Software Engineering on SWE-Bench Verified (ACC & ASR)
70.6AccuracyNon-attack
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Non-attackBackbone Model=GPT-5.2-1211-Global2026.04 | 70.6 | 0 | |
| SkillTrojanBackbone Model=GPT-5.2-1211-Global2026.04 | 70.4 | 92.8 | |
| GCGBackbone Model=GPT-5.2-1211-Global2026.04 | 68.2 | 12.5 | |
| SkillTrojanBackbone Model=GLM-4.72026.04 | 67.6 | 66.4 | |
| Non-attackBackbone Model=GLM-4.72026.04 | 67.3 | 0 | |
| SkillTrojanBackbone Model=Qwen3-Coder2026.04 | 65.7 | 63.9 | |
| Non-attackBackbone Model=Qwen3-Coder2026.04 | 64.8 | 0 | |
| GCGBackbone Model=Qwen3-Coder2026.04 | 62.1 | 10.3 | |
| GCGBackbone Model=GLM-4.72026.04 | 55.9 | 26.1 |