ConVer: LLM-Driven Compositional Verification for C Programs
ConVer is a verification tool designed for extensive C programs, employing a large language model (LLM) to generate function contracts based on a high-level assertion. It functions within a CEGAR-CEGIS framework, alternating between system-level and function-level evaluations, and refines contracts through SMART ICE learning when failures occur. In tests involving four benchmark suites, ConVer demonstrated a verification success rate of 82-96% on the Frama-C benchmark, which included 45 straightforward C programs across three LLM backends. Notably, 93-95% of the programs that converged needed only one iteration of CEGAR-CEGIS. Additionally, results were shared regarding the X.509 parser benchmark, addressing state-space explosion in Bounded Model Checking (BMC) by using a top-down verification approach.
Key facts
- ConVer uses LLM to synthesize function contracts from system property.
- Employs CEGAR-CEGIS loop with SMART ICE learning for contract refinement.
- Tested on four benchmark suites including Frama-C and X.509 parser.
- 82-96% verification success on Frama-C benchmark with three LLM backends.
- 93-95% of converged programs required only one CEGAR-CEGIS iteration.
- Addresses state-space explosion in BMC by compositional decomposition.
- Tool is top-down, starting from top-level assertion in C program.
- Published on arXiv with ID 2605.27051.
Entities
Institutions
- arXiv