Activation Steering for Smart Contract Vulnerability Detection
UT Dallas, 2025-2026
This project evolved into my first first-author paper at ISI’26: IEEE International Conference on Intelligence and Security Informatics.
Steering vectors were computed and applied at inference time, without fine-tuning, and then evaluated for six LLMs on an 11-class Solidity vulnerability classification task using the SmartBugs Curated dataset.
The core finding of this work is that activation steering can improve multi-class vulnerability classification in Solidity smart contracts beyond simple few-shot prompting through direction-specific confidence calibration.
Methodology
