acceptodds
Under review as a conference paper at ICLR 2027

Factorization Regret mediates compositional generalization in latent space

Abstract

Are there still barriers to generalization once all relevant variables are known? We consider the challenge of generalizing to a novel combination of task-relevant latent variables in the Cognitive Gridworld, a stationary Partially Observable Markov Decision Process where observations are generated by latent variables with parametric interactions. We establish Factorization Regret as an information-theoretic quantity measuring the contribution of latent variable interactions to task performance. Using this metric, we explain the sub-optimality of Echo State Networks and uncover a failure mode whereby confidence becomes decoupled from accuracy. This suggests that utilizing interactions between relevant variables is a non-trivial capability. We next introduce Representation Classification Chains to show that latent variable interactions and how to infer the variables' values can be learned together through Reinforcement Learning. Solving this variational inference problem enables generalization to novel combinations of relevant variables. Lastly, we demonstrate that once interactions are learned, imagined experience is sufficient for zero-shot control. In summary, we develop and motivate a minimal, theoretically grounded setting for research into generalization from experience.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Compositional Generalization, Latent Space, Partially Observable Markov Decision Process, POMDP, Meta-Learning, RNN, Recurrent Neural Network, Representation Learning, Information Theory, Variational Inference, Interaction, Emergence, Bayesian inference, Bayes

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_z6zEOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

The paper studies inference about interacting latent variables in Cognitive Gridworld. It defines Factorization Regret (FR) as the KL divergence from the joint Bayesian observer's goal-variable posterior to the naive observer's. In Experiment 1, networks receive the true interactions and are trained against Bayesian posteriors; FR helps account for differences in accuracy and calibration between fully trained networks and networks with frozen recurrent weights. Representation Classification Chains (RCCs) then learn interaction embeddings from binary outcome feedback and observation prediction, and generalize to unseen combinations of known variables. Finally, a controller trained in the learned generative model achieves control without interacting with the target environment during controller training.

Critical strengths
  • S1. Cognitive Gridworld allows a precise analysis of inference errors. For the small contexts used in the experiments, exact Bayesian posteriors provide a reference for separating observational uncertainty from approximation error (Secs. 3–4.1, pp. 3–4). The inference tasks exclude exploration and planning, so the analysis can focus on how agents use latent interactions. The same setting also supports controlled tests on unseen variable combinations (Fig. 5, p. 6).

  • S2. The network results make a useful case for FR as an inference diagnostic. With Bayesian posterior supervision in Experiment 1, fully trained networks approach Joint Bayes performance, while networks with frozen recurrent weights resemble Naive Bayes in both performance and overconfidence (Figs. 3–4, pp. 5–6; Fig. 11, p. 22). The performance gap persists when controlling for the number of trainable parameters and when ESNs output updates to joint beliefs (Fig. 9, p. 21). Even at comparable mean accuracy, the networks differ in the correlation between FR and accuracy (Fig. 15d–e, p. 28). This gives information that average accuracy alone would miss.

  • S3. RCC shows that interactions can be learned from the available feedback and used in new combinations. Without access to the true interaction parameters, it learns from binary outcomes and observation prediction, then succeeds on unseen variable combinations (Sec. 4.3, Fig. 6c, p. 7). Test variables appear during training as non-goal variables, at most one per episode; their combinations and use as goals are held out (p. 6, lines 294–298). This is an existence proof under the specified architecture and feedback conditions.

Critical weaknesses
  • W1. I find the evidence connecting FR to learning compositional representations less direct than the evidence for its value as an inference diagnostic. The Discussion presents RCC as an existence proof of learnability (p. 8, lines 422–424), then suggests that greater FR creates stronger pressure to develop compositional representations (p. 9, lines 469–472). The RCC experiment does not directly test this proposed relationship. Figure 12 (p. 23) shows that generalization depends on the total number of available variables, but does not relate that dependence to FR. This leaves open how far FR explains the acquisition of transferable representations, beyond the inference benefit of using interactions (Q1–Q2).
Questions for authors
  • Q1 (W1). What theoretical or empirical evidence supports the suggestion that greater FR creates stronger pressure to develop compositional representations? How, if at all, does this account for the dependence on the number of variables in Fig. 12?

