Umur Yıldız^1,2 (), Burcu Ayşen Ürgen^1,2; ¹Bilkent University, Ankara, Turkey, ²Aysel Sabuncu Brain Research Center and National Magnetic Resonance Research Center (UMRAM), Bilkent University, Ankara, Turkey

Naturalistic videos are increasingly used in vision science to probe how the visual system represents events, but standard behavioral methods still rely on sequential ratings or exhaustive pairwise comparisons of individual clips. These approaches are temporally inefficient for large stimulus sets and capture little contextual information across stimuli. While the multi-arrangement method (Kriegeskorte & Mur, 2012) offers a scalable alternative, practical implementation options are limited. Existing tools typically focus on static images, require payment, or demand substantial custom implementation, leaving no free, off-the-shelf software option. To address this gap, we present Multiarrangement, a Python toolkit designed to efficiently implement the multi-arrangement method for dynamic visual stimuli. During the procedure, participants arrange subsets of videos within a circular arena such that spatial distances correspond to perceived dissimilarity. These partial arrangements are then aggregated to estimate a full representational dissimilarity matrix (RDM) reflecting the high-dimensional geometry of the stimulus space. The toolkit includes two trial schedulers: a fixed set-cover strategy that guarantees at least one co-occurrence for every unordered pair while keeping trial size small and that does not require an initial full-stimulus trial, and an adaptive lift-the-weakest strategy that preferentially samples poorly constrained regions of the similarity space. We validate the approach with two participants on a 58-item set of human action videos, comparing multi-arrangement dissimilarities against those from traditional one-by-one comparison tasks. Multiarrangement yields stable similarity structures that closely track those obtained from pairwise ratings (mean Pearson r = 0.77, concordance correlation = 0.68), while requiring fewer explicit comparisons and showing a mild compression of the dissimilarity range. The method offers a practical way to obtain dense, behaviorally grounded similarity data for large sets of naturalistic videos with minimal programming effort, providing a useful bridge between human judgments, computational models, and neural measures in vision science.