Samuele Punzo

LLMs possess a distributed animacy circuit that reveals the intricate, context-dependent nature of how they process animate and inanimate concepts.

Hierarchical planning can unlock substantial performance gains in long-horizon control tasks, but only if high-level actions align with low-level execution capabilities.

Intuitive-physics knowledge in video models is not only present but also varies dramatically with model architecture and layer depth, revealing critical insights into representation learning.