The approach, called COAR, allows us to make surgical "edits" to ML models. By intervening on the right components, we can fix errors, "forget" harmful labels, or boost performance on subpopulations that are underrepresented in data.

→ View original post on X — @mit_csail,