AI & ML
A Systematic Survey of Decompilation: Mapping Classical and Machine-Learning Approaches
arXiv cs.SE0 views
arXiv:2611.21612v1 Announce Type: new Abstract: Context: Decompilers reconstruct source-level code from compiled binaries, and analysts rely on them whenever source code is unavailable. Many recent studies apply machine learning (ML), including large language models, to parts of this task. From the primary literature alone, it is unclear where these methods sit in the classical decompilation pipeline and how well their claims are supported. Objectives: We map decompilation research onto a pipeline-centered taxonomy and ask two questions: where classical, ML-trained, and ML-prompted approaches are positioned (RQ1), and how ML-based studies vary in technical design and evaluation evidence (RQ2). Methods: We surveyed 405 decompilation studies from 1973 to 2026, collected through database search and citation tracing. We classified 394 primary studies into eight pipeline stages and two complementary categories. For the 116 ML-based studies that propose or modify decompilation methods, we extracted the model-use strategy, reusable backbones, target platform, and the evidence behind each evaluation. Results: Classical techniques still dominate front-end recovery, IR lifting, and control-flow structuring (277 of 394 primary studies are classical). ML-based work (117 studies) clusters instead around symbol and name recovery, output refinement, and end-to-end translation. Type recovery is the clearest mixed region. Most ML-based studies target native binaries (97 of 116), and most evidence is item-level matching against references (89 of 116); execution-based tests, human or LLM judgment, and downstream-task evidence are much rarer. Conclusion: ML has not displaced the classical pipeline. It has been taken up where a recovered artifact can be posed as a prediction, translation, or refinement problem, and reported gains only make sense when read against the pipeline position and the evidence used to support them.
Source:https://arxiv.org/abs/2611.21612PreviousMulti-GPU Acceleration of Interface Problems in FEniCS, with Application to Cardiac ModelingNextExploring Test Smells Across Programming Languages: A Systematic Mapping Study
Related posts
The AI Token Casino: How the Regenerate Button Drains Your Budget
0dev.to AI
The Teacher in the Age of AI: Will the Machine Take Their Place, or Give Them Back Their Time?
0dev.to AI

Skillware v0.6.2: Section-Chunked Tax PDFs, RPC Health Gates & Strict Core Typing
0dev.to AI

GPT-6.2 Astra vs GPT-6 Astra: What Changes in Cost per Accepted Task?
0dev.to AI
Jobs That Fade, Jobs That Emerge: What the Labour Market Looks Like Up to 2030
0dev.to AI
Can an AI agent read your pricing page and submit your demo form? A dev checklist for agent-ready sites
0dev.to WebDev