When LLM Decompilers Recompile More and Preserve Less
While LLM-powered decompilers excel at turning assembly into clean, recompilable C code, that cosmetic fluency often masks behavioral divergence. By evaluating decompiled outputs against dynamically generated fuzzing inputs rather than static unit tests, the authors demonstrate that LLM refinement can boost build rates from 75% to 90% while simultaneously reducing true behavioral alignment from 74% to 62%. Crucially, models frequently invent types, callees, and guard clauses to resolve missing context, silently erasing up to a tenth of known CVE crash signatures.
Why it's worth reading
It exposes a critical evaluation blind spot in code generation, showing how optimizing for clean, compilable syntax often leads LLMs to hallucinate logic and silently erase security vulnerabilities.