ResearchBenchmarksHuman Preference Evaluation (Helpfulness) on 1,172 prompts (test)Human Preference Evaluation (Helpfulness) on 1,172 prompts (test)Follow695CS WinsCS-RLHF421.48492.49563.5634.51Oct 3, 2025Evaluation resultsMethodMethodLinksCS WinsOpponent WinsTiesCS Win Rate∆ EloElo Rating (CS)Elo Rating (Opponent)CS-RLHFOpponent=SACPO (Stepwi...Opponent=SACPO (Stepwise-DPO)2025.106954611659.9870.281,270.281,200CS-RLHFOpponent=Safe-RLHFOpponent=Safe-RLHF2025.106025036754.4230.71,230.71,200CS-RLHFOpponent=P-SACPOOpponent=P-SACPO2025.1043225848257.4251.971,251.971,200