New AI System Transforms Binary Code Into Readable Source Code More Accurately

Turning compiled binary programs back into human-readable source code remains one of the hardest tasks in cybersecurity and software maintenance. While large language models have recently opened new possibilities for this process—known as decompilation—they still struggle to reliably reconstruct how a program actually works, accurately recover data structures, and handle it all without consuming massive computational resources.

The CodeInverter Approach#

Researchers have introduced the CodeInverter Suite, a system designed to overcome these limitations through three complementary components. The core innovation centers on a specialized prompt engineering method called the CodeInverter Workflow (CIW), which feeds control flow graphs and explicit data mappings directly into language models during decompilation. These additional inputs help the AI better understand the program's structure and how data moves through it.

To support this method, the team built the CodeInverter Dataset (CID)—a massive collection of 8.69 million code samples that include control flow information and data mapping annotations. This domain-specific training resource enables models to learn decompilation patterns specific to binary code recovery.

Lightweight Models, Strong Results#

Rather than scaling up to massive language models, the researchers created two efficient CodeInverter Models (CIMs) with just 1.3 billion and 6.7 billion parameters respectively. These compact models are designed to run effectively in resource-constrained and privacy-sensitive settings, where downloading code to large cloud systems isn't feasible.

Testing on standard benchmarks showed the improvements are substantial. The CIW method enhanced decompilation performance across multiple existing LLMs by an average of 13.07% in re-executability (whether the decompiled code actually runs) and 23.94% in re-compilability (whether it can be compiled back). The largest CodeInverter model achieved state-of-the-art results, surpassing competitors that are over 100 times larger by 11.03% in re-executability and 6.27% in readability.

Why It Matters#

This work addresses a critical bottleneck in reverse engineering, threat analysis, and legacy software maintenance. By making decompilation both more accurate and computationally efficient, the approach could enable security researchers and developers to work with binary code in resource-limited environments without sacrificing quality.