Science Reasoning on GPQA Diamond
71.6AUCOAAFormat-Adaptive-Answer
Evaluation Results
| Method | Links | |
|---|---|---|
| Format-Adaptive-AnswerBackbone=Qwen3-8B2026.01 | 71.6 | |
| Base modelBackbone=Qwen3-8B2026.01 | 69.4 | |
| Hard-Length 16kBackbone=Qwen3-8B2026.01 | 68.9 | |
| Adaptive-AnswerBackbone=Qwen3-8B2026.01 | 68.8 | |
| SFTBackbone=Qwen3-8B2026.01 | 68.7 | |
| TWYNBackbone=Qwen3-8B2026.01 | 67.1 | |
| Hard-Length 8kBackbone=Qwen3-8B2026.01 | 63.5 | |
| Soft-LengthBackbone=Qwen3-8B2026.01 | 62.9 | |
| Hard-Length 8k → 4kBackbone=Qwen3-8B2026.01 | 61 | |
| Normalized-LengthBackbone=Qwen3-8B2026.01 | 58.3 | |
| No-ThinkingBackbone=Qwen3-8B2026.01 | 40.8 |