STEM Reasoning on Minerva (Accuracy)
32AccuracyGeoMin
Evaluation Results
| Method | Links | |
|---|---|---|
| GeoMinBackbone=Deepseek-R1-Distill-Qwen-1.5B, Supervision Setting=semi-supervised (10% labeled data), Domain Context=In-Domain2026.06 | 32 | |
| Seq-entropyBackbone=Deepseek-R1-Distill-Qwen-1.5B, Supervision Setting=semi-supervised (10% labeled data), Domain Context=In-Domain2026.06 | 27.6 | |
| Self-certaintyBackbone=Deepseek-R1-Distill-Qwen-1.5B, Supervision Setting=semi-supervised (10% labeled data), Domain Context=In-Domain2026.06 | 27.6 | |
| Tok-entropyBackbone=Deepseek-R1-Distill-Qwen-1.5B, Supervision Setting=semi-supervised (10% labeled data), Domain Context=In-Domain2026.06 | 27.2 | |
| TTRLBackbone=Deepseek-R1-Distill-Qwen-1.5B, Supervision Setting=semi-supervised (10% labeled data), Domain Context=In-Domain2026.06 | 25.7 | |
| Co-rewardingBackbone=Deepseek-R1-Distill-Qwen-1.5B, Supervision Setting=semi-supervised (10% labeled data), Domain Context=In-Domain2026.06 | 24.6 | |
| TraPOBackbone=Deepseek-R1-Distill-Qwen-1.5B, Supervision Setting=semi-supervised (10% labeled data), Domain Context=In-Domain2026.06 | 24.3 |