PaperlayerView exact revision on arXiv

Source-linked AI summary

World Model Interpretability: From Decoding Representations to Understanding Latent Worlds

Jize Wang, Yongcheng Jing, Xinyi Le, Cailian Chen, Xinping Guan, Dacheng Tao

arXiv:2610.25296v1cs.LGcs.AI

TL;DR

JEPAs increasingly use latent states for prediction and action, but those states lack a naturally interpretable interface, and decoding them does not reveal how the predictive system works. The paper proposes the Ground–Structure–Verify–Act blueprint; current JEPA interpretability evidence is concentrated at Ground, while systematic study of structure, causal support, and intervention remains open.

  • Problem

    JEPA latents lack a naturally interpretable interface, so decoding accessible information does not show how latent variables are organized or causally contribute to prediction.

    Source: §1 · p. 1, §2.3 · p. 3

  • Method

    The paper formulates latent-world interpretability and proposes Ground, Structure, Verify, and Act as a blueprint for connecting, analyzing, testing, and using latent-world understanding.

    Source: §1 · p. 2

  • Results

    Existing JEPA interpretability evidence is concentrated primarily at Ground, while systematic study of latent structure, causal support for prediction, and targeted intervention remains largely open.

    Source: §4 · p. 6

  • Takeaways & Limitations

    The paper calls for interpretability efforts to move beyond isolated representation-level analyses toward systematically understanding, validating, and acting on predictive models’ internal worlds.

    Source: §8 · p. 9

    The proposed blueprint’s assumptions and methods require further validation.

    Source: §6 · p. 8

Abstract

from arXiv · show

Emerging Joint-Embedding Predictive Architectures (JEPAs) are moving beyond representation learning towards latent world models that support prediction, planning, and control. Yet their latent worlds naturally lack a human-interpretable interface. In this Perspective, we argue that as JEPA latents become internal states for prediction and action, interpretability must move beyond what they encode to how latent variables interact, evolve, and causally drive prediction. To this end, we envision a four-stage research agenda, involving Ground, Structure, Verify, and Act, to guide this shift toward actionable latent-world understanding. Within our framework, existing JEPA interpretability efforts provide early evidence for Ground, while systematic analysis of latent structure, causal verification, and intervention remains largely unexplored. Overall, this Perspective aims to call for a broader latent-world interpretability agenda that shifts attention from decoding representations toward understanding, validating, and acting on latent worlds.

1. Introduction

As JEPAs increasingly support prediction and action through latent states, their lack of an interpretable interface makes it difficult to understand what world they have learned. The paper proposes a four-stage blueprint for moving from decoding latent content toward understanding how latent variables interact, evolve, and drive prediction.

1. Introduction

  • JEPAs use latent representations to predict future states and, when conditioned on actions, support planning and control.Their use is expanding across embodied control, 3D scene understanding, medical analysis, and scientific discovery.Source: §1 · p. 1
  • Without a naturally interpretable interface, prediction failures are difficult to trace to state representation, learned dynamics, missing interactions, or errors accumulated over time.Predictive performance alone also cannot establish whether a model learned meaningful world structure or relies on shortcuts within its training distribution.Source: §1 · p. 1
  • Representation probing can reveal recoverable semantic or physical information, but not how that information functions in the model’s internal predictive world.For example, decoding velocity does not show how it relates to other variables, shapes latent dynamics, or causally contributes to prediction.Source: §1 · pp. 1–2
  • The Ground–Structure–Verify–Act blueprint connects latent factors to external variables, studies their relations and dynamics, tests causal involvement, and uses verified understanding to diagnose and improve behavior.Existing efforts largely probe latent content or apply isolated analysis and intervention techniques, leaving systematic understanding underexplored.Source: §1 · p. 2, Figure 1 · p. 2
  • The paper formulates latent-world interpretability, organizes existing efforts, identifies gaps, and outlines priorities for studying predictive world models.It presents the blueprint as a systematic research agenda.Source: §1 · p. 2

2. Technical Setup

The technical setup treats a JEPA world model as a predictive process over latent states and distinguishes decoding accessible information from understanding how the predictive system uses it. The gap motivates studying latent organization, evolution, and causal contribution, including how actions affect evolution in action-conditioned models.

