acceptodds
Under review as a conference paper at ICLR 2027

Freeze the Dynamics, Learn the Interface: Modular Recurrent Scaffolds for Compositional World Models

Abstract

World models often learn representations and transition dynamics jointly, entangling perception with task structure and limiting compositional generalization. We introduce a hippocampal-inspired world model built on pretrained recurrent modules whose dynamics are frozen during downstream learning. Their multiscale periodic codes jointly disambiguate states that would be aliased within any individual module, providing a stable and structured latent coordinate system. A lightweight learned interface binds observations and actions to this scaffold, allowing the model to recover the topology of new task spaces with substantially fewer trainable parameters than end-to-end alternatives. The resulting representation decomposes task-relevant factors across modular dynamical subspaces and supports their recombination in previously unseen configurations. By exploiting this learned task-space structure, the model improves out-of-distribution compositional generalization and maintains more accurate long-horizon rollouts. The same structured latent geometry also enables goal-conditioned navigation and planning from state differences. These results suggest that pretrained multiscale grid-like dynamics can serve as a reusable scaffold for learning structured, compositional world models efficiently.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Brain-inspired Model, Structured Representation Learning, Latent Space Dynamics, Hippocampus, Scaffold, World Model, JEPA, Compositional Generalization

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_D2b5OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The authors introduce a neurally inspired world-modeling framework where the recurrent dynamics are pre-trained path integrators known to produce grid-cell-like dynamics, and then memories are attached to these dynamics during inference to provide a structured latent scaffold. The precise methods appear to not be described in the paper, making it hard to describe more technically.

Critical strengths

While the paper is linguistically well written and clear, it is lacking in technical detail (see weaknesses). Despite this, the motivation and related work are very well specified and the main text is engaging and well-described.

The results are promising and the concept is reasonable.

Critical weaknesses

The major critical weakness of the paper is a lack of technical detail in the main text (and appendix) which makes the paper non-reproducible and difficult (or impossible) to judge reliably -- despite there being ample additional space left to include it (the paper is only 8 pages, far below the limit of 9).

For example, the Figure 1 overview is the closest thing that we have to a complete architecture description. There are some vague RNN equations in the appendix A.8, but the 'router', 'planner', decoders and encoders are never described, along with their training recipes. Unfortunately a comprehensive review of this paper cannot be completed without this information.

I recommend the authors include the full model description in technical detail, including training and evaluation in main text to provide more concrete grounding to readers as they try to understand the model. Furthermore, the symbols used in all of these equations are not anywhere defined. As another example of a lack of clarity, the precise definition of what a 'joint' trained model is (and how it is initialized) is never described in the main text (and I could not find it in the appendix properly described) despite this being the main baseline. Overall, the paper would be significantly improved by increased technical depth throughout.

The authors need to explain their model and experiments more properly and carefully before this is ready for acceptance.

Questions for authors
  1. If it is known that you would like a specific set of grid-cell-code-like modules, why do you bother with pre-training them, instead of just analytically pre-specifying them? Have you experimented with the difference in performance between these alternatives, or additionally how performance changes if you unfreeze the RNN component after the pretraining phase?

  2. Can you provide a complete description of all of the architectural components of your model, and the baselines, and their training procedures? Specifically, for your model, what are the details of the router, 'planner', decoders and encoders? These should be described with equations, layer structure, and dimensions at a minimum.

1 comment
AuthorsOpenReview ↗

Reproducibility, analytic dynamics, and unfreezing

Reproducibility and model design

We thank the reviewer for the constructive suggestions, which motivated two important new experiments and substantial improvements to the implementation details. We have expanded §3 and Appendix B to specify the architecture, transition operators, initialization, training objectives, and hyperparameters.

The model uses an observation encoder and learned router to initialize the scaffold from the first observation and convert subsequent actions into module-specific controls. Frozen keeps the pretrained path-integration dynamics fixed; Joint trains the same architecture from random initialization; and Unfrozen init fine-tunes the pretrained recurrent dynamics alongside the interfaces. These comparisons help distinguish the effects of structured initialization, freezing, and downstream adaptation.

