Mathematical Reasoning on AIME 25 (Accuracy and Token Usage)
57.5AccuracySC@32
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| SC@32Backbone=Qwen3-4B-Instruct-Thinking2026.06 | 57.5 | 13,100 | 215,300 | |
| ASCBackbone=Qwen3-4B-Instruct-Thinking2026.06 | 57.5 | 126,600 | 126,600 | |
| RL-guidedBackbone=Qwen3-4B-Instruct-Thinking2026.06 | 57.1 | 32,100 | 96,200 | |
| ESCBackbone=Qwen3-4B-Instruct-Thinking2026.06 | 57 | 49,700 | 167,300 | |
| RL-guidedBackbone=Qwen3-1.7B-Thinking2026.06 | 44.6 | 83,900 | 271,600 | |
| ASCBackbone=Qwen3-1.7B-Thinking2026.06 | 44.2 | 358,800 | 358,800 | |
| SC@32Backbone=Qwen3-1.7B-Thinking2026.06 | 44.1 | 25,800 | 511,500 | |
| ESCBackbone=Qwen3-1.7B-Thinking2026.06 | 44.1 | 116,600 | 421,700 | |
| SC@32Backbone=GPT-4.1-nano2026.06 | 33.5 | 5,000 | 86,500 | |
| ASCBackbone=GPT-4.1-nano2026.06 | 33.5 | 59,700 | 59,700 | |
| RL-guidedBackbone=GPT-4.1-nano2026.06 | 33.4 | 23,000 | 49,200 | |
| ESCBackbone=GPT-4.1-nano2026.06 | 33.2 | 22,400 | 75,200 | |
| SC@32Backbone=Qwen3-0.6B-Thinking2026.06 | 30.1 | 23,500 | 441,100 | |
| ASCBackbone=Qwen3-0.6B-Thinking2026.06 | 30.1 | 359,500 | 359,500 | |
| RL-guidedBackbone=Qwen3-0.6B-Thinking2026.06 | 30 | 97,300 | 298,700 | |
| ESCBackbone=Qwen3-0.6B-Thinking2026.06 | 29.8 | 121,000 | 415,200 |