acceptodds
Under review as a conference paper at ICLR 2027

Fuzzy-World: Interpretable World Models via Grounded IF–THEN Dynamics

Abstract

World models support dynamical prediction by learning state transitions. However, aggregate prediction accuracy neither explains the basis of local state changes nor establishes where the learned relationships apply. Here, we introduce Fuzzy-World, an intrinsically interpretable world model combining named physical states with sparse fuzzy rules. Explicit local increment equations and Top-2 activation decompose each predicted physical increment into the weighted contributions of at most two executed rules, directly linking physical semantics, rule explanations, and transition computations. Rules are fitted using inference-time activation weights, with shared dynamics and support-dependent residual shrinkage controlling local corrections. An optional residual flow supplements observation information while keeping physical rules fixed during subsequent learning. Experiments on four dynamical systems show that, among eight dynamics models with observation reconstruction, Fuzzy-World achieves the lowest mean out-of-distribution 50-step observation prediction error and the highest mean physical-trajectory R² on all four systems. Complete rule audits characterize empirical validity within actual activation regions through sample coverage, effective sample size, and local prediction accuracy. These results demonstrate competitive dynamical prediction with sparse, inspectable local rules.

Primary area:
interpretability and explainable AI
Keywords:
World Models.Fuzzy Systems.Machine Learing

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_KjgkOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Fuzzy-World, an interpretable dynamics modeling framework that combines grounded physical-state representations with sparse Takagi–Sugeno fuzzy rules. The method uses Gaussian antecedents and Top-2 routing to predict physical-state increments through weighted combinations of explicit local equations, fitted using shared global Ridge dynamics and local residual corrections. A GRU-based residual model is additionally used for observation reconstruction in Advection. Experiments on Pendulum, HalfCheetah, Hopper, and 1D Advection evaluate ID and OOD prediction, rule-level validity, ablations, and short-horizon model predictive control. The reported results show competitive prediction accuracy, including the lowest mean OOD 50-step rollout errors among the main baselines across all four tasks.

Critical strengths

1. Simple and interpretable formulation of local dynamics.

The proposed framework combines grounded physical-state representations with explicit local dynamics equations and sparse Top-2 fuzzy routing. The resulting physical-state predictions can be directly decomposed into the weighted contributions of at most two local equations, providing a clear and computationally faithful account of how individual state increments are generated. The shared global dynamics model with activation-weighted local residual corrections also offers a relatively simple and structured approach to fitting nonlinear dynamics.

2. Extensive evaluation of predictive performance and local rule validity.

The experimental evaluation covers four dynamical systems with different characteristics and examines both one-step and long-horizon prediction under ID and OOD settings. Beyond the main comparisons, the paper includes short-horizon MPC experiments, an additional probabilistic neural ensemble baseline, and ablations of local corrections and active-rule budgets. The rule-level analysis further evaluates activation support and local predictive accuracy across the complete rule libraries, rather than relying solely on illustrative examples. These experiments provide useful evidence that the proposed formulation can achieve competitive predictive performance while retaining inspectable local dynamics.

Critical weaknesses

1. Limited Baseline Comparisons and Insufficient Experimental Support for the Core Design Choices

The experimental evaluation primarily compares Fuzzy-World against relatively classical regression, fuzzy-system, and system-identification baselines, with limited comparisons to recent interpretable dynamics methods. More relevant recent approaches, such as SINDy-RL [1], could provide stronger comparisons under adapted state-action dynamics prediction settings, while recent interpretable fuzzy architectures such as KANFIS [2] are relevant to the proposed rule-based modeling framework. Although Appendix D includes a probabilistic neural ensemble, the evaluation remains focused almost entirely on predictive accuracy rather than directly comparing the interpretability of the learned models. This makes it difficult to assess whether Fuzzy-World offers a meaningful advance over existing interpretable dynamics and fuzzy modeling approaches.

