Reproduce “CauSight: Learning to Supersense for Visual Causal Discovery” 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
Yize Zhang, Meiqi Chen, Sirui Chen, Bo Peng, Yanxi Zhang, Tianyu Li, Chaochao Lu
CauSight formalizes the task of visual causal discovery, enabling Vision-Language Models to infer directed, predicate-labeled causal relationships between entities in single images. It achieves improved performance over leading VLMs in causal graph generation on the new VCG-32K dataset, demonstrating robust generalization.
Key insights
- Process supervision through synthesized reasoning trajectories (ToCT) is highly effective in teaching VLMs structured, multi-step causal reasoning, improving their causal comprehension.
- Reinforcement Learning with a task-specific causal reward further refines model policy, leading to substantial gains in accurate visual causal discovery and faithful graph generation.
- Specialized training for visual causal discovery can enhance reasoning capabilities and cross-domain generalization without degrading the base model's broader general perception and reasoning abilities.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.