Nova addresses critical challenges in binary code analysis by developing language models specifically designed for assembly code, overcoming limitations of traditional LLMs for security applications.

• Introduces hierarchical attention mechanisms to handle low information density in assembly code
• Employs contrastive learning to manage diverse optimizations in assembly
• Significantly improves performance on key security tasks including code decompilation and similarity detection
• Creates a specialized foundation for advanced binary code analysis in security applications

This research provides security teams with more effective tools for analyzing potentially malicious code and identifying vulnerabilities in binary files without source code access.

Nova: Generative Language Models for Assembly Code with Hierarchical Attention and Contrastive Learning