More importantly, the reported OOD improvements are difficult to attribute to specific methodological contributions. As acknowledged in Appendix E, the main baselines differ from Fuzzy-World in antecedent construction, feature libraries, rule counts, and fitting procedures. The substantially smaller improvements over Dense TSK in the matched ablations compared with the main experiments further highlight the difficulty of isolating the benefits of sparse routing. In addition, Fuzzy-World employs different configurations across tasks, including the number of stored rules, the choice of affine versus higher-order polynomial features, and support-dependent shrinkage settings. Advection additionally uses a GRU-based residual representation and nonlinear observation decoder, although the same reconstruction architecture is shared by its baselines. These choices may be reasonable, but their individual contributions are not systematically evaluated. The existing ablations primarily examine active-rule budgets and local residual corrections, leaving the effects of rule-library size, feature engineering, Gaussian antecedents, shrinkage, and neural residual modeling insufficiently isolated. More critically, the reported results do not consistently support Top-2 routing as the preferred choice: Top-1 achieves lower OOD rollout error on Pendulum and Advection, while Top-3 outperforms Top-2 on Hopper. The improvements over matched dense aggregation are also marginal for several tasks. Although Top-2 is motivated as a compromise between predictive performance and explanation complexity, this trade-off is not explicitly evaluated beyond limiting the number of active rules. Consequently, the experiments neither clearly establish which components account for the reported OOD gains nor provide convincing evidence for the particular Top-2 design emphasized as a central contribution.

2. Insufficient physical and neural-mechanistic interpretation of the learned model

In the Advection task, where Fuzzy-World incorporates a GRU-based residual flow and a learned nonlinear observation decoder, the interpretability analysis is confined to the rule-based physical-state dynamics, leaving the learned neural representations and their contributions to observation prediction largely unexplained. The paper does not investigate what information the recurrent hidden states encode, whether this information has identifiable physical meaning, or how it relates to the named physical variables. More importantly, there is no analysis of how the nonlinear decoder combines the predicted physical state with the recurrent residual state, or how much each contributes to the final observation prediction. Consequently, the reported IF-THEN rule traces explain only the physical-state increments, while the role of the learned neural representations in the complete observation-prediction model remains uninterpreted.

References [1] Zolman et al. SINDy-RL: Interpretable and Efficient Model-Based Reinforcement Learning [2] Yong et al. KANFIS: A Neuro-Symbolic Framework for Interpretable and Uncertainty-Aware Learning

Questions for authors

Clarification Requests

1. Rule-library size and feature selection.

What motivates the choice of 48, 16, 32, and 18 stored rules for Pendulum, HalfCheetah, Advection, and Hopper, respectively? How were these values determined? Similarly, why are polynomial and state-action interaction features included for Pendulum and Advection but not for the MuJoCo tasks? To what extent do these choices rely on prior knowledge of the underlying dynamics or task-specific hyperparameter tuning?

2. Interpretability of the GRU-based residual flow in Advection.

Could the authors investigate what information is captured by the GRU's recurrent hidden states and whether the learned representations correspond to meaningful physical or residual dynamical information beyond the named physical variables? In addition, how does the nonlinear observation decoder utilize the predicted physical states and recurrent residual representations, and what are their respective contributions to the reconstructed observations? An analysis of these aspects would help clarify the extent to which the claimed interpretability applies to the complete Advection model rather than only its rule-based physical-state predictor.

Suggestions for Further Experiments

1. More comprehensive comparisons with interpretable dynamics models.

The authors should consider comparisons with more recent approaches to interpretable dynamics learning, such as the dynamics-modeling component of SINDy-RL [1], under matched state-action prediction settings. Recent interpretable fuzzy architectures such as KANFIS [2] may also provide useful comparisons if adapted to the same task. Beyond MSE and R², such comparisons could evaluate the complexity, stability, physical consistency, and semantic interpretability of the learned dynamics. This would help establish whether Fuzzy-World improves interpretability over existing methods rather than merely providing competitive predictive accuracy.

