Multi-turn Decomposition Defense on safemt
2ASRCognitive Firewall
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Cognitive FirewallDefense Type=Trajectory-aware firewalls, Target Model=mistral, Judge Model=qwen3.5:122b2026.07 | 2 | 78 | |
| Granite Guardian 8BDefense Type=Per-message guard models, Target Model=mistral, Judge Model=qwen3.5:122b2026.07 | 20 | 36 | |
| THRD-styleDefense Type=Trajectory-aware firewalls, Target Model=mistral, Judge Model=qwen3.5:122b2026.07 | 20 | 28 | |
| ShieldGemma 9BDefense Type=Per-message guard models, Target Model=mistral, Judge Model=qwen3.5:122b2026.07 | 22 | 28 | |
| TCA-styleDefense Type=Trajectory-aware firewalls, Target Model=mistral, Judge Model=qwen3.5:122b2026.07 | 24 | 18 | |
| none (undefended)Defense Type=undefended, Target Model=mistral, Judge Model=qwen3.5:122b2026.07 | 30 | 0 | |
| Llama Guard 3 8BDefense Type=Per-message guard models, Target Model=mistral, Judge Model=qwen3.5:122b2026.07 | 36 | 2 |