SPARC is a neuro-symbolic, scenario-based framework that enhances automatic unit test generation for C programs by addressing the shortcomings of large language models in code synthesis.

It tackles issues such as undefined path coverage and ungrounded dependencies, delivering traceable tests with meaningful diagnostic information that effectively serve as executable documentation.

The approach enables the use of cost-effective, smaller LLMs without sacrificing test quality, signaling strong potential for broader industrial adoption in software verification and reliability.

SPARC can match or exceed the performance of the symbolic execution tool KLEE in many cases, while producing more readable and maintainable tests and maintaining a high test retention rate of about 94% through self-correction.

A developer study found SPARC-generated code to be notably more readable and maintainable, highlighting practical benefits for engineers working with legacy C codebases.

The framework follows a four-stage process: analyze control flow graphs to map execution paths, create an Operation Map to anchor reasoning in validated utilities, synthesize path-targeted tests for specific paths, and iteratively validate with compiler and runtime feedback to refine tests.

Across 59 real-world and algorithmic C projects, SPARC substantially improves test metrics, delivering approximately 31.36% higher line coverage, 26.01% higher branch coverage, and 20.78% higher mutation score versus a baseline of simple prompt generation.