World Model Interpretability: From Decoding Representations to Understanding Latent Worlds
Jize Wang, Yongcheng Jing, Xinyi Le, Cailian Chen, Xinping Guan, Dacheng Tao
Shanghai Jiao Tong University · Nanyang Technological University
Published October 18, 2026 · on the feed October 20, 2026 · ▲ 9 · arXiv 2610.25296
The 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.
The Brief
1:53A short podcast summary of the paper. Written and voiced by AI.
Transcript
World Model Interpretability. From Decoding Representations to Understanding Latent Worlds. By Jize Wang and five others, from Shanghai Jiao Tong University and Nanyang Technological University. Published on the eighteenth of October. Its code repository is a placeholder: a readme, and nothing to run yet.
Joint-Embedding Predictive Architectures, known as JEPAs, are growing beyond representation learning. They are turning into latent world models that support prediction, planning and control. But the latent worlds they build come with no interface a human can read. This piece is written as a Perspective, and its argument is that interpretability has to change along with the models. Once JEPA latents become the internal states used for prediction and action, it is no longer enough to ask what they encode. The questions become how latent variables interact, how they evolve, and how they causally drive prediction.
To guide that shift, the authors lay out a research agenda in four stages. Ground. Structure. Verify. Act. The stages are named rather than spelled out one by one, but the order carries the argument. It runs from grounding what the latents are, through the structure among latent variables, to verifying what causally drives prediction, and finally to acting on the latent world. Each step asks more of our understanding than the one before it. The direction matches the shift the authors call for: from what is encoded, toward how latent variables interact, evolve, and cause.
Placed on that map, existing interpretability work on JEPAs offers early evidence for the first stage, Ground. The later stages are where the field is thin. Systematic analysis of latent structure, causal verification, and intervention all remain largely unexplored.
The Perspective is a call to widen the agenda. The goal it sets is to move attention away from decoding representations, and toward understanding, validating and acting on the latent worlds these models carry.
More with an account
Listen to every day’s papers
Daily playlists, deep dives, quests for the papers, and quizzes on their key terms.
Join free