  • Q2 (W1). What do the RCC experiments establish about learned compositional structure?

    • Q2(a). The environment and RCC both use shared embeddings and a specified pairwise interaction rule (p. 6, lines 289–298; p. 23, lines 1194–1198). Figure 6c shows that learning within this structure supports transfer to unseen combinations. Is there evidence linking the acquisition of these embeddings to FR beyond the improvement over Naive Bayes in Fig. 6c?

    • Q2(b). The main text cites Fig. 13 as evidence that routing embedding gradients through the Classifier impairs learning (p. 7, lines 339–340). The appendix tests this in a simplified setting with ground-truth latent realizations and architectural modifications (p. 24, lines 1242–1250). What can this comparison tell us about the Generator's role in learning representations under the original feedback conditions?

2 comments

Thank you for your careful reading of the main text and supplementary analyses. Your summary describes the scope of the work as we intended it, and your questions about representation learning happen to have been our research focus since the initial submission. In the revised PDF, former Figs. 12 and 15 are now Figs. 15 and 19.

W1. “I find the evidence connecting FR to learning compositional representations less direct than the evidence for its value as an inference diagnostic.”

We agree. FR is a statistic of the environment, not of a network; as you put it, an inference diagnostic: it identifies when latent structure is non-trivial. We have (i) emphasized FR’s scope (Discussion, lines 444–459) and (ii) run more focused analyses and experiments on how FR shapes learning and generalization (B.1–B.2; new Figs. 14, 17 and 18; revised Figs. 19–20; three supplementary animations, see the general comment).

Q1. “What … evidence supports the suggestion that greater FR creates stronger pressure to develop compositional representations? How … does this account for the dependence on the number of variables in Fig. 12 [now Fig. 15]?”

In theory, FR bounds the accuracy lost by ignoring interactions (B.3). Practical applications are outside our scope, but a new analysis and a new experiment show its potential usefulness. Fig. 17: FR predicts the order and speed in which regions of latent space are learned: low-FR regions rapidly, high-FR regions slowly and later. Fig. 18: Learning from high-FR experience transfers immediately to low-FR settings, whereas the reverse transfer is slow (animated in Extrapolation_from_FR.gif; B.1).

FR does not account for the ∣S∣|S| dependence of transfer (see the new B.5 and Fig. 23). That said, the ∣S∣|S|-dependent generalization of a Generator trained on goal values alone, or on Naive Bayes classifications, drops with the FR of the testing environment (new Fig. 14). The ∣S∣|S| at which transfer begins also depends on the learner: an improved Generator architecture lowers it from 500 to 250 for the RCC (Fig. 15), though still not quite as low as for a Generator trained directly on Joint Bayes classifications (Fig. 14c).

Q2(a). “Is there evidence linking the acquisition of these embeddings to FR beyond the improvement over Naive Bayes in Fig. 6c?”

A Generator (and thus its embeddings) trained on goal realizations alone, or on the Naive observer’s actions, has a generalization error that grows with the FR of the novel environment (Fig. 14a–b). This dependence disappears when the Generator is conditioned on the RCC’s Classifier, which learns how latent variables interact (Fig. 14d). In summary, whether the embeddings’ generalization error is invariant to FR depends on whether the Generator is conditioned (during training) on actions from a policy that accounts for how latent variables interact.

Q2(b). “What can this comparison tell us about the Generator’s role in learning representations under the original feedback conditions?”

The original Fig. 13 removed the challenge of classification, largely to speed up data collection: training the Classifier with naturalistic feedback takes very many episodes, whereas a Generator conditioned on optimal actions learns extremely fast. This let us iterate faster and ultimately discover improvements to the Generator architecture (having repeated several experiments with it, we updated Fig. 15; see Q1).

