Reinforcement Learning on DMC PointMass
944.5Top Left ScoreHILP
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| HILPReward Specification=Ground-truth reward, Protocol=Zero-Shot RFRL, Data Budget=10,000 transitions2026.05 | 944.5 | 96 | 192.3 | 0.17 | 308.3 | |
| Ours-FT (FB-PbRL)Reward Specification=Preferences, Fine-tuning=true, Data Budget=2,000 preference pairs2026.05 | 928.1 | 483.5 | 541.5 | 51.4 | 501.1 | |
| FB-PbRL (Ours-FT)fine-tuning=true2026.05 | 926.6 | 597.8 | 752.8 | 6.2 | 570.8 | |
| RLDPReward Specification=Ground-truth reward, Protocol=Zero-Shot RFRL, Data Budget=10,000 transitions2026.05 | 890.4 | 795.6 | 805.2 | 193.4 | 671.1 | |
| PSMReward Specification=Ground-truth reward, Protocol=Zero-Shot RFRL, Data Budget=10,000 transitions2026.05 | 831.4 | 730.3 | 451.4 | 43.3 | 514.1 | |
| LaplaceReward Specification=Ground-truth reward, Protocol=Zero-Shot RFRL, Data Budget=10,000 transitions2026.05 | 713.5 | 581.1 | 689.1 | 21.3 | 501.2 | |
| OPRL2026.05 | 558.1 | 399.7 | 338.8 | 53.2 | 337.5 | |
| CLARIFY2026.05 | 536.6 | 287.2 | 369 | 78.2 | 317.8 | |
| FBReward Specification=Ground-truth reward, Protocol=Zero-Shot RFRL, Data Budget=10,000 transitions2026.05 | 489.7 | 761.2 | 107.2 | 236.1 | 398.5 | |
| LIRE2026.05 | 160.7 | 106.1 | 95.2 | 47.2 | 102.3 | |
| FB-PbRL (Ours-T)test-time CPTS=true2026.05 | 70.2 | 77.8 | 116.7 | 11.7 | 69.1 | |
| OPPO2026.05 | 59.8 | 19.4 | 11.9 | 5.5 | 24.1 | |
| DPPO2026.05 | 52.4 | 4.1 | 3.9 | 4.8 | 16.3 |