BODHI: LLM-Based OS Kernel Specification Inference

A new method called BODHI has been introduced by researchers to enhance the automated creation of formal specifications for operating system kernels through large language models. This approach enriches few-shot prompts with a detailed C-to-Python translation guide that encompasses 15 categories of domain-specific patterns. Drawing inspiration from Structured Chain-of-Thought prompting, it distinguishes between the extraction of pre-conditions and the generation of post-conditions. BODHI was tested on nine models from six different providers—Anthropic, Mistral, Amazon, DeepSeek, Meta, and Alibaba—addressing the OSV-Bench benchmark, which includes 245 specification generation tasks from the Hyperkernel OS kernel, achieving a highest reported Pass@1 rate of 55.10%.