The architecture the developer saw on a Lisp Machine like a Symbolics was tagged, 36 or 40 bit, mostly stack oriented and had lots of primitive Lisp functionality, type checking /dispatching, some GC support, etc.
That was all done in software on the MIT/LMI/TI machines.
I don't know of any papers on the Symbolics hardware that are not behind the ACM paywall, how much did it really differ from the other Lisp Machines ?
I would say that the main feature of a Lisp Machine is that you do everything from within Lisp.