Kuna: Decompiler Development in the Age of Coding Agents

TL;DR Highlight

An LLM-written decompiler, Kuna, nearly matches IDA Pro on control flow structuring

Who Should Read

A backend or tooling developer curious how to hand a large project to coding agents and improve quality automatically

Core Mechanics

An LLM wrote nearly all the code
- Kuna is an experimental decompiler that turns binaries back into C code.
- The author says an LLM wrote nearly every line.
Structuring score is near IDA
- On C programs, Kuna perfectly structures 44.4% of functions (control flow structuring recovers if, loop and switch).
- IDA Pro 9.2 gets 45.7%.
The secret is an autonomous refinement loop
- The LLM studies cases where it performs worse than IDA Pro.
- It learns by trial and error how other decompilers solve hard problems.
20+ angr features reimplemented
- Ghidra and the angr decompiler were also studied in the process.
- This let the LLM reimplement more than 20 fundamental angr features in Kuna.
Human-led research came first
- Starting automatic refinement required new fundamental metrics and meaningful benchmark data.

Community Reactions

Creator explains the Ghidra origin
- The creator says Kuna started as Ghidra ported to Rust and has changed considerably.
- The Switch layout and some variable simplification in the site screenshot are not Ghidra features.
Pushback: why not train an LLM to be the decompiler
- One commenter argued compilers can generate unlimited training data.
- The creator replied that this has been tried for two decades with no project beating traditional decompilers on fundamental metrics.
Interest in agentic naming of functions and variables
- A commenter hoped agents could track data flow and propose names.
- The creator said on DecBench, Codex and Claude given only assembly dominate but have significant recall problems.

How to Apply

Start a loop with a failure-case set
- Collect 5 cases today where your coding agent does worse than a reference tool.
- Attach the reference output and metric to each and ask the agent to analyze the gap.
Define a measurable metric first
- Design the metric and benchmark data yourself before handing improvement to an agent.
- Have the agent automatically verify only that the score goes up.
Use an existing tool as a teacher
- When reimplementing a feature, give the agent input/output examples from a mature open-source version.
- Check the license first.

Try It

copy & run

# Paste into Claude or Claude Code as is (simplified version of the article's method)

You are an agent improving {{our tool}}. The reference tool is {{reference tool}}.
The metric is {{metric name}}.

Below are cases where our tool scores lower than the reference.
{{failure cases and both outputs}}

Do the following in order.
1. For each case, analyze why the reference does better.
2. Pick one common pattern and propose a fix for our code.
3. Write a test that checks the metric improves after the fix.

Next step: re-measure the score after the fix and repeat with the remaining failures.

Terminology

DecompilerA tool that turns compiled binaries back into human-readable code like C. It is like restoring a translated book to something close to the original.

Control flow structuringThe step that recovers if, loop and switch structures from low-level jumps. It redraws a route as a proper road map instead of raw turn-by-turn jumps.

IDA ProA commercial decompiler and reverse engineering tool that is the industry standard.

angr / Ghidraangr is an open-source research decompiler and Ghidra is an open-source reverse engineering tool released by the NSA.