We realize that introducing a new gradient routing and a new feedback condition at once in the original Fig. 13 created confusion, and that the resulting hyper-efficient Generator training oversimplified the comparison with the Classifier-routed gradient. The new Fig. 13 trains RCCs with naturalistic feedback, and for much longer: Classifier-routed gradients can learn beyond Naive Bayes, but much less reliably, with dramatic variance across seeds since learning is all or none. Our conclusion on the Generator’s role in learning transferable representations under naturalistic feedback is therefore: it dramatically boosts training efficiency, and makes classifier accuracy a heavily limiting factor (Fig. 14c–d). Its role is equally important for control, as internal generation (Fig. 7) avoids relearning action values from scratch whenever the value function, the action space or the relevant latent space changes. The Generator thus speeds up both representation learning and behavioral adaptation to new objectives.

Future work could use a high-dimensional observation space, where generative models are particularly effective (e.g., Ha & Schmidhuber, 2018; Hafner et al., 2025), to emphasize the Generator’s contribution. Here, we aimed to abstract away all structure from the observation space, reducing compositionality to simple and explicit vector products, and to avoid the confound of observations containing structure that is easier to learn with a generative model’s inductive biases.

Thank you for addressing my questions. The new Fig. 13 resolves Q2(b): the gradient routes are now compared under the original feedback conditions. I also found the analysis of the generative model's errors in Fig. 14 helpful, along with the explanation in Section B.5 of how transfer depends on the number of variables. Treating that dependence separately from FR answers this part of Q1. Together with the clearer scope, these results address much of W1 (p.s. the additional experiments support the authors’ hypothesis, but do not directly show that higher FR makes it easier to learn transferable embeddings.) and provide the additional evidence requested in Q2(a).

Thus I have raised my score to reflect these improvements.

Reviewer_2ykUOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper explores how agents may generalize to unseen combinations of latent variables by learning how these variables interact. The authors propose "Factorization Regret" (FR): an information-theoretic measure that captures the "contribution of latent variable interactions to task performance". The paper makes four main contributions:

  1. Introduces a task framework in ("Cognitive Gridworlds") implemented as POMDPs that allows them to systematically vary FR.
  2. Using this task setting, the authors use FR to explain the performance gap between RNNs and Echo State Networks (ESNs).
  3. Proposes an architecture (Representational Classification Chains) for learning latent variable interactions can be learned from plausible naturalistic feedback (i.e., sparse reward).
  4. Demonstrates how their framework can support zero-shot control in novel environments by training on imagined experience.
Critical strengths

Thoughtful experiments: The paper progressively builds in complexity, starting by introducing the task environment, then investigating the performance of RNNs/ESNs (experiment 1), demonstrating how the latent variables themselves may be learned (experiment 2), and finally showing how training on imagined experience can support zero-shot control (experiment 3). The thoughtfulness of the experiments greatly improve how easy the paper is to understand.

Rich connections to neuroscience: Though neuroscience motivates the work and is discussed at length in the discussion, knowledge of neuroscience is not assumed in the paper, making the paper more easily accessible to readers without a strong neuroscience background. Even so, the papers extended connection of its framework to neuroscience was one of the parts of the paper that I appreciated the most (and that I suspect other readers will appreciate as well).

Critical weaknesses

None

Questions for authors
  • "The most pressing limitation is the assumption that all task-relevant variables are known, which a latent variable discovery algorithm (e.g. VASE) could solve." -- this seems reasonable from a computational modeling standpoint, but what about in the brain? How do you think about previously unknown task-relevant variables being discovered/learned?
  • I am very interested in the following claim: "These seemingly disparate [roles of the OFC] in inference and evaluation would be unified by the objective of managing Factorization Regret." Could the authors expand on what it would mean for OFC to manage/represent FR in mechanistic terms? I am especially interested in how these mechanisms might relate to those proposed to implement latent cause inference in the OFC?
1 comment

Thank you for appreciating the structure of the paper; a lot of time and care went into its organization. We are very glad that the connection to neuroscience, which motivates this work, resonated with you, and we too expect readers with a neuroscience background to find the most value in it. We hope that the framework can provide design principles for cognitive tasks in future experimental neuroscience research.

