Reproduce “ReCopilot: Reverse Engineering Copilot in Binary Analysis” 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

Guoqiang Chen, Huiqi Sun, Daguang Liu, Zhiqi Wang, Qiang Wang, Bin Yin, Lu Liu, Lingyun Ying

ReCopilot helps reverse engineers interpret stripped-binary pseudocode by recovering symbols. It combines a domain-trained 7B model with static analysis that adds related functions, call chains, and cross-function variable flow. On 1,038 functions from 60 binaries, it scores 43.50 versus 23.76 for BinaryNinja Sidekick on variable-name ROUGE, a recall-oriented metric, not exact-name accuracy or proof of dependable downstream analysis.

Key insights

  • Domain-specific training via a three-stage approach (Continued Pretraining, Supervised Fine-Tuning, Direct Preference Optimization) is effective for building expertise in binary analysis tasks.
  • Context enhancement, specifically through call chain and data flow analysis, significantly improves the accuracy of binary analysis by providing richer semantic information to the model.
  • While excelling in semantic correctness, generating syntactically precise outputs in specific formats remains a challenge for the model.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.