Energy-Efficient LLM-Based Decompilation via a Two-Stage Pipeline on Constrained Hardware

Recent advances in large language models (LLMs) have enabled substantial progress in decompiling binary executables. Yet, current approaches remain highly resource-intensive and mainly target single-function programs. In this work, we present a lightweight two-stage decompilation pipeline designed to operate on constrained hardware while extending LLM-based decompilation to multi-function programs. We construct a new dataset of synthetic yet structurally diverse C programs and their corresponding assembly code, generated using two distinct LLMs to enforce a deliberate distributional shift between the training and testing data. Our system relies on a fine-tuned 7B-parameter model for primary decompilation and a secondary, inference-only 24B-parameter correction model for targeted compilation-error repair, enabling efficient processing under strict memory and compute constraints. To enable a direct comparison with prior work, we adapt the LLM4Decompile framework for multi-function settings by automatically reconstructing unified headers while preserving cross-function dependencies. Experimental evaluations using AST similarity, lexical similarity, compilation rate, and execution similarity show that our method achieves competitive structural and textual fidelity while surpassing LLM4Decompile baselines in compilation and execution success, despite operating on commodity-class hardware. Moreover, we demonstrate that our approach reduces energy consumption by 85% compared to existing refine-based methods, making it highly suitable for sustainable deployment in resource-constrained environments.