Haining Wang

LLMs can find more real-world firmware vulnerabilities (and with higher precision) when structured as a feedback-driven system that interleaves reasoning and tool interaction, rather than a one-pass static analysis.