SoK: Are LLMs Reliable at Source Code Recovery? A Taxonomy and Empirical Evaluation
An SoK evaluates LLM binary-to-source recovery on 45,000 samples across architectures and languages.
This Systematization of Knowledge examines whether LLMs can reliably recover source code from assembly or decompiler pseudo-C for malware analysis, vulnerability assessment, and legacy maintenance. It presents a design-centric taxonomy and evaluates methods with seven metrics across six dimensions on 45,000 samples spanning four standard and five embedded architectures, five optimization levels, and symbol stripping. Ablations cover input representation, contextual enrichment, model scale, iteration, and review roles using in-house pipelines and three unnamed off-the-shelf models. The study also compares same-language and cross-language recovery across four mature and two legacy languages.
- First SoK dedicated to LLM-assisted binary-to-source recovery.
- Evaluates 45,000 samples across nine architectures and five optimization levels.
- Ablates input form, context, model scale, iteration, and review roles.
- Tests same-language and cross-language recovery, including legacy languages.
From the original article
1 min read at arxiv.orgEffective source recovery is critical to security applications such as malware analysis, vulnerability assessment, and legacy maintenance. Large Language Models (LLMs) are reshaping this field, shifting the paradigm away from rule-based heuristics to probabilistic and high fidelity semantic recovery of source code from assembly or classical decompiler-derived pseudo-C.
By Varun Kohli · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.