Researchers Fine-Tune LLMs to Reduce Vulnerabilities in Auto-Completed Smart Contract Code
Researchers from the Norwegian University of Science and Technology in Nanjing have unveiled a novel method, “vulnerability-constrained decoding,” to tackle vulnerabilities in auto-completed smart contract code, especially within Ethereum Blockchain. This technique harnesses a curated dataset of known vulnerable code lines to fine-tune large language models (LLMs), reducing potential code breaches. Remarkably, their efficient fine-tuning process takes only an hour, compared to traditional methods which might last a week. When tested on Ethereum smart contracts, the modified model demonstrated a significant 30% reduction in vulnerabilities. This groundbreaking research promises enhanced digital security in coding and paves the way for future security-focused studies in the tech domain.