Multi-objective molecular optimization (BDQ) on MuMOInstruct (unseen instruction)
34.2Success RateSFT
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| SFT2026.03 | 34.2 | 25.7 | 8.8 | |
| RePO2026.03 | 17 | 32.2 | 5.5 | |
| GRPOinitialization=None2026.03 | 7.8 | 45.7 | 3.6 | |
| Base Model2026.03 | 4.2 | 10.4 | 0.4 | |
| GRPOinitialization=SFT init2026.03 | 2.6 | 5.8 | 0.2 | |
| Distill-SFT2026.03 | 2 | 7.7 | 0.2 |