Experiments on large-scale vision & language models show COAR yields accurate component attributions, outperforming prior methods. This makes it a valuable tool for understanding & debugging complex ML systems.

→ View original post on X — @mit_csail,