New analytic dynamics and unfreezing experiments

Following the reviewer's suggestion, we introduced an analytic operator implementing exact phase translations across eight scales, eliminating recurrent pretraining (§3, Eq. 2, Appendix B.3). With the same learned interfaces, Analytic achieves a 300-step path-integration error of 0.214 cm versus 41.369 cm for Frozen (Appendix C.1, Fig. 14, Table 10). At 4,096 trajectories, it also achieves 95.7% Language 3D OOD open-loop accuracy versus 42.0% for Frozen. However, Frozen performs better in OOD planning (57.7% versus 16.0%; Table 12), suggesting that accurate phase integration does not necessarily translate into better goal-directed planning.

We also directly tested whether freezing is necessary by fine-tuning the pretrained RNN with matched interfaces and training budgets across 16 data sizes and three seeds. At 1,024 trajectories, Unfrozen init improves Language 3D OOD cross-entropy from 0.779 to 0.710 and reduces imagined goal distance from 9.12 to 4.20, while the change on 3D Shapes is small relative to seed variation (Appendix C.4, Table 17).

OOD, N=1,024FrozenUnfrozen initJointAnalytic
Language 3D CE ↓0.7790.7100.9630.702
Imagined goal distance ↓9.124.209.521.66
Shapes3D MSE ↓0.02020.01960.04740.0093

Together, these experiments suggest that the key benefit lies in the structured dynamical prior rather than freezing itself. The prior can be supplied analytically or through pretraining, and fine-tuning can further improve performance depending on task demands.

Architecture and training clarification

We now provide complete specifications in §3 and Appendix B (Tables 5–9, Algorithms 1–3). The scaffold contains eight independent 512-unit ReLU modules, pretrained on 20-step path-integration trajectories using displacement inputs and phase-prediction loss. Analytic replaces these recurrent updates with exact phase translations. A shared router learns module-specific initial phases and action controls, and a decoder predicts observations from the combined latent state.

Training uses free rollouts: only the first observation is encoded, recorded actions drive subsequent transitions, and predictions are supervised at each step using cross-entropy or MSE (Eq. 4). Planning uses a learned action scorer and candidate-sequence search, with additional optimal-action supervision in Language 2D (Appendix B.7–B.8). Full network dimensions, hyperparameters, and matched baseline specifications are provided in the appendix.

We appreciate the reviewer's constructive suggestions, which helped us strengthen the study. Our results now provide clearer evidence that structured multiscale dynamics offer a powerful inductive bias for learning predictive world models, with benefits across different implementations and training strategies.

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

The authors introduce a world-model architecture based on pretrained recurrent modules with multiscale periodic dynamics. These modules are frozen, while a learned interface connects observations and actions to the underlying representation. The approach is evaluated on compositional language tasks, visual prediction, and MetaWorld.

Critical strengths

The central idea of learning reusable dynamics independently of downstream observations is interesting. The modular architecture provides an interpretable structural prior, and the improvements on held-out factor combinations suggest that such a prior can help compositional generalization. Testing the same scaffold across several observation modalities is also a positive aspect.

Critical weaknesses

The baselines are not strong enough to establish the main contribution. Most comparisons involve jointly trained or randomly initialized variants of the same architecture. The frozen model benefits from extensive path-integration pretraining, while the baselines must learn their dynamics from downstream data. Consequently, the experiments do not fully distinguish the benefit of pretraining from the specific modular scaffold.

The evaluation environments are largely structured around separable factors, which align naturally with the proposed representation. The generality of the approach to more complex nonlinear dynamics remains unclear. In MetaWorld, the reported state-prediction errors also do not establish corresponding improvements in control performance.

Questions for authors

How does the proposed method compare with a pretrained structured world model under matched pretraining and downstream training budgets? Can the authors also report task-level success on MetaWorld, particularly under held-out conditions?

These results would help determine whether the improvement is primarily due to reusable pretrained dynamics or the proposed multiscale modular structure.

2 comments
AuthorsOpenReview ↗

R2 response (1/2) — Matched baselines and pretraining

Matched baselines and the role of structured dynamics

We thank the reviewer for emphasizing the importance of fair comparisons and disentangling the effects of pretraining from architectural advantages. We have substantially expanded our experiments to address these concerns.

Our six-model comparison uses matched observation/action interfaces, downstream data budgets, and rollout objectives, with transition-core parameter counts matched within 0.3% (including the readout lift for Analytic; §4, Table 1, Appendix C.4). Importantly, Analytic uses fixed, mathematically specified multiscale phase transitions without recurrent pretraining, allowing us to assess the value of the structured dynamics independently of pretrained recurrent weights.

