Math Reasoning on AIME25 (Accuracy, Tokens)
70.42AccuracyCOPT
Evaluation results
| Method | Links | ||
|---|---|---|---|
| COPTBackbone=Qwen3-8B, Reasoning Effort=High (Peak Accuracy)2026.05 | 70.42 | 12,801 | |
| CoTBackbone=Qwen3-8B2026.05 | 67.5 | 12,924 | |
| CoT (Greedy)Backbone=Qwen3-8B2026.05 | 60 | 13,292 | |
| StaticBudget=40962026.04 | 30 | 12,701 | |
| TABBudget=10k2026.04 | 30 | 9,429 | |
| TAB All-SubQBudget=10k2026.04 | 30 | 8,567 | |
| TABModel Variant=4B, Budget=10k2026.04 | 30 | 9,222 | |
| TAB All-SubQBudget=8k2026.04 | 26.7 | 7,921 | |
| TABModel Variant=4B, Budget=8k2026.04 | 26.7 | 6,813 | |
| TABBudget=8k2026.04 | 23.3 | 6,304 | |
| L-JProtocol=Multi-Turn, Strategy=All-SubQ2026.04 | 23.3 | 6,837 | |
| TAB All-SubQBudget=5k2026.04 | 23.3 | 4,116 | |
| TABModel Variant=4B, Budget=5k2026.04 | 23.3 | 5,018 | |
| StaticBudget=20482026.04 | 22.5 | 6,784 | |
| L-JProtocol=Individual2026.04 | 22.5 | 7,084 | |
| L-JProtocol=Multi-Turn2026.04 | 22.5 | 6,795 | |
| L-JProtocol=Multi-Turn, Model Variant=4B2026.04 | 22.5 | 6,140 | |
| TABBudget=5k2026.04 | 20 | 4,202 | |
| L-JProtocol=Individual, Model Variant=4B2026.04 | 20 | 6,286 | |
| TABModel Variant=4B, Budget=3k2026.04 | 20 | 4,089 | |
| StaticBudget=10242026.04 | 17.5 | 4,616 | |
| TABBudget=3k2026.04 | 16.7 | 3,063 | |
| TAB All-SubQBudget=3k2026.04 | 16.7 | 3,203 | |
| StaticBudget=5122026.04 | 15 | 3,119 | |
| StaticBudget=2562026.04 | 12.5 | 2,483 |