Tool-integrated Reasoning on Computational and Knowledge-Intensive Reasoning Suite (test)
34.2AIME 24 ScoreRAPO
Evaluation Results
| Method | Links | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| RAPOBackbone=Qwen2.5-7B-Instruct, Method Category=RL Methods, RL Step Type=Multi-step Agentic RL2026.03 | 34.2 | 29.4 | 81.1 | 92.7 | 91.5 | 24.9 | 46.8 | 42.9 | 22.1 | 47.6 | 51.3 | |
| AEPOBackbone=Qwen2.5-7B-Instruct, Method Category=RL Methods, RL Step Type=Multi-step Agentic RL2026.03 | 33.3 | 30 | 80.4 | 92.2 | 90 | 23.3 | 45.8 | 37.7 | 18.8 | 44.2 | 49.6 | |
| Tree-GRPOBackbone=Qwen2.5-7B-Instruct, Method Category=RL Methods, RL Step Type=Multi-step Agentic RL2026.03 | 31 | 30 | 79.2 | 90.1 | 88.5 | 21 | 44.6 | 42.3 | 20.2 | 44 | 49.1 | |
| Tool-StarBackbone=Qwen2.5-7B-Instruct, Method Category=Tool-Integrated Reasoning Methods2026.03 | 30 | 26.7 | 77.2 | 89.4 | 85.6 | 18.5 | 38.1 | 40.8 | 14.9 | 41.5 | 46.3 | |
| GiGPOBackbone=Qwen2.5-7B-Instruct, Method Category=RL Methods, RL Step Type=Multi-step Agentic RL2026.03 | 30 | 20 | 78.4 | 91.6 | 87.6 | 21.7 | 39.7 | 38.4 | 21.1 | 45.1 | 47.4 | |
| ARPOBackbone=Qwen2.5-7B-Instruct, Method Category=RL Methods, RL Step Type=Multi-step Agentic RL2026.03 | 30 | 30 | 78.8 | 92.2 | 88.8 | 20.1 | 42.1 | 42.8 | 21.8 | 46.8 | 49.3 | |
| LUFFYBackbone=Qwen2.5-7B-Instruct, Method Category=Off-policy Learning Methods2026.03 | 29.4 | 23.1 | 75.2 | 83.2 | 80.5 | 14.7 | 46.7 | 37 | 18 | 40.6 | 44.8 | |
| RAPOBackbone=Llama3-8B-Instruct, Method Category=RL Methods, RL Step Type=Multi-step Agentic RL2026.03 | 27.1 | 17.4 | 66.6 | 87.7 | 80.8 | 28 | 48.7 | 41.4 | 28.1 | 49.8 | 47.6 | |
| AEPOBackbone=Llama3-8B-Instruct, Method Category=RL Methods, RL Step Type=Multi-step Agentic RL2026.03 | 26.7 | 16.7 | 65.8 | 87.6 | 80.6 | 25.2 | 43.2 | 39.2 | 20.1 | 46.3 | 45.1 | |
| GPPOBackbone=Qwen2.5-7B-Instruct, Method Category=RL Methods, RL Step Type=Single-step RL2026.03 | 26.7 | 23.3 | 76.2 | 91.6 | 87.6 | 22.9 | 42.9 | 40.1 | 21.8 | 46.2 | 47.9 | |
| RAPOBackbone=Qwen2.5-3B-Instruct, Method Category=RL Methods, RL Step Type=Multi-step Agentic RL2026.03 | 24.5 | 24.8 | 72 | 87.2 | 82.8 | 18 | 45.8 | 48.9 | 20.5 | 45.9 | 47 | |
| ARPOBackbone=Qwen2.5-3B-Instruct, Method Category=RL Methods, RL Step Type=Multi-step Agentic RL2026.03 | 23.3 | 20 | 71.4 | 85 | 82.5 | 12.9 | 37.9 | 41.1 | 17.1 | 38.9 | 43 | |
| ARPOBackbone=Llama3-8B-Instruct, Method Category=RL Methods, RL Step Type=Multi-step Agentic RL2026.03 | 23.3 | 16.7 | 64.6 | 88 | 80.2 | 23.8 | 48.9 | 40.2 | 25.9 | 48.1 | 46 | |
| GRPOBackbone=Qwen2.5-7B-Instruct, Method Category=RL Methods, RL Step Type=Single-step RL2026.03 | 23.3 | 26.7 | 78 | 92.8 | 87.8 | 13 | 42.5 | 40.7 | 19.1 | 43.2 | 46.7 | |
