The value restriction is a rule that governs when type inference is allowed to polymorphically generalize a value declaration. In short, the value restriction says that generalization can only occur if the right-hand side of an expression is syntactically a value. [...] The value restriction prevents a ref cell (or an array) from holding values of different types, which would allow a value of one type to be cast to another and hence would break type safety
It can be annoying to work around this, because there are potentially ML programs that would be type-safe and would otherwise compile if they were not hindered by the value restriction.
I got my answer, though:
Ii is interesting that in ML, the presence of mutable ref cells and parametric polymorphism requires the whole language to be dominated by a "value restriction"  to ensure that the type system remains sound, whereas in Haskell, because IORef's can only be created (and used) in
the IO monad, no such restriction is necessary.