2.1. JEPA as a Latent Predictive World Model

  • A JEPA world model uses an encoder and predictive dynamics model to represent observations as latent states and predict their evolution rather than reconstructing future observations in pixel space.The latent world comprises internal variables, relations, and dynamics that unfold over time and may depend on actions.Source: §2.1 · p. 3

2.2. Interpretability under the Decoding Paradigm

  • The decoding paradigm asks what semantic or physical information can be recovered from latent representations, using probes and methods such as concept analysis, clustering, and sparse representations.It provides a starting point for JEPA interpretability by characterizing accessible content.Source: §2.2 · p. 3

2.3. The Gap: Decoding Is Not Understanding

  • Decoding a variable such as velocity does not show how it relates to position or contact, evolves jointly over time, or contributes causally to prediction; the predictor may instead use correlated cues or shortcuts.World-model understanding concerns organization, evolution through learned dynamics, and causal contribution; for action-conditioned models, it also concerns how actions influence evolution.Source: §2.3 · p. 3

3. Blueprint: From Decoding to Understanding Latent Worlds

The blueprint shifts interpretability from identifying what latent factors represent to understanding their relationships, dynamics, and causal roles in prediction. It progresses through Ground, Structure, Verify, and Act, using verified understanding to diagnose and improve model behavior.

3. Blueprint: From Decoding to Understanding Latent Worlds

  • The four-stage blueprint moves from grounding latent factors in world variables, through recovering relations and dynamics and testing causal roles, to using verified understanding to diagnose and improve behavior.The stages are Ground, Structure, Verify, and Act.Source: §3 · p. 3, Figure 2 · p. 4

3.1. Ground: Connecting Latents to World Variables

  • Ground establishes correspondences between latent representations and interpretable world variables; probes, clustering, sparse autoencoders, and related methods are examples of this stage.Each mapping associates a latent factor or subspace with a world variable.Source: §3.1 · p. 4

3.2. Structure: Recovering Relations and Dynamics

  • Structure recovers relations among grounded factors and their evolution, including how actions influence dynamics in action-conditioned models.For example, it examines how position, velocity, and contact interact during motion or collision.Source: §3.2 · p. 4

3.3. Verify: Testing Causal Roles

  • Verify uses controlled interventions or counterfactuals to test whether grounded factors, relations, or dynamics causally participate in prediction.Changing a factor grounded as velocity while holding other relevant factors fixed should systematically alter predicted motion, distinguishing predictive participation from correlation.Source: §3.3 · pp. 4–5

3.4. Act: Using Interpretations to Change Model Behavior

  • Act derives targeted interventions from verified mechanisms, including latent steering, activation editing, retraining, or fine-tuning, to diagnose and improve model behavior.The proposed evaluation criterion J may measure prediction accuracy, out-of-distribution robustness, planning success, safety, or other relevant outcomes.Source: §3.4 · p. 5
  • Act evaluates an intervention by whether its task-dependent criterion improves, expressed as J(M ′) > J(M).The criterion may measure prediction accuracy, out-of-distribution robustness, planning success, safety, or other relevant outcomes.Source: §3.4 · p. 5

4. Mapping Existing Work onto the Framework

Existing JEPA interpretability work mainly grounds world-relevant information in latent representations, while systematic study of latent structure, causal support, and intervention remains largely open. The framework also draws on broader interpretability tools and emerging work in neighboring model classes.

4. Mapping Existing Work onto the Framework

  • XAI and mechanistic-interpretability methods offer tools for moving from feature identification toward causal testing, but latent world models require analysis of relations and predictive dynamics too.The proposed framework includes attribution, counterfactuals, probing, sparse autoencoders, circuit analysis, and activation interventions.Source: §4 · p. 5
  • The latent variables and interactions to interpret depend on the application, including objects, contacts, and motion in manipulation or instruments and procedural phases in surgical video (Assran et al., 2025).The framework is common across domains, but the variables, relations, and dynamics are domain-defined.Source: §4 · p. 6
  • V-JEPA 2 probing and passive discrete probing provide early evidence for grounding appearance, motion, action, and physical variations in JEPA latents (Assran et al., 2025; Ming, 2026).These studies examine recoverable information and correspondences with unsupervised symbolic structure.Source: §4 · p. 6
  • WorldModelLens offers analysis and intervention infrastructure across world-model architectures, while World Action Models provide a non-JEPA precedent for modeling and manipulating activations (Challagundla et al., 2026; Hong et al., 2026).Direct causal-intervention evidence in JEPA-based world models remains limited.Source: §4 · p. 6
  • Direct JEPA evidence is concentrated at Ground, while systematic analysis of relational and dynamical structure, causal support, and targeted intervention remains largely open.The table maps representative work onto Ground, Structure, Verify, and Act; its caption defines the coverage symbols as explicit, partial, or not addressed.Source: §4 · p. 6, Table 1 · p. 6

5. Implications: What Understanding Latent Worlds Enables

Understanding latent worlds could help localize prediction failures, distinguish learned structure from shortcuts, and examine generalization under environmental change. For models used in planning and control, it could also inform whether predictions should be trusted before they guide decisions.

