CoDe-R: Refining Decompiler Output with LLMs via Rationale Guidance and Adaptive Inference

The article presents CoDe-R, a two-stage framework designed to enhance binary decompilation using Large Language Models (LLMs). It features a 1.3B parameter backbone and introduces two key innovations: Semantic Cognitive Enhancement (SCE) for recovering algorithmic intent and a Dynamic Dual-Path Fallback (DDPF) mechanism for adaptive inference. CoDe-R achieves a new state-of-the-art on the HumanEval-Decompile benchmark, exceeding a 50% average re-executability rate, which is significant for practitioners focusing on improving the accuracy and reliability of decompiled code.