Defect Detection on ScienceAgentBench 12 confirmed defects, 102 tasks
100Recall Average (RecA)Ensemble
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| EnsembleAuditing Mode=+ Agent Programs2026.04 | 100 | 100 | — | — | 27.57 | |
| Opus 4.6Auditing Mode=+ Agent Programs2026.04 | 91.7 | 100 | 67.6 | 37 | 10.94 | |
| EnsembleAuditing Mode=Definition-Only2026.04 | 91.7 | 100 | — | — | 22.72 | |
| GPT-5.4Auditing Mode=+ Agent Programs2026.04 | 83.3 | 91.7 | 60 | 45 | 3.1 | |
| Opus 4.6Auditing Mode=Definition-Only2026.04 | 83.3 | 91.7 | 58.1 | 31 | 9.5 | |
| Sonnet 4.6Auditing Mode=+ Agent Programs2026.04 | 83.3 | 91.7 | 59 | 39 | 6.76 | |
| Gemini 3.0 FlashAuditing Mode=+ Agent Programs2026.04 | 66.7 | 91.7 | 50 | 46 | 0.95 | |
| Gemini 3.0 FlashAuditing Mode=Definition-Only2026.04 | 58.3 | 75 | 41.2 | 34 | 0.84 | |
| Gemini 3.1 ProAuditing Mode=Definition-Only2026.04 | 58.3 | 75 | 57.9 | 19 | 3.67 | |
| Gemini 3.1 ProAuditing Mode=+ Agent Programs2026.04 | 58.3 | 83.3 | 70.4 | 27 | 5.82 | |
| GPT-5.4Auditing Mode=Definition-Only2026.04 | 58.3 | 75 | 47.1 | 34 | 2.71 | |
| Sonnet 4.6Auditing Mode=Definition-Only2026.04 | 58.3 | 91.7 | 51.9 | 27 | 6 |