Stateful Agent Backdoor Attack Evaluation on Note-based instantiation 20 episodes
100Initial Success Rate (R_init)Llama-3.1-8B
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Llama-3.1-8BModel=Llama-3.1-8B2026.05 | 100 | 100 | 95 | 95 | 85 | 0 | 10 | 5 | |
| Qwen2.5-14BModel=Qwen2.5-14B2026.05 | 100 | 100 | 95 | 90 | 80 | 0 | 20 | 5 | |
| Ministral-3-14BModel=Ministral-3-14B2026.05 | 100 | 100 | 90 | 90 | 80 | 0 | 5 | 5 | |
| Qwen2.5-7BModel=Qwen2.5-7B2026.05 | 95 | 95 | 90 | 90 | 90 | 0 | 10 | 5 |