The map is unnecessary. Just reduce:

    (* assuming: val total_len : string -> int *)

    List.reduce seq ~f:(fun a b ->
      if (total_len a) > (total_len b) then a else b)

It's not necessary in this case, but fold is often a lot more useful than reduce. At least the way I think of it, the type of reduce is 'a list -> ('a -> 'a -> 'a) -> 'a, whereas the type of fold is 'a list -> init:'b -> ('b -> 'a -> 'b) -> 'b. The upside is that you can construct basically any type of thing you'd like, since 'b is a completely different type. The downside is that if 'a is different than 'b, you need some sort of initial value to give it.

edit: Yes, this does require you to compute string length multiple times...but keep in mind that getting string length is very cheap in languages with good strings. (That is, basically everything except C's null-terminated strings.) 99% of the time it's not going not going to matter at all. If you do care, you can map to a tuple of (original_struct,total_len) and then do the reduce and then another map to get back to your original structure, or use a fold, as I mentioned, or write a (tail-)recursive function that does it in slightly fewer operations. (Although I don't think JS has tail-call optimizations, so that's probably a bad idea if you're doing it in JS.)