Jize-W / Awesome-World-Model-Interpretability

# Awesome World Model Interpretability
A curated list of papers on **world model interpretability**, **Joint-Embedding Predictive Architectures (JEPAs)**, and related explainable AI methods.
> 📢 We have released our perspective paper, **World Model Interpretability: From Decoding Representations to Understanding Latent Worlds**. This Perspective aims to establish latent-world understanding as a central goal for interpreting predictive world models. We note that decodability is not understanding: accessible information does not reveal how it is organized, evolves, or causally supports prediction. Building on this observation, we envision the Ground–Structure–Verify–Act blueprint as a path toward this goal, and as a framework for shaping future research on latent-world interpretability. We call for future interpretability efforts to move beyond isolated representation-level analyses toward systematically understanding, validating, and acting on the internal worlds learned by predictive models. This repository collects related work supporting this agenda.
Contributions and corrections are welcome through issues and pull requests.
- Awesome World Model Interpretability
 - Category Overview
 - General XAI Toolbox
 - Latent Space World Models
 - General Models and Predictive Architectures
 - Downstream JEPA Applications
 - Robotics and Embodied Control
 - 3D Perception and Autonomous Driving
 - Medicine and Physiological Sensing
 - Genomics, Cells, and Collider Physics
 - Graphs, Time Series, and Physical Fields
 - Remote Sensing, Audio, and Language
 - Related Work in XAI Involving World Models
 - Explicit Structure and Interpretable Model Design
 - World Representations in Transformers and LLMs
 - World Models for Explaining Agents
 - Transferable Methods and Behavioral Evaluation
 - Interpreting Learned World Models
## Category Overview
| Category | What It Covers | Relationship to World Model Interpretability |
| :--- | :--- | :--- |
| **General XAI Toolbox** | Attribution, concepts, probes, sparse features, circuits, causal interventions, and steering. | **Methodological foundations.** Provides tools that can be adapted to world models; their effectiveness for latent predictive dynamics requires validation. |
| **Latent Space World Models** | General predictive architectures and domain-specific JEPA applications. | **Models to interpret.** Defines the model families and application settings that motivate interpretation; downstream performance alone does not explain internal mechanisms. |
| **Related Work in XAI Involving World Models** | Explicitly structured model design, emergent world representations in Transformers and LLMs, and world models used to explain agents. | **Weakly related / indirect relevance.** Primarily studies structured model design, emergent world representations, or agent explanations, rather than directly interpreting the internal predictive mechanisms of learned world models, especially JEPAs. Provides contextual references