Square$χ$PO: Differentially Private and Robust $χ^2$-Preference Optimization in Offline Direct Alignment

Open in new window