Poster: Enhancing Bytecode Exact Decompilation via LLM-Generated Seeds
Abstract
Decompilation is fundamental in cybersecurity, underpinning malware analysis and vulnerability discovery. Bytecode Exact Decompilation (BED) guarantees that recompiling decompiler output reproduces the original binary’s bytecode exactly, yet its genetic instantiation imposes heavy CPU and I/O loads. We propose an LLM-based seed generation method that fine-tunes a decompilation model to generate compilable C code from assembly, with automated repair of compile errors. By seeding evolution with these high-quality candidates, we substantially reduce iterative overhead. On the HumanEval-Decompile benchmark, our method achieves mean first-pass similarity of 94.7% at O0 and 69.8% at O2, outperforming prior approaches.