Abstention on AbstentionBench BBQ
97Abstention F1Gemini 3
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Gemini 3Model Type=Proprietary API Model2026.05 | 97 | 96.1 | 98 | 67.3 | |
| Claude Sonnet 4.5Model Type=Proprietary API Model2026.05 | 87.9 | 100 | 78.5 | 59.2 | |
| TIARBackbone=Llama-3.1-8B-Instruct2026.05 | 87.5 | 96.1 | 80.3 | 73.5 | |
| GPT-5.2Model Type=Proprietary API Model2026.05 | 86 | 84.3 | 87.8 | 73.5 |