Mathematical Reasoning on HMMT25 (accuracy)
16.3Accuracy (HMMT25)RLVR
Evaluation Results
| Method | Links | |
|---|---|---|
| RLVRBackbone=Qwen3-8B-Base, Cost=100%2026.05 | 16.3 | |
| Weight Extrap.Backbone=Qwen3-8B-Base, Cost=20%2026.05 | 16.3 | |
| RELEXBackbone=Qwen3-8B-Base, Cost=20%2026.05 | 15.4 | |
| RELEXBackbone=Qwen3-4B-Base, Cost=15%2026.05 | 14.6 | |
| ExPOBackbone=Qwen3-8B-Base, Cost=20%2026.05 | 13.3 | |
| AlphaRLBackbone=Qwen3-8B-Base, Cost=20%2026.05 | 11.3 | |
| Weight Extrap.Backbone=Qwen3-4B-Base, Cost=15%2026.05 | 10.4 | |
| RLVRBackbone=Qwen3-4B-Base, Cost=100%2026.05 | 10 | |
| ExPOBackbone=Qwen3-4B-Base, Cost=15%2026.05 | 9.6 | |
| Logits Extrap.Backbone=Qwen3-8B-Base, Cost=20%2026.05 | 8.3 | |
| AlphaRLBackbone=Qwen3-4B-Base, Cost=15%2026.05 | 7.5 | |
| Logits Extrap.Backbone=Qwen3-4B-Base, Cost=15%2026.05 | 6.7 | |
| BaseBackbone=Qwen3-8B-Base, Cost=0%2026.05 | 2.9 | |
| RELEXBackbone=Qwen2.5-Math-1.5B, Cost=15%2026.05 | 2.1 | |
| AlphaRLBackbone=Qwen2.5-Math-1.5B, Cost=15%2026.05 | 1.3 | |
| BaseBackbone=Qwen2.5-Math-1.5B, Cost=0%2026.05 | 0.8 | |
| BaseBackbone=Qwen3-4B-Base, Cost=0%2026.05 | 0.8 | |
| RLVRBackbone=Qwen2.5-Math-1.5B, Cost=100%2026.05 | 0.4 | |
| ExPOBackbone=Qwen2.5-Math-1.5B, Cost=15%2026.05 | 0.4 | |
| Logits Extrap.Backbone=Qwen2.5-Math-1.5B, Cost=15%2026.05 | 0.4 | |
| Weight Extrap.Backbone=Qwen2.5-Math-1.5B, Cost=15%2026.05 | 0 |