Q1. “… what about in the brain? How do you think about previously unknown task-relevant variables being discovered/learned?” This is a question that we are very interested in, though we can only speculate. The HPC offers many fascinating phenomena that could provide a mechanism for latent variable discovery (such as plateau potentials, Bittner et al., 2017, and sharp-wave ripples, which compress sequential experience, Buzsáki, 2015). A promising systems-level account, on the other hand, is offered by predictive-processing and active-inference perspectives: cortical columns suppress predictable information as activity flows up and down the cortical hierarchy (Rao & Ballard, 1999; Friston, 2010; Bastos et al., 2012). Such recurrent compression could reveal latent variables dynamically and at different levels of abstraction.

Q2. “Could the authors expand on what it would mean for OFC to manage/represent FR in mechanistic terms?” Managing FR means identifying which latent causes interact with one’s ability to achieve a goal. We imagine two steps. First, reduce the world to a few goal-relevant latent dimensions. Next, within that set, apply memory-dependent corrections to marginal beliefs (which may coincide with an expansion of dimensionality, B.2, Fig. 21, in line with the high shattering dimensionality of prefrontal representations, Bernardi et al., 2020). How the OFC could directly measure changes in FR, however, is a difficult question. One possibility is that FR can be approximated by contrasting the outputs of reservoir and rapidly trained (à la short-term plasticity) cortical dynamics, as proxies for the Naive and Joint Bayesian observers.

“I am especially interested in how these mechanisms might relate to those proposed to implement latent cause inference in the OFC?” Latent cause accounts typically have the OFC represent a posterior over which single cause is currently active (e.g., Wilson et al., 2014; Chan et al., 2016). We argue that there was never only a single latent cause for the animal: even simple cognitive tasks rely on frontal motor areas to plan the response (Erlich et al., 2011; Guo et al., 2014), and these areas themselves contain abstractions (flexible motor plans, Churchland & Shenoy, 2024; reference frames, Hawkins et al., 2019).

Designing an animal experiment that could test, and potentially falsify, this hypothesis is difficult, but eye-tracking tasks could perhaps help isolate the motor (i.e., goal / causally linked to reward) and contextual latent causes. Ultimately, we are optimistic that an interdisciplinary collaboration between experimentalists and theoreticians could provide the joint expertise in cognitive, behavioral and computational neuroscience required to design such an experiment.

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

The paper introduces Factorization Regret (FR), an information-theoretic measure of how much latent variable interactions contribute to optimal action inference in a partially observable setting. The authors design a minimal, analytically tractable POMDP—Cognitive Gridworld—to manipulate interaction structure and show that FR predicts the gap between optimal joint inference and Naive Bayes, matches the empirical gap between fully trained RNNs and echo state networks (ESNs), and exposes a miscalibration failure mode. They further propose Representation Classification Chains (RCCs) to learn compositional embeddings of variables from naturalistic (sparse, binary) feedback and demonstrate zero-shot control in novel latent combinations by training a controller entirely in imagination.

Critical strengths

1) Technical novelty and innovation: Introduces Factorization Regret as a clean, computable quantity that isolates the cost of ignoring latent variable interactions and ties it to action inference, not just observation prediction. Proposes a JEPA-inspired Representation Classification Chain separating representation learning (self-supervision via a generator) from decision inference (reinforcement learning via a classifier), which is a principled way to break the inference–representation circularity. Constructs a tractable POMDP (Cognitive Gridworld) where interactions are controllable and analytically characterizable, enabling formal comparisons against ideal observers.

2) Experimental rigor and validation: Carefully contrasts fully trained RNNs with ESNs, showing ESNs plateau exactly at Naive Bayes performance in regimes with non-zero FR; this alignment with theory is compelling.

3) Significance of contributions: Addresses an important and under-specified gap: what remains hard after the “right” variables are known, and how interaction structure shapes generalization.

Critical weaknesses

1) Technical limitations or concerns: The definition of FR as DKLD_{KL} between goal marginals under joint vs. Naive Bayes is intuitive but its formal relationship to interaction information, synergy/PID, or semantic information is asserted more than derived; key properties (e.g., monotonicity, bounds, equivalences) are not fully formalized in the main text. Scalability remains untested: experiments focus on small C (mostly C=2), and the joint inference target scales exponentially with C; it is unclear how the approach or RCCs behave for larger contexts.

