Multi-hop Reasoning on MuSiQue (Recall@2, Recall@5, SR)
42.67Recall@2HeadRank
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| HeadRankModel=Qwen3-4B, Method=HeadRank2026.04 | 42.67 | 55.1 | 100 | |
| RankGPTModel=Qwen3-4B, Method=RankGPT2026.04 | 41.25 | 52.99 | 98.9 | |
| ICRModel=Qwen3-1.7B, Method=ICR2026.04 | 39.61 | 51.07 | 100 | |
| RankGPTModel=Llama-3.1 8B, Method=RankGPT2026.04 | 39.3 | 51.4 | 98.2 | |
| HeadRankModel=Qwen3-1.7B, Method=HeadRank2026.04 | 39.05 | 54.47 | 100 | |
| RankGPTModel=GPT-3.5 Turbo, Method=RankGPT2026.04 | 38.7 | 54.1 | 98.6 | |
| ColBERTv2Method=ColBERTv22026.04 | 37.94 | 49.52 | — | |
| HeadRankModel=Qwen3-0.6B, Method=HeadRank2026.04 | 36.72 | 51.02 | 100 | |
| RankGPTModel=Mistral 7B, Method=RankGPT2026.04 | 36.2 | 48.7 | 20.5 | |
| RankGPTModel=Qwen3-1.7B, Method=RankGPT2026.04 | 35.02 | 47.83 | 93.7 | |
| ICRModel=Qwen3-4B, Method=ICR2026.04 | 34.23 | 46.52 | 100 | |
| ICRModel=Qwen3-0.6B, Method=ICR2026.04 | 31.23 | 43.69 | 100 | |
| RankGPTModel=Qwen3-0.6B, Method=RankGPT2026.04 | 30.42 | 37.49 | 80.1 |