When LLM Decompilers Recompile More and Preserve Less
Summary
LLM-based decompilers are evaluated not just by whether generated code recompiles but by how their output behaves on unseen inputs. The paper introduces Decompile-Diverge, a behavioral oracle that synthesizes a driver and fuzzes inputs to detect divergence between the original and decompiled code, revealing that some systems show behavior changes or reveal vulnerabilities despite passing shipped tests. It reports non-trivial divergence rates across eight systems and CVE-related functions, highlighting risks for vulnerability analysis and malware research.