World Model Interpretability: From Decoding Representations to Understanding Latent Worlds

AI summary

Overview

Research area: Interpretability of latent-space world models, specifically Joint-Embedding Predictive Architectures (JEPAs) and their use for prediction, planning, and control.

Technical level: Intermediate. The paper is a conceptual Perspective rather than an empirical study; it assumes familiarity with latent representation learning, probing methods, and causal intervention, but introduces its framework formally enough to be followed by readers with a machine learning background.

Scope: A position paper arguing that JEPA interpretability must shift from decoding what latent representations encode toward understanding how latent variables are organized, evolve, and causally drive prediction, organized as a four-stage research agenda (Ground, Structure, Verify, Act).

What This Paper Is About

JEPAs such as V-JEPA 2 predict in an abstract representation space rather than reconstructing pixels, which makes them useful as latent world models for planning and control but leaves no human-interpretable interface. Existing interpretability work mostly probes what information can be decoded from a latent state, which does not reveal how that information is organized, how it evolves over time, or whether the predictor actually relies on it. The paper's goal is to define "latent-world interpretability" as a distinct research target and lay out a staged blueprint for pursuing it.

Key Contributions

1. A formal framing of latent-world interpretability. The paper defines a JEPA-based latent world model as an encoder-predictor pair, with latents produced by the encoder and future latents produced by an action-conditioned predictor, and defines a corresponding set of interpretable external world variables. It distinguishes representational accessibility (what can be recovered from a latent state) from world-model understanding (how information is organized, evolves, and causally contributes to prediction).

2. A four-stage blueprint: Ground, Structure, Verify, Act. Each stage is given a formal definition: Ground maps latent factors or subspaces to world variables; Structure recovers relations and dynamics among grounded factors; Verify tests causal roles through controlled latent interventions; Act uses verified mechanisms to diagnose and change model behavior, closing the loop between interpretation and behavior.

3. A mapping of existing work onto the blueprint. The paper surveys JEPA interpretability efforts and neighboring XAI and mechanistic interpretability tools, and tabulates representative work (V-JEPA 2 probing, AIM probing, WorldModelLens, WAM Steering) by how much of each stage they cover. Its conclusion is that direct evidence in JEPAs clusters at Ground, with Structure, Verify, and Act largely unaddressed.

4. A discussion of implications, counterarguments, and a research call. The paper argues that latent-world understanding enables diagnosing where prediction failures arise, distinguishing learned world structure from shortcuts, explaining generalization limits, and supporting decision-bounded interpretation before predictions are used for action. It also addresses objections (decoding is sometimes sufficient, human-defined ontologies may be the wrong reference frame, latent interventions may be invalid, interpretability may not be necessary for reliability) and proposes four research priorities.

Main Findings

• Decodability is not understanding. Successfully recovering a variable such as velocity from a latent state shows the information is accessible, but not how it relates to other variables, how it shapes latent dynamics, or whether the predictor uses it; the same prediction could be driven by correlated visual cues or shortcuts.

• Existing JEPA interpretability is concentrated at the Ground stage. Probing in V-JEPA 2 shows frozen representations contain information about appearance, motion, and action, and passive discrete probing examines correspondences between physical variations and unsupervised symbolic structure; both provide grounding evidence rather than evidence about structure, causality, or intervention.

• Tools for later stages come mainly from outside JEPAs. WorldModelLens provides infrastructure for probing, attribution, sparse representation analysis, activation patching, and intervention replay across world-model architectures including JEPAs, while work on World Action Models provides a non-JEPA precedent in which structured activation patterns are analyzed, dynamically modeled, and manipulated to affect downstream behavior. Direct evidence in JEPA-based world models remains limited.

• Predictive performance alone cannot establish world structure. Similar performance may arise from meaningful physical structure or from correlated cues such as background texture and camera motion, so internal analysis is needed to distinguish a learned world model from a shortcut.

• Latent worlds are domain-dependent. The relevant variables, relations, and dynamics differ by application: objects, contacts, and motion in robot manipulation; objects, positions, and spatial relations in 3D localization; instruments, anatomical structures, interactions, and procedural phases in surgical video; domain-specific variables in scientific settings. No universal set of interpretable features can be assumed.