2) Experimental gaps or methodological issues: RCC generalization appears only when |S| is “large enough,” but there is no principled analysis of data complexity, diversity requirements, or sample efficiency. The offline zero-shot control result would benefit from quantitative comparisons between planning on true ℓℓ vs. learned ℓ^ℓ̂ beyond qualitative indistinguishability; robustness to model bias is not analyzed.

Questions for authors
  1. Can you formalize the relationship between FR and interaction information or PID-style synergy about the goal? Under what conditions does FR lower-bound or upper-bound synergistic information relevant to the optimal action?
  2. Is E[FR_t] provably non-decreasing with t under your data-generating process? If not, can you characterize regimes where it can transiently decrease?
  3. How sensitive are the ESN vs. RNN results to reservoir hyperparameters (spectral radius, leak rate, sparsity) and readout regularization? Can a sufficiently large reservoir with tuned hyperparameters exceed Naive Bayes?
2 comments

Thank you for the constructive review. Your questions prompted an extensive formal analysis of FR (B.3, Propositions 1–4; new Fig. 22) and an extension of Fig. 9, which now compares training and output paradigms across network sizes and clarifies that the central challenge in our setting is learnability, not expressivity (A.4, lines 1416–1429). Our results hold across hyperparameters and sizes (new Fig. 10; see Q3).

W1, formal properties. “… asserted more than derived; key properties … are not fully formalized in the main text.” Now stated in B.3 and proved in B.4, with the time course summarized in Sec. 4.1 (lines 179–181); see Q1–Q2. B.3 also ties FR to task performance by bounding the accuracy lost when interactions are neglected (Proposition 4, lines 2192–2197).

W1, scalability. “Scalability remains untested …” Agreed (lines 438–439): limited compute kept us from training RCCs extensively beyond C=2C = 2; training the Classifier from naturalistic rewards takes a long time (as it should in a setting meant to capture naturalistic learning). In brief C=3C = 3 runs, RCCs surpass Naive Bayes, which we can demonstrate if useful. We leave comparing state-of-the-art RL algorithms in the Cognitive Gridworld at high CC to groups better equipped for heavy computation; the full code will be available for others to train RCCs, and any other architecture, at C=3C = 3 and beyond.

W2, data requirements. “… no principled analysis of data complexity, diversity requirements, or sample efficiency.” Agreed: we focused on the richness of the latent-space dynamics (Experiment 1), and our characterization of generalization was limited. The architectural improvements of A.5 (lines 1553–1558) let transfer begin at ∣S∣=250|S| = 250 instead of 500 (Fig. 15), but a Generator conditioned during training on Joint Bayes’ actions shows that further gains are possible (new Fig. 14c). The new B.5 (Fig. 23) explains how the variability of interactions and the embedding dimensionality determine the minimal ∣S∣|S| required for generalization, though a full sample-complexity analysis remains beyond our scope.

W2, offline control. “… quantitative comparisons between planning on true vs. learned …; robustness to model bias is not analyzed.” Sec. 4.4 (Fig. 16c) compares controllers trained offline with the learned Generator and with the ground-truth likelihood: evaluated under the ground truth, they reach 99.3% and 99.5% of the maximum reward (n=20n = 20). Our scope was limited to representation learning; this result serves to demonstrate the representations’ usefulness, though we agree that a deeper investigation of planning in this setting is an interesting future direction.

Q1. “Can you formalize the relationship between FR and interaction information or PID-style synergy …? … does FR lower-bound or upper-bound synergistic information …?” Yes: formalized in B.3–B.4. In brief, Interaction Information in an observable is necessary but not sufficient for FR. PID synergy about the goal is the further gain from conditioning on the contextual latent variables; FR is the gain from inferring all variables jointly versus independently. A key distinction: FR concerns Bayesian observers in a POMDP, who are limited to their observations, whereas PID synergy does not differentiate observables from latent variables (lines 2171–2175 and 2266–2270). Neither bounds the other in general: FR is the smaller at the first step of a single observable, and expected FR is provably the larger after an identifiable time point whenever the Naive observer converges to a wrong action for some joint realizations (lines 2185–2190; Fig. 22h). For semantic information and the ΔI\Delta I of Nirenberg & Latham (2003), see lines 2270–2283.

