Reproduce “GenDSR: Transferring Representations From Video Generation Models for Dynamic Spatial Reasoning” 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
Ke Yang, Zhenyu Zhang, Jun Li
The authors introduce GenDSR, which adds video-generation-model features to a vision-language model for dynamic spatial reasoning. It combines raw features that retain spatiotemporal context with differences between neighboring features, then adds their gated fusion to visual tokens. On DSR-Bench, Qwen3-VL-8B reaches 67.2% overall sample-level accuracy, versus 62.0% with supervised fine-tuning alone; this benchmark comparison does not establish real-world performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.