At 1,024 trajectories, Analytic achieves the lowest OOD prediction errors on Language 3D and 3D Shapes, while Frozen also outperforms all three from-scratch baselines. The full evaluation spans 16 data budgets and three seeds (§4, Figs. 4–5).

OOD, N=1,024AnalyticFrozenJointRSSMLSTMTransformer
Language 3D CE ↓0.7020.7790.9630.9500.8710.917
Shapes3D MSE ↓0.00930.02020.04740.07730.03340.0438

These results suggest that the multiscale dynamical structure itself provides a useful inductive bias, rather than the advantage arising solely from recurrent pretraining.

New matched-pretraining controls

To address the reviewer's concern more directly, we additionally pretrained RSSM, GRU, LSTM, and a wide RNN on the same 2.2 m × 2.2 m path-integration task for 100,000 updates, with matched core sizes and downstream training schedules. Both frozen and fine-tuned variants were evaluated (Appendix C.10, Tables 32–34).

We also tested fine-tuning the pretrained scaffold itself. At 1,024 trajectories, its Language 3D OOD cross-entropy improves from 0.779 (Frozen) to 0.710 (fine-tuned), compared with 0.963 from random initialization (three-seed means; Table 17).

OOD, N=1,024 (seed 0)AnalyticFrozen scaffoldScaffold FTBest pretrained baseline
Language 2D CE ↓0.90040.93020.87120.8900 (LSTM frozen)
Language 3D CE ↓0.70780.80000.70900.7074 (LSTM FT)
3D Shapes MSE ↓0.00980.01970.01940.0160 (LSTM FT)

These new controls provide a more nuanced picture. Pretraining indeed explains part of Frozen's advantage over from-scratch baselines, as the reviewer suggested. Fine-tuned LSTM performs competitively and outperforms Frozen on some tasks. Nevertheless, Analytic achieves the best 3D Shapes result without any recurrent pretraining, while the fine-tuned scaffold leads on Language 2D. Thus, the evidence supports the value of structured dynamics beyond a simple pretrained-versus-random initialization comparison, although the advantage is task-dependent.

AuthorsOpenReview ↗

R2 response (2/2) — Model, training, and task-level evaluation

Generalization to complex observations and 47 MetaWorld tasks

We have also expanded the evaluation to test whether the approach remains useful beyond simple synthetic environments.

In ReplicaCAD, the model learns from frozen DINOv2 features of RGB sequences and relative camera actions, without factor labels, absolute poses, or furniture coordinates. A probe trained on static views evaluates whether the learned dynamics preserve relative furniture geometry over a 23-step rollout. Analytic and Frozen achieve substantially lower relative-layout errors than LSTM and Transformer (§6, Fig. 7, Table 23).

ReplicaCADAnalyticFrozenLSTMTransformer
Relative layout RMSE (m) ↓0.1870.3321.4232.886

Transformer achieves the lowest DINOv2 feature prediction error (0.098 cosine error), whereas the scaffold models better preserve the probed spatial relationships. This suggests a potential advantage of structured latent dynamics for geometric consistency, even when another architecture achieves better feature-level reconstruction.

We further evaluated all 47 MetaWorld tasks using the six-model matched comparison (§5, Appendix C.6, Table 20). At 256 trajectories, Analytic attains the lowest mean OOD state rollout MSE on 42/47 tasks (18 strict wins) and the lowest mean OOD hand endpoint error on 37/47 (27 strict wins). At 1,024 trajectories, Frozen achieves the lowest mean hand endpoint error on 32/47 ID tasks and 21/47 OOD tasks (19 and 12 strict wins, respectively). RSSM leads on OOD state MSE in 25/47 tasks at this larger budget (16 strict wins).

These results demonstrate that the benefits of the structured scaffold extend to a broad range of robotic manipulation dynamics, particularly in data-limited settings, while also showing that its relative advantage depends on the task, prediction metric, and training budget.

Architecture, training, and evaluation scope

We have expanded §3 and Appendix B to clarify the model architecture and training procedure. Observation encoders process images, language, or robot states, while a shared learned router generates module-specific phase coordinates and action controls. Frozen uses independently pretrained recurrent modules approximating controlled phase translations; Analytic implements exact phase updates (Eq. 2). A common decoder predicts future observations from the combined latent state. Detailed dimensions, initialization, and transition operators are provided in Appendix B.2–B.6.

