BODHI: Precise OS Kernel Specification Inference with Large Language Models

A new arXiv preprint introduces BODHI, a system that aims to automate the generation of precise operating‑system kernel specifications using large language models (LLMs). Formal verification of kernels — the mathematical proof that system calls behave as intended — depends on detailed, correct specifications. Writing those specs by hand requires deep systems expertise. Can generative models close that gap?

Paper and approach

arXiv:2605.23931v1 presents BODHI and evaluates it on OSV‑Bench, a benchmark of 245 specification‑generation tasks. The paper frames the problem around the gap between LLM fluency and the stringent accuracy required for verification. BODHI combines model prompting, task decomposition, and verification‑aware checks to produce specifications that are both precise and semantically aligned with kernel source and documentation. The authors report improvements over prior, more naive LLM approaches on the OSV‑Bench suite; those gains are framed as a step toward reducing the expert effort needed for formal verification.

Implications and context

The work touches on issues beyond academic tooling. Kernel specifications live at the intersection of software reliability and security; errors can yield subtle, exploitable behavior. Advances that let LLMs write higher‑fidelity specs could accelerate verification of critical infrastructure software used in cloud services, telecommunications gear, and embedded devices. At the same time, such capabilities raise dual‑use questions. It has been reported that AI export controls and trade policies are increasingly relevant to model access and deployment, and governments in the U.S., China, and Europe are watching advances in code‑capable models closely.

BODHI is a preprint; community validation, replication, and adversarial testing will be necessary before its outputs can be trusted in high‑assurance contexts. Still, the paper marks a concrete step in a fast‑moving area: can LLMs reach the precision required by formal methods? BODHI suggests they might be getting closer.