Multi-Modal Spatial Reasoning with Human Motion
Multi-modal language models still struggle with the 3D world — judging directions, taking another viewpoint, estimating distances — and prior work traces most of these failures to egocentric–allocentric frame transformation, not perception. Humans learn this by moving through space. This thesis will test whether full-body human motion can supply that signal. The plan is to extract a spatial-reasoning dataset coupled to real human motion, train an MLLM on it, and measure whether the learned ability transfers to existing benchmarks.