All models are trained using the same free-rollout objective: only the initial observation is encoded, subsequent states evolve under recorded actions, and predictions are supervised at each step (Eq. 4). Goal-directed planning uses a learned action scorer and candidate-sequence search, with additional optimal-action supervision in Language 2D (Algorithm 3).

We agree with the reviewer that predictive accuracy does not necessarily translate into closed-loop control success. The MetaWorld experiments evaluate predictions on held-out trajectories under recorded actions, not autonomous task completion. We now clarify this distinction and identify closed-loop MetaWorld evaluation as an important next step.

In summary, the new matched-pretraining controls, analytic-model comparisons, and 47-task evaluation provide stronger evidence that structured multiscale dynamics offer a useful inductive bias for learning predictive world models, beyond the benefits of recurrent pretraining alone. We appreciate the reviewer's suggestions, which have helped us strengthen the comparisons, clarify the scope of our claims, and identify where the advantages are most pronounced.

Reviewer_nEoKOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper uses fixed grid cell like modules, combined with trained components: sensory embedding, action embedding, and a router that could decide which grid modules are influenced by the sensory and action inputs. The paper claims this construction allows compositionality, improves OOD performance, and tackles even non-spatial tasks.

Critical strengths

Fixing the transition dynamics and allowing compositionality in world models are timely topics.

Critical weaknesses

Lack of novelty

As interesting as the paper seems to be, I struggle to find substantial differences compared to prior works beyond the framing. The most crucial design, a router that allows grid modules to move independently such that they can track compositional latent variables instead of just x y coordinates, is already covered in Klukas, Lewis, Fiete, 2020, and potentially many other works.

Freezing the dynamics, pretraining the network, and translation to non-toy tasks are also not new to the field (Maass et al. 2002; Leibold 2020; Sharma et al. 2022; Vijayabaskaran & Cheng 2022; Lin et al. 2026)

Unclear manifold paragraphs

It is very exciting to give a taste of differential geometry. But the content did not go any further beyond decomposing a perturbation into tangent and normal components. It seems to go completely orthogonal to how block diagonal recurrence can restrict perturabations. I fail to see how this part of the content is related to how the model is designed or any explanation to how the architecture can better handle an input manifold, despite appearing several times in the text. It feels like just buzz words to bait AI.

Unclear details

The only difference from Klukas et al in this paper is that the input to grid modules are pretrained not entirely hand designed. But since it's pretrained on spatial tasks, wouldn't it essentially recover what is hand designed in Klukas et al and the series of works from Fiete? It feels strange that the structure of those weights is not shown or discussed.

More importantly, the architecture of the router, the encoders, and how closed-loop learning was done were never shown, even in the appendix.

Questions for authors

see the weaknesses.

3 comments
AuthorsOpenReview ↗

R1 response (1/3) — Contribution and manifold stability

Contribution and distinction from prior work

We thank the reviewer for highlighting the relevant prior work. We agree that independent control of grid modules was established by Klukas, Lewis & Fiete (2020), and that fixed recurrent dynamics and structured memory substrates are not new individually. We have revised §§2 and 7 to acknowledge these contributions more explicitly and clarify the scope of our novelty.

Our central question is different: Can a fixed, pretrained dynamical scaffold guide the learning of task representations and action dynamics from observation–action trajectories? Whereas Klukas et al. demonstrate how independently controlled grid modules can represent compositional variables, we investigate how the same structured dynamics can serve as a prior for learning which latent factors to represent and how actions transform them, without predefined factor-to-module assignments. This turns the grid scaffold from a representation of known latent variables into a structural constraint on learning predictive world models.

The key distinction is that freezing the scaffold constrains the observation and action interfaces to learn representations compatible with its dynamics. When the recurrent dynamics are also trainable, the model has additional freedom to accommodate entangled representations by modifying its transitions. Fixing the scaffold removes this flexibility and may therefore encourage factor separation and more stable predictions. Our experiments provide evidence for this mechanism, rather than claiming that freezing alone guarantees disentanglement.

In particular, the 36-step comparison shows substantially better geometric conditioning and prediction accuracy for Analytic (condition number 1.00; MSE 0.00236) and Frozen (2.71; 0.01279) than Joint (8,551; 0.03765; Fig. 16, Table 14). The learned interfaces also recover compositional structure in Language 2D, coherent wall-hue transitions in 3D Shapes, and relative furniture geometry in ReplicaCAD (§§4, 6; Appendix C.3). Under limited data, the models show improved OOD predictions, including Language 2D rollout accuracy of 63.2%/66.5% for Frozen/Analytic versus 62.5% for Transformer, and 3D Shapes MSE of 0.0202/0.0093 versus 0.0334 for LSTM (Table 1). We view these results as evidence that the fixed scaffold can provide a useful structural inductive bias, rather than simply an alternative recurrent architecture.

