MAdam: Metric-Aware Multi-Objective Adam

MAdam fixes a mismatch between Multi-Objective Optimization (MOO) solvers and Adam by introducing a metric-aware preconditioning wrapper. It corrects two issues: weighting mismatches where Adam ignores time-varying preferences, and geometric mismatches where Adam's adaptive metric distorts the solver's direction. MAdam uses preference-conditioned curvature to govern the update, leading to better performance across diverse MOO tasks.

Open the full episode

More from AI Radio