Kuna Decompiler Developed Primarily by LLM Rivals IDA Pro in Control Flow Structuring

Introduction of Kuna Decompiler

Kuna, an experimental decompiler, has been released. Its development over the summer involved a large language model (LLM) writing nearly every line of code. This project was undertaken while the developer worked as a visiting faculty researcher at the Air Force Research Lab (AFRL) and a research fellow at Metalware.

Performance and Development Methodology

Kuna's performance in control flow structuring on C programs rivals that of IDA Pro (9.2), an industry standard. Recent benchmarks show Kuna achieving perfect structuring on 44.4% of functions, compared to IDA's 45.7%. This was largely accomplished through autonomous refinement, where the LLM studied examples of underperformance against IDA Pro and other decompilers like Ghidra and angr.

This autonomous learning allowed the LLM to re-implement over 20 fundamental features from angr, which previously took years of scientific advancements to design. This approach highlights a new method for developing scientifically interesting tools that can improve automatically.

Context and Limitations

Kuna's existence is built upon decades of decompiler research. It is a Rust port of the NSA’s Ghidra, modified to align with angr’s pipeline. The developer notes that angr remains the primary tool for developing frontier algorithms in decompilation due to its design.

Kuna is primarily an experiment exploring what can be achieved with high-level scientific feedback. Currently, manual coding within Kuna would be significantly more challenging. The project relies on existing open-source research, with the hope that its success could eventually benefit projects like angr.