Taste, not procedure, becomes the training target

A small model trained by reinforcement learning to acquire research taste beats far larger ones at reproducing science. What is being learned is which questions are worth asking, the thing the field assumed could only be apprenticed.