Tag
This paper probes whether visual foundation models maintain persistent spatial states by evaluating four capability levels—spatial structure, object persistence, predictive spatial update, and actionability—using frozen-feature probes on ten representations under streaming egocentric observations, finding that readout accuracy does not guarantee history-based updates or reliable goal-directed action.
This position paper argues that LLM self-explanations can be plausible, questionably faithful, but highly actionable, and proposes evaluation guidelines beyond traditional metrics.
This position paper argues that interpretability research should be evaluated based on actionability—the extent to which insights enable concrete decisions and interventions. The authors propose a framework with evaluation criteria aligned with practical outcomes to address the lack of real-world impact in current interpretability work.