I have also found this — once you get out into some really novel workspaces they will revert back to replication of adjacencies that already exist. Fable 5 and 5.1 in particular will talk around the issue in a way that no model prior seemed to.

I think the issue is fairly fundamental, perhaps because of tuning, perhaps even beyond what they are tuned or trained on, because the bulk of what they can go find and ground into seems to drag them heavily toward implicit biases that split around categories I can’t quite discern. The category boundaries sometimes look like academic fields, but sometimes they seem like more fundamental attractors.