Now spending most of my time training agents at Jinesis AI Lab, I've been revisiting this approach as an alternative to elaborate RL environments and large-scale SFT. We’re experimenting with hand-crafted trajectory pairs to see how precisely we can steer behaviour from limited feedback, particularly in our medical LLM work, where continual-learning signals are sparse and targeted corrections are difficult to express as RL rewards. I had largely moved on from this work, but the conversations at COLM made me realize there might be a much broader question hiding in one of its results :)