Long-form factuality evaluation on FactBench
84.4AccuracyCURE
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| CUREPipeline Stage=Factuality Optimization (Full Pipeline)2026.04 | 84.4 | 0.136 | 0.195 | 0.667 | |
| SFT + F-RL + COPipeline Stage=Calibration Optimization2026.04 | 82.5 | 0.159 | 0.214 | 0.648 | |
| SFTPipeline Stage=Supervised Fine-Tuning2026.04 | 82 | 0.137 | 0.184 | 0.563 | |
| SFT + F-RLPipeline Stage=Feasibility RL2026.04 | 78.1 | 0.157 | 0.2 | 0.605 | |
| L2RF2026.04 | 77.1 | — | — | — | |
| Base LLMBackbone=Llama3.1-8B-Instruct2026.04 | 73.6 | 0.162 | 0.219 | 0.541 | |
| LitCab2026.04 | 73.6 | 0.212 | 0.276 | 0.575 |