Multi-hop Question Answering on Bamboogle multi-hop (test)
40.1Avg@16GRPO
Evaluation results
| Method | Links | ||
|---|---|---|---|
| GRPOBackbone=Qwen2.5-7B, Agent Configuration=Multi-Agent & LLM Sharing2026.02 | 40.1 | 58.4 | |
| Dr. MASBackbone=Qwen2.5-7B, Agent Configuration=Multi-Agent & LLM Sharing2026.02 | 40.1 | 59.2 | |
| Dr. MASBackbone=Qwen2.5-7B, Agent Configuration=Multi-Agent & LLM Non-Sharing2026.02 | 39.8 | 57.6 | |
| GRPOBackbone=Qwen2.5-7B, Agent Configuration=Single-Agent2026.02 | 36.7 | 54.4 | |
| GRPOBackbone=Qwen2.5-7B, Agent Configuration=Multi-Agent & LLM Non-Sharing2026.02 | 31.9 | 46.4 | |
| Dr. MASBackbone=Qwen2.5-3B, Agent Configuration=Multi-Agent & LLM Sharing2026.02 | 28.6 | 49.6 | |
| GRPOBackbone=Qwen2.5-3B, Agent Configuration=Multi-Agent & LLM Sharing2026.02 | 26.4 | 46.4 | |
| Dr. MASBackbone=Qwen2.5-3B, Agent Configuration=Multi-Agent & LLM Non-Sharing2026.02 | 25.4 | 46.4 | |
| GRPOBackbone=Qwen2.5-3B, Agent Configuration=Multi-Agent & LLM Non-Sharing2026.02 | 21 | 33.6 | |
| GRPOBackbone=Qwen2.5-3B, Agent Configuration=Single-Agent2026.02 | 15.3 | 27.2 |