Everybody is working on (or waiting for) world models. Language models are not sufficient for doing work in the meatspace.

It's wild to me that this is even a question, because of how AI brained the entire world has become. I mean, you have a tool that you know works, and a tool that you know doesn't, but you still feel the need to find a way to make the square peg fit into the round hole.

I'm sure there are ways to "correct the spatial reasoning of LLMs" (whatever that even means) but I'd question whether you even need an LLM to begin with. I know that isn't the answer you're looking for, I'm just a slack-jawed knuckle-dragging Luddite. But this seems like the kind of simple, common problem for which established solutions already existed, like "talk to an actual engineer and not a chatbot that can't even do basic math."

And if it's a matter of not wanting to pay someone, consider that if you're already paying for an inadequate simulation of knowledge, you should be willing to pay for actual knowledge.

I would consider asking it to model its proposed designs in openscad or build123d (ideally something query-able). Then have it render and examine plausibility / suitability from different angles. Get it to render the part in use also and give instructions to think about forces and motion.

Recommend doing this in a coding harness not a chat box.

The reason I think you might have more success with this is that the model is mostly thinking about the part in words, which it can convert to a part design in CAD in code, which in turn means it can use relative positioning and relationships.

You will be able to iterate more easily, compare things, compute properties, commit to git etc and the process is more steerable than generative production of images.

Once you can get renders most current models should be good enough to discriminate when they are producing nonsense, misaligned parts, things that don't fit etc. It's not going to be perfect, but I suspect it'll be much better. The whole process of code -> render -> inspect forces way more grounding than your current workflow.

As far as I know models don't have a "visual imagination" but a process like this could sort of be a slow approximation of one. They clearly do have SOME spatial understanding (pelican tests show us that!) but it feels really non-human.