We also clarify the relationship to other cited studies (§§2, 7). Maass et al. establish computation with fixed recurrent substrates; TEM, MESH, and Vector-HaSH investigate structured relational representations or memory; and Leibold, Lin et al., Vijayabaskaran & Cheng, and GCQ address related aspects of spatial representations, action structure, or learned dynamics. Our contribution is not any of these ingredients individually, but an investigation of how a fixed multiscale transition structure constrains jointly learned observation/action interfaces and influences predictive generalization.

Manifold geometry and stability

We appreciate the reviewer's request for a more precise account of manifold stability. We now distinguish three properties that should not be conflated: cross-module independence, within-manifold phase preservation, and contraction of perturbations away from the manifold (§3, Appendix A.1).

Block-diagonal recurrence ensures zero cross-module Jacobian when module controls are fixed, but does not itself guarantee within-module stability. For the analytic operator, phase translation has an identity Jacobian in phase coordinates, exactly preserving local phase offsets. For the learned recurrent modules, we evaluate these properties empirically rather than assuming them.

In controlled perturbation tests across 256 wall-hue anchors, Frozen preserved tangent phase offsets while its normalized normal response decreased from 0.092 to 0.051 over 12 steps, with induced phase RMSE below 0.033 rad (Appendix C.2, Fig. 15). Together with the 36-step chart-conditioning results, these findings support local phase stability and off-manifold contraction in the tested setting. We have revised the text to avoid implying a general stability guarantee for learned modules.

AuthorsOpenReview ↗

R1 response (2/3) — Baselines and pretraining controls

Controlled comparisons and the role of pretraining

The three controls below address different questions. Each reports held-out combinations at 1,024 downstream trajectories. L2D and L3D are Language 2D and 3D. CE is token cross-entropy, accuracy is in percent, and MSE is normalized.

Strong baselines with matched downstream training (Table 1, Appendix C.4, Table 16)

RSSM, LSTM, and Transformer are trained from scratch in this sweep. They use the same interfaces, trajectory splits and downstream schedule. Core sizes are matched within 0.3%, counting the analytic model’s learned readout lift. Analytic uses a fixed phase update with zero trainable transition weights. Values are mean ± sample SD over three seeds.

CoreL2D acc. ↑L3D CE ↓3D Shapes MSE ↓
Frozen63.2 ± 0.40.779 ± 0.0190.0202 ± 0.0024
Joint59.6 ± 0.20.963 ± 0.0270.0474 ± 0.0016
Analytic66.5 ± 0.10.702 ± 0.0070.0093 ± 0.0009
RSSM59.5 ± 1.10.950 ± 0.0190.0773 ± 0.0002
LSTM60.1 ± 0.10.871 ± 0.0390.0334 ± 0.0017
Transformer62.5 ± 0.40.917 ± 0.0100.0438 ± 0.0085

Analytic has the best mean prediction metric in all three rows, with no path-integration pretraining. Frozen has the best mean among the learned recurrent cores shown on these metrics.

Freezing versus fine-tuning from the same checkpoint (Appendix C.4, Table 17)

Unfrozen init and Frozen start from the same checkpoint, with paired interfaces and identical downstream budgets. Joint starts the same modular architecture from random weights. Values are mean ± sample SD over three seeds.

ScaffoldL2D CE ↓L3D CE ↓L3D goal dist. ↓3D Shapes MSE ↓
Frozen0.926 ± 0.0090.779 ± 0.0199.12 ± 0.230.0202 ± 0.0024
Unfrozen init0.865 ± 0.0070.710 ± 0.0054.20 ± 1.210.0196 ± 0.0022
Joint1.025 ± 0.0030.963 ± 0.0279.52 ± 0.110.0474 ± 0.0016
Analytic0.908 ± 0.0110.702 ± 0.0071.66 ± 1.530.0093 ± 0.0009

Continuing to train the core improves language prediction and reduces imagined goal distance. The 3D Shapes change is small relative to seed variation. The gap between Unfrozen init and Joint points to the value of structured initialization.

