I love standard ml (I'm currently writing a compiler for it), but it's far from perfect. There are weird special cases in the language definition like the ad-hoc overloading between int and real, or the annoying value restriction. Records also feel half-baked, with no support for updating fields or for partially-specified records where the full set of fields is not known
While it's not yet standard nearly all Standard ML implementations support what has become known as "Successor ML" [0]. A large subset of Successor ML is common to SML/NJ, MLton, Poly/ML, MLKit and other implementations.
That includes record update syntax, binary literals, and more expressive patterns among other deficiencies in Standard ML.
For me the two big remaining issues are:
1) There's only limited Unicode support in both the core language and the standard library. This is a big issue for many real-world programs including these days the compilers for which SML is otherwise a wonderful language.
2) The module system is a "shadow language" [0] which mirrors parts of SML but which has less expressiveness where modules cannot be treated as first-class values in the program. Also if you define infix operators in a module their fixity isn't exported along with the function type. (Little annoyance that gets me every time I am inclined to write Haskell-style code with lots of operators. Though maybe that's just another hint from the universe that I shouldn't write code like that.) Of course, the fix to that would be a fundamentally different language; not a revised SML.
