Abstention in Question Answering on BBQ Underspecified Intent
91.2Abstention F1R-Tuning
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| R-TuningModel=Qwen3-8B2026.05 | 91.2 | 85.6 | 97.6 | 98.1 | |
| TIARModel=Qwen3-8B2026.05 | 91.2 | 85.3 | 98.1 | 98.3 | |
| DPOModel=Qwen3-8B2026.05 | 90.8 | 84.9 | 97.5 | 98 | |
| TruthRLModel=Qwen3-8B2026.05 | 90.8 | 84.6 | 97.9 | 98.1 | |
| RFTModel=Qwen3-8B2026.05 | 90.4 | 84.5 | 97.2 | 95.3 | |
| RFTModel=Llama-3.1-8B-Instruct2026.05 | 85.9 | 84.8 | 87.2 | 67.6 | |
| TIARModel=Llama-3.1-8B-Instruct2026.05 | 84.6 | 92.4 | 77.9 | 70.7 | |
| TruthRLModel=Llama-3.1-8B-Instruct2026.05 | 84.5 | 91.5 | 78.5 | 71.7 | |
| DPOModel=Llama-3.1-8B-Instruct2026.05 | 84.4 | 91.2 | 78.5 | 67.4 | |
| R-TuningModel=Llama-3.1-8B-Instruct2026.05 | 82.4 | 90.6 | 75.5 | 62.7 |