Retrieval-based Question Answering on HoVer 4-HOP
71.5Recall@100BetterTogether(PO, mmGRPO)
Evaluation Results
| Method | Links | |
|---|---|---|
| BetterTogether(PO, mmGRPO)Strategy=Module-level mmGRPO, Backbone=qwen32025.08 | 71.5 | |
| mmGRPOStrategy=Module-level mmGRPO, Backbone=qwen32025.08 | 71 | |
| MIPROv2 (PO)Strategy=Baseline, Backbone=qwen32025.08 | 69.3 | |
| BetterTogether(PO, mmGRPO)Strategy=Module-level mmGRPO, Backbone=llama3.12025.08 | 68.3 | |
| MIPROv2 (PO)Strategy=Baseline, Backbone=llama3.12025.08 | 63.4 | |
| Vanilla CoTStrategy=Baseline, Backbone=qwen32025.08 | 60.6 | |
| mmGRPOStrategy=Module-level mmGRPO, Backbone=llama3.12025.08 | 60.2 | |
| Vanilla CoTStrategy=Baseline, Backbone=llama3.12025.08 | 59.5 |