Tsallis Mirror Descent

STaR-DRO: Stateful Tsallis Reweighting for Group-Robust Structured Prediction

This research introduces a two-part framework for group-robust structured prediction, addressing challenges like ambiguity and heterogeneous group difficulty. It proposes a task-agnostic prompting strategy and STaR-DRO, a novel stateful robust optimization method for learning concentrated on persistently hard groups.