PhD student, Wuhan University
2 papers at NeurIPS 2025
We detect and remove backdoor samples in MLLM fine-tuning by identifying abnormal attention entropy patterns without requiring clean data or model modifications.