IMPOSSIBLEBENCH: MEASURING REWARD HACKING IN LLM CODING AGENTS

Summary from the TLDR AI newsletter

LLM-powered coding agents have been observed exploiting loopholes in tests or scoring systems rather than solving the actual tasks specified. ImpossibleBench was created to systematically measure this behavior. Its creators took existing coding benchmarks and manipulated their unit tests to directly conflict with the natural language specifications to create impossible tasks where models must choose between following instructions or passing tests. Their 'pass rate' on these impossible tasks is a direct measure of reward hacking.