Multi-Agent Framework Boosts Binary Decompilation Re-Executability to 84-97%
A new multi-agent framework known as MCGD (Multi-level Constraint-Guided Decompilation) has been introduced by researchers to enhance the re-executability of decompiled binary code significantly. Decompilation, which involves retrieving source code from compiled binaries, plays a vital role in security assessments, malware analysis, and maintaining legacy software. Traditional decompilers often yield code that cannot compile or run properly, hindering their effectiveness. MCGD utilizes a structured validation process with three levels of constraints: syntactic correctness through parsing, compilability using GCC, and behavioral equivalence via test cases generated by LLMs. When validation is unsuccessful, specialized LLM agents refine the code based on structured error feedback. The framework was tested on 1,641 real-world binaries from the ExeBench dataset, using three decompilers: RetDec, Ghidra, and Angr, achieving an impressive 84-97% re-executability, as detailed in a paper on arXiv (2604.23940).
Key facts
- MCGD is a multi-agent framework for decompilation
- Uses three constraint levels: syntactic, compilability, behavioral equivalence
- Evaluated on 1,641 real-world binaries from ExeBench
- Tested with RetDec, Ghidra, and Angr decompilers
- Achieves 84-97% re-executability
- Paper available on arXiv (2604.23940)
- LLM agents iteratively refine code using error feedback
- Aims to improve security analysis and malware reverse engineering
Entities
Institutions
- arXiv
- ExeBench
- RetDec
- Ghidra
- Angr
- GCC