MoTIF Wins Best Paper Award at ICML 2026 Workshop

The paper “Concepts in Motion: Temporal Concept Bottleneck Model for Interpretable Video Classification” has received the Best Paper Award at the 2nd Workshop on Compositional Learning: Safety, Interpretability, and Agents at ICML 2026. The paper was also selected for a spotlight presentation. The workshop took place in Seoul on July 11, 2026.

Authored by Patrick Knab, Sascha Marton, Philipp Johannes Schubert, Drago Andres Guggiana Nilo, and Christian Bartelt, the paper introduces MoTIF (Moving Temporal Interpretable Framework).

MoTIF helps explain video predictions over time. It shows which concepts appear, when they recur, and how they shape the final result. Its explanations cover the whole video, specific time windows, and the temporal behavior of individual concepts.

The framework also uses a vision-language model to discover object- and action-based concepts from training videos. This reduces the need for manual concept labels while producing concepts that remain meaningful over time.

Across several video benchmarks, MoTIF improves on global concept bottlenecks and remains competitive with other interpretable video models. The Best Paper Award recognizes both the technical contribution and its value for making video models easier to understand.

The workshop explored compositional AI as a path toward safer systems, stronger reasoning, and better generalization beyond training data.

The full paper is available on arXiv.