Complex Reasoning on Arena-Hard 2.0 (test)
52.9Overall AccuracyTCR
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| TCROptimizer=GRPO, Foundation Model=Qwen3-14B2025.10 | 52.9 | 53.2 | 47.2 | 58.3 | |
| TCROptimizer=GSPO, Foundation Model=Qwen3-8B2025.10 | 52.5 | 51.7 | 46 | 59.7 | |
| ELOOptimizer=GRPO, Foundation Model=Qwen3-14B2025.10 | 52 | 51 | 47.9 | 57.1 | |
| PREFOptimizer=GSPO, Foundation Model=Qwen3-8B2025.10 | 51.9 | 49.6 | 47 | 59.1 | |
| PREFOptimizer=GRPO, Foundation Model=Qwen3-14B2025.10 | 51.6 | 49 | 47 | 58.8 | |
| ELOOptimizer=GSPO, Foundation Model=Qwen3-8B2025.10 | 50.9 | 49.4 | 47.6 | 55.7 | |
| LWOptimizer=GSPO, Foundation Model=Qwen3-8B2025.10 | 50.8 | 48.8 | 45.5 | 58.1 | |
| PWOptimizer=GSPO, Foundation Model=Qwen3-8B2025.10 | 50.2 | 49.6 | 45.1 | 55.9 | |
| LWOptimizer=GRPO, Foundation Model=Qwen3-14B2025.10 | 49.4 | 49.1 | 43.6 | 55.5 | |
| Base ModelOptimizer=GSPO, Foundation Model=Qwen3-8B2025.10 | 49.2 | 48.4 | 46.2 | 53 | |
| Base ModelOptimizer=GRPO, Foundation Model=Qwen3-14B2025.10 | 49.1 | 47 | 45 | 55.3 | |
| PWOptimizer=GRPO, Foundation Model=Qwen3-14B2025.10 | 48.4 | 49.3 | 43 | 52.9 |