Reproduce “FlowBench: Counterfactual Imagination Training for Spatial Intelligence” 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

Yifan Liu, Baihan Zhang, Khai Loong Aw, Klemen Kotar, Daniel L.K. Yamins

The authors introduce FlowBench, a benchmark for predicting spatial changes after hypothetical object motion from an initial image, motion, and question. It pairs these questions with outcome supervision, using simulator metadata for exact simulated answers. Simulation-only fine-tuning raises Qwen3-VL-8B accuracy from 42.7% to 59.6% on real tests, but a 14-point gap from outcome-visible answering remains.

Summary by alphaXiv.

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