Molecular Property Optimization Ranking and Generation on ChemCoTBench
-0.01LogP DeltaGPT-4o
Evaluation Results
| Method | Links | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| GPT-4oFamily=Text-only2026.05 | -0.01 | 37 | 0.92 | 80 | 0.13 | 70 | 0.07 | 48 | -0.02 | 30 | 0 | 39 | |
| Qwen-3-8BFamily=Text-only2026.05 | 0 | 3 | 0 | 4 | 0 | 4 | 0 | 4 | -0.01 | 0 | 0 | 2 | |
| Qwen-3-8B (SFT)Family=Text-only, SFT=true2026.05 | 0.15 | 47 | 0.48 | 52 | 0.1 | 48 | 0.04 | 38 | -0.02 | 20 | 0.02 | 36 | |
| Coconut-ChemFamily=Latent2026.05 | 0.17 | 44 | 0.57 | 58 | 0.15 | 67 | 0.04 | 45 | 0 | 35 | 0.06 | 47 | |
| QwenAtom + S2 + FTFamily=Our (0.6B), Model Size=0.6B, Stage 2 replacement (S2)=true, Fine-Tuning (FT)=true2026.05 | 1.02 | 93 | 1.04 | 95 | 0.19 | 88 | 0.36 | 87 | 0.24 | 83 | 0.33 | 89 | |
| FORGEFamily=Our (0.6B), Model Size=0.6B, Stage 1 ranking (S1)=true, Stage 2 replacement (S2)=true, Fine-Tuning (FT)=true, Backbone=QwenAtom2026.05 | 1.06 | 94 | 1.03 | 95 | 0.25 | 93 | 0.37 | 91 | 0.27 | 86 | 0.34 | 94 | |
| QwenAtom + FTFamily=Our (0.6B), Model Size=0.6B, Fine-Tuning (FT)=true2026.05 | 1.07 | 90 | 0.81 | 82 | 0.18 | 85 | 0.32 | 82 | 0.13 | 67 | 0.28 | 82 | |
| LatentChemFamily=Latent2026.05 | 1.37 | 96 | 1.53 | 89 | 0.18 | 83 | 0.26 | 74 | 0.08 | 60 | 0.17 | 82 |