LLAMA: A Multi-Feedback Smart Contract Fuzzing Framework with LLM-Guided Seed Generation

The smart contract fuzzing framework LLAMA combines large language models, evolutionary mutation strategies and hybrid testing, outperforming state-of-the-art fuzzing tools in both coverage and vulnerability detection. The framework uses a hierarchical prompting strategy to guide the LLM in generating semantically valid initial seeds, and leverages runtime coverage and dependency feedback to simultaneously optimize seed generation, selection and mutation scheduling. In experiments, LLAMA achieved 91% instruction coverage and 90% branch coverage, detecting 132 of 148 known vulnerabilities.