Multi-hop Question Answering on HotpotQA (F1, Tokens)
69.9F1 ScorePACT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| PACTBackbone=Qwen3-8B2026.06 | 69.9 | 6,704 | |
| Multi-Agent DebateBackbone=Qwen3-8B2026.06 | 69.6 | 13,520 | |
| CoABackbone=Qwen3-8B2026.06 | 68.6 | 7,281 | |
| PACTBackbone=Qwen3-32B2026.06 | 61.5 | 4,821 | |
| Multi-Agent DebateBackbone=Qwen3-32B2026.06 | 60.3 | 12,258 | |
| TextMASBackbone=Qwen3-8B2026.06 | 59.5 | 9,384 | |
| PACTBackbone=Qwen3-14B2026.06 | 56.8 | 5,996 | |
| CoABackbone=Qwen3-14B2026.06 | 54.8 | 6,782 | |
| CoABackbone=Qwen3-32B2026.06 | 54.2 | 5,871 | |
| TextMASBackbone=Qwen3-32B2026.06 | 50.1 | 8,857 | |
| Multi-Agent DebateBackbone=Qwen3-14B2026.06 | 48.4 | 12,133 | |
| TextMASBackbone=Qwen3-14B2026.06 | 45.8 | 8,652 |