Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

SmartCoder-R1: Towards Secure and Explainable Smart Contract Generation with Security-Aware Group Relative Policy Optimization

Created by
  • Haebom

Author

Lei Yu, Jingyuan Zhang, Xin Wang, Jiajia Ma, Li Yang, and Fengjun Zhang

Outline

This paper proposes the SmartCoder-R1 framework to address security and explainability challenges in smart contract generation using large-scale language models (LLMs). SmartCoder-R1 is developed through three steps: continuous pretraining (CPT), long-chain reasoning supervised learning (L-CoT SFT), and security-aware group-relative policy optimization (S-GRPO). CPT specializes the model, and L-CoT SFT trains the model to mimic human security analysis using 7,998 expert-verified reason and code samples. Finally, S-GRPO refines the generation policy by considering security compliance, formal correctness, and compilation success. Evaluation results on 756 real-world function benchmarks show that SmartCoder-R1 achieves top performance in five key metrics (ComPass, VulRate, SafeAval, FuncRate, and FullRate), with FullRate improving by 45.79% over the previous best-performing model, DeepSeek-R1. Human evaluations of the generated inferences also scored highly in terms of functionality, security, and clarity.

Takeaways, Limitations

•
Takeaways:
◦
Presenting an effective solution to the security and explainability issues of LLM-based smart contract creation.
◦
SmartCoder-R1 achieves a new state-of-the-art by demonstrating performance improvements over existing models.
◦
Increased reliability through improved security and explainability of generated code.
◦
High scores in human evaluations demonstrate practical applicability
•
Limitations:
◦
The current benchmark dataset lacks sufficient scale and diversity. Evaluation using a broader dataset is needed.
◦
Lack of analysis on the computational cost and training time of reinforcement learning techniques such as S-GRPO.
◦
Lack of performance and stability verification in actual smart contract deployment and operation environments.
◦
Lack of discussion about the possibility of discovering potential new vulnerabilities.
👍
Made with Slashpage