Instruction alignment training sharpens binary code embeddings for reverse engineering

Binary code representation learning is a foundational challenge in software security and reverse engineering. Existing approaches typically learn function-level embeddings that capture coarse-grained semantic relationships between binary functions, but they largely overlook the fine-grained instruction-level correspondences that could provide richer supervision. Huaijin Wang and Shuai Wang, in a new arXiv paper (2608.11766) accepted at ASE 2026, propose integrating instruction alignment as an auxiliary training objective. The key insight is that models already finetuned for function-level binary code similarity display substantially better instruction alignment than their pretrained base models, implying a strong correlation between instruction-level alignment and overall embedding quality. By explicitly teaching the model to align corresponding instructions across binaries—using supervision signals derived from compiler debug information—the training process encourages more accurate and interpretable code representations.

The authors' experiments demonstrate that instruction alignment training yields measurable gains: improved retrieval accuracy in binary code search and more discriminative signals for similarity judgments. In practical terms, the model can better distinguish between binaries that are merely similar versus those that are semantically identical, which matters for tasks like cross-architecture binary diffing, vulnerability hunting, and malware clustering. The work also suggests that leveraging debug data can act as a form of weak supervision to help models learn instruction-level semantics without heavy manual annotation. Since the paper has been accepted at a top software engineering venue, it aligns with a wider trend of using alignment techniques—popularized in LLM fine-tuning—to improve specialized representations. For security engineers, this means more robust binary analysis tools that can reduce false positives when scanning for vulnerable code patterns across compiled binaries.