LLM4Decompile

Why we included this project

Security analysts, malware researchers, vulnerability hunters, and maintainers trying to recover lost source all have a reason to look at this project: it turns disassembled Linux x86_64 functions into readable C using large language models. You get two paths, a direct end-to-end decompile and a refinement pass that cleans up the pseudo-code Ghidra already produces, and model sizes from 1.3B to 33B parameters let you pick something that fits your GPU budget. The benchmark treats re-executability, not syntax similarity, as the measure of success, so decompiled code has to actually compile and pass the original test cases. The project is research-led, with a paper and training data behind it, but the released checkpoints and Colab notebook mean you can run it on your own binaries without training a model yourself.