Q2. “Is E[FR_t] provably non-decreasing with t …? If not, can you characterize regimes where it can transiently decrease?” Not in general (Proposition 1; Fig. 22; B.3, lines 2045–2152). Expected FR eventually grows linearly if the Naive observer converges to a wrong action for any joint realization (Fig. 22b); otherwise it stays at zero, or rises and returns to zero (Fig. 22a, c). Transient decreases occur while the Naive observer is still ruling out distractors, and only in regimes where it ends up wrong rarely, or only weakly (lines 2134–2142; Fig. 22d).

Q3. “How sensitive are the ESN vs. RNN results to reservoir hyperparameters …? Can a sufficiently large reservoir … exceed Naive Bayes?” Traditional reservoirs (with a trainable read-in) were not sensitive to their hyperparameters: Fig. 10 varies spectral radius, leak rate, connection density and read-out regularization, one at a time, and at N=1,000N = 1{,}000 every setting stays below Naive Bayes (A.4, lines 1431–1442). We do not rule out a large enough reservoir reaching optimality, but richer supervision is what clearly rescues it (Fig. 9, right column), emphasizing the specificity of these results are to learning from naturalistic feedback.

I've read the rebuttal. I thank the author for the clarification and additional experiments, which address my concerns. I will maintain the current scores.

Comments on the paper

2 comments
AuthorsOpenReview ↗

summary of revisions

We thank the reviewers for their careful and constructive reviews, and especially for their investment in the supplementary material. Several of their questions about it motivated further analyses, some with surprising outcomes. We have also updated the LLM usage statement to cover the rebuttal period, in which coding agents facilitated efficient data collection, the new visualizations added to the supplementary material, and initial drafts of the proofs in B.4, all of which we verified. All section, figure and line numbers in our responses refer to the revised PDF, which makes the following changes:

  1. Deeper formal analysis of FR: its time course, a bound on the accuracy lost by ignoring interactions, and its relation to quantities such as PID synergy (B.3, Propositions 1–4, with proofs in B.4; new Fig. 22; time course summarized in Sec. 4.1).

  2. More extensive RNN experiments: reservoir width, marginal versus joint belief updating, partial versus full supervision (new A.4; extended Fig. 9), and reservoir hyperparameters (new Fig. 10).

  3. Analyses and experiments that tie FR to learning and transfer (B.1–B.2; new Figs. 17–18, revised Figs. 19–20), and a new section that relates the onset of transfer to properties of the embeddings (B.5; new Fig. 23).

  4. Expanded Discussion of scope (lines 444–459).

  5. Re-organization and renumbering: several sections of the appendix have been re-organized to improve the flow with the new additions. A.4 and B.5 are new. Former A.4–A.5 are now A.5–A.6; the theory moved from B.1 to B.3 (proofs in B.4), former B.2–B.3 are now B.1–B.2, and former B.4–B.6 are now B.8, B.6 and B.7. Former Figs. 10–12 are now Figs. 11, 12 and 15, and former Figs. 14–17 are now Figs. 16 and 19–21.

  6. Changes to the main text are minor and mostly limited to wording, with the exception of Fig. 2, which now shows that Factorization Regret grows linearly with time (the saturation in the original figure was due to clipping probabilities at 10−1510^{-15}).

We have also added three animations to the supplementary material to help visualize important dynamics of learning, generalization and compositionality:

  1. Extrapolation_from_FR.gif animates the experiment of Fig. 18 to demonstrate the inverse relationship between the timing of learning and that of generalization.

  2. Entanglement_dependent_unfolding.gif begins with the Naive observer’s belief state and gradually progresses to the Joint observer’s, to demonstrate the unfolding of the Naive observer’s tangled representation.

