Beyond Static Mechanistic Interpretability: Agentic Long-Horizon Tasks as the Next Frontier

Martian argues that interpretability research needs to move past static mechanistic analysis toward understanding agentic, long-horizon task behavior in models.