Mathematical Reasoning on Olympiadbench (Average@16)
60.6Average@16Ours (Shorten Prefix)
Evaluation Results
| Method | Links | |
|---|---|---|
| Ours (Shorten Prefix)Backbone=Qwen3-4B (non-think)2026.01 | 60.6 | |
| GRPOBackbone=Qwen3-4B (non-think)2026.01 | 60 | |
| RTSBackbone=Qwen3-4B (non-think)2026.01 | 59.3 | |
| MixedBackbone=Qwen3-4B (non-think)2026.01 | 55.9 | |
| Ours (Shorten Prefix)Backbone=Qwen3-8B-Base2026.01 | 52.3 | |
| GRPOBackbone=Qwen3-8B-Base2026.01 | 51.5 | |
| BaseBackbone=Qwen3-4B (non-think)2026.01 | 51.3 | |
| MixedBackbone=Qwen3-8B-Base2026.01 | 48.6 | |
| RTSBackbone=Qwen3-8B-Base2026.01 | 48.6 | |
| Ours (Shorten Prefix)Backbone=Qwen3-4B-Base2026.01 | 47.7 | |
| GRPOBackbone=Qwen3-4B-Base2026.01 | 47.5 | |
| RTSBackbone=Qwen3-4B-Base2026.01 | 46.4 | |
| MixedBackbone=Qwen3-4B-Base2026.01 | 45.4 | |
| BaseBackbone=Qwen3-4B-Base2026.01 | 28.7 | |
| BaseBackbone=Qwen3-8B-Base2026.01 | 22.4 |