5.1. Diagnosing where prediction failures arise

  • Grounded factors, recovered relations and transitions, and causal analysis could distinguish whether prediction errors arise from missing state information, inaccurate dynamics, interactions, or rollout drift.The cup-falling example illustrates how distinct internal failures can produce the same observed error.Source: §5.1 · pp. 6–7

5.2. Distinguishing learned world models from shortcuts

  • Similar predictive performance may reflect meaningful physical structure or reliance on correlated cues, so examining verified mechanisms as correlations change could distinguish world structure from shortcuts.Stable dependence on world-relevant factors across objects, scenes, and environments would provide stronger evidence of transferable structure than dependence on incidental training-distribution properties.Source: §5.2 · p. 7

5.3. Explaining generalization and its limits

  • Tracking which latent factors remain stable across environmental changes could connect internal organization to reliable prediction and reveal likely sources of out-of-distribution failure.Factors stable to appearance changes but varying with geometry or dynamics may support transfer; factors tied to background or viewpoint may signal likely failure sources (Wang et al., 2025; Sun et al., 2024; Poudel et al., 2024).Source: §5.3 · p. 7

5.4. From post-hoc explanation to decision-bounded interpretation

  • Before predictions guide planning or control, interpretability could assess whether state variables are represented reliably and predicted dynamics match verified mechanisms.This could provide a signal for deciding when predictions should be trusted, rather than limiting interpretation to post-action explanations.Source: §5.4 · p. 7

6. Alternative Views and Limitations

The blueprint is intended for questions about predictive organization and causal reliance, not every representation-level question. The authors also identify challenges in mapping human concepts to latent structure and validating interventions, and position latent-world analysis as complementary to behavioral evaluation.

6. Alternative Views and Limitations

  • A probe may suffice to test whether a representation contains object identity or motion, but predictive organization and causal reliance require evidence beyond decodability.The blueprint targets questions about how represented information participates in prediction.Source: §6 · p. 7
  • Human concepts need not map one-to-one onto neurons or latent directions because predictive factors may be distributed, nonlinear, relational, or task-dependent (Chawla et al., 2026).Grounding can use human concepts as reference points for testable correspondences involving subspaces or distributed representations.Source: §6 · pp. 7–8
  • Changing a latent factor may also change other properties or create an implausible state, making the resulting prediction change difficult to attribute to the intended intervention.Verify therefore needs interventions that preserve relevant information and keep the modified latent state plausible.Source: §6 · p. 8
  • Latent-world interpretability should complement, not replace, behavioral evaluation, uncertainty estimation, robustness training, and out-of-distribution testing.Behavioral evaluation can show whether and when a model fails, while internal analysis may help identify mechanisms behind failures and distinguish different internal strategies.Source: §6 · p. 8

7. Call to Action

The paper calls for a systematic research program centered on latent-world variables, structures, dynamics, and causal mechanisms. It identifies priorities spanning controlled evaluation, comparisons across models, behavior and design, and realistic settings.

7. Call to Action

  • Build controlled testbeds with known world properties to test whether recovered structure reflects the system, remains stable across contexts, and supports valid causal claims.Such settings cannot capture the full complexity of real-world prediction, but can provide a foundation for systematic evaluation.Source: §7 · p. 8
  • Compare models across architecture, objective, data, scale, and training stage to study when predictive structure emerges and which design choices yield stable or reusable structure.This extends interpretability to how latent worlds form and evolve.Source: §7 · p. 8
  • Test interpretations against prediction, generalization, failure, and planning or control behavior, and use associations with reliability to inform comparisons of model design choices.Interpretability could help analyze trained models and examine which choices lead to better predictive world models.Source: §7 · p. 8
  • Scale analysis to natural video, long-horizon prediction, and embodied agents in complex environments using tools for state comparison, temporal tracing, interventions, and rollout analysis.Shared infrastructure and evaluation protocols could improve reproducibility and comparisons across models and studies.Source: §7 · p. 8

8. Conclusion

The Perspective makes latent-world understanding a central goal for interpreting predictive world models. It proposes Ground–Structure–Verify–Act as a research blueprint for moving beyond isolated representation-level analyses.

  • The Perspective establishes latent-world understanding as a central goal for interpreting predictive world models.Source: §8 · p. 9
  • The Ground–Structure–Verify–Act blueprint offers a path toward understanding latent worlds.Source: §8 · p. 9
  • Future efforts should move beyond isolated representation-level analyses toward systematically understanding, validating, and acting on predictive models’ internal worlds.Source: §8 · p. 9
Loading 2610.25296v1…