What's this about?
The method derives exactly checkable spatial and topological questions from solved protein structures, then trains language and 3D-geometry outputs from shared representations.
What exactly is post-training supervision?
After a base model learns general language, it receives curated examples and feedback that shape particular capabilities—here, structural reasoning rather than ordinary text imitation.
Why does this matter?
The reported +3.23 percentage-point macro gain across ten benchmarks suggests scientific structure can teach transferable reasoning. Controls with random, synthetic and shuffled structures were weaker, though independent replication is still needed.