Reproduce “Rethinking Neural Decompilation: Candidate Selection as the Hidden Bottleneck” 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
Felix Mächtle, Gautham Sriram, Manjil Neupane, Nils Loose, Thomas Eisenbarth
The authors evaluate neural decompilation by generating multiple source candidates from binary-derived inputs and selecting among them. Execution-guided selection reaches 86.3% pass@1 on ExeBench with Ghidra pseudo-code input, near the candidate-pool oracle at 86.8%. On obfuscated inputs, fine-tuning plus execution-guided selection reaches 88.8%, versus 79.5% for the original checkpoint with the same selector; this result requires a test harness.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.