang et al. (2023), to show that competing state and action definitions make distinguishable and testable predictions.
10 comments
AuthorsOpenReview ↗

Response (Part 1)

Thank you for recognizing the value of our work and for your insightful suggestions. In response to your questions, we conducted additional experiments and provided detailed explanations of the methodology and visualizations. Please see the “Reviewer HG1L” folder in the recently submitted supplementary material.

Weakness1. Fragility of dtaskd_{task}

The central quantity is fragile and describes only a thin slice of the representation.

We thank the reviewer for raising this concern and agree that non-balanced visitation, together with other potential factors, easily breaks the exact eigenvalue equality in our original theorem. We therefore developed an additional theorem C.1 in Appendix C to characterize the stability of the task-determined PCA subspace under perturbations. The intuition is as follows.

Under balanced visitation, the centered sample Gram matrix G0G_0 has a clean block structure: all dtaskd_{\mathrm{task}} task eigenvalues are exactly σmax⁡2\sigma_{\max}^2, and they sit above a well-separated block of non-task eigenvalues. Now assume a perturbed Gram matrix G1G_1, it can be decomposed into three parts when expressed in the original task/non-task basis: an upper-left block A1A_1 (the effect of non-uniform visitation within the original task subspace), a lower-right block D1D_1 (the effect within the orthogonal complement), and an off-diagonal block B1B_1 that couples the two subspaces. Crucially, only B1B_1 rotates the subspace; changes within A1A_1 or D1D_1 alone merely rescale variances and leave the task subspace invariant.

Two scalar quantities control the stability. First, δ1:=λmin⁡(A1)−λmax⁡(D1)\delta_1 := \lambda_{\min}(A_1) - \lambda_{\max}(D_1) is the “surviving” spectral gap that would remain between the two blocks if the coupling were removed—i.e., after accounting for within-subspace effects of non-uniform visitation. Second, b1:=∥B1∥2b_1 := \lVert B_1\rVert_2 is the strength of the cross-subspace coupling. Weyl's perturbation inequality then says each eigenvalue can shift by at most b1b_1 under the perturbation B1B_1. So if δ1>2b1\delta_1 > 2b_1, even in the worst case the smallest task eigenvalue drops by b1b_1 and the largest non-task eigenvalue rises by b1b_1, and they still do not cross. This guarantees a positive eigengap at the boundary of PC dtaskd_{\mathrm{task}} and PC dtask+1d_{\mathrm{task}}+1; the leading dtaskd_{\mathrm{task}}-dimensional eigenspace therefore stays a well-defined object. Finally, a residual Davis–Kahan argument bounds the rotation between this perturbed leading subspace and the original balanced-visitation subspace: ∥sin⁡Θ(U0,U1)∥2≤b1/(δ1−b1)\lVert \sin\Theta(U_0,U_1)\rVert_2 \leq b_1/(\delta_1-b_1).

In short, the quantity dtaskd_{task} is sensitive to deviations from balanced visitation and other perturbations, the task-subspace geometry remains stable provided that the cross-subspace coupling is less than half the surviving spectral gap, b1<δ1/2b_1 < \delta_1/2. Theorem C.1 makes this robustness precise by giving a sufficient condition for the leading dtaskd_{\mathrm{task}}-dimensional PCA subspace to remain spectrally separated and an explicit bound on its rotation relative to the original task subspace.

Weakness2.1. The broader insights are largely established by construction.

The broader insights are largely established.

We agree that these insights follow from the construction. However, as noted in our clarification, this may be precisely what makes the theory compelling. The action-consistent coordinate system is determined jointly by the state space SS and the action space AA. The former is largely determined by the environment and behavior. The action space AA, by contrast, reflects cognition by specifying which relational operations are treated as equivalent. Together, these constructions determine the neural population geometry predicted by our theory. As many previous studies have reported, different ways of training animals can produce distinct neural population geometries and dynamics while leaving task performance largely unchanged (e.g., Latimer and Freedman, 2023). These findings by construction highlight the importance of cognition in shaping neural representations. We therefore believe that the insight provided by our theory is both logically grounded and intuitively plausible.

Nevertheless, despite the arguments above, we acknowledge that the empirical validity of the specific construction used in this paper remains to be established. We believe that this is precisely the most important direction for future work: investigating how animals and humans actually group actions into equivalence classes and how these action groupings are represented in the brain. We thank the reviewer for highlighting this important direction.

AuthorsOpenReview ↗

Response (Part 2)

Weakness 2.2: novelty clarification.

Technical novelty over Braun et al. (2025) is modest, resting on a Rayleigh quotient argument and rank nullity counting.