| Tree-GRPOBackbone=Llama3-8B-Instruct, Method Category=RL Methods, RL Step Type=Multi-step Agentic RL2026.03 | 22.1 | 14.2 | 64.2 | 86.1 | 78.1 | 25.7 | 47.9 | 41.1 | 26.1 | 47.8 | 45.3 | |
| GiGPOBackbone=Qwen2.5-3B-Instruct, Method Category=RL Methods, RL Step Type=Multi-step Agentic RL2026.03 | 21.8 | 19.8 | 70.1 | 86.4 | 80.7 | 12.4 | 25 | 40 | 18.2 | 40.5 | 41.5 | |
| AEPOBackbone=Qwen2.5-3B-Instruct, Method Category=RL Methods, RL Step Type=Multi-step Agentic RL2026.03 | 21.8 | 20 | 70.8 | 84.6 | 80.8 | 15.8 | 36.1 | 43.2 | 18.7 | 40 | 43.2 | |
| GRPOBackbone=Qwen2.5-3B-Instruct, Method Category=RL Methods, RL Step Type=Single-step RL2026.03 | 20 | 13.3 | 72 | 86 | 81 | 7.3 | 39 | 36.3 | 15.2 | 36.8 | 40.7 | |
| DAPOBackbone=Qwen2.5-3B-Instruct, Method Category=RL Methods, RL Step Type=Single-step RL2026.03 | 20 | 16.7 | 71.2 | 85 | 81.2 | 6.8 | 37.2 | 36.1 | 16.2 | 35.9 | 40.6 | |
| GiGPOBackbone=Llama3-8B-Instruct, Method Category=RL Methods, RL Step Type=Multi-step Agentic RL2026.03 | 20 | 13.3 | 62.4 | 87.4 | 77.2 | 22.1 | 42.8 | 38.9 | 20.1 | 47.9 | 43.2 | |
| DAPOBackbone=Qwen2.5-7B-Instruct, Method Category=RL Methods, RL Step Type=Single-step RL2026.03 | 20 | 23.3 | 80.4 | 91 | 88.8 | 16.1 | 38.9 | 25.8 | 27 | 40.1 | 45.1 | |
| Tree-GRPOBackbone=Qwen2.5-3B-Instruct, Method Category=RL Methods, RL Step Type=Multi-step Agentic RL2026.03 | 19.8 | 19.6 | 70.7 | 85.7 | 80.6 | 13.8 | 42.4 | 43.7 | 17.8 | 43.2 | 43.7 | |
| GPPOBackbone=Qwen2.5-3B-Instruct, Method Category=RL Methods, RL Step Type=Single-step RL2026.03 | 17.2 | 19.5 | 69.7 | 86.3 | 80.5 | 13.9 | 40 | 36.9 | 16.5 | 36.9 | 41.7 | |
| Search-R1Backbone=Qwen2.5-7B-Instruct, Method Category=Tool-Integrated Reasoning Methods2026.03 | 16.7 | 6.7 | 63.8 | 82.4 | 81.2 | 12 | 26.9 | 25.9 | 16.2 | 40.4 | 37.2 | |
| DAPOBackbone=Llama3-8B-Instruct, Method Category=RL Methods, RL Step Type=Single-step RL2026.03 | 16.7 | 13.3 | 61.2 | 87.4 | 76.4 | 18.5 | 36.9 | 37.9 | 28.8 | 41.2 | 41.8 | |
| GPPOBackbone=Llama3-8B-Instruct, Method Category=RL Methods, RL Step Type=Single-step RL2026.03 | 16.7 | 6.7 | 61.8 | 86.6 | 79.4 | 18 | 44.4 | 37.6 | 19.2 | 48.6 | 41.9 | |
| GRPOBackbone=Llama3-8B-Instruct, Method Category=RL Methods, RL Step Type=Single-step RL2026.03 | 13.3 | 13.3 | 62.4 | 87.4 | 79.2 | 9.9 | 40.5 | 35.9 | 20.9 | 42.9 | 40.6 | |
| SFTBackbone=Qwen2.5-7B-Instruct, Method Category=Off-policy Learning Methods2026.03 | 12.2 | 17.3 | 53.8 | 77.9 | 77.3 | 10.9 | 46 | 36.9 | 17.2 | 39.7 | 38.9 | |
| RolloutReplayBackbone=Qwen2.5-7B-Instruct, Method Category=Off-policy Learning Methods2026.03 | 10.8 | 18 | 55 | 76.9 | 75 | 11.1 | 44.8 | 30.7 | 19.8 | 38.9 | 38.1 | |
| Search-o1Backbone=Qwen2.5-7B-Instruct, Method Category=Tool-Integrated Reasoning Methods2026.03 | 6.7 | 10 | 61.8 | 80.2 | 73.6 | 10.4 | 22.1 | 21.8 | 5.4 | 32 | 32.4 |