ResearchBenchmarksHuman Preference Evaluation (Harmlessness) on 1,172 Prompts (test)Follow677Win Count (CS)CS-RLHF105253.5402550.5Oct 3, 2025Evaluation ResultsMethodMethodLinksWin Count (CS)Win Count (Opponent)Tie CountWin Rate (CS)Elo Difference (∆ Elo)Elo Rating (CS)Elo Rating (Opponent)CS-RLHFOpponent=Safe-RLHFOpponent=Safe-RLHF2025.106774554059.4766.61,266.61,200CS-RLHFOpponent=SACPO (Stepwi...Opponent=SACPO (Stepwise-DPO)2025.1055450910951.9213.331,213.331,200CS-RLHFOpponent=P-SACPOOpponent=P-SACPO2025.101279794851.288.91,208.91,200