Sequential Experimental Design on ChemLex Acid–Amine Wetlab (complete matched replay suite)
92.1Final Best PerformanceCARE
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| CARE2026.06 | 92.1 | 81.6 | 7.9 | 13 | |
| Gryffin-style categorical BO2026.06 | 89.8 | 77.2 | 10.2 | 3 | |
| Fixed public heuristic2026.06 | 89.5 | 73.8 | 10.5 | 0 | |
| EDBO-style descriptor GP-EI2026.06 | 88.9 | 80.5 | 11.1 | 0 | |
| CAREablation=w/o frontier proposal mode2026.06 | 88.6 | 79.2 | 11.4 | 0 | |
| CAREablation=w/o intervention gate2026.06 | 87.6 | 79.3 | 12.4 | 0 | |
| CAREablation=w/o Policy Planner2026.06 | 87.4 | 78.7 | 12.6 | 0 | |
| SMAC-style RF-EI2026.06 | 86.3 | 71 | 13.7 | 0 | |
| GP-EI2026.06 | 86.2 | 75 | 13.8 | 7 | |
| CAREablation=w/o recovery proposal mode2026.06 | 85.3 | 78.4 | 14.7 | 10 | |
| TPE-style BO2026.06 | 85.1 | 78.6 | 14.9 | 0 | |
| Public incumbent2026.06 | 83.9 | 77.2 | 16.1 | 10 | |
| BayBE/BoFire-style mixed BO2026.06 | 83.1 | 74.7 | 16.9 | 3 | |
| GP-UCB2026.06 | 83.1 | 75.5 | 16.9 | 0 | |
| Chemistry descriptor BO2026.06 | 82.4 | 73.5 | 17.6 | 0 | |
| Random2026.06 | 81.9 | 73.7 | 18.1 | 0 | |
| Non-evolving LLM policy2026.06 | 81.4 | 76.1 | 18.6 | 7 | |
| BoTorch-style GP-LogEI2026.06 | 81.3 | 72.8 | 18.7 | 7 | |
| Ungated LLM-evolving policy2026.06 | 81.2 | 71.2 | 18.8 | 10 | |
| LMABO-style LLM-BO2026.06 | 79 | 73.3 | 21 | 0 | |
| Stratified random2026.06 | 78.6 | 71.1 | 21.4 | 0 | |
| Generic surrogate BO2026.06 | 76.8 | 70.4 | 23.2 | 10 | |
| Categorical empirical-Bayes UCB2026.06 | 75.4 | 68.2 | 24.6 | 3 |