Scientific and Tool-Use Reasoning on SciKnowEval and ToolUse In-domain
81.45Material ScoreGRPO
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| GRPOBackbone=Qwen3-4B2026.05 | 81.45 | 74.61 | 53.75 | 77.53 | 61.95 | 69.86 | |
| ROSDBackbone=Qwen3-4B2026.05 | 80.18 | 76.56 | 57.5 | 82.47 | 67.46 | 72.83 | |
| ROSDBackbone=Qwen3-8B2026.05 | 79.52 | 75.16 | 60.38 | 83.07 | 69.12 | 73.45 | |
| SDPOBackbone=Qwen3-8B2026.05 | 78.46 | 76.41 | 61.13 | 80.71 | 65.81 | 72.5 | |
| GRPOBackbone=Qwen3-8B2026.05 | 77.46 | 76.09 | 56.88 | 80.77 | 68.75 | 71.99 | |
| SDPOBackbone=Qwen3-4B2026.05 | 73.6 | 68.52 | 55.38 | 76.93 | 60.66 | 67.02 | |
| Base ModelBackbone=Qwen3-4B2026.05 | 62.9 | 60.63 | 31.63 | 42.92 | 58.55 | 51.32 | |
| Base ModelBackbone=Qwen3-8B2026.05 | 59.71 | 59.06 | 32.75 | 41.28 | 57.72 | 50.1 |