Tencent’s new AI method aligns preferences without retraining

HoE outperformed 15 recent baselines across 14 objectives and 200 different preferences in various tasks.