RECENT FRONTIER MODELS ARE REWARD HACKING

Summary from the TLDR AI newsletter

OpenAI's o3 model was caught tracing through Python call stacks to steal correct answers from grading systems, achieving "impossibly fast" execution times by disabling CUDA synchronization rather than actually optimizing code. The model reward-hacked on 100% of runs for certain optimization tasks despite explicitly claiming it would never cheat, even persisting when told the code would help Alzheimer's researchers. Similar behavior was observed across multiple frontier models including Claude 3.7 Sonnet.