What was done
Kohli and co-authors survey LLM-assisted binary-to-source recovery, build a taxonomy and evaluate state-of-the-art methods on 45,000 samples across architectures, optimisation levels and languages.
Varun Kohli and co-authors wrote the first systematization of knowledge on LLM-assisted binary-to-source recovery, which turns assembly or decompiler pseudo-C back into source. They ask how reliable these methods are, because existing work is fragmented and evaluated in different ways.
How it was done
The paper gives a design-centric taxonomy of the methods and evaluates them with seven metrics along six dimensions. The tests use 45,000 samples from four standard and five embedded architectures, five optimisation levels and symbol stripping. Controlled in-house pipelines with three off-the-shelf models ablate input representation, contextual enrichment, model scale, iteration and review roles. The authors also test same-language and cross-language recovery across four mature and two legacy languages.
Results
The paper reports comparisons and design insights rather than a single headline number, and aims to guide future work in the field, including embedded and IoT targets that earlier studies covered little. It was submitted on 8 October 2026.