Out-of-Distribution Reasoning on OOD Reasoning Suite GPQA-D, Connect4 (test)
49.1GPQA-D ScoreRFT
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| RFTBackbone=Qwen3-8b2026.03 | 49.1 | 3.1 | 26.1 | |
| Qwen3-8bSampling=avg@52026.03 | 48.9 | 10.9 | 29.9 | |
| SFT+Backbone=Qwen3-8b2026.03 | 47.3 | 14.1 | 30.7 | |
| SPOTBackbone=Qwen3-8b2026.03 | 46.8 | 36 | 41.4 | |
| SFTBackbone=Qwen3-8b2026.03 | 39 | 12 | 25.5 | |
| RFTBackbone=Llama3.1-8b-Instruct2026.03 | 32.3 | 2 | 17.2 | |
| SFT+Backbone=Llama3.1-8b-Instruct2026.03 | 31.9 | 1.5 | 16.7 | |
| SPOTBackbone=Llama3.1-8b-Instruct2026.03 | 30 | 7 | 18.5 | |
| SFTBackbone=Llama3.1-8b-Instruct2026.03 | 29.6 | 1.7 | 15.7 | |
| Llama3.1-8b-InstructSampling=avg@52026.03 | 28.5 | 5 | 16.8 |