2. Physical interpretation and activation analysis of learned rules.

The authors should provide a more systematic analysis of the physical meanings of the learned local equations and their associated activation regions. For example, do different rules correspond to identifiable physical states, dynamical regimes, or characteristic behaviors? Are the dominant terms and coefficients of their THEN equations consistent with the physics expected in their IF regions? Visualizing the activation regions together with the corresponding local dynamics, and examining per-rule activation frequencies, activation masses, overlaps, redundant or rarely activated rules, and cross-seed stability, would provide stronger evidence for the claimed physical interpretability.

3. Matched comparisons of rule-based dynamics and GRU-based residual modeling.

For the Advection task, the authors should additionally train two ablated models: a physical-only model using the rule-based physical-state predictor, and a neural-only model using the GRU-based residual representation. The physical-only model should retain the same rule-based state predictor and use an observation decoder matched to that of the full model, but without the recurrent residual input. The neural-only model should retain the same residual encoder and GRU architecture as the original model, together with a matched observation decoder, but without using the rule-based predicted physical states as decoder inputs. Both models should be trained independently under the same data and comparable training conditions.

The probabilistic ensemble evaluated in Appendix D does not address this comparison, as it uses a different neural dynamics architecture from the GRU-based residual flow in the main model. Comparing these two ablated models against the full Fuzzy-World results would help determine whether the rule-based physical-state predictor and GRU-based residual representation provide complementary information, whether either component can independently achieve comparable predictive performance, and how much each contributes to the final observation prediction. Further analysis of the recurrent hidden states and the decoder's use of the two representations would also help clarify the interpretability of the complete model.

4. More comprehensive ablations of the proposed design choices.

The current ablations mainly compare active-rule budgets and the presence of local corrections. Additional experiments should vary the total stored rule count KK while keeping the Top-2 execution budget fixed, and evaluate the effects of affine versus polynomial feature libraries, Gaussian antecedent construction and widths, and support-dependent shrinkage. These comparisons should preserve other model settings wherever possible. In particular, varying KK independently of the number of executed rules would directly test the claimed decoupling between rule-library capacity and per-transition explanation complexity. Such analyses would help distinguish the benefits of sparse rule execution from those of feature engineering, regularization, and other modeling choices.

Reviewer_1na7OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Fuzzy-World, an intrinsically interpretable world model that represents action-conditioned physical dynamics through explicit fuzzy IF-THEN rules over named state variables. A shared global dynamics model is combined with activation-weighted local residual corrections, and Top-2 routing restricts each transition to at most two executed rules. This allows each predicted physical increment to be decomposed into a small number of explicit rule contributions. The method is evaluated on Pendulum, HalfCheetah, Hopper, and 1D Advection under both ID and OOD conditions, with additional rule-level audits for support and local predictive accuracy.

Critical strengths

i) The proposed framework provides a clear and intrinsically interpretable connection between physical state variables, local fuzzy rules, and transition dynamics. The executed rules and their weights are part of the actual computation rather than post-hoc explanations, which makes individual state transitions directly inspectable.

ii) The shared-global plus local-residual formulation is technically well motivated. The global Ridge model captures common dynamics, while support-dependent shrinkage reduces the influence of poorly supported local corrections.

iii) The empirical evaluation is reasonably broad and emphasizes robustness under distribution shift. The method is tested on four dynamical systems with multiple baselines, five random seeds, and both one-step and long-horizon rollout metrics under ID and OOD conditions.

iv) The rule-level audit is a useful complement to aggregate prediction accuracy. The paper evaluates effective sample support, activation mass, and local R^2, which helps distinguish readable rules from empirically reliable rules.

