Evaluating LLM-Generated ACSL Annotations for Formal Verification

Formal specifications are crucial for building verifiable and dependable software systems, yet generating accurate and verifiable specifications for real-world C programs remains challenging. This paper presents an empirical evaluation of automated ACSL annotation generation strategies for C programs, comparing a rule-based Python script, Frama-C's RTE plugin, and three large language models (DeepSeek-V3.2, GPT-5.2, and OLMo 3.1 32B Instruct). The study focuses on one-shot annotation generation, assessing how these approaches perform when directly applied to verification tasks. Using a filtered subset of the CASP benchmark, we evaluate generated annotations through Frama-C's WP plugin with multiple SMT solvers, analyzing proof success rates, solver timeouts, and internal processing time. Our results show that rule-based approaches remain more reliable for verification success, while LLM-based methods exhibit more variable performance. These findings highlight both the current limitations and the potential of LLMs as complementary tools for automated specification generation.