Using DeepSeek and Qwen Local Models for Binary Reverse Engineering
Researchers implemented a hierarchical multi-agent system where a DeepSeek lead model coordinates four smaller Qwen3.5-9B local models to successfully reverse engineer binaries into tested Python code. The system utilizes an oracle to run test cases and verify correctness before final submission. This experiment demonstrates a cost-effective, practical approach to complex software engineering tasks by leveraging a capable frontier model to orchestrate cheaper, smaller local models. It opens up advanced reverse engineering capabilities to a broader range of developers and researchers. During the process, the lead model dynamically modified worker prompts and even wrote a custom lookup tool when workers struggled with function signatures and memory layouts. However, on smaller tasks, the team coordination created overhead, taking more time and money than running the lead model alone.
## BACKGROUND
Binary reverse engineering is the process of analyzing compiled machine code to understand its functionality, often for security auditing or interoperability purposes. Hierarchical multi-agent LLM systems use a top-level model to plan and supervise lower-level models, combining high-level reasoning with decentralized execution.