Philip Torr

Featured Work

official workshop page — 2025

Why Now

He belongs in the set because he is helping define the evaluation and robustness side of the frontier, especially around whether world models are actually controllable and understandable.

Key Ideas

• -World models matter only if they are robust under intervention, not merely visually impressive.
• -Interpretability for vision is becoming operational rather than philosophical because deployed systems need debuggable failure modes.
• -The field needs harder tests for interaction, robustness, and causal structure.

Open Questions

• ?What makes a world model interactable in practice rather than only generative?
• ?How do we evaluate mechanistic understanding in vision systems without reducing it to saliency theater?
• ?Can robustness become a design principle instead of a post hoc patch cycle?

Research Theme

robust perception and uncertainty