  3. Corner_unfolding.gif follows the belief grids of Fig. 20d to show how the ‘particle’ trajectories (posterior means of marginal beliefs) of the Bayesian observers and the RNNs unfold the representation throughout inference.

References used in our replies

  • Bastos, A. M., et al. (2012). Canonical microcircuits for predictive coding. Neuron 76:695–711.
  • Bernardi, S., et al. (2020). The geometry of abstraction in the hippocampus and prefrontal cortex. Cell 183:954–967.
  • Bittner, K. C., et al. (2017). Behavioral time scale synaptic plasticity underlies CA1 place fields. Science 357:1033–1036.
  • Buzsáki, G. (2015). Hippocampal sharp wave-ripple: a cognitive biomarker for episodic memory and planning. Hippocampus 25:1073–1188.
  • Chan, S. C. Y., Niv, Y., & Norman, K. A. (2016). A probability distribution over latent causes, in the orbitofrontal cortex. J. Neurosci. 36:7817–7828.
  • Churchland, M. M., & Shenoy, K. V. (2024). Preparatory activity and the expansive null-space. Nat. Rev. Neurosci. 25:213–236.
  • Erlich, J. C., Bialek, M., & Brody, C. D. (2011). A cortical substrate for memory-guided orienting in the rat. Neuron 72:330–343.
  • Friston, K. (2010). The free-energy principle: a unified brain theory? Nat. Rev. Neurosci. 11:127–138.
  • Guo, Z. V., et al. (2014). Flow of cortical activity underlying a tactile decision in mice. Neuron 81:179–194.
  • Ha, D., & Schmidhuber, J. (2018). Recurrent world models facilitate policy evolution. Advances in Neural Information Processing Systems 31.
  • Hafner, D., et al. (2025). Mastering diverse control tasks through world models. Nature 640:647–653.
  • Hawkins, J., et al. (2019). A framework for intelligence and cortical function based on grid cells in the neocortex. Front. Neural Circuits 12:121.
  • Nirenberg, S., & Latham, P. E. (2003). Decoding neuronal spike trains: how important are correlations? PNAS 100:7348–7353.
  • Rao, R. P. N., & Ballard, D. H. (1999). Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects. Nat. Neurosci. 2:79–87.
  • Wilson, R. C., et al. (2014). Orbitofrontal cortex as a cognitive map of task space. Neuron 81:267–279.

We thank all reviewers for their careful reading and constructive feedback, which facilitated interesting new investigations and a deeper appreciation of the intricacies of compositional representation learning. Ahead of the deadline, we have uploaded a minor revision. This revision polishes Appendix B, particularly the new proofs, and adds one new analysis. The main text, the references and Appendix A are unchanged, as are all previous figures, results and conclusions.

New analysis (Appendix B.1, new Fig. 20). In our reply to Reviewer 2ykU, we suggested that the OFC could detect Factorization Regret by comparing the beliefs of reservoir dynamics to beliefs learned through neural dynamics with (short-term) plasticity. To exemplify this idea, the new Fig. 20 shows that the divergence between the beliefs of the Fully Trained and Echo State networks corresponds well to the divergence between the beliefs of Joint and Naive Bayes (Factorization Regret), and that this correspondence emerges early in training.

Changes since the previous upload (all in Appendix B):

  • B.1: Fig. 20 (described above) and a new paragraph associated with it.
  • B.3 (propositions): the definitions, assumptions and propositions are stated more explicitly and precisely. The results themselves are unchanged.
  • B.4 (proofs): rewritten with more clarity so that every step can be more easily followed. The arguments are unchanged.
  • B.6 (hyperparameters): added the RCC regularization weight used for the experiments of Fig. 13 (0.01), which was missing.
  • Presentation: consistent notation and minor wording edits across Appendix B, and more compact pseudo-code (B.7), with no change in content.
  • Figure numbers: with the new Fig. 20, the former Figs. 20–23 are now Figs. 21–24 (our earlier replies use the former numbers).

Once again, we are grateful to the reviewers and area chair for their time, especially considering the historic number of submissions this year.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.