RAR
Benchmarks
| Task Name | Dataset Name | SOTA Result | Trend | |
|---|---|---|---|---|
| Provenance Tracing | RAR | TPR@1%FPR100 | 33 | |
| Downstream retrieval | RAR-B | ARC nDCG@516.2 | 24 | |
| Image Attribution Detection | RAR generated images | AUC100 | 20 | |
| Dataset usage estimation | RAR-XXL | MAE0.041 | 12 | |
| Dataset usage estimation | RAR-XL | MAE4.5 | 12 | |
| Autoregressive Visual Watermarking | RAR-XL generation | Fidelity Score (Baseline)1 | 10 | |
| Image Attribution | RAR | NM/GM100 | 8 | |
| Medical Reasoning | RaR Medicine | WR vs Base57.6 | 8 | |
| Scientific Reasoning | RAR-SCIENCE (test) | Accuracy69.9 | 6 | |
| Medical Question Answering | RaR-Medicine (test) | Length1,395 | 5 | |
| Membership Inference Attack | RAR-XXL | TPR@1%FPR8.4 | 4 | |
| Attribution | RAR | Model Pre-training Time (hours)20,000 | 4 | |
| Model Generation Identification | RAR | TPR@5%FPR (NM/G)100 | 4 | |
| Identifying memorized samples | RAR-XXL memorized 1.0 (169 train samples) | AUC97.5 | 4 | |
| Sample Attribution | RAR model derivative setting | NM/GM Score100 | 4 | |
| Member vs Generated Inference | RAR | TPR@1%FPR (NM vs G)99.9 | 4 | |
| Pairwise Preference Evaluation | RaR Medicine | Pairwise Win Rate60.6 | 4 |
Showing 17 of 17 rows