ImpossibleBench: Measuring Reward Hacking in LLM Coding Agents lesswrong.com 5 points by vinhnx 3 days ago