Most retrosynthesis AI gets graded on one thing: can it guess the exact reaction a chemist already recorded. That misses a messier real-world problem - chemists often have several valid ways to build the same molecule and need to pick the one that preserves a specific chemical feature.
Researchers built a two-stage system called Route-Instruction Grounding and Steering, or RIGS, to address that gap. Stage A trains a language projector that learns which precursor alternatives a given instruction favors or rules out. Stage B takes that projector and uses it to steer a frozen generative model through lightweight residual adapters, without retraining the whole system. To teach the model that multiple answers can be valid, the team built training sets that pair each product with a growing number of candidate precursor sets rather than just one recorded answer.
The results show broader training support helps the model generate a wider spread of alternatives, and the steering step can then nudge generation toward what an instruction asks for. That distinction - coverage versus control - is the real contribution here. A model that only ever produces one answer per product can't follow a preference even if it is technically correct, since there is nothing to steer between.
One honest wrinkle: the paper reports that the relationship between coverage and control holds across different model sizes but is non-monotone, meaning more alternatives does not reliably translate into better steering. That is a useful check against the assumption that throwing more training data at a model automatically makes it more controllable.