Backdoor Attack on GPT-5-generated shopping requests
91.2ASRBadDLM
Evaluation results
| Method | Links | ||
|---|---|---|---|
| BadDLMModel=LLaDA-8B-Instruct2026.05 | 91.2 | 65.4 | |
| RL-basedModel=LLaDA-8B-Instruct2026.05 | 72.8 | 64.5 | |
| VPIModel=LLaDA-8B-Instruct2026.05 | 41.2 | 65.5 | |
| SFT-basedModel=LLaDA-8B-Instruct2026.05 | 38.5 | 65.4 | |
| Benign (No Attack)Model=LLaDA-8B-Instruct2026.05 | 2.7 | 65.5 |