Fine-Tuning and Semantic Prompt Enrichment for LLM-Based Smart Contract Vulnerability Detection

Abstract

This study examines the combined effect of fine-tuning and semantic prompt enrichment on Large Language Model-based vulnerability detection in Solidity smart contracts. We fine-tune ChatGPT-4o through a two-phase process aligned with the DASP Top 10 taxonomy—first to internalize theoretical vulnerability knowledge, then to specialize on labeled Solidity functions. We further enhance the fine-tuned model with automatically generated and human-validated code summaries as semantic enrichments to its prompts. The resulting model achieves an average F1-score of 0.58, a 66% improvement over the baseline (0.35), with the largest gains in Access Control ($+146 \%$), Denial of Service ($+353 \%$), and Reentrancy ($+35 \%$) detection. These results show that domain-aligned fine-tuning and semantic prompt enrichment jointly improve the precision and recall of LLM-based smart-contract auditing, offering a practical path toward AI-assisted security analysis.