Reproduce “Intervene, Calibrate, Discover: Learning Causal Structure from Pixels” 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
Ziyi Ding, Xiaodi Huang, Xianxin Lai, Haijier Chen, Zhang Zhang, Xiao-Ping Zhang, Jiayu Chen
The authors calibrate visual readouts for known variables to verified environment values using dimension-wise monotone maps, then learn graph directions from interventions and fit causal mechanisms. In a three-seed Domino test, calibration lowers mean non-target counterfactual-change error from 5.463 ± 0.946 to 4.968 ± 0.123, with errors reported ×10³. However, affine calibration performs better on in-distribution effects, so gains are conditional, not universal.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.