Our distinct contribution is the action-consistent coordinate space (F\mathcal{F}), introduced in Section 2.2, which is defined independently of the MRNS framework and characterizes the task structure that can be represented by action-consistent coordinates. We use MRNS as a linear network-based method for recovering (F\mathcal{F}). We establish that, under balanced visitation—a condition that has received limited attention in prior work—the learned MRNS representation aligns with this space. We will revise the paper to distinguish these conceptual and analytical contributions more clearly from the mathematical tools inherited from Braun et al. and to moderate our novelty claims accordingly.

Weakness 2.3: Relevance to relational embedding literature.

the paper misses connections that would substantially situate the contribution.

We thank the reviewer for highlighting these connections. We agree that the Introduction should better situate our contribution, and we will expand it to discuss TransE (Bordes et al., 2013), RotatE (Sun et al., 2019), the Tolman–Eichenbaum Machine (Whittington et al., 2020), and actionable representations (Dorrell et al., 2023).

Our contribution is to connect action consistency to the dominant PCA geometry of minimum-representation-norm predictive networks, with exact subspace and dimensionality guarantees under stated assumptions.

These approaches share the principle that relational operations constrain representations. In particular, Dorrell et al.’s actionable representations and our space F\mathcal{F} formalize closely related requirements: actions must have consistent effects across states. Dorrell et al. express these effects through state-independent linear transformations, whereas F\mathcal{F} comprises coordinates with state-independent additive action increments. The key distinction is their purpose. Dorrell et al. combine actionability with biological and functional constraints to explain grid-cell tuning and modular organization; we use action consistency to characterize the dominant PCA subspace and dimensionality of minimum-representation-norm predictive networks. Our contribution thus builds on a shared representational principle to address a different question: how action-labeled task structure determines population geometry under a specified neural objective. We will clarify this connection, alongside comparisons with TransE, RotatE, and the Tolman–Eichenbaum Machine, in the Introduction.

Q1.1. Robust characterization of task subspace and dimensionality through eigengap criteria

Can the task determined subspace and its dimensionality be characterized robustly, for example through the null space formulation in Appendix B or an eigengap criterion, rather than through exact eigenvalue multiplicity?

In an ideal scenario with known action labels, the dimensionality can bedirectly from the null-space formation, exactly as the dtask=G−1+dincd_{task} = G-1+d_{inc}. A more challenging question is how transitions are internally grouped into actions: this assignment may reflect the participant’s interpretation of the task and need not be uniquely specified by the observed transitions.

For an arbitrary network representation, exact eigenvalue multiplicity is unlikely to hold in practice. Theorem C.1 in Appendix C instead provides a sufficient criterion for assessing whether its leading subspace remains close to a candidate task subspace. Specifically, we can construct candidate models with different action assignments and derive a predicted subspace U0U_0 and dimensionality for each. We then compare each candidate with the observed Gram matrix G1G_1 of the network. Specifically, relative to U0U_0, we define A1=U0⊤G1U0A_1=U_0^\top G_1U_0, B1=U0⊤G1U0,⊥B_1=U_0^\top G_1U_{0,\perp}, and D1=U0,⊥⊤G1U0,⊥D_1=U_{0,\perp}^\top G_1U_{0,\perp}, together with δ1=λmin⁡(A1)−λmax⁡(D1)\delta_1=\lambda_{\min}(A_1)-\lambda_{\max}(D_1) and b1=∥B1∥2b_1=\lVert B_1\rVert_2. If δ1>2b1\delta_1>2b_1, Theorem C.1 guarantees a positive eigengap at the predicted dimensionality and bounds the rotation of the observed leading subspace away from the candidate task subspace. Thus, the null-space formulation determines dimensionality from a specified task model, while the eigengap criterion assesses the robustness of its predicted subspace in an observed representation. This supports evaluating candidate internal action assignments, although it does not by itself establish that a candidate is uniquely identifiable.

We validated this procedure on representations learned through network training (see Q3.1), rather than obtained directly from the analytical MRNS solution, and successfully recovered the predicted dimensionality using the eigengap criterion.

AuthorsOpenReview ↗

Response (Part 3)

Q1.2. What fraction of variance does the predicted subspace explain?

Please report, for every task, the fraction of variance in the dominant subspace and the gap to the next eigenvalue, and show how the subspace behaves under non uniform visitation and stochastic transitions.

The predicted subspace explains 25.6525.65%, 7.887.88%, 48.1348.13%, 16.3216.32%, 26.1626.16%, 5.915.91%, 39.4939.49%, and 100100% of the total variance in the 2D grid, 3D grid, family-tree, disconnected-maze with shared actions, disconnected-maze with map-specific actions, Cartesian-composition, Garvert action-merged, and Garvert edge-specific tasks, respectively. The corresponding relative eigengaps are 0.2380.238, 0.0640.064, 0.1850.185, 0.0220.022, 0.0330.033, 0.0590.059, 0.4170.417, and 1.0001.000. For the superposition-composition task, dtask=0d_{\mathrm{task}}=0, so no nonconstant dominant subspace is predicted and these quantities are not applicable.