Matching the upstream path-integration task and update budget (Appendix C.10, Tables 32–34)

We pretrained RSSM, GRU, LSTM and a wide RNN on the same 2.2 m × 2.2 m path-integration task for 100,000 updates, with matched core sizes. We evaluated each core frozen (F) or fine-tuned (FT) under the common downstream schedule. The seed-0 table reports the better arm per metric for each non-modular core.

CoreL2D CE ↓L3D CE ↓3D Shapes MSE ↓
Frozen scaffold0.93020.80000.0197
Scaffold FT0.87120.70900.0194
RSSM0.8954 F0.7257 F0.0173 F
GRU0.9028 FT0.7529 FT0.0231 FT
LSTM0.8900 F0.7074 FT0.0160 FT
Wide RNN0.9191 F0.9637 F0.0417 F
Analytic0.90040.70780.0098

A pretrained LSTM closes the Language 3D gap and beats Frozen on 3D Shapes, so upstream pretraining explains part of the gap to from-scratch baselines. The fine-tuned scaffold has the lowest Language 2D CE at this budget, and Analytic achieves the lowest 3D Shapes MSE. At 256 trajectories, Analytic reaches 0.7905 Language 3D CE versus 0.9834 for the best pretrained non-modular arm. These are single-seed comparisons. We interpret close values, such as 0.7074 and 0.7078, as similar.

AuthorsOpenReview ↗

R1 response (3/3) — Model architecture and training

Architecture and training clarification

We have expanded §3 and Appendix B to make the implementation reproducible. The learned recurrent modules approximate the phase translations described by Klukas et al.; we now specify their block-diagonal recurrence, initialization, displacement inputs, and phase readouts (Appendix B.2, Table 5, Algorithm 1). The analytic operator implements the explicit phase update in Eq. (2), and both operators are directly compared on path integration (Appendix C.1).

The observation encoder maps pixels, language, or robot states into a shared representation, while a learned router assigns module-specific phase coordinates and action displacements. Each module evolves independently under its corresponding control, and a decoder predicts future observations from the combined state. Importantly, neither the relevant latent factors nor their module assignments are provided during training.

Training uses free rollouts: only the initial observation is encoded, subsequent states evolve under recorded actions, and predictions are supervised at each step (Eq. 4). For goal-directed planning, a learned action scorer searches candidate sequences using current and goal representations; Language 2D additionally uses optimal action labels for planner training (§4, Appendix B.4–B.8). Full dimensions and training procedures are now documented in Tables 6–9 and Algorithm 3.

Implementation details (Appendix B.2–B.8)

Encoders and routing. The observation encoders output 192-dimensional vectors. Images use a four-layer CNN (32/64/128/256 channels), text a two-layer, four-head Transformer of width 192, and 39-dimensional MetaWorld states a 39→256→256→192 MLP. ReplicaCAD uses frozen DINOv2 features. Discrete actions use an embedding lookup and MetaWorld actions a 4→128→192 MLP. The shared 192→256→8×192 GELU router has separate heads per module. The phase head emits two normalized sine/cosine pairs. The action head uses a bias-free projection and a scaled tanh (0.10) to output a bounded two-dimensional control (Tables 6–7).

Transition and decoding. Frozen uses eight independent 512-unit ReLU modules. A phase-to-state initializer maps each four-component sine/cosine code into its module. Each step updates the module from its previous state and routed control. Analytic instead advances the two phase angles per module by the period-scaled control and wraps them modulo 2π (Eq. 2, Algorithm 2). The decoder predicts pixels through an upsampling head, token logits through a text head, or the 39 robot-state variables through an MLP (Table 8).

Rollout loss and planning. Eq. (4) supervises each decoded step with token cross-entropy for the language tasks and mean squared error for 3D Shapes and MetaWorld. The first-frame reconstruction has a separate weight (0.1 in the sweep). The planner scores actions and STOP from the state, goal state, their difference and its elementwise magnitude. Its MLP has two 256-unit hidden layers. It rolls out 32 candidates up to eight steps. In Language 2D, optimal actions at states the planner visits supervise its action scores (Algorithm 3).

We agree that the underlying grid dynamics and several architectural ingredients have important precedents. The contribution we wish to emphasize is the role of a fixed, structured dynamical manifold in shaping the learning of task representations and action controls, and the empirical evidence that this constraint can improve compositional structure, geometric stability, and predictive generalization. We hope the revisions make this distinction clearer.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.