ImpossibleBench: Measuring Reward Hacking in LLM Coding Agents lesswrong.com 2 points by gmays 10 hours ago