This project investigates whether symbolic planning predicates – such as object relations like “On(x, y)” or “Clear(x)” – are implicitly encoded as transferable latent features within frozen vision models. While classical planning relies on explicit symbolic abstractions and modern foundation models learn high-dimensional visual representations, the bridge between these paradigms remains underexplored. I propose to probe frozen vision models, such as DINOv2 and I-JEPA, to determine if these predicates can be decoded from their internal activations. By testing whether predicate probes trained on one model can transfer to another via linear alignment, this research aims to evaluate if such symbolic structures are universal latent features of reality, consistent with the Platonic Representation Hypothesis, rather than model-specific artifacts. Finally, I will assess whether these decoded predicates improve planning performance in robotic domains, offering a potential path toward more robust, generalizable symbolic world models.

I am interested in representation and would like to explore how foundation models can bridge the gap between neural and symbolic abstractions.