We can see that the explained-variance fraction and the eigengap of the nvis\sqrt{n_{\mathrm{vis}}} eigenvalue block can be small, particularly when the task contains more states and actions. We believe that this arises from the constraints imposed by the linear architecture and one-hot encoding: the network must devote substantial representational capacity to encoding and decoding individual state identities. Consequently, the fraction of variance captured by the task-determined coordinates decreases sharply as the numbers of states and actions increase.

For non-uniform visitation and stochastic transitions, we measured subspace recovery relative to the deterministic balanced-visitation reference (see supplementary material) and found that it remained nearly perfect under non-uniform visitation (0.995) and high even with 50% transition noise (0.942/0.938 for balanced/non-uniform visitation). The sufficient condition in Theorem C.1 was satisfied for non-uniform visitation alone (δ1=0.00952>2b1=0.00370\delta_1=0.00952>2b_1=0.00370), thereby certifying the predicted two-dimensional leading subspace, whereas at transition-noise levels of 2525% or higher it was not satisfied under either visitation regime.

Q2.1.Limited Application to Periodic Structure.

For periodic environments the theory yields even though the representation clearly encodes the cyclic structure in its leading components. How should the theory be read in this case?

We were genuinely surprised and impressed by your independent discovery of the ring structure, as it closely parallels a direction that we are currently investigating. Our intuition is that an action-consistent coordinate system defines an additive Euclidean structure: each action induces a fixed displacement in the represented coordinates. So periodic or non-Euclidean environments generally need not admit such coordinates exactly. Nevertheless, even outside the exact regime in which our F\mathcal{F} theory applies, MRNs may still recover meaningful approximate geometric structure.

To explore this possibility, we designed two supplementary navigation experiments on a spherical environment (see supplementary material). The first is closely related to your ring control: states are nodes on a latitude–longitude grid, and actions correspond to movements along meridians and parallels. The exact analysis gives dtask=1d_{\mathrm{task}}=1, with the action-consistent coordinate corresponding to latitude. However, PC2 additionally captures the periodic longitude variable, analogous to the ring structure in your control. Consequently, the subspace spanned by PC1 and PC2 resembles a planar world-map projection of the sphere.

In the second experiment, states are distributed approximately uniformly over the sphere, and actions are defined by 88, 2020, or 8080 three-dimensional direction sectors centered at the origin, corresponding to progressively finer spatial resolution. Although this task is not compatible with a nonconstant action-consistent coordinate system (dtask=0d_{\mathrm{task}}=0), the first three principal components nevertheless form an approximately spherical representation. Moreover, the recovered geometry becomes increasingly sphere-like as the action resolution increases.

Together, these supplementary experiments suggest that the framework introduced in our paper may also provide a useful starting point for studying periodic and non-Euclidean environments. Although the exact mathematical characterization does not directly extend to these settings, the empirical results reveal nontrivial and interpretable approximate structures. We are currently investigating the mechanisms that produce these representations and the conditions under which they emerge.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Clarification

Dear Reviewers,

We would like to clarify that our theory extends beyond to neural population geometry. It is intended as a general framework for cognitive maps that links task demands, cognitive representations of states and relational operations, and neural implementation.

At its core, the action-consistent coordinate space FF (Section 2.2) offers a candidate mathematical definition of cognitive maps: states are assigned coordinates such that each action produces the same displacement wherever it occurs. This definition depends only on state identities and action-labeled transitions and requires neither neural networks nor balanced visitation. Its functional motivation is relational compression: reusable action increments express task-relevant relations compactly and support computation by vector addition. Summing these increments yields the displacement along a path, enabling inference about indirect relations, while sharing action identities across environments permits coordinate reuse.

Crucially, these action identities reflect an agent’s interpretation of the task. Predicting transitions alone does not require relational abstraction; an agent could simply memorize individual transitions. Treating different transitions as instances of the same action introduces a shared relational operation and constrains the map’s coordinates. The resulting geometry therefore reflects both task demands and how the agent organizes experience.

Our theory make this connection precise by characterizing FF and deriving its dimensionality, dim⁡F=G+dinc\dim F=G+d_{\mathrm{inc}}, where GG is the number of connected components in the transition graph and dincd_{\mathrm{inc}} counts the independent action increments compatible with all observed paths. Under the stated assumptions, we further prove that the nonconstant subspace of FF determines the dominant PCA subspace of minimum-representation-norm solutions in linear predictive networks. This establishes a quantitative link between relational abstraction and neural representation geometry, whose biological realization remains open to empirical testing.

To make the mathematical arguments in our paper easier to follow, we have reorganized the proofs in the Appendix A and B to improve their presentation. We also develop another theorem in Appendix C characterizing the neural geometry when the balanced-visitation assumption is not satisfied.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.