Multi-hop Question Answering on MuSiQue (dev)
65.073F1 ScoreDense-aware HKVM controller
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Dense-aware HKVM controllervariant_type=Dense-aware controller, training_signals=out-of-fold HKVM2026.06 | 65.073 | 64.722 | 64.115 | |
| ColBERTv2variant_type=Reference predictor2026.06 | 58.309 | 57.758 | 49.317 | |
| ColBERTv22026.06 | 58.309 | 57.758 | 49.317 | |
| Controller dense-onlyvariant_type=Source-only controller control, training_signals=out-of-fold HKVM2026.06 | 58.247 | 57.758 | 52.172 | |
| Controller HKVM-onlyvariant_type=Source-only controller control, training_signals=out-of-fold HKVM2026.06 | 56.76 | 56.213 | 58.64 | |
| Learned HKVM calibrationvariant_type=Reference predictor, training_signals=out-of-fold HKVM2026.06 | 56.754 | 56.213 | 53.303 | |
| Contriever2026.06 | 55.565 | 54.903 | 54.199 | |
| BM252026.06 | 46.045 | 45.345 | 41.746 | |
| Weighted-HG-KV2026.06 | 39.925 | 39.222 | 38.643 | |
| Static-HG2026.06 | 37.714 | 36.905 | 36.74 | |
| RaRBackbone=Qwen3-4B2026.06 | 36.46 | 26.6 | — | |
| Weighted-KG-PPR2026.06 | 36.378 | 35.581 | 32.727 | |
| KG-PPR2026.06 | 36.332 | 35.54 | 32.478 | |
| R1-SearcherBackbone=Qwen3-4B2026.06 | 36.02 | 26 | — | |
| Co-occurrence-HG2026.06 | 35.632 | 34.712 | 31.734 | |
| ARCOBackbone=Qwen3-4B2026.06 | 35.55 | 27.4 | — | |
| AgentPRMBackbone=Qwen3-4B2026.06 | 34.81 | 25.8 | — | |
| Search-R1Backbone=Qwen3-4B2026.06 | 33.66 | 24.8 | — | |
| RLERBackbone=Qwen3-4B2026.06 | 33.35 | 24.4 | — | |
| CARMOBackbone=Qwen3-4B2026.06 | 32.26 | 22.2 | — | |
| ARCOBackbone=Llama-3.2-3B2026.06 | 29.7 | 21.6 | — | |
| Base ModelBackbone=Qwen3-4B, Prompting=zero-shot2026.06 | 28.05 | 19.4 | — | |
| Search-R1Backbone=Llama-3.2-3B2026.06 | 26.99 | 18.8 | — | |
| R1-SearcherBackbone=Llama-3.2-3B2026.06 | 26.41 | 17.4 | — | |
| AgentPRMBackbone=Llama-3.2-3B2026.06 | 25.83 | 16.6 | — | |
| RaRBackbone=Llama-3.2-3B2026.06 | 25.26 | 16 | — | |
| RLERBackbone=Llama-3.2-3B2026.06 | 23.02 | 14.6 | — | |
| CARMOBackbone=Llama-3.2-3B2026.06 | 21.49 | 13.4 | — | |
| Base ModelBackbone=Llama-3.2-3B, Prompting=zero-shot2026.06 | 15.12 | 9.4 | — |