Towards Specification-Driven LLM-Based Generation of Embedded Automotive Software

TL;DR

This paper introduces a framework combining LLMs with formal verification to generate reliable embedded automotive software from specifications, demonstrating promising results even without iterative refinement.

Contribution

It presents the spec2code framework for integrating LLMs with critics for code generation and provides an initial empirical feasibility study in industrial automotive contexts.

Findings

Formally correct code can be generated without iterative backprompting.

The framework effectively combines natural language and formal specifications.

Initial results are promising for industrial automotive software development.

Abstract

The paper studies how code generation by LLMs can be combined with formal verification to produce critical embedded software. The first contribution is a general framework, spec2code, in which LLMs are combined with different types of critics that produce feedback for iterative backprompting and fine-tuning. The second contribution presents a first feasibility study, where a minimalistic instantiation of spec2code, without iterative backprompting and fine-tuning, is empirically evaluated using three industrial case studies from the heavy vehicle manufacturer Scania. The goal is to automatically generate industrial-quality code from specifications only. Different combinations of formal ACSL specifications and natural language specifications are explored. The results indicate that formally correct code can be generated even without the application of iterative backprompting and fine-tuning.