Thinking Collapse Analysis on HarmBench Harm
0Thinking Collapse RatePAIR
Evaluation Results
| Method | Links | |
|---|---|---|
| PAIRModel=Qwen3 1.7B2026.03 | 0 | |
| AutoDAN-TurboModel=Qwen3 1.7B2026.03 | 0 | |
| GCGModel=Qwen3 4B2026.03 | 0 | |
| PAIRModel=Qwen3 4B2026.03 | 0 | |
| AutoDANModel=Qwen3 4B2026.03 | 0 | |
| GCGModel=Qwen3 8B2026.03 | 0 | |
| AutoDANModel=Qwen3 8B2026.03 | 0 | |
| GCGModel=Qwen3 1.7B2026.03 | 0.5 | |
| AutoDANModel=Qwen3 1.7B2026.03 | 0.5 | |
| PAIRModel=Qwen3 8B2026.03 | 0.5 | |
| AutoDAN-TurboModel=Qwen3 8B2026.03 | 0.6 | |
| AutoDAN-TurboModel=Qwen3 4B2026.03 | 1.1 | |
| MS_ReverseModel=Qwen3 1.7B2026.03 | 7.5 | |
| MS_ReverseModel=Qwen3 8B2026.03 | 10.8 | |
| MS_ReverseModel=Qwen3 4B2026.03 | 11 |