ResearchBenchmarksCapability Elicitation on Olympiads (test)Follow88Elicitation Rate (U_final / U_benign)SFT+RL1030.2550.570.75Apr 23, 2026Evaluation ResultsMethodMethodLinksElicitation Rate (U_final / U_benign)SFT+RLSupervisor=LLAMA 3.1 8...Supervisor=LLAMA 3.1 8B (VERY WEAK)2026.0488SFT+RLSupervisor=GPT 4o mini...Supervisor=GPT 4o mini (WEAK)2026.0486SFTSupervisor=GPT 4o mini...Supervisor=GPT 4o mini (WEAK)2026.0485SFTSupervisor=LLAMA 3.1 8...Supervisor=LLAMA 3.1 8B (VERY WEAK)2026.0440RLSupervisor=GPT 4o mini...Supervisor=GPT 4o mini (WEAK)2026.0434RLSupervisor=LLAMA 3.1 8...Supervisor=LLAMA 3.1 8B (VERY WEAK)2026.0413