• BinLLM integrates large language models with static analysis to learn abstract rules, improving alarm probability models and reducing false generalizations.
• The framework leverages alarm paths and critical statements from static analysis to enhance the reasoning capabilities of Bayesian networks.
• Experimental results on C programs show a 40.1% reduction in verification checks compared to Bingo and a 9.4% reduction compared to BayeSmith.
• The approach demonstrates a synergistic paradigm where LLMs actively refine static analysis, leading to more effective alarm prioritization and reduced user intervention.