What does the preference test capture?
RetroChimera combines two models through a learned ensemble to work backward from a target molecule toward possible starting materials. The Nature paper says the components contribute different inductive biases, aiming to reduce missed rare but strategically important reactions and chemically incorrect suggestions.[1]
In Microsoft's account of the 10-molecule evaluation, chemists fully accepted 9 RetroChimera routes, compared with 2 to 5 for other models. In a separate preference test, 9 PhD organic chemists chose the model's suggestion over a previously published route about 64% of the time. The result shows that the system more often presents chemical choices specialists find useful.[1]
A selected route still has to work
This setup measures expert judgment; it does not measure successful laboratory completion, product yield, time spent, or failed attempts. Expert preference still carries information because it may capture the strategic value of an uncommon transformation better than one historical answer in a dataset. The competing explanation is that 10 molecules and evaluators drawn from the same research network may not represent broader chemistry work.[1]
A strong next test would prospectively select unseen targets for different teams and have them execute the proposed routes. Reporting route-completion rate, failed reactions, chemist time, and starting-material cost together would show whether RetroChimera moves from a preferred suggestion to a working research tool.[1]