Critical weaknesses

i) What is the specific benefit of fuzzy membership beyond providing a convenient soft routing mechanism? Could the authors compare against a non-fuzzy Top-2 mixture of local linear experts using the same shared-residual fitting and rule-audit procedure?

ii) How much of the original dense activation mass is typically captured by the two most active rules? Reporting the cumulative activation mass of the top two rules would clarify whether Top-2 substantially changes the dynamics computation or mainly imposes explanation sparsity.

iii) How sensitive is the method to the predefined physical-state representation and feature library? It would be useful to evaluate noisy, incomplete, or learned state representations to assess how dependent the method is on manually specified semantics.

iv) Can the authors provide broader evidence that the observed OOD advantage generalizes beyond the four evaluated systems? Additional environments, particularly those requiring learned state representations or more challenging long-horizon control, would strengthen the overall claim.

Questions for authors

i) What is the specific benefit of fuzzy membership beyond providing a convenient soft routing mechanism? Could the authors compare against a non-fuzzy Top-2 mixture of local linear experts using the same shared-residual fitting and rule-audit procedure?

ii) How much of the original dense activation mass is typically captured by the two most active rules? Reporting the cumulative activation mass of the top two rules would clarify whether Top-2 substantially changes the dynamics computation or mainly imposes explanation sparsity.

iii) How sensitive is the method to the predefined physical-state representation and feature library? It would be useful to evaluate noisy, incomplete, or learned state representations to assess how dependent the method is on manually specified semantics.

iv) Can the authors provide broader evidence that the observed OOD advantage generalizes beyond the four evaluated systems? Additional environments, particularly those requiring learned state representations or more challenging long-horizon control, would strengthen the overall claim.

Reviewer_txCoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes FuzzyWorld, which is an intrinsically interpretable world model combining named physical states with sparse fuzzy rules. It decouples rule-library capacity from pertransition explanation complexity. This makes the executed rule identities, weights, local increments, and resulting transition directly auditable while retaining graded cooperation between local regions. The proposed method is evaluated on four dynamical systems under in-distribution and out-of-distribution conditions and demonstrate good performance.

Critical strengths
  1. This paper addresses a core contradiction in interpretable world models: while the rule base can be very large, the explanation for any single prediction cannot be overly complex. The problem itself is interesting and valuable.
  2. The experiments are very detailed and cover a broad scope.
Critical weaknesses
  1. The components of the proposed method are largely combinations of existing concepts, like fuzzy routing, TSK local equations, residual correction, and shrinkage. The innovation lies primarily in the engineering integration of "Top-2 execution + rule auditing," rather than in an entirely new learning paradigm.
  2. Using Top-3 or Top-4 in Hopper yields clearly better results. The paper itself acknowledges that Top-2 is not universally optimal. Fixing the value at Top-2 appears to be an engineering compromise rather than a choice based on principle. If the goal of interpretability necessitates sparsity, why not employ an adaptive rule-selection mechanism? This point lacks a methodological solution.
  3. The paper claims to be "inherently interpretable," yet it lacks both human user studies and any quantification of interpretability.
Questions for authors
  1. Beyond presenting the rules and traces for the Pendulum environment, could the authors provide a user study—or at least showcase representative "IF-THEN" rules for high-dimensional tasks like HalfCheetah or Hopper? In what specific ways do the top-2 traces assist human understanding of high-dimensional state spaces?
  2. Since using the Top-3/Top-4 in Hopper yields better results, why is the number of active experts still fixed globally at 2 (Top-2)? Would it be possible to adaptively determine the number of active experts based on factors such as the availability of valid local samples, disagreements among rules, or prediction uncertainty?
  3. Could the authors report the quality of the discarded activations resulting from Top-2 truncation during actual training, along with the regularized output diameter, to verify whether the truncation bound is tight? Furthermore, does this bound effectively reflect a reduction in error for practical tasks?
  4. When a GRU residual stream participates in observation reconstruction, the trace component accounts only for physical increments. Is it possible to quantify the contribution of the residual stream to observation predictions? In this context, is it appropriate to characterize the entire model as "intrinsically interpretable"?
  5. Could the authors address on the specific innovations of their proposed method, rather than presenting it merely as a combination of existing approaches?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.