Position: Advocating for using Motion Figures in your ML Submissions

Zhenyuan Chen, Feng Zhang

Abstract: The rapid advancement of machine learning research---particularly in areas producing dynamic results such as video generation, 3D/4D scene synthesis, robotics, and world modeling---has created a fundamental disconnect between the dynamic nature of results and the static medium of traditional paper submissions. While researchers often resort to external websites or supplementary materials to display video results, this practice introduces friction for reviewers, risks archival decay through link rot, and raises concerns about double-blind review integrity. In this position paper, we propose a minimal specification for embedded dynamic figures (\DFIG) that leverages the \LaTeX{} \texttt{animate} package to present frame-sequence animations directly within PDF documents. We discuss technical feasibility, viewer compatibility constraints, and concrete guidelines for resolution, frame rate, file size, accessibility, and security. Rather than advocating immediate standardization, we present this as a call for community discussion and propose a pilot program for future venues. We argue the benefits of self-contained, archivally robust dynamic figures merit serious consideration as modern machine learning research increasingly depends on temporal evaluation.