• Prediction failures have multiple internally distinguishable causes. Using a robot pushing a cup near a table edge as an example, the paper separates failure in state representation, in learned dynamics, in modeled interaction, and in drift during multi-step rollout, noting that an observed error alone does not identify which occurred.

• No quantitative results are reported. The paper is a Perspective and presents no experiments, datasets, benchmarks, or measured numbers; its claims are conceptual and organizational.

Methodology in Plain English

The paper is a position piece, so its "method" is conceptual scaffolding rather than experiments. It first writes down a minimal mathematical description of a JEPA world model: an encoder turns an observation into a latent state, and a predictor turns the history of latent states (and actions, if available) into the next latent state. It then contrasts this with the standard "decoding paradigm," in which a probe or decoder is trained to recover interpretable world variables from latents. The central move is to argue that this decoding test answers only an accessibility question, and to define a sequence of progressively stronger questions: which latent factors correspond to which world variables, how those factors relate and evolve, whether the predictor causally depends on them (tested by altering a factor in latent space and checking whether prediction changes in the expected way), and finally whether verified understanding can be used to fix or steer the model. The authors then sort existing papers into this four-stage scheme, discuss what the gaps imply, and respond to likely objections before listing research priorities. A curated list of related work accompanies the paper on a public GitHub repository.

Why This Matters

Impact on research. The paper reframes the object of interpretation for world models: not individual representations, but the predictive system that transforms latent states over time. It gives the field shared vocabulary and a staged progression from correspondence to causality to intervention, which makes claims about what a world model "understands" testable rather than rhetorical.

Real-world applications (domains the paper identifies as JEPA application areas where interpretability would matter):

• Robot manipulation and embodied control, where contact and motion predictions feed directly into actions and failures propagate into decisions.
• 3D scene understanding and localization, where objects, positions, and spatial relations must be represented reliably.
• Medical analysis, including surgical video where instruments, anatomical structures, and procedural phase progression are the relevant latent world.
• Scientific discovery, where the variables and interactions under study are domain-specific and cannot be assumed in advance.

Industry relevance. Systems that use JEPA predictions for planning and control need a way to judge whether a prediction rests on verified internal mechanisms or on patterns known to be unstable. The paper's decision-bounded interpretation proposal — assessing predictions before they are acted on, not just explaining failures afterward — speaks directly to safety-critical deployment, and its emphasis on separating state, dynamics, interaction, and rollout failures maps onto practical debugging of deployed predictive models.

Future Directions

• Build controlled testbeds and evaluation criteria. Environments with known objects, physical variables, interactions, and dynamics would allow interpretations to be checked against ground truth, with evaluation beyond decoding accuracy to include stability across contexts, consistency of recovered relations and dynamics with observed behavior, and whether causal claims survive controlled changes.

• Study how latent worlds emerge and differ across models. Architecture, training objective, data, model scale, and training stage may all shape the internal predictive structure a JEPA learns; comparing along these axes could show when world-relevant factors and dynamics emerge and which design choices yield more stable or reusable structure.

• Develop valid latent interventions. The paper flags that changing a factor associated with velocity may inadvertently change other properties, and that the modified state may no longer be a plausible world state, making causal attribution difficult; constructing interventions that change the target while preserving other information and plausibility is an open problem for the Verify stage.

• Scale analysis to realistic settings. Methods validated in controlled environments must be tested on natural video, long-horizon prediction, and embodied agents in complex environments, which requires shared infrastructure and protocols for extracting and comparing internal states, tracing predictive computation over time, performing interventions, and analyzing rollouts across model families.

Target Audience

Researchers and graduate students working on world models, self-supervised representation learning, and JEPA-style architectures who want a research roadmap for interpretability; mechanistic interpretability and XAI researchers looking to extend their toolbox from representations to predictive dynamics; and practitioners deploying predictive world models in robotics, embodied control, medical video, or scientific applications who need to know when a model's predictions can be trusted. Readers seeking empirical results or benchmark comparisons will not find them here, since the paper reports no experiments.

Authors’ abstract

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.