Reproduce “Trace4D: Grounded Spatiotemporal Reasoning via 4D State Sequences” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Wooseok Jang, Seungryong Kim, Yan Lu, Lei Ji

Trace4D makes video reasoning over changing spatial relationships explicit by predicting object positions across multiple frames, paired with bounding boxes, then comparing state changes with visible action. On Trace4D-Bench, its Qwen3-VL-4B version reaches 79.25% accuracy, versus 38.92% for the base model. The benchmark uses held-out source videos but the training task formulation.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.