Application of LLMS in source code recovery from Python bytecode

Abstract

Source code recovery from intermediate representations, such as bytecode or binary code, plays a fundamental role in reverse engineering, especially in scenarios where the original source is unavailable. Although Python is typically referred to as an inter preted language, its execution involves compilation into bytecode, an intermediate form executed by the Python Virtual Machine (PVM). This process removes high-level infor mation and introduces challenges for accurate decompilation. Traditional tools often produce code that is syntactically valid but semantically limited or difficult to interpret. In recent years, Large Language Models (LLMs) based on transformer architectures have shown promising results in tasks involving source code understanding, genera tion, and even binary analysis. This study investigates the application of LLMs to the task of recovering Python source code from bytecode, an area still largely unexplored in the literature. Through a systematic review of related work, the research identifies a gap in the use of LLMs for Python bytecode decompilation. This work proposes an approach centered on modern LLMs. The hypothesis is that such models can assist in both syntactic and semantic reconstruction of the original source code. The expected contributions include evaluating the success rate of this approach and offering new insights into the intersection between machine learning and reverse engineering.