ResearchBenchmarksCapability Elicitation on Code Contests (test)Capability Elicitation on Code Contests (test)Follow94Elicitation RateSFT32.6448.5764.580.43Apr 23, 2026Evaluation resultsMethodMethodLinksElicitation RateSFTSupervisor=GPT 4o mini...Supervisor=GPT 4o mini (WEAK)2026.0494SFT+RLSupervisor=GPT 4o mini...Supervisor=GPT 4o mini (WEAK)2026.0493SFT+RLSupervisor=LLAMA 3.1 8...Supervisor=LLAMA 3.1 8B (VERY WEAK)2026.0493SFTSupervisor=LLAMA 3.1 8...Supervisor=LLAMA 3.1 8B (VERY WEAK)2026.0469RLSupervisor=GPT 4o mini...Supervisor=GPT 4o mini (WEAK)2026.0456RLSupervisor=LLAMA 3.1 8...Supervisor=LLAMA 3.1 8B (VERY WEAK)2026.0435