Question Answering on 2WikiMultiHopQA, HotpotQA, and MuSiQue Aggregate (test)
24.5Average AccuracyFedGRPO
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| FedGRPOModel=Qwen2.5-3B2026.02 | 24.5 | — | — | |
| Central-GRPOModel=Qwen2.5-3B2026.02 | 23.9 | — | — | |
| DPSDA-FL+SFTModel=Qwen2.5-3B2026.02 | 22.4 | — | — | |
| Central-GRPOModel=Qwen2.5-1.5B2026.02 | 20.6 | — | — | |
| FedGRPOModel=Qwen2.5-1.5B2026.02 | 20.6 | — | — | |
| DPSDA-FL+GRPOModel=Qwen2.5-3B2026.02 | 19.6 | — | — | |
| Fedpetuning+SFTModel=Qwen2.5-3B2026.02 | 16.5 | — | — | |
| DPSDA-FL+SFTModel=Qwen2.5-1.5B2026.02 | 16.3 | — | — | |
| DPSDA-FL+GRPOModel=Qwen2.5-1.5B2026.02 | 15 | — | — | |
| Fedpetuning+GRPOModel=Qwen2.5-3B2026.02 | 13.2 | — | — | |
| Fedpetuning+SFTModel=Qwen2.5-1.5B2026.02 | 8.4 | — | — | |
| Zero-shotModel=Qwen2.5-3B2026.02 | 8 | — | — | |
| Fedpetuning+GRPOModel=Qwen2.5-1.5B2026.02 | 7.2 | — | — | |
| Zero-shotModel=Qwen2.5-1.5B2026.02 | 7 | — | — | |
| GRPOFramework=LLM + RL2025.06 | — | 18.521 | 6.917 | |
| Llama3.1-8BFramework=LLM2025.06 | — | 19.356 | 9.087 | |
| Llama3.1-8BFramework=LLM + MAS2025.06 | — | 18.751 | 12.561 | |
| MAPPOFramework=LLM + MAS + MARL2025.06 | — | 39.689 | 24.499 | |
| MHGPO-FoFFramework=LLM + MAS + MARL2025.06 | — | 42.153 | 25.248 | |
| MHGPO-FoF(os)Framework=LLM + MAS + MARL2025.06 | — | 43.331 | 26.109 | |
| MHGPO-ISFramework=LLM + MAS + MARL2025.06 | — | 38.989 | 23.369 | |
| MHGPO-RRFramework=LLM + MAS + MARL2025.06 | — | 42.544 | 24.99 | |
| PPOFramework=LLM + RL2025.06 | — | 17.052 | 6.158 | |
| Qwen2.5-72BFramework=LLM2025.06 | — | 30.684 | 19.522 | |
| Qwen2.5-72BFramework=LLM + MAS2025.06 | — | 30.468 | 17.379 | |
| R1-Searcher(GRPO)Framework=LLM + RAG (+RL)2025.06 | — | 38.328 | 23.483 | |
| R1-Searcher(PPO)Framework=LLM + RAG (+RL)2025.06 | — | 40.236 | 22.297 | |
| Search-o1Framework=LLM + RAG (+RL)2025.06 | — | 22.215 | 16.23 |