Reproduce “SLaDe: A Portable Small Language Model Decompiler for Optimized Assembly” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jordi Armengol-Estapé, Jackson Woodruff, Chris Cummins, Michael F. P. O'Boyle
SLaDe, a small language model decompiler developed by the University of Edinburgh and Meta AI, generates human-readable and semantically correct C code from optimized assembly. It is the first neural decompiler to demonstrate portability across x86 and ARM ISAs and different optimization levels (O0, O3) on real-world code.
Key insights
- A hybrid approach, combining a neural translation model with a symbolic type inference engine, is essential for generating accurate and compilable decompiled code.
- Task-specific training and architectural design with a small language model (200M parameters) can outperform much larger general-purpose LLMs for specialized code tasks like decompilation.
- Neural decompilation can achieve portability across diverse Instruction Set Architectures (ISAs) like x86 and ARM, and across different compiler optimization levels (O0 and O3), on real-world binaries.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.