LLM decompilers can produce code that passes all standard tests yet behaves incorrectly on other inputs or removes security vulnerabilities—recompilability is not a reliable measure of correctness.
This paper reveals a critical flaw in how we evaluate LLM-based decompilers: they're judged by whether their output recompiles and passes tests, but this misses cases where the decompiled code behaves differently on other inputs or hides security vulnerabilities.