ResearchBenchmarksDialogue Preference Alignment on HH (40% Flip Rate Test Set)Dialogue Preference Alignment on HH (40% Flip Rate Test Set)Follow71Win Rate (GPT-5)UDPO61.6464.0766.568.93Jul 3, 2026Evaluation resultsMethodMethodLinksWin Rate (GPT-5)Win Rate (Human)UDPOBase Model=Qwen-3-1.7B...Base Model=Qwen-3-1.7B, Noise Rate=40%2026.077167.6α-UDPOBase Model=Qwen-3-1.7B...Base Model=Qwen-3-1.7B, Noise Rate=40%2026.076869DPOBase Model=Qwen-3-1.7B...Base Model=Qwen-3-1.7B, Noise Rate=40%2026.076262.6