Question Answering on 2WikiMultihopQA
62.5AccuracyGRPO
Evaluation results
| Method | Links | ||
|---|---|---|---|
| GRPOBackbone=Llama3.1-8B-Instruct2026.02 | 62.5 | — | |
| TEPOdenseBackbone=Llama3.1-8B-Instruct2026.02 | 61.7 | — | |
| ARPOBackbone=Llama3.1-8B-Instruct2026.02 | 61.1 | — | |
| TEPOsparseBackbone=Llama3.1-8B-Instruct2026.02 | 60.7 | — | |
| GRPOBackbone=Qwen2.5-7B-Instruct2026.02 | 59.3 | — | |
| TEPOsparseBackbone=Qwen2.5-7B-Instruct2026.02 | 57.5 | — | |
| TEPOdenseBackbone=Qwen2.5-7B-Instruct2026.02 | 57 | — | |
| SFTBackbone=Qwen2.5-7B-Instruct2026.02 | 56.2 | — | |
| ARPOBackbone=Qwen2.5-7B-Instruct2026.02 | 55.17 | — | |
| SFTBackbone=Llama3.1-8B-Instruct2026.02 | 54.3 | — | |
| AEPOBackbone=Qwen2.5-7B-Instruct2026.02 | 54.2 | — | |
| Qwen2.5-7B-InstructBackbone=Qwen2.5-7B-Instruct2026.02 | 38.5 | — | |
| ToTBackbone=LLAMA2-13B2024.06 | 36.3 | 1,097.1 | |
| CPOBackbone=Mistral-7B2024.06 | 31.7 | 46.3 | |
| TS-SFTBackbone=Mistral-7B2024.06 | 31 | 44.2 | |
| CPOBackbone=LLAMA2-13B2024.06 | 30.3 | 60.4 | |
| CPOBackbone=LLAMA2-7B2024.06 | 29.7 | 35.7 | |
| TS-SFTBackbone=LLAMA2-13B2024.06 | 29 | 66.5 | |
| ToTBackbone=LLAMA2-7B2024.06 | 28.6 | 847.6 | |
| ToTBackbone=Mistral-7B2024.06 | 28.4 | 2,356.9 | |
| CoTBackbone=LLAMA2-13B2024.06 | 28.3 | 67 | |
| TS-SFTBackbone=LLAMA2-7B2024.06 | 27.6 | 35.5 | |
| CoTBackbone=Mistral-7B2024.06 | 26.7 | 45.1 | |
| CoTBackbone=LLAMA2-7B2024.06 | 26.3 | 35.7 | |
| Llama3.1-8B-InstructBackbone=Llama3.1-8B-Instruct2026.02 | 25.6 | — |