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

Researchers propose CoDe-R, a two-stage framework using Large Language Models to improve binary decompilation by reducing logical errors and semantic misalignment. A 1.3B model using this approach achieves state-of-the-art performance on the HumanEval-Decompile benchmark, becoming the first lightweight model to exceed 50% re-executability rates.

Binary decompilation—converting stripped executables back into readable source code—represents a fundamental challenge in reverse engineering and security research. Traditional approaches struggle with the irreversible information loss that occurs during compilation, but recent LLM advances have shown potential for this task. CoDe-R addresses a critical limitation of current LLM-based decompilers: they generate code that appears plausible but fails to execute properly due to "logical hallucinations" and semantic drift.

The framework introduces two innovations that target different failure modes. Semantic Cognitive Enhancement trains models to explicitly reason about algorithmic intent, not just syntax, during refinement. The Dynamic Dual-Path Fallback mechanism during inference creates a verification layer that prevents the model from committing to problematic outputs, instead adopting fallback strategies when confidence is low. This dual-stage approach reflects a maturation in how researchers apply LLMs to structured code tasks.

The breakthrough metric—a 1.3B model exceeding 50% re-executability on a standard benchmark—has implications for resource-constrained deployments. Security researchers, malware analysts, and reverse engineers working on commodity hardware gain access to more capable tools. The efficiency-versus-accuracy tradeoff has historically limited adoption of sophisticated decompilation techniques in field conditions.

Looking ahead, the open-source release should accelerate iteration on decompilation methods. Standardized benchmarks like HumanEval-Decompile enable comparative progress tracking. The next frontier likely involves domain-specific models trained on particular instruction sets or architectures, as well as integration with symbolic execution engines that can validate inferred semantics against runtime behavior.

• →CoDe-R achieves state-of-the-art binary decompilation performance using only a 1.3B parameter model, enabling efficient reverse engineering on resource-limited systems.
• →The framework combines semantic reasoning training with adaptive dual-path inference to reduce both logical errors and syntactic failures in decompiled code.
• →First lightweight model to exceed 50% average re-executability rate on HumanEval-Decompile benchmark, significantly outperforming prior baselines.
• →Open-source release accelerates research in LLM-based code recovery for security analysis, malware analysis, and vulnerability research workflows.
• →Approach demonstrates how specialized training and verification mechanisms improve LLM reliability for structured tasks where execution validation is possible.